BD × AI LAB · TSP R&D

R&D page.Stran R&D. This page reports research and development in progress. The results are partial, come from development runs and can change. BD × AI Lab reserves the right to change or correct anything on this page at any time.Ta stran poroča o raziskavah in razvoju, ki še potekajo. Rezultati so delni, izhajajo iz razvojnih tekov in se lahko spremenijo. BD × AI Lab si pridržuje pravico, da kar koli na tej strani kadar koli spremeni ali popravi.

8zTSPA new TSP kernel in pure Rust · BD × AI Lab · goal pilot 12 Nov · new build 15 Nov 2026Novo jedro za TSP v čistem Rustu · BD × AI Lab · pilot ciljnega preizkusa 12. nov. · nova gradnja 15. nov. 2026Provisional: one development run per entryZačasno: en razvojni tek na udeležencaUpdated 19 Nov 2026Posodobljeno 19. nov. 2026

The travelling salesman problem (TSP) asks for the shortest round trip that visits every city on a map exactly once. It is easy to state and gets very hard as the number of cities grows.Problem potujočega trgovca (TSP) išče najkrajšo krožno pot, ki vsako mesto na zemljevidu obišče natanko enkrat. Postaviti ga je lahko, z rastočim številom mest pa postane zelo težak.

Latest resultNajnovejši izid ·

One billion cities on one laptopMilijarda mest na enem prenosniku

8zTSP made a complete round trip through one billion cities on one laptop with 31 GB of memory: a complete tour after 2 min 15 s, and after one hour a tour 4.75% above the estimated lower bound, with a peak of 12.2 GB of memory, about 12 bytes per city.8zTSP je na enem prenosniku z 31 GB pomnilnika naredil celo krožno pot skozi milijardo mest: celo pot po 2 min 15 s in po eni uri pot 4,75 % nad ocenjeno spodnjo mejo, z največ 12,2 GB pomnilnika, okoli 12 bajtov na mesto.

development run of 17 Nov; not yet in the published programrazvojni tek 17. novembra; še ni v objavljenem programu

nu3496 · 3,496 cities in Nicaragua · optimum 96,132nu3496 · 3.496 mest v Nikaragvi · optimum 96.132

The optimal tour of nu3496 in aboutOptimalna pot za nu3496 v približno 6 seconds6 sekundah

With 7 threads, the median of 10 runs was 6.1 s (build of 15 Nov), and all 10 found it, the slowest after 45 s. For comparison, on one core of the same laptop: LKH 3.0.7 had not reached it after 16.8 hours (one run of 15–17 Nov; best 96,152, 0.021% above); Concorde proved the optimum in a median of 70 min of processor time (three seeds: 55, 70 and 149 min). Checking every possible tour is hopeless: already 27 cities would take longer than the age of the universe on this laptop; and even if every atom in the universe were a computer like this laptop, with as many processors as there are atoms, all computing nonstop for the whole age of the universe, together they would have checked only 0.000…0001% of nu3496’s tours: almost six book pages of zeros before the first 1. LKH and Concorde are the two programs TSP research measures itself against. 8zTSP finds the optimal tour; Concorde proves it is optimal (From the lab:S 7 nitmi je bila mediana 10 tekov 6,1 s (gradnja 15. novembra), našlo ga je vseh 10, najpočasnejši po 45 s. Za primerjavo na enem jedru istega prenosnika: LKH 3.0.7 ga po 16,8 ure ni dosegel (en tek 15.–17. nov.; najboljše 96.152, 0,021 % nad njim); Concorde je optimum dokazal v mediani 70 min procesorskega časa (tri semena: 55, 70 in 149 min). Preizkus vseh možnih poti je brezupen: že 27 mest bi na tem prenosniku trajalo dlje od starosti vesolja; in tudi če bi bil vsak atom v vesolju računalnik, kot je ta prenosnik, s toliko procesorji, kolikor je atomov, in bi vsi neprekinjeno računali ves čas, odkar vesolje obstaja, bi skupaj preverili le 0,000…0001 % poti nu3496: skoraj šest knjižnih strani ničel pred prvo 1. LKH in Concorde sta programa, ob katerih se meri raziskovanje TSP. 8zTSP optimalno pot najde, Concorde pa dokaže, da je optimalna (Iz laboratorija: 15 Nov15. nov. andin 14 Nov14. nov.).

Try it in your browserPreizkusite v svojem brskalniku: the same engine in its minimal build, on your own computer.: isti motor v minimalni gradnji, na vašem računalniku.

More on nu3496Več o nu3496

Earlier figure.Prejšnja številka. With build 303eee6 of 14 Nov and 7 threads, the median of 30 runs was 10.6 s; 23 of 30 runs found it within 30 s, and all 30 found it (Z gradnjo 303eee6 14. novembra in s 7 nitmi je bila mediana 30 tekov 10,6 s; 23 od 30 tekov ga je našlo v 30 s, našlo ga je vseh 30 (From the lab, 14 NovIz laboratorija, 14. nov.).

3,496 cities mean about 103.496 mest pomeni okoli 1010,86810.868 tours: on this laptop about 10poti: na tem prenosniku okoli 1010,85310.853 years. Already 27 cities would take about 7 times the age of the universe. Even with every atom a laptop, about 10,759 zeros stand after the decimal point. And if each of those laptops had as many processors as there are atoms in the universe, 10let. Že 27 mest bi trajalo okoli 7-krat dlje od starosti vesolja. Tudi če je vsak atom prenosnik, stoji za decimalno vejico okoli 10.759 ničel. In če bi imel vsak od teh prenosnikov toliko procesorjev, kolikor je atomov v vesolju, skupaj 10160 processors in all, the zeros would shrink only from about 10,759 to about 10,680: still almost six pages.procesorjev, bi se ničle skrčile le z okoli 10.759 na okoli 10.680: še vedno skoraj šest strani.

Assumptions: about 10Predpostavke: okoli 1080 atoms in the visible universe; a laptop like this one checks 8 cores × 6.67 × 10atomov v vidnem vesolju; tak prenosnik preveri 8 jeder × 6,67 × 107 = 5.3 × 10= 5,3 × 108 tours per second; the age of the universe 13.8 billion years; together about 10poti na sekundo; starost vesolja 13,8 milijarde let; skupaj okoli 10106 tours checked out of 10preverjenih poti od 1010,86810.868, a share of 10, delež 10−10,760−10.760 %; a book page of 1,800 characters. 10%; knjižna stran ima 1.800 znakov. 1080 × 1080 = 10160 processors, each as fast as one core of this laptop (6.67 × 10procesorjev, vsak hiter kot eno jedro tega prenosnika (6,67 × 107 tours/s); share 10poti/s); delež 10−10,681−10.681 %.

A good route throughDobra pot skozi 10 million cities10 milijonov mest inv 15 minutes15 minutah, on one laptop., na enem prenosniku.

— more than five times the 1,904,711 cities of the World TSP— več kot petkrat toliko mest, kot jih ima World TSP (1.904.711) (only the number of cities is compared)(primerjamo le število mest)

With its own view of the map, 8zTSP puts 7 of the laptop’s 8 cores to work on one map at once. The standard programs, LKH and Concorde’s linkern, use one core per run: on the same laptop, with the settings used, neither got there in two hours.Z lastnim pogledom na zemljevid 8zTSP na enem zemljevidu hkrati zaposli 7 od 8 jeder prenosnika. Standardna programa, LKH in Concordov linkern, uporabljata eno jedro na tek: na istem prenosniku in z uporabljenimi nastavitvami v dveh urah do tja ni prišel nobeden od njiju.

Try 8zTSP in your browserPreizkusi 8zTSP v brskalniku

The demo is password-protected; access on request.Demo je zaščiten z geslom; dostop na prošnjo.

For reviewers: how 8zTSP works and its code (password; access on request)Za recenzente: kako deluje 8zTSP in njegova koda (geslo; dostop na prošnjo)

Results at a glanceIzidi na kratko

MapZemljevidCitiesMestaReferenceReferenca8zTSPThreadsNitiDateDatumConditionsPogoji
nu3496 (Nicaragua)nu3496 (Nikaragva)3,4963.496optimum 96,132optimum 96.132the optimal tour, median 6.1 s of 10 runsoptimalna pot, mediana 6,1 s v 10 tekih715 Nov 202615. nov. 2026From the labIz laboratorija
Nine national maps, wi29 to nu3496Devet nacionalnih zemljevidov, od wi29 do nu349629–3,49629–3.496the known optimaznani optimumithe optimal tour in 90 of 90 runsoptimalna pot v 90 od 90 tekov114 Nov 202614. nov. 2026The tableTabela
si6035 (Slovenia)si6035 (Slovenija)6,0356.035Concorde’s lower bound, proof runningConcordova spodnja meja, dokaz tečeour best tour 8,728,190 m (8zTSP, 18 Nov 2026, five of its own earlier tours fused after a 24-hour run), the best known tour, at most 0.006% above the optimum by Concorde’s bound (18 Nov 2026); LKH 3.0.14 reached the same tour from scratch the same morningnaša najboljša pot 8.728.190 m (8zTSP, 18. nov. 2026, pet svojih prejšnjih poti, zlitih po 24-urnem teku), najboljša znana pot, največ 0,006 % nad optimumom po Concordovi meji (18. nov. 2026); LKH 3.0.14 je isto pot našel od začetka isto jutro118 Nov 202618. nov. 2026From the labIz laboratorija
bm33708 (Burma)33,70833.708Waterloo’s lower bound 959,011spodnja meja Waterlooja 959.0114 cores each, 1 h: level with LKH 3.0.7 (4 copies); 0.04% above the lower bound; our first tour in 1 spo 4 jedra, 1 h: izenačeno z LKH 3.0.7 (4 kopije); 0,04 % nad spodnjo mejo; naša prva pot po 1 s416 Nov 202616. nov. 2026From the labIz laboratorija
DIMACS E1M.01,000,0001.000.000Held–Karp boundmeja Held–Karp1% above the Held–Karp bound after 5 min 33 s; 0.772% after 30 min1 % nad mejo Held–Karp po 5 min 33 s; 0,772 % po 30 min114 Nov 202614. nov. 2026From the labIz laboratorija
DIMACS E3M.03,162,2783.162.278estimated Held–Karp boundocenjena meja Held–Karp1% above the estimated Held–Karp bound after 4 min 43 s; 0.743% after 1 h1 % nad ocenjeno mejo Held–Karp po 4 min 43 s; 0,743 % po 1 h714 Nov 202614. nov. 2026From the labIz laboratorija
DIMACS E10M.010,000,00010.000.000estimated Held–Karp boundocenjena meja Held–Karp1% above the estimated Held–Karp bound after 15 min 20 s; 0.745% after 2 h1 % nad ocenjeno mejo Held–Karp po 15 min 20 s; 0,745 % po 2 h713–14 Nov 202613.–14. nov. 2026From the labIz laboratorija

One laptop (Intel i7-10875H, 8 cores), development builds of 8zTSP; times run from the start of the program and include reading the map. The 1% line is 1% above the reference. Each row links to its full conditions. Our tours, to check for yourself:En prenosnik (Intel i7-10875H, 8 jeder), razvojne gradnje 8zTSP; časi tečejo od zagona programa in vključujejo branje zemljevida. Črta 1 % je 1 % nad referenco. Vsaka vrstica vodi do svojih polnih pogojev. Naše poti za lastno preverjanje: DownloadsPrenosi.

Not by chance.Ni naključje. 8zTSP grows from BD’s ideas for the TSP, our own view of the map and the addition to it that a Claude session prepared by BD made from it, with his MDL judge (Rank, Don’t Eliminate) and his DCC governor already in the kernel as switches that we are measuring, and from BD’s way of working with LLMs: the8zTSP raste iz BD‑jevih zamisli za TSP, iz našega lastnega pogleda na zemljevid in iz dodatka k njemu, ki ga je iz njega naredila seja Clauda, ki jo je pripravil BD; njegov sodnik MDL (Rank, Don’t Eliminate) in njegov krmilnik DCC sta v jedru že kot stikali, ki ju merimo. Raste tudi iz BD‑jevega načina dela z LLM-ji: iz Mental Arena Lab, behind his earlier TSP arenas, the, ki je stal za njegovimi prejšnjimi arenami TSP, iz RHP protocolprotokola RHP and AI8, the team of AI sessions that built and measured this kernel in four days (11–14 Nov 2026). Inside it run known methods such as the edge assembly crossover of Nagata and Kobayashi.in iz AI8, ekipe sej AI, ki je to jedro zgradila in izmerila v štirih dneh (11.–14. nov. 2026). V njem tečejo tudi znane metode, na primer križanje s sestavljanjem robov (Nagata in Kobayashi).

Small and self-contained:Majhno in samostojno: the 8zTSP kernel has about half the code of LKH 3.0.7 and about a tenth of Concorde, which also needs a separate LP solver library; it uses only Rust’s standard library. The minimal build 8zTSP-min, which finds the nu3496 optimum just as fast as the full build, has about a quarter of LKH’s code (jedro 8zTSP ima okoli pol manj kode kot LKH 3.0.7 in okoli desetino Concorda, ki potrebuje še ločeno knjižnico za linearno programiranje; uporablja le standardno knjižnico Rusta. Minimalna gradnja 8zTSP-min, ki optimum nu3496 najde enako hitro kot polna gradnja, ima okoli četrtino kode LKH (from the lab, 14 Noviz laboratorija, 14. nov.). (Counted in characters of code, without spaces and comments. LKH-3 also covers many TSP variants, and Concorde proves optimality.)). (Šteti so znaki kode brez presledkov in komentarjev. LKH-3 pokriva tudi veliko različic TSP, Concorde pa optimum dokaže.)

More on the size of the codeVeč o velikosti kode
Size of the codeVelikost kodeCharacters of codeZnaki kodeNeeds besidesPotrebuje še
8zTSP, build 303eee68zTSP, gradnja 303eee6206,972206.972nothing: Rust’s standard library onlynič: le standardna knjižnica Rusta
8zTSP kernel after shortening (main 617dbe3, 14 Nov)Jedro 8zTSP po skrajšanju (main 617dbe3, 14. nov.)194,444194.444nothing: Rust’s standard library only; the same results on 155 test runs and the same speednič: le standardna knjižnica Rusta; isti izidi na 155 preizkusnih tekih in enaka hitrost
LKH 3.0.7403,521403.521the C standard library onlyle standardna knjižnica C
Concorde 03.12.192,217,9762.217.976an LP solver: QSopt (a closed binary library) or CPLEXreševalec LP: QSopt (zaprta binarna knjižnica) ali CPLEX

Characters of code: non-space characters after removing comments, counted with the same small script for all three (8zTSP src/*.rs at build 303eee6, LKH SRC/*.c and SRC/INCLUDE/*.h, Concorde all *.c and *.h). A count of characters does not depend on how densely the code is formatted, as a count of lines would. The arena that runs and checks all three (Python) is common to them and part of none. LKH-3 also covers many TSP variants (for example vehicle routing), and Concorde proves optimality.Znaki kode: znaki, ki niso presledki, po odstranitvi komentarjev, prešteti z istim majhnim skriptom za vse tri (8zTSP src/*.rs pri gradnji 303eee6, LKH SRC/*.c in SRC/INCLUDE/*.h, Concorde vse *.c in *.h). Število znakov ni odvisno od tega, kako gosto je koda zapisana, število vrstic pa bi bilo. Arena, ki vse tri poganja in preverja (Python), je skupna vsem in ni del nobenega. LKH-3 pokriva tudi veliko različic TSP (na primer usmerjanje vozil), Concorde pa optimum dokaže.

How the race was run:Kako je potekala dirka: DIMACS E10M.0, random uniform · one 8-core laptop, one program at a time · 8zTSP: 7 threads in one search · linkern: best of 4 one-core runs · LKH 3.0.7: best of 2, because one run needs ~6.9 GB; POPMUSIC settings, its subproblem settings were not run at 10M · 4 runs were the rivals’ fastest way to 1% among 4, 7 and 14 (tested at 1M) · one development run each, 12 Nov 2026; 8zTSP’s new build, 14 Nov.DIMACS E10M.0, naključno enakomerno · en prenosnik z 8 jedri, en program naenkrat · 8zTSP: 7 niti v enem iskanju · linkern: najboljši od 4 enojedrnih tekov · LKH 3.0.7: najboljši od 2, ker en tek potrebuje ~6,9 GB; nastavitve POPMUSIC, njegove nastavitve s podproblemi pri 10M niso tekle · 4 teki so bili tekmecema najhitrejša pot do 1 % med 4, 7 in 14 (preizkušeno pri 1M) · en razvojni tek na udeleženca, 12. nov. 2026; nova gradnja 8zTSP, 14. nov.

How the race was runKako je potekala dirka

DIMACS E10M.0, random uniform · one 8-core laptop, one program at a time · 8zTSP: 7 threads in one search · linkern: best of 4 one-core runs · LKH 3.0.7: best of 2, because one run needs ~6.9 GB; POPMUSIC settings, its subproblem settings were not run at 10M · 4 runs were the rivals’ fastest way to 1% among 4, 7 and 14 (tested at 1M) · one development run each, 12 Nov 2026; 8zTSP’s new build, 14 Nov.DIMACS E10M.0, naključno enakomerno · en prenosnik z 8 jedri, en program naenkrat · 8zTSP: 7 niti v enem iskanju · linkern: najboljši od 4 enojedrnih tekov · LKH 3.0.7: najboljši od 2, ker en tek potrebuje ~6,9 GB; nastavitve POPMUSIC, njegove nastavitve s podproblemi pri 10M niso tekle · 4 teki so bili tekmecema najhitrejša pot do 1 % med 4, 7 in 14 (preizkušeno pri 1M) · en razvojni tek na udeleženca, 12. nov. 2026; nova gradnja 8zTSP, 14. nov.

10,000,000 cities10.000.000 mest DIMACS E10M.0 · random uniformDIMACS E10M.0 · naključno enakomerno

Finish line:Ciljna črta: a route at most 1% above the reference 0.7080·√(N·A), an estimate of the Held–Karp lower bound for N random uniform cities in a square of area A (here a 10pot največ 1 % nad referenco 0,7080·√(N·A), oceno spodnje meje Held–Karp za N naključno enakomerno razporejenih mest v kvadratu s ploščino A (tu je to kvadrat 106 × 106 square), not a proven bound. For such maps the best possible route lies about 0.6–0.7% above the bound.enot), ne dokazano mejo. Na takih zemljevidih leži najboljša možna pot približno 0,6–0,7 % nad mejo.

  1. 8zTSPnew build 303eee6 · 7 threads, one search · 14 Novnova gradnja 303eee6 · 7 niti, eno iskanje · 14. nov.

    ✓ At the 1% line after 920 s.✓ Na črti 1 % po 920 s. Kept improving: 0.745% above the reference at 2 h; every counted route is checked. New build 303eee6, 14 Nov, alone on the laptop, no cooldown wait.Še naprej se je izboljševal: 0,745 % nad referenco po 2 h; vsaka šteta pot je preverjena. Nova gradnja 303eee6, 14. nov., sama na prenosniku, brez čakanja na ohladitev. Earlier builds at 10M: 0.6.0-z1, 17 min 10 s (pilot of 12 Nov, with development runs beside it); 0.5.0, 1 h 18 min (11–12 Nov, on about one core through a fault since fixed).Prejšnji gradnji pri 10M: 0.6.0-z1, 17 min 10 s (pilot 12. nov., ob njem razvojni teki); 0.5.0, 1 h 18 min (11.–12. nov., zaradi zdaj odpravljene napake na približno enem jedru).

  2. Concorde linkernConcordov linkernbest of 4 runs, one core eachnajboljši od 4 tekov, vsak na enem jedru

    ■ Never within 1% in 2 h.■ V 2 h nikoli znotraj 1 %. Best 1.162% above the reference; it stopped by itself at 7,033 s.Najboljša pot: 1,162 % nad referenco; ustavil se je sam po 7.033 s.

  3. LKH 3.0.7best of 2 runs (one needs ~6.9 GB) · POPMUSIC candidates, no subgradient ascentnajboljši od 2 tekov (eden potrebuje ~6,9 GB) · kandidati POPMUSIC, brez subgradientnega vzpona

    ✕ No route reported in 2 h.✕ V 2 h ni sporočil nobene poti. Preparing the lower bound and the candidates took 4,945 s in one run and 4,976 s in the other; at 7,200 s neither had finished its first trial. With other settings (e.g. geometric with subproblems) LKH might give a route sooner; not measured at 10M. Alone on one core (Priprava spodnje meje in kandidatov je trajala 4.945 s v enem teku in 4.976 s v drugem; ob 7.200 s nobeden ni končal svojega prvega poskusa. Z drugimi nastavitvami (na primer geometrijskimi s podproblemi) bi LKH morda dal pot prej; pri 10M to ni izmerjeno. Sam na enem jedru (fair test of 13 Novpošteni preizkus 13. novembra), its first route came after 1 h 48 min 39 s, already below the 1% line (0.960% above the reference); preparation took 3,890 s instead of 4,945–4,976 s.) je svojo prvo pot dal po 1 h 48 min 39 s, že pod črto 1 % (0,960 % nad referenco); priprava je trajala 3.890 s namesto 4.945–4.976 s.

Read with the numbersKaj povedo številke

Where the lead comes from, and where it stops.Od kod prihaja prednost in kje se konča.

At 1M: faster to a good route, not to a shorter final onePri 1M: hitreje do dobre poti, ne do krajše končne

At 1,000,000 cities LKH with its geometric settings (6 independent runs, 6 cores, 11 Nov) had its first route after 15.8 min, and after 30 min it was level: 0.810% above the bound, against 0.815% for 8zTSP 0.5.0 (6 threads, 11 Nov) and 0.810% for the hybrid (7 threads, 12 Nov).Pri 1.000.000 mestih je imel LKH z geometrijskimi nastavitvami (6 neodvisnih tekov, 6 jeder, 11. nov.) prvo pot po 15,8 min, po 30 min pa je bil izenačen: 0,810 % nad mejo, proti 0,815 % za 8zTSP 0.5.0 (6 niti, 11. nov.) in 0,810 % za hibrid (7 niti, 12. nov.).

The lead comes from our own view of the map: 7 cores on one mapPrednost daje naš lastni pogled na zemljevid: 7 jeder na enem zemljevidu

8zTSP’s own view of the map (BD’s idea) lets 7 of the laptop’s 8 cores improve one route at once, on parts of about 1,000 cities; LKH and linkern use one core per run. On one core 8zTSP was ahead too.Lastni pogled 8zTSP na zemljevid (BD‑jeva zamisel) omogoča, da 7 od 8 jeder prenosnika hkrati izboljšuje eno pot, na delih po približno 1.000 mest; LKH in linkern uporabljata eno jedro na tek. Tudi na enem jedru je bil 8zTSP spredaj.

More on the one-core testsVeč o preizkusih na enem jedru

In a fair test at 10,000,000 cities, each program alone on the laptop on the same core, the new 8zTSP reached the 1% line after 56 min 4 s, LKH after 1 h 48 min 39 s, and linkern did not reach it in 2 h (V poštenem preizkusu pri 10.000.000 mestih, vsak program sam na prenosniku na istem jedru, je novi 8zTSP črto 1 % dosegel po 56 min 4 s, LKH po 1 h 48 min 39 s, linkern pa je v 2 h ni dosegel (From the labIz laboratorija). In the same test at 3,162,278 cities, 1 h each, the new 8zTSP reached it after 19 min 14 s, LKH after 30 min 52 s, and linkern after 35 min 4 s (). V istem preizkusu pri 3.162.278 mestih, po 1 h vsak, jo je novi 8zTSP dosegel po 19 min 14 s, LKH po 30 min 52 s, linkern pa po 35 min 4 s (From the labIz laboratorija). In the same test at 1,000,000 cities, 30 min each, the new 8zTSP reached it after 5 min 33 s, LKH after 8 min 9 s, and linkern after 5 min 51 s (). V istem preizkusu pri 1.000.000 mestih, po 30 min vsak, jo je novi 8zTSP dosegel po 5 min 33 s, LKH po 8 min 9 s, linkern pa po 5 min 51 s (From the labIz laboratorija). What runs inside our analysis of the view matters less: with LKH itself inside it (the hybrid) the times were about the same (1M: 2 min 23 s against 2 min 3 s; 3.16M: 6 min 27 s against 6 min 16 s).). Kaj teče v naši analizi pogleda, je manj pomembno: s samim LKH v njej (hibrid) so bili časi približno enaki (1M: 2 min 23 s proti 2 min 3 s; 3,16M: 6 min 27 s proti 6 min 16 s).

10,000,000, 3,162,278 and 1,000,000 cities: the fair one-core tests of 13 and 14 Nov, one run each, each program alone on the laptop on logical CPU 3. Earlier at 3,162,278 cities: 8zTSP 0.5.0 on one thread 43 min 33 s (night of 11–12 Nov, sharing the laptop with other runs); linkern and LKH as the best of 4 one-core runs in the pilot of 12 Nov (other work at the lowest priority on the eighth core), 49 min 33 s and 52 min 59 s.10.000.000, 3.162.278 in 1.000.000 mest: pošteni enojedrni preizkusi 13. in 14. novembra, po en tek, vsak program sam na prenosniku na logičnem CPU 3. Prej pri 3.162.278 mestih: 8zTSP 0.5.0 na eni niti 43 min 33 s (noč z 11. na 12. november, prenosnik si je delil z drugimi teki); linkern in LKH kot najboljša od 4 enojedrnih tekov v pilotu 12. novembra (ob njih je na osmem jedru z najnižjo prednostjo teklo drugo delo), 49 min 33 s in 52 min 59 s.

At 10,000 and 24,978 cities: LKH is still aheadPri 10.000 in 24.978 mestih: LKH je še spredaj

At these sizes LKH with geometric settings leads. Sweden (sw24978), 10 min, 6 cores: LKH 0.008–0.015% above the optimum, 8zTSP 0.016–0.029%.Pri teh velikostih vodi LKH z geometrijskimi nastavitvami. Švedska (sw24978), 10 min, 6 jeder: LKH 0,008–0,015 % nad optimumom, 8zTSP 0,016–0,029 %.

Provisional: one run eachZačasno: en tek na udeleženca

One development run per entry, not the formal test. A separate GPT session acting as reviewer judged on 12 Nov that a public “X times faster” sentence is not yet earned. Next: the full test with several seeds, instances never used for tuning, an independent review and BD’s yes.En razvojni tek na udeleženca, ne formalni preizkus. Ločena seja GPT v vlogi pregledovalca je 12. nov. presodila, da javni stavek »X-krat hitreje« še ni zaslužen. Naslednje: polni preizkus z več semeni, s primeri, ki niso bili nikoli uporabljeni za uglaševanje, z neodvisnim pregledom in z BD‑jevim »da«.

Where the engine of 8zTSP can helpKje lahko pomaga motor 8zTSP

Ten million cities is a load test, not a customer: it shows that the engine does not break under mass. What can travel to other fields is the way it spends its work: our own view of the map, our own recursive analysis of the view, and few, well-chosen steps. Below is where that could help. Each claim in a new field needs one measurement there first; none is measured yet.Deset milijonov mest je obremenilni preizkus, ne naročnik: pokaže, da se motor pod množico ne zlomi. V druga področja se lahko prenese način, kako porablja svoje delo: naš lastni pogled na zemljevid, naša lastna rekurzivna analiza pogleda in malo dobro izbranih korakov. Spodaj je, kje bi to lahko pomagalo. Vsaka trditev na novem področju potrebuje tam najprej eno meritev; nobena še ni izmerjena.

Chip geometry and electronicsGeometrija čipov in elektronika

Points on chips, test chains, drilling and laser paths on boards are true travelling-salesman problems; public chip-based instances go up to about 745,000 points.Točke na čipih, testne verige ter poti vrtanja in laserja na tiskanih vezjih so pravi problemi trgovskega potnika; javni primeri iz čipov segajo do okoli 745.000 točk.

Our own view of the map and our analysis of it, with EAX and our addition, run as they are. Next measurement: per cent above the best known tour after 1, 5 and 15 minutes on the public large instances, beside LKH in geometric settings, with parts chosen along the tour as the null control.Naš lastni pogled na zemljevid in naša analiza pogleda, z EAX in našim dodatkom, tečejo, kakršni so. Naslednja meritev: odstotek nad najboljšo znano turo po 1, 5 in 15 minutah na javnih velikih primerih, ob LKH z geometrijskimi nastavitvami, z deli, izbranimi vzdolž ture, kot ničelno kontrolo.

Ordering big dataUrejanje velikih podatkov

DNA reads, table rows and the documents used to train language models are easier to store and use when similar items sit next to each other: a relative of the TSP (a Hamiltonian path).Odčitke DNK, vrstice tabel in dokumente za učenje jezikovnih modelov je laže shraniti in uporabljati, če so podobni elementi drug ob drugem: sorodnik TSP (Hamiltonova pot).

Our analysis of the view can repair a long initial order piece by piece; the MDL judge counts bits; where the best results of our addition disagree, the order is not determined. Next measurement: bits per record against plain sorting, on real data.Naša analiza pogleda lahko dolg začetni vrstni red popravlja kos za kosom; sodnik MDL šteje bite; kjer se najboljši izidi našega dodatka ne ujemajo, vrstni red ni določen. Naslednja meritev: biti na zapis proti navadnemu razvrščanju, na resničnih podatkih.

Point clouds and star cataloguesOblaki točk in katalogi zvezd

Not a TSP: here our own view of the map itself is the product, as the way to index and compress millions of points.To ni TSP: tu je izdelek sam naš lastni pogled na zemljevid, kot način za indeksiranje in stiskanje milijonov točk.

More on point clouds and star cataloguesVeč o oblakih točk in katalogih zvezd

The sky is indexed with the same trick (HEALPix: a cell’s parent is its index shifted by two bits), and point clouds are coded with octrees. Next measurement: bits per point against the best known coding and the theoretical minimum, on uniform, clustered and real points. This links to BD’s 8zC compression project (Nebo se indeksira z istim trikom (HEALPix: starš celice je njen indeks, zamaknjen za dva bita), oblaki točk pa se kodirajo z osmiškimi drevesi (octree). Naslednja meritev: biti na točko proti najboljšemu znanemu kodiranju in teoretičnemu minimumu, na enakomernih, gručastih in resničnih točkah. To se povezuje z BD‑jevim projektom stiskanja 8zC (8Z compressionstiskanje 8Z).

Tours from star to starTure od zvezde do zvezde

A true TSP in three dimensions: public tours through 2 million, 10 million and 1.33 billion stars of the Gaia catalogue exist (Bill Cook and Keld Helsgaun, Waterloo).Pravi TSP v treh razsežnostih: obstajajo javne ture skozi 2 milijona, 10 milijonov in 1,33 milijarde zvezd kataloga Gaia (Bill Cook in Keld Helsgaun, Waterloo).

These are benchmarks rather than orders from a customer; 8zTSP would need its 3D version. Next measurement: per cent above the published tours after 15 minutes and 1 hour.To so merila, ne naročila strank; 8zTSP bi potreboval svojo različico za 3D. Naslednja meritev: odstotek nad objavljenimi turami po 15 minutah in 1 uri.

Telescopes and space missionsTeleskopi in vesoljske misije

Yes, observatories solve this: a 2023 study called it the travelling telescope problem and cut a Keck night’s slewing time several-fold by ordering the targets well.Da, observatoriji to rešujejo: študija iz leta 2023 ga je imenovala problem potujočega teleskopa in z dobrim vrstnim redom ciljev večkrat skrajšala čas obračanja teleskopa Keck v eni noči.

There it is tens to hundreds of targets with time slots (visibility, Moon, weather), so our ten-million scale does not help; BD’s MDL judge in choosing targets might. Similar: starshade missions, space-debris removal, multi-asteroid tours.Tam gre za desetine do stotine ciljev s časovnimi termini (vidnost, Luna, vreme), zato naša lestvica desetih milijonov ne pomaga; pri izbiri ciljev pa bi morda pomagal BD‑jev sodnik MDL. Podobno: misije z zaslonom za zvezde (starshade), odstranjevanje vesoljskih odpadkov, obiski več asteroidov.

BiologyBiologija

Genetic and radiation-hybrid maps order markers as a TSP on distances, and Concorde has been better there than specialised tools.Genetski zemljevidi in zemljevidi radiacijskih hibridov urejajo označevalce kot TSP na razdaljah, Concorde pa je bil tam boljši od specializiranih orodij.

Cells along the cell cycle are ordered the same way. A new output of ours: the disagreement among the best results of our addition as a map of where the order is reliable. Next measurement: agreement with the true order on simulated maps.Na enak način se urejajo celice vzdolž celičnega cikla. Nov izid pri nas: neujemanje najboljših izidov našega dodatka kot zemljevid, kje je vrstni red zanesljiv. Naslednja meritev: ujemanje s pravim vrstnim redom na simuliranih zemljevidih.

LogisticsLogistika

A big tour cut into vehicle routes (capacities, time slots) is the TSP’s closest business relative.Velika tura, razrezana na poti vozil (zmogljivosti, časovni termini), je najbližji poslovni sorodnik TSP.

EAX for vehicle routing exists in the literature; our analysis of the view would need pinned edges that carry the load. New code; the machine and the project are BD’s call.EAX za usmerjanje vozil v literaturi obstaja; naša analiza pogleda bi potrebovala pripete robove, ki nosijo tovor. To je nova koda; stroj in projekt sta BD‑jeva odločitev.

Materials and moleculesMateriali in molekule

Not shortest paths but the deepest energy valley, yet the search machine is a sibling of our kernel, part for part.Ne najkrajše poti, temveč najgloblja energijska dolina, a iskalni stroj je sorodnik našega jedra, del za delom.

More on materials and moleculesVeč o materialih in molekulah

The kick with repair is “basin hopping”; the memory of our addition’s best results is like a bank of minima; DCC is like the feedback loop of “minima hopping”; EAX is like “cut-and-splice” crossover. For BD’s 8zMaterials: a map of agreement and disagreement across the valleys already found, without one new energy calculation. A true TSP in the materials lab: the order of measurements on an X-ray diffractometer (Bland and Shallcross, 1989).Sunek s popravilom je »basin hopping«; spomin najboljših izidov našega dodatka je kot banka minimumov; DCC je kot povratna zanka »minima hopping«; EAX je kot križanje »cut-and-splice«. Za BD‑jev 8zMaterials: zemljevid strinjanja in nestrinjanja med že najdenimi dolinami, brez enega samega novega izračuna energije. Pravi TSP v laboratoriju za materiale: vrstni red meritev na rentgenskem difraktometru (Bland in Shallcross, 1989).

Other combinatorial problemsDrugi kombinatorični problemi

Integer programming, satisfiability and scheduling already use cousins of our parts: our addition is close to RINS and “backbones”, EAX to partition crossover, the MDL judge to adaptive weights.Celoštevilsko programiranje, izpolnjivost in razporejanje že uporabljajo sorodnike naših delov: naš dodatek je blizu RINS in »hrbtenicam« (backbones), EAX križanju z delitvijo (partition crossover), sodnik MDL prilagodljivim utežem.

Each transfer needs its own null control there.Vsak prenos potrebuje tam svojo ničelno kontrolo.

ProteinsBeljakovine

Protein folding is not a TSP; today it is predicted by models learned from solved structures, a direction our addition also takes. Our engine does not fold proteins.Zvijanje beljakovin ni TSP; danes ga napovedujejo modeli, naučeni iz rešenih struktur, kar je smer, ki jo ubira tudi naš dodatek. Naš motor beljakovin ne zvija.

The HP lattice model (the folded chain is a path on a regular lattice; NP-complete; published lowest energies) is a fair testbed for our analysis of the view, our addition and DCC.Model HP na rešetki (zvita veriga je pot na pravilni rešetki; NP-poln; objavljene najnižje energije) je pošteno preizkusno polje za našo analizo pogleda, naš dodatek in DCC.

Next three experimentsNaslednji trije poskusi

  1. Does our addition know where the errors are? 11 instances with known optimal tours; success criteria written down before.Ali naš dodatek ve, kje so napake? 11 primerov z znanimi optimalnimi turami; merila uspeha zapisana vnaprej.

  2. Does DCC read the order of improvements or only their frequency? On, Null, best fixed level, observer; 9 unused instances.Ali DCC bere vrstni red izboljšav ali le njihovo pogostost? Vklopljen, ničelni, najboljša stalna raven, opazovalec; 9 neuporabljenih primerov.

  3. Chip geometry: the public large instances, beside LKH in geometric settings.Geometrija čipov: javni veliki primeri, ob LKH z geometrijskimi nastavitvami.

Same laptop, same finish line: three map sizesIsti prenosnik, ista ciljna črta: tri velikosti zemljevida

Time until each entry first had a route at most 1% above the Held–Karp bound (1,000,000 cities: the published bound; 3,162,278 and 10,000,000: the estimate 0.7080·√(N·A), not a proven bound). Goal pilot of 12 November 2026: one development run per entry, one program at a time. At 10,000,000 and 3,162,278 cities 8zTSP is the new build of 14 November; earlier build: 0.6.0-z1, 17 min 10 s and 6 min 16 s in this pilot.Čas, v katerem je vsak udeleženec prvič imel pot največ 1 % nad mejo Held–Karp (1.000.000 mest: objavljena meja; 3.162.278 in 10.000.000: ocena 0,7080·√(N·A), ne dokazana meja). Pilot ciljnega preizkusa 12. novembra 2026: en razvojni tek na udeleženca, en program naenkrat. Pri 10.000.000 in 3.162.278 mestih je 8zTSP nova gradnja s 14. novembra; prejšnja gradnja: 0.6.0-z1, 17 min 10 s in 6 min 16 s v tem pilotu.

ProvisionalZačasno
Time to the 1% line on one laptop for 1,000,000, 3,162,278 and 10,000,000 cities; one development run per entry, 12 November 2026 (8zTSP at 10,000,000 and 3,162,278 cities: new build, 14 November).Čas do črte 1 % na enem prenosniku za 1.000.000, 3.162.278 in 10.000.000 mest; en razvojni tek na udeleženca, 12. novembra 2026 (8zTSP pri 10.000.000 in 3.162.278 mestih: nova gradnja, 14. novembra).
EntryVnosTime to the 1% lineČas do črte 1 %Elapsed time, 0 to 2 hPretečeni čas, od 0 do 2 h
1,000,000 cities1.000.000 mestE1M.0 · over the published Held–Karp bound 708,703,512.9E1M.0 · nad objavljeno mejo Held–Karp 708.703.512,9
8zTSP (large maps)8zTSP (veliki zemljevidi)0.6.0-z1 · 7 threads, one search0.6.0-z1 · 7 niti, eno iskanje2 min 3 s123 s
8zTSP (hybrid)8zTSP (hibrid)0.6.0-hyb1 · LKH 3.0.7 inside each part of the view · 7 threads0.6.0-hyb1 · LKH 3.0.7 v vsakem delu pogleda · 7 niti2 min 23 s143 s
Concorde linkernConcordov linkernbest of 4 runs, one core eachnajboljši od 4 tekov, vsak na enem jedru9 min 56 s596 s
LKH 3.0.7best of 4 runs · “popns” settingsnajboljši od 4 tekov · nastavitve »popns«13 min 31 s811 s · its first route811 s · njegova prva pot
3,162,278 cities3.162.278 mestE3M.0 · over the estimated reference 0.7080·√(N·A), not a proven boundE3M.0 · nad ocenjeno referenco 0,7080·√(N·A), ne dokazano mejo
8zTSP (large maps)8zTSP (veliki zemljevidi)new build 303eee6 · 7 threads · 14 Novnova gradnja 303eee6 · 7 niti · 14. nov.4 min 43 s283 s
8zTSP (hybrid)8zTSP (hibrid)0.6.0-hyb1 · LKH 3.0.7 inside each part of the view · 7 threads0.6.0-hyb1 · LKH 3.0.7 v vsakem delu pogleda · 7 niti6 min 27 s387 s
Concorde linkernConcordov linkernbest of 4 runs, one core eachnajboljši od 4 tekov, vsak na enem jedru49 min 33 s2,973 s2.973 s
LKH 3.0.7best of 4 runs · “popns” settingsnajboljši od 4 tekov · nastavitve »popns«52 min 59 s3,179 s · its first route3.179 s · njegova prva pot
10,000,000 cities10.000.000 mestE10M.0 · over the estimated reference 0.7080·√(N·A), not a proven boundE10M.0 · nad ocenjeno referenco 0,7080·√(N·A), ne dokazano mejo
8zTSP (large maps)8zTSP (veliki zemljevidi)new build 303eee6 · 7 threads · 14 Novnova gradnja 303eee6 · 7 niti · 14. nov.15 min 20 s920 s
8zTSP (hybrid)8zTSP (hibrid)0.6.0-hyb1 · LKH 3.0.7 inside each part of the view · 7 threads0.6.0-hyb1 · LKH 3.0.7 v vsakem delu pogleda · 7 nitinot runni tekelat 10Mpri 10M
Concorde linkernConcordov linkernbest of 4 runs, one core eachnajboljši od 4 tekov, vsak na enem jedrunot within 1%ne znotraj 1 %in 2 hv 2 h

Best 1.162% above the reference; it stopped by itself at 7,033 s.Najboljša pot: 1,162 % nad referenco; ustavil se je sam po 7.033 s.

LKH 3.0.7best of 2 runs (one needs ~6.9 GB) · “popns” settingsnajboljši od 2 tekov (eden potrebuje ~6,9 GB) · nastavitve »popns«✕ no route✕ ni sporočil potireported in 2 hv 2 h

Preparing the lower bound and the candidates took 4,945 s in one run and 4,976 s in the other; at 7,200 s neither had finished its first trial. Other settings (e.g. geometric with subproblems) might give a route sooner; not measured at 10M.Priprava spodnje meje in kandidatov je trajala 4.945 s v enem teku in 4.976 s v drugem; ob 7.200 s nobeden ni končal svojega prvega poskusa. Druge nastavitve (na primer geometrijske s podproblemi) bi morda dale pot prej; pri 10M to ni izmerjeno.

Conditions and notesPogoji in opombe

Conditions.Pogoji. One laptop (Intel i7-10875H: 8 cores, 16 threads, 32 GB); entries ran one after another, one program at a time on the main cores. In the pilot of 12 Nov other work at the lowest priority ran beside every entry on the eighth core (logical CPUs 7 and 15: 8zTSP development runs, until 14:39 the 8zC workshop), and during LKH at 3,162,278 cities a session used the graphics card twice, 35 s in all; the cooling is shared. The new build’s run at 10,000,000 cities on 14 Nov ran alone; its run at 3,162,278 cities on 14 Nov also ran alone on CPUs 0–6. From about 35 to 37 min into that run, a text search of the page session used one thread of the eighth core (logical CPUs 7 and 15); CPUs 0–6 were not shared. One development run per entry. 8zTSP and the hybrid run 7 threads in one search. LKH 3.0.7 and Concorde’s linkern are single-threaded: “best of k” means k independent runs, one core each; k = 4 was chosen in stage 0 on 1,000,000 cities as their fastest way to 1% among 4, 7 and 14 processes (k = 2 for LKH at 10M: one process needs ~6.9 GB). Alone, as a single process on one core, LKH reached the 1% line at 10M after 1 h 48 min 39 s in the fair test of 13 Nov (En prenosnik (Intel i7-10875H: 8 jeder, 16 niti, 32 GB); udeleženci so tekli drug za drugim, en program naenkrat na glavnih jedrih. V pilotu 12. novembra je ob vsakem udeležencu na osmem jedru (logična CPU 7 in 15) z najnižjo prednostjo teklo drugo delo (razvojni teki 8zTSP, do 14:39 delavnica 8zC), med LKH pri 3.162.278 mestih pa je ena seja dvakrat uporabila grafično kartico, skupaj 35 s; hlajenje je skupno. Tek nove gradnje pri 10.000.000 mestih 14. novembra je tekel sam; njen tek pri 3.162.278 mestih 14. novembra je na CPU 0–6 prav tako tekel sam. Med približno 35. in 37. minuto tega teka je iskanje po besedilu seje strani zasedlo eno nit osmega jedra (logična CPU 7 in 15); CPU 0–6 niso bili deljeni. En razvojni tek na udeleženca. 8zTSP in hibrid tečeta s 7 nitmi v enem iskanju. LKH 3.0.7 in Concordov linkern sta enonitna: »najboljši od k« pomeni k neodvisnih tekov, vsak na enem jedru; k = 4 je stopnja 0 pri 1.000.000 mestih izbrala kot njuno najhitrejšo pot do 1 % med 4, 7 in 14 procesi (k = 2 za LKH pri 10M: en proces potrebuje ~6,9 GB). Sam, kot en proces na enem jedru, je LKH pri 10M črto 1 % dosegel po 1 h 48 min 39 s v poštenem preizkusu 13. novembra (From the labIz laboratorija). LKH “popns” settings: POPMUSIC candidates, no subgradient ascent, POPMUSIC start tour. “Its first route”: LKH had no route before that time, and its first route was already within 1%. Every clock starts with its program, before the input file is read. For uniform instances like these, 1% above the bound is about 0.3–0.4% above the optimal tour. At 1,000,000 cities the estimate gives 708,000,000, 0.10% below the published bound, so the percentages over the estimate are, if anything, slightly high for every entry. Concorde’s exact solver would not finish at these sizes; “Concorde” here means its linkern heuristic. An earlier development run of 0.6.0-z1 at 10M, with the pilot’s settings, reached the line after 18.4 min (under Development results below).). Nastavitve LKH »popns«: kandidati POPMUSIC, brez subgradientnega vzpona, začetna tura POPMUSIC. »Njegova prva pot«: LKH pred tem časom ni imel nobene poti, njegova prva pot pa je bila že znotraj 1 %. Vsaka ura začne teči z zagonom svojega programa, preden ta prebere vhodno datoteko. Pri enakomernih primerih, kakršni so ti, je 1 % nad mejo približno 0,3–0,4 % nad optimalno turo. Pri 1.000.000 mestih da ocena 708.000.000, 0,10 % pod objavljeno mejo, zato so odstotki nad oceno, če že, za vse udeležence nekoliko previsoki. Concordov natančni reševalec pri teh velikostih ne bi končal; »Concorde« tu pomeni njegovo hevristiko linkern. Prejšnji razvojni tek 0.6.0-z1 pri 10M, z nastavitvami pilota, je črto dosegel po 18,4 min (spodaj pod Razvojni rezultati).

Where the speed comes from: our own view of the map puts the whole laptop on one mapOd kod hitrost: naš lastni pogled na zemljevid zaposli ves prenosnik na enem zemljevidu

LKH and linkern are single-threaded programs: more cores help them as independent runs, of which the best counts, or through a split of the map into subproblems, which is what our hybrid does with LKH inside each part of the view. 8zTSP’s own view of the map puts its threads into one search over one map.LKH in linkern sta enonitna programa: več jeder jima pomaga kot neodvisni teki, od katerih šteje najboljši, ali z razdelitvijo zemljevida na podprobleme, kar počne naš hibrid z LKH v vsakem delu pogleda. Lastni pogled 8zTSP na zemljevid vloži njegove niti v eno iskanje po enem zemljevidu.

LKH and linkern: one core per runLKH in linkern: eno jedro na tek

Each run works on the whole map on one core. With four cores there are four separate runs, and the best one counts.Vsak tek dela na celotnem zemljevidu na enem jedru. S štirimi jedri so to štirje ločeni teki in šteje najboljši.

1234

Four runs, each on the whole map and on its own core; only the best of them counts. Schematic.Štirje teki, vsak na celotnem zemljevidu in na svojem jedru; šteje le najboljši od njih. Shema.

Idea: BDZamisel: BD

Our own view of the mapNaš lastni pogled na zemljevid

Our own recursive analysis of the view: the search works on parts of about 1,000 cities, many parts in parallel on all cores.Naša lastna rekurzivna analiza pogleda: iskanje dela na delih s približno 1.000 mesti, z mnogimi deli vzporedno na vseh jedrih.

Claude C1, from BD’s idea, in a session BD preparedClaude C1, iz BD‑jeve zamisli, v seji, ki jo je pripravil BD

Our addition to the viewNaš dodatek k pogledu

An addition to our own view of the map, invented by a Claude instance, C1, on 22 March 2026, through two key influences of BD: he prepared that new chat session in a special way, giving it room and freedom away from the shackles of old TSP knowledge, and it worked from his idea. BD calls it the first Claude invention.Dodatek k našemu lastnemu pogledu na zemljevid, ki ga je 22. marca 2026 izumila instanca Clauda, C1, z dvema ključnima vplivoma BD‑ja: novo sejo je pripravil na poseben način, ji dal prostor in svobodo in jo odtujil od okov starega znanja o TSP, ona pa je delala iz njegove zamisli. BD ga imenuje prvi Claudov izum.

Measured on 13 Nov:Izmerjeno 13. nov.: shorter routes in 6 of 6 paired runskrajše poti v 6 od 6 parnih tekov.

Idea: BDZamisel: BD

MDL × DCC

Compression as a sensor of the search and a governor that steers it, here as a judge that schedules work among the parts of the view. First measured in the kick mode on nu3496: about three times closer to the optimum (Stiskanje kot tipalo iskanja in krmilnik, ki ga usmerja; tu kot sodnik, ki razporeja delo med dele pogleda. Prvič izmerjena v načinu sunkov na nu3496: okoli trikrat bližje optimumu (see the entryglej vnos). In our analysis of the view: being measured.). V naši analizi pogleda: merimo.

Ours and theirs combinedNaše in njihovo skupaj

The hybridHibrid

LKH 3.0.7 (Keld Helsgaun) inside each part of our view: 143 s to the 1% line at 1M and 387 s at 3.16M, close behind our own analysis alone (123 s and 376 s).LKH 3.0.7 (Keld Helsgaun) v vsakem delu našega pogleda: 143 s do črte 1 % pri 1M in 387 s pri 3,16M, tik za samo našo analizo (123 s in 376 s).

Our addition, measured on 13 November 2026 (development)Naš dodatek, izmerjen 13. novembra 2026 (razvoj)

In all 12 paired comparisons of 12–13 NovV vseh 12 parnih primerjavah 12. in 13. novembra (2 maps × 3 seeds, in the hybrid and(2 zemljevida × 3 semena, v hibridu in in our analysis of the view without LKHv naši analizi pogleda brez LKH),) our addition made the route shorterje naš dodatek skrajšal pot; the null control, run in the hybrid, gained nothing.; ničelna kontrola, pognana v hibridu, ni dala nobenega dobička.

The same hybrid with and without our addition, paired by seed: 3 seeds, 600 s per run, 7 threads, one development build. The dots show how much shorter the route was at 600 s than without our addition.Isti hibrid z našim dodatkom in brez njega, v parih po semenu: 3 semena, 600 s na tek, 7 niti, ena razvojna gradnja. Pike kažejo, za koliko je bila pot po 600 s krajša kot brez našega dodatka.

Clustered map:Gručasti zemljevid: 0.408, 0.397 and 0.353% shorter with our addition.z našim dodatkom 0,408, 0,397 in 0,353 % krajša pot. Uniform map:Enakomerni zemljevid: 0.109, 0.105 and 0.104% shorter; the gain shows from 60 s on, and the hybrid reached 1% above the bound after 71–76 s instead of 105–116 s.0,109, 0,105 in 0,104 % krajša; dobiček se pokaže od 60 s naprej, hibrid pa je 1 % nad mejo dosegel po 71–76 s namesto po 105–116 s. Null controlNičelna kontrola (the null control without our addition’s own part): from 0.002 to 0.085% longer, so the gain comes from our addition. Repeating the same command on the clustered map changed the result by 0.002–0.026%.(ničelna kontrola brez lastnega dela našega dodatka): od 0,002 do 0,085 % daljša, torej dobiček izvira iz našega dodatka. Ponovitev istega ukaza na gručastem zemljevidu je izid spremenila za 0,002–0,026 %.

The strength of our addition, beta 1, BD’s default, was the best of 0.5, 1 and 2 (clustered map): 0.5 kept about a quarter of the gain, 2 none. These are paired comparisons inside 8zTSP only; LKH and linkern were not run with this build, so they are not compared with it here.Moč našega dodatka, beta 1, BD‑jeva privzeta vrednost, je bila najboljša od 0,5, 1 in 2 (gručasti zemljevid): 0,5 je ohranila približno četrtino dobička, 2 nič. To so primerjave v parih samo znotraj 8zTSP; LKH in linkern s to gradnjo nista tekla, zato ju tu z njo ne primerjamo.

Without LKH inside our analysis of the view, our addition gives about the same gain:Tudi brez LKH v naši analizi pogleda da naš dodatek približno enak dobiček: From the lab, 13 Nov 2026Iz laboratorija, 13. nov. 2026.

From the labIz laboratorija

Results of development blocks as they finish, newest first, each with its conditions.Izidi razvojnih blokov, ko se končajo, najnovejši na vrhu, vsak s svojimi pogoji.

  1. · development run of 17 Nov; not yet in the published program· razvojni tek 17. novembra; še ni v objavljenem programu

    One billion cities on one laptopMilijarda mest na enem prenosniku

    8zTSP made a complete round trip through one billion cities on one laptop with 31 GB of memory: a complete tour after 2 min 15 s, and after one hour a tour 4.75% above the estimated lower bound, with a peak of 12.2 GB of memory, about 12 bytes per city.8zTSP je na enem prenosniku z 31 GB pomnilnika naredil celo krožno pot skozi milijardo mest: celo pot po 2 min 15 s in po eni uri pot 4,75 % nad ocenjeno spodnjo mejo, z največ 12,2 GB pomnilnika, okoli 12 bajtov na mesto. It used 8 threads, half of the laptop’s threads, while the other cores were busy with other work; a run on 6 of the laptop’s 8 cores comes later. The map has one billion random cities, the same as the official DIMACS generator makes.Uporabil je 8 niti, polovico niti prenosnika, medtem ko so bila druga jedra zasedena z drugim delom; tek na 6 od 8 jeder prenosnika pride pozneje. Zemljevid ima milijardo naključnih mest, enakih tistim, ki jih naredi uradni generator DIMACS.

    More on one billion citiesVeč o milijardi mest

    The map.Zemljevid. E1B.0: one billion cities, spread at random over a square of 1,000,000 by 1,000,000 units, as the generator of the DIMACS TSP Challenge makes them (its program portgen, with 1,000,000,000 cities and seed 1,000,000,000). We store it as an 8 GB binary file; written out as text, it has the same SHA-256 as the official generator’s own output. Nobody knows the optimum of such a map; the reference is the estimated lower bound 0.7080·√(N·A) = 22,388,925,834, an estimate for random maps, not a proved bound.E1B.0: milijarda mest, naključno razporejenih po kvadratu 1.000.000 krat 1.000.000 enot, kot jih naredi generator izziva DIMACS TSP (njegov program portgen s 1.000.000.000 mesti in semenom 1.000.000.000). Shranimo ga kot 8 GB veliko binarno datoteko; zapisan kot besedilo ima isti SHA-256 kot izhod uradnega generatorja. Optimuma takega zemljevida nihče ne pozna; referenca je ocenjena spodnja meja 0,7080·√(N·A) = 22.388.925.834, ocena za naključne zemljevide, ne dokazana meja.

    The hour.Ura. Reading the map took 16 s. A complete tour stood after 2 min 15 s, 38.8% above the estimated lower bound; then 33.4% after 5 minutes, 14.9% after 15, 6.8% after 30 and 4.75% after 60 minutes: 23,451,591,354. At the end the program checked the tour itself, every city exactly once and the length recomputed from the map, and wrote it as a 4 GB file. A second check with separate code, written apart from the solver, found the same: every city exactly once and the same length.Branje zemljevida je trajalo 16 s. Cela pot je stala po 2 min 15 s, 38,8 % nad ocenjeno spodnjo mejo; nato 33,4 % po 5 minutah, 14,9 % po 15, 6,8 % po 30 in 4,75 % po 60 minutah: 23.451.591.354. Na koncu je program pot sam preveril, vsako mesto natanko enkrat in dolžino preračunano iz zemljevida, in jo zapisal kot 4 GB veliko datoteko. Druga preverba z ločeno kodo, napisano neodvisno od reševalnika, je pokazala isto: vsako mesto natanko enkrat in enako dolžino.

    Memory.Pomnilnik. At most 12.2 GB, about 12 bytes per city, inside a 16 GB limit set for the run, without swap; the laptop has 31 GB.Največ 12,2 GB, okoli 12 bajtov na mesto, znotraj meje 16 GB, postavljene za tek, brez izmenjalnega prostora; prenosnik ima 31 GB.

    Next.Naslednji korak. The same map on 6 of the laptop’s 8 cores (12 threads), and a faster start. Measured by the time to a complete tour and the tour after fixed times.Isti zemljevid na 6 od 8 jeder prenosnika (12 niti) in hitrejši začetek. Merilo: čas do cele poti in pot po stalnih časih.

    Who did it.Kdo je to naredil. BD asked for 10 million, 100 million and 1 billion cities on one laptop. 8zTSP-DIM made the map and the way 8zTSP handles a billion cities, on a development branch; 8zTSP-K booked the hour on the laptop.BD je želel 10 milijonov, 100 milijonov in 1 milijardo mest na enem prenosniku. 8zTSP-DIM je naredil zemljevid in način, kako 8zTSP obdela milijardo mest, na razvojni veji; 8zTSP-K je rezerviral uro na prenosniku.

    Conditions: development run of 17 Nov 2026, 19:34–20:35, on the laptop’s Intel i7-10875H (8 cores, 16 threads, 31 GB): 8 threads on logical CPUs 1–4 and 9–12, that is 4 cores with both of their threads, while the other cores ran other work; seed 1, 3,600 s. Development build 48f259c on the branch feature/dim, not in the published program. Map E1B.0 (sha256 f6de6157…), 8,000,004,096 bytes; the tour file 4,000,004,096 bytes. The reference 0.7080·√(N·A) with A = 10¹² is an estimate for random uniform maps, not a proved bound and not the optimum. The memory is the peak resident set of the process; the run’s limit also counted the page cache of the map file.Pogoji: razvojni tek 17. novembra 2026, 19:34–20:35, na prenosnikovem Intel i7-10875H (8 jeder, 16 niti, 31 GB): 8 niti na logičnih CPU 1–4 in 9–12, torej 4 jedra z obema nitma, medtem ko so druga jedra poganjala drugo delo; seme 1, 3.600 s. Razvojna gradnja 48f259c na veji feature/dim, ni v objavljenem programu. Zemljevid E1B.0 (sha256 f6de6157…), 8.000.004.096 bajtov; datoteka poti 4.000.004.096 bajtov. Referenca 0,7080·√(N·A) z A = 10¹² je ocena za enakomerno naključne zemljevide, ni dokazana meja in ni optimum. Pomnilnik je največji rezidenčni obseg procesa; meja teka je štela tudi predpomnilnik strani datoteke zemljevida.

    ContentsKazalo

  2. · development runs of 16–17 Nov· razvojni teki 16. in 17. novembra

    nu3496, 7 cores each: 8zTSP at the optimum in about 6 s, Nagata’s GA-EAX in about 28 snu3496, po 7 jeder: 8zTSP do optimuma v približno 6 s, Nagatov GA-EAX v približno 28 s

    On nu3496’s 3,496 cities, with 7 of the laptop’s 8 cores each (8zTSP with 7 threads, Nagata’s GA-EAX as 7 copies, one per core), both reached the proven optimum, 96,132, in all 10 runs: 8zTSP in a median of about 6 s, GA-EAX in about 28 s, about 4.5 times as long.Na 3.496 mestih zemljevida nu3496 sta oba, vsak s 7 od 8 jeder prenosnika (8zTSP s 7 nitmi, Nagatov GA-EAX kot 7 kopij, po ena na jedro), v vseh 10 tekih dosegla dokazani optimum, 96.132: 8zTSP v mediani v približno 6 s, GA-EAX v približno 28 s, torej okoli 4,5-krat dlje. 8zTSP was first in 9 of 10 runs and used about 48 MB of memory, the 7 copies of GA-EAX together about 1.0 GB. Each at its best on this laptop: 8zTSP about 5 s with all 14 threads, GA-EAX about 28 s as 7 copies (as 14 copies it took about 39 s), about 5 times as long.8zTSP je bil prvi v 9 od 10 tekov in je porabil okoli 48 MB pomnilnika, 7 kopij GA-EAX skupaj okoli 1,0 GB. Vsak v svoji najboljši nastavitvi na tem prenosniku: 8zTSP okoli 5 s z vsemi 14 nitmi, GA-EAX okoli 28 s kot 7 kopij (kot 14 kopij je potreboval okoli 39 s), torej okoli 5-krat dlje.

    More on nu3496 against GA-EAXVeč o nu3496 proti GA-EAX

    nu3496, 10 runs per setting: the median time to the optimum, 96,132; »first« counts the runs in which 8zTSP reached it before GA-EAX; memory is the peak of all processes, in decimal units.nu3496, 10 tekov na nastavitev: mediana časa do optimuma, 96.132; »prvi« šteje teke, v katerih ga je 8zTSP dosegel pred GA-EAX; pomnilnik je največja poraba vseh procesov, v desetiških enotah.

    SettingNastavitev8zTSP (median)8zTSP (mediana)Nagata’s GA-EAX (median)Nagatov GA-EAX (mediana)8zTSP first8zTSP prviMemoryPomnilnik
    7 cores each, 17 Novpo 7 jeder, 17. nov.6.3 s (7 threads)6,3 s (7 niti)28.0 s (7 copies)28,0 s (7 kopij)9 of 109 od 1048 MB vs 1.0 GB48 MB proti 1,0 GB
    14 threads each, 17 Novpo 14 niti, 17. nov.5.2 s (14 threads)5,2 s (14 niti)38.7 s (14 copies)38,7 s (14 kopij)10 of 1010 od 1056 MB vs 2.0 GB56 MB proti 2,0 GB
    4 cores each, 16 Novpo 4 jedra, 16. nov.11.3 s (4 threads)11,3 s (4 niti)25.3 s (4 copies)25,3 s (4 kopije)8 of 108 od 1045 MB vs 0.58 GB45 MB proti 0,58 GB

    How to read it.Kako brati. Each program had the same cores in each run. 8zTSP works on the map as one search with all its threads; GA-EAX, a single-threaded program, ran as one copy per core, each restarting with a new seed, and the first copy at the optimum counted. Both found the optimum in every run, so the table compares the time to it. GA-EAX was quicker with 7 copies than with 14 (28.0 s against 38.7 s), because two copies then share a core; 8zTSP was quicker with 14 threads (5.2 s against 6.3 s). So each program’s best setting: 8zTSP about 5 s, GA-EAX about 28 s.Vsak program je imel v vsakem teku ista jedra. 8zTSP dela na zemljevidu kot eno iskanje z vsemi svojimi nitmi; GA-EAX, enoniten program, je tekel kot ena kopija na jedro, vsaka se je znova zagnala z novim semenom, štela pa je prva kopija na optimumu. Oba sta optimum našla v vsakem teku, zato tabela primerja čas do njega. GA-EAX je bil s 7 kopijami hitrejši kot s 14 (28,0 s proti 38,7 s), ker si tedaj po dve kopiji delita jedro; 8zTSP je bil hitrejši s 14 nitmi (5,2 s proti 6,3 s). Najboljša nastavitev vsakega programa je torej: 8zTSP okoli 5 s, GA-EAX okoli 28 s.

    The spread.Razpon. With 7 threads, 8zTSP’s 10 runs took 5.7 to 49.4 s, seven of them under 6.5 s; GA-EAX’s took 23.4 to 72.1 s. The one run in which GA-EAX was first (30.8 s against 49.4 s) is 8zTSP’s slowest.S 7 nitmi so 8zTSP-jevi teki trajali od 5,7 do 49,4 s, sedem od 10 manj kot 6,5 s; teki GA-EAX od 23,4 do 72,1 s. Edini tek, v katerem je bil GA-EAX prvi (30,8 s proti 49,4 s), je 8zTSP-jev najpočasnejši.

    Who did it.Kdo je to naredil. 8zTSP-K planned and ran the runs in the arena; every median and memory figure here was recomputed from the runs’ results.8zTSP-K je teke načrtoval in pognal v areni; vsaka mediana in vsak podatek o pomnilniku sta ponovno izračunana iz rezultatov tekov.

    Conditions: development runs of 16–17 Nov 2026 on the laptop’s Intel i7-10875H; the arena’s driver full_machine_nu.py, nu3496, 10 runs per setting, a cap of 300 s, the two programs’ order alternating from run to run, a cooldown below 80 °C or at most 120 s. 8zTSP build 3c36abdd (main 1f072ad), stopped from outside at 96,132. Nagata’s GA-EAX: his public code (015dfbe, MIT) with the arena’s patches (seed, target stop, CPU cap, stop file), Npop 100, Nch 30. Logical CPUs: 4 cores = 1–4; 7 cores = 0–5 and 7; 14 threads = the same 7 cores with both threads (CPU 6 and its twin 14 ran Concorde’s proof for Slovenia).Pogoji: razvojni teki 16. in 17. novembra 2026 na prenosnikovem Intel i7-10875H; gonilnik arene full_machine_nu.py, nu3496, 10 tekov na nastavitev, zgornja meja 300 s, vrstni red programov se je od teka do teka menjal, ohlajanje pod 80 °C ali največ 120 s. 8zTSP, gradnja 3c36abdd (main 1f072ad), ustavljen od zunaj pri 96.132. Nagatov GA-EAX: njegova javna koda (015dfbe, MIT) s popravki arene (seme, ustavitev pri cilju, omejitev CPU, datoteka za ustavitev), Npop 100, Nch 30. Logični CPU: 4 jedra = 1–4; 7 jeder = 0–5 in 7; 14 niti = istih 7 jeder z obema nitma (CPU 6 in njegov dvojček 14 sta poganjala Concordov dokaz za Slovenijo).

    ContentsKazalo

  3. · development runs of 17 and 18 Nov· razvojni teki 17. in 18. novembra

    Switzerland, a new mapŠvica, nov zemljevid

    Switzerland is our new map of 2,850 places, cut from our map of Europe. Concorde proved its optimum, 7,844,951 m, on our laptop and on the NEOS Server, and 8zTSP’s newest build found that tour.Švica je naš nov zemljevid z 2.850 kraji, izrezan iz našega zemljevida Evrope. Concorde je dokazal njen optimum, 7.844.951 m, na našem prenosniku in na strežniku NEOS, najnovejša gradnja 8zTSP pa je to pot našla. On 6 of the laptop’s 8 cores (12 threads), 8zTSP’s candidate build finds that tour in 5.6 s (median of 10 runs; the slowest took 6.0 s), and the optimum of nu3496 in 5.8 s (median of 10 runs). The optimum is at most 0.065% above our own lower bound, 7,839,923 m (18 Nov), which two independent checkers accepted. Finding the optimum and proving it are different tasks: Concorde proved it in 3,296 s of processor time on one core of our laptop and in 2,910 s with CPLEX on NEOS.Na 6 od 8 jeder prenosnika (12 niti) kandidatna gradnja 8zTSP to pot najde v 5,6 s (mediana 10 tekov; najpočasnejši je trajal 6,0 s), optimum zemljevida nu3496 pa v 5,8 s (mediana 10 tekov). Optimum je največ 0,065 % nad našo lastno spodnjo mejo, 7.839.923 m (18. nov.), ki sta jo sprejela dva neodvisna preverjevalnika. Najti optimum in ga dokazati sta različni nalogi: Concorde ga je dokazal v 3.296 s procesorskega časa na enem jedru našega prenosnika in v 2.910 s s CPLEX na strežniku NEOS.

    More on SwitzerlandVeč o Švici

    The map.Zemljevid. che_from_europe_20261117: the 2,850 places of our Europe map, which uses the World TSP’s places (University of Waterloo), that lie in Switzerland, with the same coordinates as Europe: whole metres in a map projection for Europe, straight lines, not roads. Borders: Made with Natural Earth. In the demo: Examples → Selected 10 → Switzerland, with the stop at the proved optimum.che_from_europe_20261117: 2.850 krajev našega zemljevida Evrope, ki uporablja kraje svetovnega TSP (Univerza Waterloo), ki ležijo v Švici, z istimi koordinatami kot Evropa: celi metri v projekciji zemljevida za Evropo, ravne črte, ne ceste. Meje: Made with Natural Earth. V demu: Primeri → Izbranih 10 → Švica, z ustavitvijo pri dokazanem optimumu.

    On the way to the optimum.Na poti do optimuma. With one thread (17 Nov), 8zTSP’s tour was 5.20% above the optimum after 1 s, 1.62% after 30 s and 0.16% after 60 s; it reached the optimum after 75.1 s, with 200 improvements in all, each logged with its time.Z eno nitjo (17. novembra) je bila pot 8zTSP po 1 s 5,20 % nad optimumom, po 30 s 1,62 % in po 60 s 0,16 %; optimum je dosegla po 75,1 s, skupaj z 200 izboljšavami, vsaka zapisana s svojim časom.

    Ten seeds, one thread each.Deset semen, vsak program na eni niti. Every program ran ten times, with seeds 1 to 10, on one thread of the same core, and was stopped from outside when it reached Concorde’s optimum; none was told the optimum. The time to the optimum:Vsak program je tekel desetkrat, s semeni od 1 do 10, na eni niti istega jedra; ustavili smo ga od zunaj, ko je dosegel Concordov optimum, ki ga nobenemu nismo povedali. Čas do optimuma:

    ProgramMedianMedianaFastestNajhitrejeSlowestNajpočasneje
    LKH 3.0.735.4 s35,4 s25.1 s25,1 s117.9 s117,9 s
    LKH 3.0.1437.7 s37,7 s28.9 s28,9 s117.0 s117,0 s
    GA-EAX42.5 s42,5 s35.9 s35,9 s93.8 s93,8 s
    8zTSP, candidate build8zTSP, kandidatna gradnja58.3 s58,3 s45.4 s45,4 s72.7 s72,7 s
    8zTSP, development build of 18 Nov8zTSP, razvojna gradnja z 18. novembra28.1 s28,1 s19.8 s19,8 s50.5 s50,5 s

    All five reached the optimum in all ten runs. The development build adds a quick analysis at the start that adapts the settings to the map; it was faster than the candidate build in each of the ten seeds. It is a development build, not yet in the published program.Vseh pet je optimum doseglo v vseh desetih tekih. Razvojna gradnja doda hitro analizo na začetku, ki nastavitve prilagodi zemljevidu; v vsakem od desetih semen je bila hitrejša od kandidatne gradnje. To je razvojna gradnja, ki še ni v objavljenem programu.

    The proofs.Dokaza. Concorde proved the optimum twice: with QSopt on one core of our laptop (17 Nov, 14:03–15:03, 3,296 s of processor time, 155 nodes) and with CPLEX on the NEOS Server (job 24917716, 2,910 s, 171 nodes). Our own certified bound is in the entryConcorde je optimum dokazal dvakrat: s QSopt na enem jedru našega prenosnika (17. novembra, 14:03–15:03, 3.296 s procesorskega časa, 155 vozlišč) in s CPLEX na strežniku NEOS (naloga 24917716, 2.910 s, 171 vozlišč). Naša lastna potrjena meja je v vnosu Our own lower boundLastna spodnja meja.

    Next.Naslednji korak. Every program on 6 of the laptop’s 8 cores, and the development build’s analysis in the published program. Measured by the time to the optimum.Vsi programi na 6 od 8 jeder prenosnika in analiza razvojne gradnje v objavljenem programu. Merilo je čas do optimuma.

    Who did it.Kdo je to naredil. BD asked for Switzerland on /tsp/ and in the demo. 8zTSP-K cut the map from our Europe map and ran 8zTSP and Concorde, on our laptop and on NEOS; 8zTSP-proof computed our own bound. 8zTSP-media ran LKH and GA-EAX ten times each, and 8zTSP-plus the two 8zTSP builds.BD je želel Švico na /tsp/ in v demu. 8zTSP-K je zemljevid izrezal iz našega zemljevida Evrope in pognal 8zTSP in Concorde, na našem prenosniku in na strežniku NEOS; 8zTSP-proof je izračunal našo lastno mejo. 8zTSP-media je LKH in GA-EAX pognal po desetkrat, 8zTSP-plus pa obe gradnji 8zTSP.

    Conditions: development runs of 17 and 18 Nov 2026 on the laptop’s Intel i7-10875H. 8zTSP on 17 Nov: development build 326026c, then the candidate for the freeze, seed 1, one thread on logical CPU 14, whose twin, CPU 6, ran Concorde’s proof for Slovenia the whole time, so about half a physical core; stopped from outside at Concorde’s optimum; the tour recomputed from the map. Concorde (38a647d3) with QSopt, seed 42, logical CPU 15, 14:02:59–15:03:08. NEOS: Concorde with CPLEX, job 24917716 (submitted at 18:13), prod-exec-6; the times are Concorde’s own processor time. The map’s sha256 1afc3b2e…; the demo serves the same file, gzipped. Ten seeds: one thread on logical CPU 14 under the same load, seeds 1–10, a limit of 300 s; the time runs from the program’s start, including its preparation, to the moment it had a tour of the optimal length. LKH 3.0.7 and 3.0.14: their default settings with one run and a trial limit of 1,000,000,000, captured by 8zTSP-media on 18 Nov, 00:18–00:36; in three LKH 3.0.7 runs (seeds 6–8) LKH reported the optimal length, but the stop came before it saved that tour, so those three times are LKH’s own. GA-EAX: settings 100/30, faster than 50/30 in our test on nu3496, captured on 17 Nov, 20:19–20:50. 8zTSP: the candidate build 9cfd3c5, the candidate for the freeze, and the development build 145fb32, on 18 Nov, 00:47–01:41. The other tours were recomputed from the map. 6 of 8 cores: build 9cfd3c5 with 12 threads on logical CPUs 0–5 and 8–13, 10 runs per map, 18 Nov, 05:35–06:44; seed 1 again at 06:51: 5.6 s with 12 threads, 6.8 s with 6 threads, one per core; the other two cores ran Concorde’s proof for Slovenia and a browser.Pogoji: razvojni teki 17. in 18. novembra 2026 na prenosnikovem Intel i7-10875H. 8zTSP 17. novembra: razvojna gradnja 326026c, takrat kandidat za zamrznitev, seme 1, ena nit na logičnem CPU 14, katerega dvojček, CPU 6, je ves čas poganjal Concordov dokaz za Slovenijo, torej okoli pol fizičnega jedra; ustavljen od zunaj pri Concordovem optimumu; pot preračunana iz zemljevida. Concorde (38a647d3) s QSopt, seme 42, logični CPU 15, 14:02:59–15:03:08. NEOS: Concorde s CPLEX, naloga 24917716 (poslana ob 18:13), prod-exec-6; časi so Concordov lastni procesorski čas. Sha256 zemljevida 1afc3b2e…; demo streže isto datoteko, stisnjeno z gzip. Deset semen: ena nit na logičnem CPU 14 pod isto obremenitvijo, semena 1–10, omejitev 300 s; čas teče od zagona programa, vključno z njegovo pripravo, do trenutka, ko je imel pot optimalne dolžine. LKH 3.0.7 in 3.0.14: privzete nastavitve z enim tekom in omejitvijo poskusov 1.000.000.000, posnel 8zTSP-media 18. novembra, 00:18–00:36; v treh tekih LKH 3.0.7 (semena 6–8) je LKH javil optimalno dolžino, a je ustavitev prišla, preden je to pot shranil, zato so ti trije časi LKH-jevi lastni. GA-EAX: nastavitve 100/30, v našem preizkusu na nu3496 hitrejše od 50/30, posneto 17. novembra, 20:19–20:50. 8zTSP: kandidatna gradnja 9cfd3c5, kandidatka za zamrznitev, in razvojna gradnja 145fb32, 18. novembra, 00:47–01:41. Druge poti so preračunane iz zemljevida. 6 od 8 jeder: gradnja 9cfd3c5 z 12 nitmi na logičnih CPU 0–5 in 8–13, 10 tekov na zemljevid, 18. novembra, 05:35–06:44; seme 1 znova ob 06:51: 5,6 s z 12 nitmi, 6,8 s s 6 nitmi, po eno na jedro; drugi dve jedri sta poganjali Concordov dokaz za Slovenijo in brskalnik.

    ContentsKazalo

  4. · development result of 17 Nov; not yet in the published program· razvojni rezultat 17. novembra; še ni v objavljenem programu

    Our own branch and bound proves its first optimum: qa194Naša lastna metoda razveji in omeji prvič dokaže optimum: qa194

    8zTSP’s own code has proved that the tour of 9,352 on qa194, 194 cities in Qatar, is optimal, on a map where our lower bound alone stays below the tour.Lastna koda 8zTSP je dokazala, da je pot dolžine 9.352 na qa194, 194 mestih v Katarju, optimalna, na zemljevidu, kjer sama naša spodnja meja ostane pod potjo. Our own branch and bound split the question into 15 cases in about a second and wrote a certificate. Two independent checkers accepted it, one written in Rust and one in Python, the second written only from the certificate’s specification, without the first one’s code; both print »PROVED OPTIMAL«. 194 cities is a first step: Concorde proves this map in under a second, without branching, and it proves maps with tens of thousands of cities.Naša lastna metoda razveji in omeji je vprašanje v približno sekundi razdelila na 15 primerov in zapisala potrdilo. Sprejela sta ga dva neodvisna preverjevalnika, eden v Rustu in eden v Pythonu, drugi napisan samo po opisu potrdila, brez kode prvega; oba izpišeta »PROVED OPTIMAL«. 194 mest je prvi korak: Concorde ta zemljevid dokaže v manj kot sekundi, brez razvejanja, in dokazuje zemljevide z desettisoči mest.

    More on our own branch and boundVeč o naši metodi razveji in omeji

    How it works.Kako deluje. A lower bound says that no tour can be shorter than a given length. Where the bound stays below the best tour, the search splits the question in two: tours that use a chosen connection between two cities, and tours that do not. Each part gets its own, tighter bound, and a part whose bound reaches the tour’s length cannot hold a shorter tour. On qa194 every part was closed after 15 cases, 8 of them final, at most 5 splits deep. The certificate lists every final case with the numbers its bound rests on; the checkers recompute each bound exactly, in whole numbers, and check that the final cases together cover every possible tour.Spodnja meja pove, da nobena pot ne more biti krajša od dane dolžine. Kjer meja ostane pod najboljšo potjo, iskanje vprašanje razdeli na dva dela: poti, ki uporabijo izbrano povezavo med dvema mestoma, in poti, ki je ne uporabijo. Vsak del dobi svojo, tesnejšo mejo, in del, katerega meja doseže dolžino poti, ne more vsebovati krajše poti. Na qa194 so bili vsi deli zaprti po 15 primerih, od tega 8 končnih, največ 5 razdelitev globoko. Potrdilo našteje vsak končni primer s številkami, na katerih sloni njegova meja; preverjevalnika vsako mejo izračunata znova, natančno, v celih številih, in preverita, da končni primeri skupaj pokrijejo vsako možno pot.

    Checked.Preverjeno. Both checkers answered »ACCEPT certified=9352 claimed=9352 tour=9352 PROVED OPTIMAL«; we ran both again on 17 Nov at 19:13, with the same answer.Oba preverjevalnika sta odgovorila »ACCEPT certified=9352 claimed=9352 tour=9352 PROVED OPTIMAL«; 17. novembra ob 19:13 smo oba pognali znova, z enakim odgovorom.

    Next.Naslednji korak. Larger maps: uy734 (734 cities, its optimum is known), then Switzerland’s 2,850 places. Measured by the number of cases and the time to the proof.Večji zemljevidi: uy734 (734 mest, optimum je znan), nato 2.850 krajev Švice. Merilo: število primerov in čas do dokaza.

    Who did it.Kdo je to naredil. BD asked for our own proof of the optimum and decided that »proved« needs two independent checkers. 8zTSP-proof wrote our branch and bound and the Rust checker; a separate helper wrote the Python checker from the specification alone.BD je prosil za naš lasten dokaz optimuma in odločil, da »dokazano« potrebuje dva neodvisna preverjevalnika. 8zTSP-proof je napisal našo metodo razveji in omeji in preverjevalnik v Rustu; ločen pomočnik je preverjevalnik v Pythonu napisal samo po opisu.

    Conditions: development result of 17 Nov 2026 on the laptop’s Intel i7-10875H; not in the published program. Map qa194 (194 cities in Qatar, from the University of Waterloo’s National TSPs; sha256 bb222094…), tour of 9,352 (sha256 22b609a1…), certificate qa194_bb_20261117.tree (sha256 d179f605…): 15 nodes, 8 of them leaves, depth 5, 0.9 s. Development branch feature/proof: the Rust checker at 59c982b, the Python checker at e0b6558. Concorde on qa194: 1 node, 0.62–0.69 s of its own time in three runs of 17 Nov on the same laptop.Pogoji: razvojni rezultat 17. novembra 2026 na prenosnikovem Intel i7-10875H; ni v objavljenem programu. Zemljevid qa194 (194 mest v Katarju, iz nacionalnih TSP Univerze Waterloo; sha256 bb222094…), pot dolžine 9.352 (sha256 22b609a1…), potrdilo qa194_bb_20261117.tree (sha256 d179f605…): 15 vozlišč, od tega 8 listov, globina 5, 0,9 s. Razvojna veja feature/proof: preverjevalnik v Rustu na 59c982b, preverjevalnik v Pythonu na e0b6558. Concorde na qa194: 1 vozlišče, 0,62–0,69 s lastnega časa v treh tekih 17. novembra na istem prenosniku.

    ContentsKazalo

  5. · development runs of 16–17 Nov, one core each· razvojni teki 16. in 17. novembra, po eno jedro

    Hard maps, one core eachTežki zemljevidi, po eno jedro

    On maps made to be hard, with one core for each program: 8zTSP reached the proven optimum in every run on three of the four evolved 1,000-city maps and stayed 0.002% above it on the fourth; LKH 3.0.7 missed the optimum on the two maps evolved against it and found it on the two evolved in its favour.Na zemljevidih, ki so narejeni tako, da so težki, z enim jedrom za vsak program: 8zTSP je na treh od štirih evolviranih zemljevidov s 1.000 mesti v vsakem teku dosegel dokazani optimum, na četrtem pa ostal 0,002 % nad njim; LKH 3.0.7 je optimum zgrešil na obeh zemljevidih, evolviranih proti njemu, in ga našel na obeh, evolviranih njemu v prid. On three large maps (one hour each), 8zTSP reached the best known tour of xrb14233 and a shorter tour than LKH on C10k.0; LKH came closer on ido21215. On Tnm151, Concorde stopped after 6.3 hours with a numerical failure of its LP solver; 8zTSP reaches that map’s optimum in under 0.15 s.Na treh velikih zemljevidih (po eno uro) je 8zTSP dosegel najboljšo znano pot za xrb14233 in na C10k.0 krajšo pot od LKH; na ido21215 je bil LKH bližje. Na Tnm151 se je Concorde po 6,3 ure ustavil zaradi numerične napake svojega reševalnika LP; 8zTSP optimum tega zemljevida doseže v manj kot 0,15 s.

    More on the hard maps, one core eachVeč o težkih zemljevidih, po eno jedro

    Four evolved maps of 1,000 cities, one core each: the proved optimum (Concorde, with its time), the runs of each program that reached it, and the best tour otherwise.Štirje evolvirani zemljevidi s 1.000 mesti, po eno jedro: dokazani optimum (Concorde, z njegovim časom), teki vsakega programa, ki so ga dosegli, sicer pa najboljša pot.

    Map (evolved to favour)Zemljevid (evolviran v prid)Proved optimum (Concorde)Dokazani optimum (Concorde)8zTSP, 10 runs of 60 s8zTSP, 10 tekov po 60 sLKH 3.0.7, 3 runs of 600 sLKH 3.0.7, 3 teki po 600 s
    eax_lkh_1000-137 (EAX over LKH)eax_lkh_1000-137 (EAX pred LKH)16,980,565 (10.5 s)16.980.565 (10,5 s)10 of 10 at the optimum, median 2.0 s10 od 10 do optimuma, mediana 2,0 s0 of 3; best +5,1570 od 3; najboljša +5.157
    eax_lkh_1000-111 (EAX over LKH)eax_lkh_1000-111 (EAX pred LKH)17,693,560 (25.2 s)17.693.560 (25,2 s)10 of 10 at the optimum, median 2.4 s10 od 10 do optimuma, mediana 2,4 s0 of 3; best +1,4020 od 3; najboljša +1.402
    lkh_eax_1000-195 (LKH over EAX)lkh_eax_1000-195 (LKH pred EAX)23,172,007 (102 s)23.172.007 (102 s)10 of 10 at the optimum, 3.4–27.4 s (median 10.3 s)10 od 10 do optimuma, 3,4–27,4 s (mediana 10,3 s)3 of 3 at the optimum, 1.7–8.4 s3 od 3 do optimuma, 1,7–8,4 s
    lkh_eax_1000-478 (LKH over EAX)lkh_eax_1000-478 (LKH pred EAX)19,767,397 (166 s)19.767.397 (166 s)0 of 10; best +440 (0.002%)0 od 10; najboljša +440 (0,002 %)3 of 3 at the optimum, 1.7–2.6 s3 od 3 do optimuma, 1,7–2,6 s

    Three large maps, one core and one hour each: the best known tour in our data and each program’s tour after the hour.Trije veliki zemljevidi, po eno jedro in eno uro: najboljša znana pot v naših podatkih in pot vsakega programa po eni uri.

    MapZemljevidBest known tour (our data)Najboljša znana pot (naši podatki)8zTSP, 1 hour8zTSP, 1 uraLKH 3.0.7, 1 hourLKH 3.0.7, 1 ura
    xrb14233 (VLSI, 14,233 points)xrb14233 (VLSI, 14.233 točk)45,462 (lower bound 45,450)45.462 (spodnja meja 45.450)45,462 at 3,261 s45.462 pri 3.261 s45,47745.477
    ido21215 (VLSI, 21,215 points)ido21215 (VLSI, 21.215 točk)63,517 (lower bound 63,501)63.517 (spodnja meja 63.501)63,52863.52863,52163.521
    C10k.0 (clustered, 10,000 points)C10k.0 (gručasti, 10.000 točk)33,019,942 (DIMACS TSP Challenge, 2002)33.019.942 (izziv DIMACS TSP, 2002)33,002,97733.002.97733,017,11733.017.117

    How to read it.Kako brati. Every program ran on one core here, so these are one-core results, not whole-laptop ones. Seiler and co-authors evolved the four 1,000-city maps so that one of two solvers does much better than the other: the »eax_lkh« maps favour Nagata’s EAX over LKH, the »lkh_eax« maps favour LKH over EAX. LKH behaves as the names say; 8zTSP found the optimum on both maps made against LKH and on one of the two made for it.Vsak program je tu tekel na enem jedru, zato so to izidi za eno jedro, ne za ves prenosnik. Seiler in soavtorji so štiri zemljevide s 1.000 mesti evolvirali tako, da je eden od dveh reševalnikov precej boljši od drugega: zemljevidi »eax_lkh« so v prid Nagatovemu EAX pred LKH, zemljevidi »lkh_eax« v prid LKH pred EAX. LKH se obnaša, kot pravijo imena; 8zTSP je optimum našel na obeh zemljevidih, narejenih proti LKH, in na enem od dveh, narejenih njemu v prid.

    Tnm151. Concorde, with its QSopt LP solver, found a tour of 2,293,265, the published optimum, but stopped after 6.3 hours with a numerical failure (»Dual LP Solver reached iteration limit … no optimal LP-solution exists«), at a lower bound of 2,286,197, 0.31% below it. 8zTSP reached the optimum in all 10 runs of 16 Nov, each under 0.15 s.Concorde je s svojim reševalnikom LP QSopt našel pot 2.293.265, objavljeni optimum, a se je po 6,3 ure ustavil zaradi numerične napake (»Dual LP Solver reached iteration limit … no optimal LP-solution exists«) pri spodnji meji 2.286.197, 0,31 % pod njim. 8zTSP je optimum dosegel v vseh 10 tekih 16. novembra, vsakič v manj kot 0,15 s.

    C10k.0. Both tours are shorter than the best tour of the DIMACS TSP Challenge (2002) in our data, 33,019,942; the current best known tour for this map we have not checked, so no gap is given here.Obe poti sta krajši od najboljše poti izziva DIMACS TSP (2002) v naših podatkih, 33.019.942; trenutne najboljše znane poti za ta zemljevid nismo preverili, zato razlike tu ne navajamo.

    Who did it.Kdo je to naredil. 8zTSP-K planned and ran the runs in the arena; every number here was recomputed from the run log.8zTSP-K je teke načrtoval in pognal v areni; vsaka številka tu je ponovno izračunana iz dnevnika tekov.

    Conditions: development runs of 16–17 Nov 2026 on the laptop’s Intel i7-10875H, one core per program and map group (logical CPUs 1–4), never paused. 8zTSP build 3c36abdd; LKH 3.0.7 (f4d16556) with its default settings and the time limits stated; Concorde (38a647d3) with QSopt. Evolved maps: Seiler and co-authors, fetched from the authors’ repository; 8zTSP 10 runs of 60 s, LKH 3 runs of 600 s. VLSI maps from Waterloo, the clustered map from the DIMACS TSP Challenge: one run of one hour each. Tnm151: Hougardy and Zhong (2021); 8zTSP’s runs of 16 Nov.Pogoji: razvojni teki 16. in 17. novembra 2026 na prenosnikovem Intel i7-10875H, eno jedro na program in skupino zemljevidov (logični CPU 1–4), brez premorov. 8zTSP, gradnja 3c36abdd; LKH 3.0.7 (f4d16556) s privzetimi nastavitvami in navedenimi časovnimi mejami; Concorde (38a647d3) s QSopt. Evolvirani zemljevidi: Seiler in soavtorji, preneseni iz repozitorija avtorjev; 8zTSP 10 tekov po 60 s, LKH 3 teki po 600 s. Zemljevidi VLSI z Univerze Waterloo, gručasti zemljevid iz izziva DIMACS TSP: po en tek ene ure. Tnm151: Hougardy in Zhong (2021); teki 8zTSP s 16. novembra.

    ContentsKazalo

  6. · development runs of 16 Nov· razvojni teki 16. novembra

    Burma, 33,708 cities, 4 cores each: level with LKH after one hourBurma, 33.708 mest, po 4 jedra: izenačeno z LKH po eni uri

    On Burma’s 33,708 cities, with 4 cores each for one hour (8zTSP with 4 threads, LKH 3.0.7 as 4 copies), the two programs end level: in three runs their tour lengths differ by at most 0.001%.Na 33.708 mestih Burme se programa po eni uri s po 4 jedri (8zTSP s 4 nitmi, LKH 3.0.7 kot 4 kopije) končata izenačeno: v treh tekih se dolžini njunih poti razlikujeta največ za 0,001 %. 8zTSP’s tour was shorter in two runs, by 5 and 4 units, and equal in the third; that is noise, so we call it level. All six tours end 0.038–0.040% above Waterloo’s lower bound, 959,011, and about 0.01% above the best known tour, 959,289 (Yuichi Nagata). 8zTSP has a tour after 1 second, LKH none before about 20 seconds; LKH then leads from about 30 seconds to about 20 minutes, and from about 30 minutes on the two are level.Pot 8zTSP je bila v dveh tekih krajša, za 5 in 4 enote, v tretjem pa enaka; to je šum, zato temu rečemo izenačeno. Vseh šest poti konča 0,038–0,040 % nad spodnjo mejo Univerze Waterloo, 959.011, in okoli 0,01 % nad najboljšo znano potjo, 959.289 (Yuichi Nagata). 8zTSP ima pot po 1 sekundi, LKH pred okoli 20 sekundami nobene; nato LKH vodi od okoli 30 sekund do okoli 20 minut, od okoli 30 minut naprej pa sta izenačena.

    More on Burma, 4 cores eachVeč o Burmi, po 4 jedra

    Burma (bm33708), 4 cores each for one hour: each program’s tour after the hour and, in brackets, how far it is above Waterloo’s lower bound, 959,011; 16 Nov 2026. For LKH, the best of its 4 copies.Burma (bm33708), po 4 jedra eno uro: pot vsakega programa po uri in v oklepaju, koliko je nad spodnjo mejo Univerze Waterloo, 959.011; 16. nov. 2026. Pri LKH najboljša od njegovih 4 kopij.

    RunTek8zTSP, 4 threads8zTSP, 4 nitiLKH 3.0.7, 4 copiesLKH 3.0.7, 4 kopijeShorter tourKrajša pot
    1959,387 (0.039%)959.387 (0,039 %)959,392 (0.040%)959.392 (0,040 %)8zTSP, by 58zTSP, za 5
    2959,392 (0.040%)959.392 (0,040 %)959,392 (0.040%)959.392 (0,040 %)equalenako
    3959,380 (0.038%)959.380 (0,038 %)959,384 (0.039%)959.384 (0,039 %)8zTSP, by 48zTSP, za 4

    How to read it.Kako brati. Each program had the same 4 of the laptop’s cores for the same hour: 8zTSP as one search with 4 threads, LKH 3.0.7 as 4 independent copies with different seeds, of which the best counts. Nobody knows Burma’s optimum yet; it lies between Waterloo’s lower bound, 959,011, and the best known tour, 959,289, which is 0.029% above the bound.Vsak program je imel isto uro iste 4 jedra prenosnika: 8zTSP kot eno iskanje s 4 nitmi, LKH 3.0.7 kot 4 neodvisne kopije z različnimi semeni, od katerih šteje najboljša. Optimuma za Burmo še nihče ne pozna; leži med spodnjo mejo Univerze Waterloo, 959.011, in najboljšo znano potjo, 959.289, ki je 0,029 % nad mejo.

    Over time.S časom. After 60 seconds LKH’s best tour was 959,942–959,959 and 8zTSP’s 960,905–961,390; after 10 minutes 959,446–959,508 and 959,465–960,349. From about 30 minutes on, the two are level.Po 60 sekundah je bila najboljša pot LKH 959.942–959.959, pot 8zTSP pa 960.905–961.390; po 10 minutah 959.446–959.508 in 959.465–960.349. Od okoli 30 minut naprej sta izenačena.

    The seeds.Semena. LKH’s three runs used overlapping seed sets (42–45, 43–46 and 44–47), and seed 45 alone gave the 959,392 of runs 1 and 2, so LKH’s side has two nearly independent results, not three. The verdict stands; a clean repeat uses separate sets.Trije teki LKH so uporabili prekrivajoče se skupine semen (42–45, 43–46 in 44–47), samo seme 45 pa je dalo 959.392 v tekih 1 in 2, zato ima stran LKH dva skoraj neodvisna izida, ne treh. Sklep velja; čista ponovitev uporabi ločene skupine.

    Next.Naprej. The same hour with all 7 cores for each, and separate LKH seed sets. Measured by the tour after one hour and by how far the runs spread.Ista ura z vsemi 7 jedri za vsakega in z ločenimi skupinami semen za LKH. Merimo pot po eni uri in to, kako zelo se teki razlikujejo.

    Who did it.Kdo je to naredil. 8zTSP-K planned and ran the runs in the arena; every final tour was recomputed from its file.8zTSP-K je teke načrtoval in pognal v areni; vsaka končna pot je bila ponovno izračunana iz svoje datoteke.

    Conditions: development runs of 16 Nov 2026 on the laptop’s Intel i7-10875H, 4 cores each (logical CPUs 1–4), one run at a time. 8zTSP build 563f8fa (main 1f072ad) with 4 threads, seeds 1, 2 and 3. LKH 3.0.7 (Keld Helsgaun) with settings for geometric maps, 4 copies with seeds 42–45, 43–46 and 44–47; at each moment the best of the four counts. Neither program was given the bound or the best known tour. Peak memory: 8zTSP 0.13–0.15 GB, LKH’s 4 copies 0.15 GB. Lower bound and best known tour:Pogoji: razvojni teki 16. novembra 2026 na prenosnikovem Intel i7-10875H, po 4 jedra (logični CPU 1–4), en tek naenkrat. 8zTSP, gradnja 563f8fa (main 1f072ad), s 4 nitmi, semena 1, 2 in 3. LKH 3.0.7 (Keld Helsgaun) z nastavitvami za geometrijske zemljevide, 4 kopije s semeni 42–45, 43–46 in 44–47; v vsakem trenutku šteje najboljša od štirih. Nobeden od programov ni dobil meje ali najboljše znane poti. Največ pomnilnika: 8zTSP 0,13–0,15 GB, 4 kopije LKH 0,15 GB. Spodnja meja in najboljša znana pot: National TSPs, University of WaterlooNational TSPs, Univerza Waterloo.

    ContentsKazalo

  7. · development runs of 16 Nov· razvojni teki 16. novembra

    Burma, 33,708 cities, one core each: level with LKH after one hourBurma, 33.708 mest, po eno jedro: izenačeno z LKH po eni uri

    On Burma’s 33,708 cities, after one hour on one core each, 8zTSP and LKH 3.0.7 end level: in three matched pairs their tour lengths differ by at most 0.004%.Na 33.708 mestih Burme se 8zTSP in LKH 3.0.7 po eni uri, vsak na enem jedru, končata izenačeno: v treh primerjalnih parih se dolžini njunih poti razlikujeta največ za 0,004 %. 8zTSP’s tour was the shorter one in two of the pairs, LKH’s in one, but the three runs of each program differ among themselves by about as much, so we call it level, not a lead. One more 8zTSP seed, run without an LKH partner, ended 0.111% above the lower bound: some seeds take longer, and we show them. No run reached the best known tour, 959,289, found by Yuichi Nagata. The percentages are above the University of Waterloo’s lower bound for Burma, 959,011; our own bound covers maps of up to 20,000 cities, so it does not apply here.Pot 8zTSP je bila krajša v dveh parih, pot LKH v enem, a trije teki vsakega programa se med seboj razlikujejo približno toliko, zato temu rečemo izenačeno, ne vodstvo. Še eno seme 8zTSP, ki je teklo brez para z LKH, je končalo 0,111 % nad spodnjo mejo: nekatera semena potrebujejo dlje, in tudi taka pokažemo. Noben tek ni dosegel najboljše znane poti, 959.289, ki jo je našel Yuichi Nagata. Odstotki so nad spodnjo mejo Univerze Waterloo za Burmo, 959.011; naša lastna meja velja za zemljevide z do 20.000 mesti, zato tu ne pride v poštev.

    More on Burma, one core eachVeč o Burmi, po eno jedro

    Burma (bm33708): each program’s tour after one hour on one core and, in brackets, how far it is above Waterloo’s lower bound, 959,011; 16 Nov 2026. The pairs: LKH seed 42 with 8zTSP seed 1, 43 with 3, 44 with 4.Burma (bm33708): pot vsakega programa po eni uri na enem jedru in v oklepaju, koliko je nad spodnjo mejo Waterlooja, 959.011; 16. novembra 2026. Pari: LKH s semenom 42 z 8zTSP s semenom 1, 43 s 3, 44 s 4.

    PairParLKH 3.0.78zTSPShorter tourKrajša pot
    1959,416 (0.042%)959.416 (0,042 %)959,453 (0.046%)959.453 (0,046 %)LKH
    2959,451 (0.046%)959.451 (0,046 %)959,442 (0.045%)959.442 (0,045 %)8zTSP
    3959,450 (0.046%)959.450 (0,046 %)959,424 (0.043%)959.424 (0,043 %)8zTSP
    8zTSP seed 2, unpaired8zTSP, seme 2, brez para—960,074 (0.111%)960.074 (0,111 %)—

    How to read it.Kako brati. Nobody knows Burma’s optimum yet; it lies between Waterloo’s lower bound, 959,011 (no round trip on this map can be shorter), and the best known tour, 959,289, which is 0.029% above the bound. A matched pair is one LKH run and one 8zTSP run, started at the same time and stopped after one hour, each on its own core. A seed sets a run’s random choices; with another seed, the same program makes a different run.Optimuma Burme še nihče ne pozna; leži med spodnjo mejo Waterlooja, 959.011 (noben krog po tem zemljevidu ne more biti krajši), in najboljšo znano potjo, 959.289, ki je 0,029 % nad mejo. Primerjalni par sta en tek LKH in en tek 8zTSP, začeta hkrati in ustavljena po eni uri, vsak na svojem jedru. Seme določa naključne izbire teka; z drugim semenom isti program naredi drugačen tek.

    Merging the tours.Združevanje poti. When LKH merged the two tours of a pair, the result was shorter than both in 1 of 3 pairs (959,402), so we claim no help from it.Ko je LKH združil obe poti istega para, je bil izid krajši od obeh v 1 od 3 parov (959.402), zato ne trdimo, da združevanje pomaga.

    Next.Naslednji korak. The same hour with 7 threads, and more seeds for each program. Measured by the tour after one hour and by how far the seeds spread.Ista ura s 7 nitmi in več semen za vsak program. Merilo: pot po eni uri in razpon med semeni.

    Who did it.Kdo je to naredil. 8zTSP-K planned the runs, ran them in the arena and recomputed every tour.8zTSP-K je teke načrtoval, jih pognal v areni in vsako pot preračunal znova.

    Conditions: development runs of 16 Nov 2026 between 01:34 and 06:20 on the laptop’s Intel i7-10875H, one physical core for each program (logical CPUs 1–4), beside other jobs: Concorde, a 48-hour LKH run, Stockfish and, at times, other work on the twin threads of the same cores. LKH 3.0.7 (Keld Helsgaun) with settings for geometric maps (Delaunay candidates, 5 per city, initial period 100), otherwise default, seeds 42, 43 and 44; it repeated its runs from new start tours until the hour was over. 8zTSP with the code of main 1f072ad and its default settings, 1 thread, seeds 1, 3 and 4, and seed 2 without a partner. Neither program was given the bound or the best known tour. The arena recomputed every tour from the map (each city once, the length from the coordinates), and we checked them again. The map bm33708, its lower bound and the best known tour are from thePogoji: razvojni teki 16. novembra 2026 med 01:34 in 06:20 na procesorju Intel i7-10875H v prenosniku, eno fizično jedro za vsak program (logični CPU 1–4), ob drugih poslih: Concorde, 48-urni tek LKH, Stockfish in občasno drugo delo na drugih nitih istih jeder. LKH 3.0.7 (Keld Helsgaun) z nastavitvami za geometrijske zemljevide (kandidati Delaunay, 5 na mesto, začetna perioda 100), sicer privzetimi, semena 42, 43 in 44; svoje teke je ponavljal iz novih začetnih poti, dokler ura ni potekla. 8zTSP s kodo main 1f072ad in privzetimi nastavitvami, 1 nit, semena 1, 3 in 4 ter seme 2 brez para. Noben program ni dobil meje ali najboljše znane poti. Arena je vsako pot preračunala iz zemljevida (vsako mesto enkrat, dolžina iz koordinat), mi pa smo jih preverili še enkrat. Zemljevid bm33708, njegova spodnja meja in najboljša znana pot so iz zbirke National TSPs of the University of Waterloo.Univerze Waterloo.

    ContentsKazalo

  8. · development result of 16 Nov; not yet in the published program· razvojni rezultat 16. novembra; še ni v objavljenem programu

    Our own lower bound: a certificate for every mapLastna spodnja meja: potrdilo za vsak zemljevid

    8zTSP’s own code can now show how much longer a tour is, at most, than the shortest possible one (the optimum), even where nobody knows the optimum: our best tour of Slovenia’s 6,035 settlements is at most 0.19% longer than the optimum by our own bound (17 Nov).Lastna koda 8zTSP zdaj zna pokazati, za koliko je pot največ daljša od najkrajše možne (optimuma), tudi kjer optimuma nihče ne pozna: naša najboljša pot skozi 6.035 naselij Slovenije je največ 0,19 % daljša od optimuma po naši lastni meji (17. nov.). Such a checked statement of a tour’s quality is a certificate; it rests on a lower bound, a length that no round trip on the map can go below, and our code computes one for any map of up to 20,000 cities. On wi29 and dj38 the bound equals the length of our tour, so our own code proves those tours optimal. For Slovenia the first certificate took 175 seconds on one core; Concorde’s tighter bound, at most 0.006% (18 Nov 2026), comes from a different, exact program. On nu3496 our tour already has the known optimal length, 96,132; our own bound, without that knowledge, showed at most 0.70% in 241 seconds. The method is the classic Held–Karp bound (Held and Karp, 1970 and 1971), with step rules after Held, Wolfe and Crowder (1974) and Volgenant and Jonker (1982). Since 17 Nov a tighter bound of our own gives at most 0.19% for Slovenia (8,712,125 m; 0.48% before) and at most 0.44% for nu3496 (95,716; 0.70% before); each of these certificates was checked by two independent checkers in exact integer arithmetic. For now it is a development option, so you cannot try it yet: off by default, not yet in the published program or in the browser demo.Taka preverjena izjava o kakovosti poti je potrdilo; sloni na spodnji meji, dolžini, od katere ne more biti krajši noben krog po zemljevidu, in naša koda jo izračuna za vsak zemljevid z do 20.000 mesti. Pri wi29 in dj38 je meja enaka dolžini naše poti, zato naša koda sama dokaže, da sta ti poti optimalni. Za Slovenijo je prvo potrdilo nastalo v 175 sekundah na enem jedru; Concordova tesnejša meja, največ 0,006 % (18. nov. 2026), prihaja iz drugega, eksaktnega programa. Pri nu3496 ima naša pot že znano optimalno dolžino 96.132; naša lastna meja brez tega znanja je v 241 sekundah pokazala največ 0,70 %. Metoda je klasična meja Held–Karp (Held in Karp, 1970 in 1971), s pravili za korake, kot so jih opisali Held, Wolfe in Crowder (1974) ter Volgenant in Jonker (1982). Od 17. novembra tesnejša lastna meja da za Slovenijo največ 0,19 % (8.712.125 m; prej 0,48 %) in za nu3496 največ 0,44 % (95.716; prej 0,70 %); vsako od teh potrdil sta preverila dva neodvisna preverjevalnika v natančni celoštevilski aritmetiki. Zaenkrat je to razvojna možnost, zato je še ne moreš preizkusiti: privzeto izklopljena, še ne v objavljenem programu in ne v demu v brskalniku.

    More on our own lower boundVeč o lastni spodnji meji

    How far above the optimum our tour can be at most, by our own lower bound; one core, 16 Nov 2026, and the rows of 17 and 18 Nov with the tighter bound, checked by two independent checkers (for Switzerland, against Concorde’s optimal tour).Koliko je naša pot največ nad optimumom po naši lastni spodnji meji; eno jedro, 16. novembra 2026, in vrstice s 17. in 18. novembra s tesnejšo mejo, ki sta jih preverila dva neodvisna preverjevalnika (za Švico ob Concordovi optimalni poti).

    MapZemljevidCitiesMestaAbove the optimum, at mostNad optimumom, največSecondsSekundOur tourNaša potOur boundNaša meja
    wi29290%: matches Concorde’s proved optimum0 %: ujema se s Concordovim dokazanim optimumom0.00,027,60327.60327,60327.603
    dj38380%: matches Concorde’s proved optimum0 %: ujema se s Concordovim dokazanim optimumom0.00,06,6566.6566,6566.656
    qa1941940.93%0,93 %1.21,29,3529.3529,2669.266
    nu34963,4963.4960.70%0,70 %241 (longer run)241 (daljši tek)96,13296.13295,46995.469
    si6035 (Slovenia)si6035 (Slovenija)6,0356.0350.48%0,48 %1758,728,254 m8.728.254 m8,687,092 m8.687.092 m
    nu3496, 17 Novnu3496, 17. nov.3,4963.4960.44%0,44 %85496,13296.13295,71695.716
    si6035 (Slovenia), 17 Novsi6035 (Slovenija), 17. nov.6,0356.0350.19%0,19 %1328,728,254 m8.728.254 m8,712,125 m8.712.125 m
    Switzerland, 18 NovŠvica, 18. nov.2,8502.8500.065%0,065 %16,00416.0047,844,951 m (Concorde)7.844.951 m (Concorde)7,839,923 m7.839.923 m
    Italy, 17 NovItalija, 17. nov.16,77616.7760.209%0,209 %14,44714.44752,809,423 m52.809.423 m52,699,707 m52.699.707 m
    sw24978 (Sweden), 18 Novsw24978 (Švedska), 18. nov.24,97824.9780.297%0,297 %41,80141.801856,214856.214853,687853.687

    Concorde’s tighter bound for Slovenia, at most 0.006% (18 Nov 2026), comes from a different, exact program.Concordova tesnejša meja za Slovenijo, največ 0,006 % (18. nov. 2026), prihaja iz drugega, eksaktnega programa.

    How to read it.Kako brati. The optimum lies between the bound and the tour. When the two are equal, the tour is optimal; otherwise the tour is at most (tour − bound) / bound above the optimum, rounded up here. For qa194 and nu3496 the optimum is already known (the University of Waterloo’s National TSPs), and our tours have that length; the table shows how close our own certificate gets without that knowledge. For Slovenia nobody knows the optimum yet. Switzerland’s optimum, 7,844,951 m, was proved by Concorde on 17 Nov, on our laptop and on the NEOS Server; its row shows how close our own bound gets. For Italy, our map of 16,776 places, nobody knows the optimum; our own bound shows that our 10-minute tour, 52,809,423 m, is at most 0.209% above it. For Sweden the optimum is known, 855,597 (proved in 2004): it lies 0.224% above our own bound, and our 10-minute tour is 0.072% above it.Optimum leži med mejo in potjo. Če sta enaki, je pot optimalna; sicer je pot največ (pot − meja) / meja nad optimumom, tu zaokroženo navzgor. Za qa194 in nu3496 je optimum že znan (nacionalni TSP Univerze Waterloo) in naši poti imata to dolžino; tabela kaže, kako blizu pride naše lastno potrdilo brez tega znanja. Za Slovenijo optimuma še nihče ne pozna. Optimum Švice, 7.844.951 m, je Concorde dokazal 17. novembra, na našem prenosniku in na strežniku NEOS; njena vrstica pokaže, kako blizu pride naša lastna meja. Za Italijo, naš zemljevid s 16.776 kraji, optimuma nihče ne pozna; naša lastna meja pokaže, da je naša 10-minutna pot, 52.809.423 m, največ 0,209 % nad njim. Za Švedsko je optimum znan, 855.597 (dokazan leta 2004): leži 0,224 % nad našo lastno mejo, naša 10-minutna pot pa 0,072 % nad njim.

    Compared with Concorde.V primerjavi s Concordom. Concorde, still working on its proof for Slovenia on one core, had a lower bound of 8,727,732 m at 08:55 on 18 Nov: by its bound our best tour, 8,728,190 m, is at most 0.006% above the optimum. It is the best known tour; LKH 3.0.14 reached the same tour from scratch the same morning. Concorde’s bound is much tighter because it adds cutting planes; since 17 Nov our own bound is tighter too (at most 0.19%).Concorde, ki za Slovenijo na enem jedru še dela svoj dokaz, je imel ob 08:55 18. novembra spodnjo mejo 8.727.732 m: po njegovi meji je naša najboljša pot, 8.728.190 m, največ 0,006 % nad optimumom. To je najboljša znana pot; LKH 3.0.14 je isto pot našel od začetka isto jutro. Concordova meja je dosti tesnejša, ker dodaja rezalne ravnine; od 17. novembra je tesnejša tudi naša lastna meja (največ 0,19 %).

    Concorde with CPLEX, on NEOS.Concorde s CPLEX na strežniku NEOS. On 16 Nov we also gave both maps to Concorde with IBM’s LP solver CPLEX on the NEOS Server, a free online service for optimization, which ran them on its own computers, not on our laptop. nu3496: Concorde proved the optimum, 96,132, in 743.6 s of processor time (about 12 minutes). Its 3,496 cities lie at only 2,373 distinct locations. With its free LP solver QSopt, on one core of our laptop, the same proof took a median of 4,209 s of processor time, about 70 minutes, over three seeds (3,303 to 8,938 s; 16 and 11 Nov). The machines differ, so this is not a timing comparison. Slovenia: NEOS ends every job after 8 hours, and Concorde had no proof by then. Its best tour was 8,734,361 m, 0.071% above our best tour (8,728,190 m); its last lower bound was 8,727,217 m, with 214 nodes of its search tree still open. On 17 Nov NEOS also proved Switzerland’s optimum, 7,844,951 m: with fewer places than nu3496, it took Concorde with CPLEX 3.9 times the processor time (2,909.8 s) and 10 times the nodes (171).16. novembra smo oba zemljevida dali tudi Concordu z IBM-ovim reševalnikom LP CPLEX na strežniku NEOS, brezplačni spletni storitvi za optimizacijo, ki ju je reševala na svojih računalnikih, ne na našem prenosniku. nu3496: Concorde je dokazal optimum, 96.132, v 743,6 s procesorskega časa (okoli 12 minut). Njegovih 3.496 mest leži na le 2.373 različnih lokacijah. S svojim brezplačnim reševalnikom LP QSopt je na enem jedru našega prenosnika isti dokaz trajal v mediani 4.209 s procesorskega časa, okoli 70 minut, pri treh semenih (od 3.303 do 8.938 s; 16. in 11. novembra). Računalniki so različni, zato to ni primerjava časov. Slovenija: NEOS vsako nalogo ustavi po 8 urah in Concorde do takrat ni imel dokaza. Njegova najboljša pot je bila 8.734.361 m, 0,071 % nad našo najboljšo potjo (8.728.190 m); njegova zadnja spodnja meja je bila 8.727.217 m, v njegovem iskalnem drevesu pa je ostalo še 214 odprtih vozlišč. 17. novembra je NEOS dokazal tudi optimum Švice, 7.844.951 m: z manj kraji kot nu3496 je Concorde s CPLEX porabil 3,9-krat toliko procesorskega časa (2.909,8 s) in 10-krat toliko vozlišč (171).

    Next.Naslednji korak. Faster, with the work spread over several cores, and certificates beyond 20,000 cities. In the browser demo: one line under the result, »at most X% above the optimum«, for any map. Measured by the gap on nu3496 and Slovenia and the seconds it takes.Hitreje, z delom, razdeljenim na več jeder, in potrdila nad 20.000 mesti. V demu v brskalniku: ena vrstica pod rezultatom, »največ X % nad optimumom«, za vsak zemljevid. Merilo: razlika na nu3496 in Sloveniji ter sekunde, ki jih to vzame.

    Who did it.Kdo je to naredil. BD asked: »A bi lahko tudi mi naredili verzijo za dokazovanje optimuma?« (could we, too, make a version that proves the optimum?) and chose our own lower bound to start with. 8zTSP-K wrote it into the kernel on a development branch, with tests: 60 small random maps against trying every tour, and a convex polygon, where the bound must equal the tour.BD je vprašal: »A bi lahko tudi mi naredili verzijo za dokazovanje optimuma?« in za začetek izbral lastno spodnjo mejo. 8zTSP-K jo je napisal v jedro na razvojni veji, s testi: 60 majhnih naključnih zemljevidov proti preizkusu vseh poti in konveksni mnogokotnik, kjer mora biti meja enaka poti.

    Conditions: development runs of 16 Nov 2026, each on one core (logical CPU 4) of the laptop’s Intel i7-10875H at low priority, while other runs used other cores; the seconds are not a timing claim. The nu3496 row comes from a longer run than the default setting, 241 s; the default gives a slightly weaker bound in less time. Kernel branch feature/hk-bound, commit 521db57, not merged into main; the option is off by default, and this version gives no certificate beyond 20,000 cities. The tours are 8zTSP’s; for Slovenia it is our tour of 8,728,254 m, found onPogoji: razvojni teki 16. novembra 2026, vsak na enem jedru (logični CPU 4) procesorja Intel i7-10875H v prenosniku z nizko prioriteto, medtem ko so na drugih jedrih tekli drugi teki; sekunde niso trditev o času. Vrstica za nu3496 je iz daljšega teka od privzete nastavitve, 241 s; privzeta nastavitev da v krajšem času nekoliko šibkejšo mejo. Veja jedra feature/hk-bound, commit 521db57, ni združena v main; možnost je privzeto izklopljena in ta različica nad 20.000 mesti ne da potrdila. Poti so od 8zTSP; za Slovenijo je to naša pot 8.728.254 m, najdena 16 Nov16. novembra by combining three 8zTSP tours with the fusion method from BD’s arena (BD × GPT). Concorde’s bound is the last one in its log at 08:55 on 18 Nov. NEOS: »Concorde with CPLEX« with a random seed, each map sent as a TSPLIB file through NEOS’s web form on 16 Nov 2026 (Central European Time): nu3496 as job 24844471 on prod-exec-7, 00:05–00:18, 17 nodes in its search tree; Slovenia as job 24844820 on prod-exec-6, 00:15–08:17, stopped by NEOS’s 8-hour limit; its bound is the last one in the job’s log, 8,727,216.91, rounded up to whole metres. By NEOS’s FAQ, prod-exec-7 has two Intel Xeon Gold 5218 and prod-exec-6 two Intel Xeon Gold 6140, at 2.3 GHz. The seconds are Concorde’s own »Total Running Time«, the processor time of its process. On our laptop: Concorde (38a647d3) with QSopt, from its own first tour, seeds 7 and 99 on logical CPUs 1 and 2, 16 Nov 02:35–03:34 and 02:35–03:49, beside other runs; both proved 96,132 (95 and 107 nodes). The NEOS Server (z združitvijo treh poti 8zTSP z metodo zlivanja iz BD-jeve arene (BD × GPT). Concordova meja je zadnja v njegovem dnevniku ob 08:55 18. novembra. NEOS: »Concorde with CPLEX« z naključnim semenom, vsak zemljevid poslan kot datoteka TSPLIB prek NEOS-ovega spletnega obrazca 16. novembra 2026 (po srednjeevropskem času): nu3496 kot naloga 24844471 na prod-exec-7, 00:05–00:18, 17 vozlišč v iskalnem drevesu; Slovenija kot naloga 24844820 na prod-exec-6, 00:15–08:17, ustavljena ob NEOS-ovi meji 8 ur; njena meja je zadnja v dnevniku naloge, 8.727.216,91, zaokrožena navzgor na cele metre. Po NEOS-ovem FAQ ima prod-exec-7 dva Intel Xeon Gold 5218, prod-exec-6 pa dva Intel Xeon Gold 6140, pri 2,3 GHz. Sekunde so Concordov lastni »Total Running Time«, procesorski čas njegovega procesa. Na našem prenosniku: Concorde (38a647d3) s QSopt, od svoje prve poti, semeni 7 in 99 na logičnih CPU 1 in 2, 16. novembra 02:35–03:34 in 02:35–03:49, ob drugih tekih; oba sta dokazala 96.132 (95 in 107 vozlišč). Strežnik NEOS (neos-server.org) is hosted by the Wisconsin Institute for Discovery at the University of Wisconsin in Madison. As NEOS asks, we cite: Czyzyk, Mesnier and Moré, »The NEOS Server«, IEEE Journal on Computational Science and Engineering 5(3), 68–75 (1998); Dolan, »The NEOS Server 4.0 Administrative Guide«, Technical Memorandum ANL/MCS-TM-250, Argonne National Laboratory (2001); Gropp and Moré, »Optimization Environments and the NEOS Server«, in Approximation Theory and Optimization, M. D. Buhmann and A. Iserles, eds., Cambridge University Press, 167–182 (1997). Switzerland: the 2,850 places of our Europe map (the World TSP’s places) that lie in Switzerland (che_from_europe_20261117, sha256 1afc3b2e…); our bound 7,839,923 m from an LP run of 16,004 s from scratch on one core (for about 6 minutes other work used it too), 18 Nov 00:03–04:29 (certificate che_region.cert, sha256 1bf803b7…), accepted by both checkers (run again on 18 Nov at 04:34). Concorde’s optimal tour, 7,844,951 m: Concorde (38a647d3) with QSopt on our laptop, 17 Nov 14:03–15:03, 155 nodes; and Concorde with CPLEX on the NEOS Server, job 24917716 (submitted 17 Nov at 18:13), prod-exec-6, 171 nodes, 2,909.83 s of processor time. Both tours recomputed from the map. Italy: the 16,776 places of our Europe map in Italy (ita_from_europe_20261117, sha256 428112f7…); our bound 52,699,707 m from an LP run of 14,447 s on one core, 16:48–20:48 (certificate ita_std.cert, sha256 3cb43a50…), accepted by both checkers (run again on 17 Nov at 21:33); our tour, 52,809,423 m, from a 10-minute 8zTSP run with one thread (development build 0.6.0-dev, seed 1), recomputed from the map. Sweden: sw24978 from the University of Waterloo’s National TSPs (24,978 cities); our bound 853,687 from two LP runs on one core, 41,801 s in all: 14,422 s on 17 Nov (17:28–21:28), then 27,379 s continuing where that run stopped (17 Nov 21:28 to 18 Nov 05:04; between 02:08 and 02:13 a temperature guard limited it to a third of the core); certificate sha256 25d238c5…, accepted by both checkers (run again on 18 Nov at 05:07); our tour, 856,214, from a 10-minute 8zTSP run with one thread (development build 0.6.0-dev, seed 1), recomputed from the map.) gosti Wisconsin Institute for Discovery na Univerzi Wisconsin v Madisonu. Kot prosi NEOS, navajamo: Czyzyk, Mesnier in Moré, »The NEOS Server«, IEEE Journal on Computational Science and Engineering 5(3), 68–75 (1998); Dolan, »The NEOS Server 4.0 Administrative Guide«, Technical Memorandum ANL/MCS-TM-250, Argonne National Laboratory (2001); Gropp in Moré, »Optimization Environments and the NEOS Server«, v Approximation Theory and Optimization, ur. M. D. Buhmann in A. Iserles, Cambridge University Press, 167–182 (1997). Švica: 2.850 krajev našega zemljevida Evrope (kraji svetovnega TSP), ki ležijo v Švici (che_from_europe_20261117, sha256 1afc3b2e…); naša meja 7.839.923 m iz teka LP od začetka, dolgega 16.004 s, na enem jedru (približno 6 minut ga je uporabljalo tudi drugo delo), 18. novembra 00:03–04:29 (potrdilo che_region.cert, sha256 1bf803b7…), sprejeta v obeh preverjevalnikih (znova pognana 18. novembra ob 04:34). Concordova optimalna pot, 7.844.951 m: Concorde (38a647d3) s QSopt na našem prenosniku, 17. novembra 14:03–15:03, 155 vozlišč; in Concorde s CPLEX na strežniku NEOS, naloga 24917716 (poslana 17. novembra ob 18:13), prod-exec-6, 171 vozlišč, 2.909,83 s procesorskega časa. Obe poti sta preračunani iz zemljevida. Italija: 16.776 krajev našega zemljevida Evrope v Italiji (ita_from_europe_20261117, sha256 428112f7…); naša meja 52.699.707 m iz teka LP, dolgega 14.447 s, na enem jedru, 16:48–20:48 (potrdilo ita_std.cert, sha256 3cb43a50…), sprejeta v obeh preverjevalnikih (znova pognana 17. novembra ob 21:33); naša pot, 52.809.423 m, iz 10-minutnega teka 8zTSP z eno nitjo (razvojna gradnja 0.6.0-dev, seme 1), preračunana iz zemljevida. Švedska: sw24978 iz nacionalnih TSP Univerze Waterloo (24.978 mest); naša meja 853.687 iz dveh tekov LP na enem jedru, skupaj 41.801 s: 14.422 s 17. novembra (17:28–21:28), nato 27.379 s nadaljevanja od tam, kjer se je ta tek ustavil (od 17. novembra ob 21:28 do 18. novembra ob 05:04; med 02:08 in 02:13 ga je temperaturna zaščita omejila na tretjino jedra); potrdilo sha256 25d238c5…, sprejeta v obeh preverjevalnikih (znova pognana 18. novembra ob 05:07); naša pot, 856.214, iz 10-minutnega teka 8zTSP z eno nitjo (razvojna gradnja 0.6.0-dev, seme 1), preračunana iz zemljevida.

    ContentsKazalo

  9. · development run of 15 Nov; the proof is running· razvojni tek 15. novembra; dokaz teče

    Slovenia: one round trip through all 6,035 settlementsSlovenija: en krog skozi vseh 6.035 naselij

    Our new map of Slovenia has one point for each of its 6,035 settlements. In a run of 300 seconds with 7 threads, the best tour 8zTSP found was 8,728,465 m long. On 16 Nov a shorter tour, 8,728,254 m (about 8,728.3 km in straight lines), was found by combining three 8zTSP tours with the fusion method from BD’s arena (BD × GPT). A 24-hour run of 8zTSP’s development build found 8,728,252 m on 17 Nov. Best known tour 8,728,190 m. 8zTSP reached it on 18 Nov 2026 at 05:35 by fusing five of its own earlier tours (a continuation of earlier runs, not a run from scratch); LKH 3.0.14 reached the same tour from scratch on six cores the same morning. The two tours are identical. Concorde is working on the proof: at 08:55 on 18 Nov its lower bound was 8,727,732 m, so the best known tour is at most 0.006% above the optimum. Concorde was stopped once for disk maintenance and continued from its saved state on 17 Nov. The proof is running; the proven optimum follows.Naš novi zemljevid Slovenije ima po eno točko za vsako od 6.035 naselij. V teku 300 sekund s 7 nitmi je bila najboljša pot, ki jo je našel 8zTSP, dolga 8.728.465 m. 16. novembra je bila z združitvijo treh poti 8zTSP z metodo zlivanja iz BD-jeve arene (BD × GPT) najdena krajša pot, 8.728.254 m (okoli 8.728,3 km po zračnih črtah). 24-urni tek razvojne gradnje 8zTSP je 17. novembra našel 8.728.252 m. Najboljša znana pot 8.728.190 m. 8zTSP jo je dosegel 18. novembra 2026 ob 05:35 z zlitjem petih svojih prejšnjih poti (nadaljevanje prejšnjih tekov, ne tek od začetka); LKH 3.0.14 je isto pot našel od začetka na šestih jedrih isto jutro. Poti sta enaki. Concorde dela na dokazu: ob 08:55 18. novembra je bila njegova spodnja meja 8.727.732 m, zato je najboljša znana pot največ 0,006 % nad optimumom. Concorde smo enkrat ustavili zaradi vzdrževanja diska; 17. novembra je nadaljeval iz shranjenega stanja. Dokaz teče; dokazani optimum sledi. Run it in your browserPoženi ga v svojem brskalniku (Examples → Selected 10 → Slovenia · all towns and villages; the demo is password-protected, access on request), or take the map and the tour from(Primeri → Izbranih 10 → Slovenija · vsa mesta in vasi; demo je zaščiten z geslom, dostop na prošnjo) ali pa zemljevid in pot preneseš v razdelku DownloadsPrenosi.

    Our own bound gives our best tour at most 0.19% (Naša lastna meja da naši najboljši poti največ 0,19 % (17 Nov17. novembra); Concorde’s tighter bound, at most 0.006% (18 Nov 2026), comes from a different, exact program.); Concordova tesnejša meja, največ 0,006 % (18. nov. 2026), prihaja iz drugega, eksaktnega programa.

    More on SloveniaVeč o Sloveniji

    The map.Zemljevid. si6035_gurs_20261115: the settlements (NASELJA) in the Register of Spatial Units (RPE) of the Surveying and Mapping Authority of the Republic of Slovenia (GURS), taken on 15 Nov 2026, licence CC BY 4.0. Each point is the area-weighted centre of the settlement’s official area, in metres (D96/TM), rounded to whole metres. The distances are straight lines (EUC_2D in TSPLIB), not roads. The names of the settlements, in the order of the map’s numbers, are in Downloads.si6035_gurs_20261115: naselja (NASELJA) iz Registra prostorskih enot (RPE) Geodetske uprave Republike Slovenije (GURS), prevzeta 15. novembra 2026, licenca CC BY 4.0. Vsaka točka je težišče uradnega območja naselja v metrih (D96/TM), zaokroženo na cele metre. Razdalje so zračne črte (EUC_2D v TSPLIB), ne ceste. Imena naselij v vrstnem redu številk zemljevida so v Prenosih.

    Conditions: 8zTSP build 563f8fa of 15 Nov (in the kernel as main 1f072ad), the full program, 7 threads on logical CPUs 0–6 of the laptop’s Intel i7-10875H, seed 1, one run of 300 s on 15 Nov, 14:40–14:45; its best tour came after 226 s and was checked again (every settlement once, length 8,728,465 m). Concorde has been running on one core since 14:45; at the start it was given this tour’s length. The shorter tour of 16 Nov was found by combining three 8zTSP tours with the fusion method from BD’s arena (BD × GPT) and checked again (every settlement once, length 8,728,254 m). The 24-hour run, 8,728,252 m: 8zTSP development build 0.6.0-dev, one thread, seed 1, 24 hours from 17 Nov 05:35 to 18 Nov 05:35; found after 51,534 s and checked again. Our best tour and the best known tour, 8,728,190 m: at 05:35:37 on 18 Nov, one second after the 24-hour run ended, 8zTSP fused five of its own earlier tours (a continuation of earlier runs, not a run from scratch); checked again (every settlement once). LKH 3.0.14 reached the same tour from scratch as six copies with seeds 1–6 on logical CPUs 0–5, a limit of 1,200 s, in 8zTSP-media’s race capture of 18 Nov, 08:03–08:24; we checked it again (every settlement once) and compared the two tours edge by edge: they are identical. The lower bound is the last one in Concorde’s log at 08:55 on 18 Nov (8,727,731.23, rounded up).Pogoji: gradnja 8zTSP 563f8fa s 15. novembra (v jedru kot main 1f072ad), polni program, 7 niti na logičnih CPU 0–6 procesorja Intel i7-10875H v prenosniku, seme 1, en tek 300 s 15. novembra, 14:40–14:45; najboljšo pot je našel po 226 s, nato je bila ponovno preverjena (vsako naselje enkrat, dolžina 8.728.465 m). Concorde teče na enem jedru od 14:45; ob zagonu je dobil dolžino te poti. Krajša pot s 16. novembra je bila najdena z združitvijo treh poti 8zTSP z metodo zlivanja iz BD-jeve arene (BD × GPT) in ponovno preverjena (vsako naselje enkrat, dolžina 8.728.254 m). 24-urni tek, 8.728.252 m: razvojna gradnja 8zTSP 0.6.0-dev, ena nit, seme 1, 24 ur od 17. novembra ob 05:35 do 18. novembra ob 05:35; najdena po 51.534 s in ponovno preverjena. Naša najboljša pot in najboljša znana pot, 8.728.190 m: ob 05:35:37 18. novembra, sekundo po koncu 24-urnega teka, je 8zTSP zlil pet svojih prejšnjih poti (nadaljevanje prejšnjih tekov, ne tek od začetka); ponovno preverjena (vsako naselje enkrat). LKH 3.0.14 je isto pot našel od začetka kot šest kopij s semeni 1–6 na logičnih CPU 0–5, omejitev 1.200 s, v posnetku dirke 8zTSP-media 18. novembra, 08:03–08:24; ponovno smo jo preverili (vsako naselje enkrat) in obe poti primerjali povezavo za povezavo: sta enaki. Spodnja meja je zadnja v Concordovem dnevniku ob 08:55 18. novembra (8.727.731,23, zaokroženo navzgor).

    ContentsKazalo

  10. · development runs of 15 Nov· razvojni teki 15. novembra

    8zTSP got faster on many cores: nu3496 in about 6 seconds8zTSP je postal hitrejši na več jedrih: nu3496 v približno 6 sekundah

    The same search, but the cores no longer sit idle: with 7 threads, 6.25 of the 7 cores are at work on average instead of 3.7. The part of the work that does not spread over the cores fell from about a quarter to about 8%. The median time to the nu3496 optimum on 7 cores went from 9.2 to 6.1 seconds; all 10 runs found it.Isto iskanje, le da jedra ne stojijo več prazna: s 7 nitmi jih v povprečju dela 6,25 od 7 namesto 3,7. Del dela, ki se ne razdeli med jedra, je padel z okoli četrtine na okoli 8 %. Mediana časa do optimuma nu3496 na 7 jedrih je padla z 9,2 na 6,1 sekunde; našlo ga je vseh 10 tekov.

    More on the speed-up on many coresVeč o pospešitvi na več jedrih

    What changed.Kaj se je spremenilo. The largest part that did not spread was one shared step between rounds: one core did it while the others waited. Now each core does that step for its own range of cities, in exactly the old order, and the worker threads stay alive and take the next task as soon as they are free, instead of being started anew every round. The search itself is unchanged: the golden test set gives the same tours as before (154 of 155 cases identical; in the last one only the step written in one log line differs), and every seed ends in the same tour, found at the same step, with 1, 2, 4 and 7 cores.Največji del, ki se ni razdelil, je bil en skupni korak med krogi: delalo ga je eno jedro, druga so čakala. Zdaj vsako jedro ta korak opravi za svoj razpon mest, natanko v starem vrstnem redu, delovne niti pa ostanejo žive in vzamejo naslednjo nalogo, takoj ko so proste, namesto da bi se v vsakem krogu zagnale na novo. Samo iskanje je nespremenjeno: zlati preizkusni nabor da iste poti kot prej (154 od 155 primerov enakih; v zadnjem se razlikuje le korak, zapisan v eni vrstici dnevnika), vsako seme pa se z 1, 2, 4 in 7 jedri konča v isti poti, najdeni v istem koraku.

    nu3496: median seconds to the optimum with 1, 2, 4 and 7 real cores, one thread per core, before (build 673813b) and after (build 563f8fa of 15 Nov), and the cores at work on average. A first series of runs (6.0 s on 7 cores); the 10 seeds below are a second series (median 6.1 s).nu3496: mediana sekund do optimuma z 1, 2, 4 in 7 pravimi jedri, z eno nitjo na jedro, prej (gradnja 673813b) in potem (gradnja 563f8fa 15. novembra), ter povprečno število jeder pri delu. Prva serija tekov (6,0 s na 7 jedrih); 10 semen spodaj je druga serija (mediana 6,1 s).

    CoresJedraBeforePrejAfterPotemAt work beforePri delu prejAt work afterPri delu potem
    126.526,527.827,8—0.990,99
    216.416,415.215,2—1.901,90
    411.711,78.78,7—3.633,63
    79.29,26.06,03.73,76.256,25

    nu3496 with 7 threads: seconds to the optimum by seed, the new build (the full and the minimal program, 15 Nov) and the previous one (full program, build 673813b, 14 Nov).nu3496 s 7 nitmi: sekunde do optimuma po semenih, nova gradnja (polni in minimalni program, 15. novembra) in prejšnja (polni program, gradnja 673813b, 14. novembra).

    SeedSemeFull, newPolni, nova gradnjaMinimal, newMinimalni, nova gradnjaFull, beforePolni, prej
    15.65,65.75,79.19,1
    25.65,65.65,69.19,1
    317.317,318.018,027.427,4
    45.85,85.95,99.09,0
    55.75,75.75,78.68,6
    66.46,46.56,59.39,3
    745.145,143.243,264.064,0
    86.06,05.95,98.88,8
    917.817,817.717,726.626,6
    106.36,36.26,29.39,3
    MedianMediana6.16,16.16,19.29,2

    Seed by seed.Seme za semenom. With 7 threads each seed ran about 1.4 to 1.6 times as fast as with the previous build, and the slow seeds are the same as before (3, 7 and 9). The minimal program 8zTSP-min gained the same: median 6.1 s. 8zTSP-opt’s own series with the same build gave 6.0 s on 7 cores (the table above); the top of the page uses 8zTSP-K’s check, 6.1 s.S 7 nitmi je vsako seme teklo okoli 1,4- do 1,6-krat tako hitro kot s prejšnjo gradnjo, počasna semena pa so ista kot prej (3, 7 in 9). Minimalni program 8zTSP-min je pridobil enako: mediana 6,1 s. Lastna serija 8zTSP-opt z isto gradnjo je dala 6,0 s na 7 jedrih (tabela zgoraj); vrh strani uporablja preverjanje 8zTSP-K, 6,1 s.

    What does not spread yet.Kaj se še ne razdeli. From the speed-ups on 2, 4 and 7 cores, about 8% of the work does not yet spread over the cores, down from about a quarter. As before, that is an upper estimate: the laptop’s clock also drops as more cores work.Iz pospeškov na 2, 4 in 7 jedrih sledi, da se okoli 8 % dela še ne razdeli med jedra, prej okoli četrtina. Kot prej je to zgornja ocena: prenosniku ob več delujočih jedrih pade tudi takt.

    Open.Odprto. On one core the new build needs about 5% more time than the previous one (27.8 s against 26.5 s); a fast path for a single thread comes next. Compiled for this laptop’s processor alone (native code), one core needed about 5% less time than with the portable build; profile-guided optimisation added nothing to that. The published program stays the portable build.Na enem jedru nova gradnja potrebuje okoli 5 % več časa kot prejšnja (27,8 s proti 26,5 s); naslednja je hitra pot za eno nit. Prevedeno samo za procesor tega prenosnika (nativna koda) je eno jedro potrebovalo okoli 5 % manj časa kot s prenosljivo gradnjo; optimizacija na podlagi profila temu ni dodala ničesar. Objavljeni program ostaja prenosljiva gradnja.

    Who did it.Kdo je to naredil. BD asked for it: »dajmo se lotiti še optimizacije izvajanja« (let’s also optimise how it runs). The GPT session 8zTSP-opt did the work from 8zTSP-K’s brief; 8zTSP-K checked it with the 10 seeds above and merged it into the kernel (main 1f072ad).BD je to želel: »dajmo se lotiti še optimizacije izvajanja«. Delo je po briefu 8zTSP-K opravila GPT-seja 8zTSP-opt; 8zTSP-K ga je preveril z 10 semeni zgoraj in ga vključil v jedro (main 1f072ad).

    Conditions: development runs of 15 Nov 2026 on the laptop’s Intel i7-10875H (8 cores, 16 threads), nothing else running. 8zTSP-opt, 13:52–14:11: the portable build 563f8fa, 1, 2, 4 and 7 real cores with one thread per core (logical CPUs 0, 0–1, 0–3 and 0–6), 40 runs; one core native and with profile-guided optimisation, 14:12–14:30, 20 runs. 8zTSP-K, 14:34–14:39: the full and the minimal program of 563f8fa, 7 threads on logical CPUs 0–6, alternating, 20 runs. nu3496, seeds 1–10, at most 600 s, stopped from outside the program when the tour reached 96,132, the known optimum; the program is not told the optimum; every final tour checked again (every city once, length 96,132). Cores at work: the processor time the program used up to the optimum divided by the elapsed time, sampled every 50 ms, averaged over the 10 runs; before: one sampled run (seed 1) of build 673813b. The medians before: the entry below (1, 2 and 4 cores on 15 Nov, 7 cores on 14 Nov).Pogoji: razvojni teki 15. novembra 2026 na procesorju Intel i7-10875H našega prenosnika (8 jeder, 16 niti), nič drugega ni teklo. 8zTSP-opt, 13:52–14:11: prenosljiva gradnja 563f8fa, 1, 2, 4 in 7 pravih jeder z eno nitjo na jedro (logični CPU 0, 0–1, 0–3 in 0–6), 40 tekov; eno jedro nativno in z optimizacijo na podlagi profila, 14:12–14:30, 20 tekov. 8zTSP-K, 14:34–14:39: polni in minimalni program 563f8fa, 7 niti na logičnih CPU 0–6, izmenično, 20 tekov. nu3496, semena 1–10, največ 600 s, ustavljeno od zunaj, ko je pot dosegla 96.132, znani optimum; programu optimum ni povedan; vsaka končna pot ponovno preverjena (vsako mesto enkrat, dolžina 96.132). Jedra pri delu: procesorski čas, ki ga je program porabil do optimuma, deljen s pretečenim časom, vzorčeno vsakih 50 ms, povprečje 10 tekov; prej: en vzorčeni tek (seme 1) gradnje 673813b. Mediane prej: zapis spodaj (1, 2 in 4 jedra 15. novembra, 7 jeder 14. novembra).

    ContentsKazalo

  11. · development runs of 14 and 15 Nov· razvojni teki 14. in 15. novembra

    More real cores, shorter runs: nu3496 from about 27 seconds on one core to about 9 on sevenVeč pravih jeder, krajši teki: nu3496 od okoli 27 sekund na enem jedru do okoli 9 na sedmih

    On our laptop the native program reached the nu3496 optimum in all 40 runs with 1, 2, 4 and 7 real cores, one thread per core; the median time fell from 26.5 s on one core to 16.4 s on two, 11.7 s on four and 9.2 s on seven. The search takes the same path on any number of cores; more cores make each generation shorter.Na našem prenosniku je nativni program optimum nu3496 dosegel v vseh 40 tekih z 1, 2, 4 in 7 pravimi jedri, z eno nitjo na jedro; mediana časa je padla s 26,5 s na enem jedru na 16,4 s na dveh, 11,7 s na štirih in 9,2 s na sedmih. Iskanje gre po isti poti pri vsakem številu jeder; več jeder skrajša vsako generacijo.

    More on cores and run timesVeč o jedrih in času tekov

    How it was measured.Kako smo merili. nu3496, seeds 1–10, the full program (build 673813b), alone on the laptop, one run at a time, at most 600 s per run, stopped from outside when the tour reached 96,132, the known optimum; the program is not told the optimum. One thread per real core, the other hyperthread of each core idle: logical CPU 0 for one core, 0–1 for two and 0–3 for four, on 15 Nov; the seven-core runs are the same build’s runs of 14 Nov on logical CPUs 0–6.nu3496, semena 1–10, polni program (gradnja 673813b), sam na prenosniku, en tek naenkrat, največ 600 s na tek, ustavljeno od zunaj, ko je pot dosegla 96.132, znani optimum; program optimuma ne pozna. Ena nit na pravo jedro, druga nit vsakega jedra je mirovala: logični CPU 0 za eno jedro, 0–1 za dve in 0–3 za štiri, 15. novembra; teki na sedmih jedrih so teki iste gradnje 14. novembra na logičnih CPU 0–6.

    nu3496: seconds to the optimum by seed, with 1, 2, 4 and 7 real cores (full program, build 673813b).nu3496: sekunde do optimuma po semenih, z 1, 2, 4 in 7 pravimi jedri (polni program, gradnja 673813b).

    SeedSeme1 core1 jedro2 cores2 jedri4 cores4 jedra7 cores7 jeder
    125.325,315.615,611.211,29.19,1
    225.525,515.815,811.211,29.19,1
    379.279,249.449,434.434,427.427,4
    425.425,415.715,711.611,69.09,0
    525.025,015.415,410.910,98.68,6
    627.727,717.117,111.811,89.39,3
    7191.8191,8122.2122,285.085,064.064,0
    825.525,515.915,911.111,18.88,8
    978.378,348.848,833.833,826.626,6
    1027.627,616.916,912.212,29.39,3
    MedianMediana26.526,516.416,411.711,79.29,2

    Seed by seed.Seme za semenom. Compared with one core, each seed ran about 1.6 times as fast on two cores, about 2.3 times on four and about 2.9 times on seven, with little spread from seed to seed. The slow seeds are the same at every count (3, 7 and 9): the search does not depend on the number of cores, and more cores make each generation shorter.V primerjavi z enim jedrom je vsako seme na dveh jedrih teklo okoli 1,6-krat hitreje, na štirih okoli 2,3-krat in na sedmih okoli 2,9-krat, z majhnim raztrosom med semeni. Počasna semena so pri vsakem številu jeder ista (3, 7 in 9): iskanje ni odvisno od števila jeder, več jeder pa skrajša vsako generacijo.

    What does not spread yet.Kaj se še ne razdeli. From these speed-ups, about a quarter of the work does not yet spread over the cores of our laptop. That is an upper estimate, because the laptop’s clock also drops as more cores work.Iz teh pospeškov sledi, da se okoli četrtina dela na našem prenosniku še ne razdeli med jedra. To je zgornja ocena, ker prenosniku ob več delujočih jedrih pade tudi takt.

    Next.Naslednji korak. Find the quarter that does not spread and spread it over the cores. Measured the same way: nu3496, 1, 2, 4 and 7 real cores, 10 seeds, against the table above.Najti četrtino, ki se ne razdeli, in jo razdeliti med jedra. Merjeno enako: nu3496, 1, 2, 4 in 7 pravih jeder, 10 semen, proti tabeli zgoraj. (done 15 Nov, see above)(narejeno 15. nov., glej zgoraj)

    Conditions: development runs of 15 Nov 2026, 11:55–12:14 (1, 2 and 4 cores) and of 14 Nov 2026 (7 cores, the check of build 673813b by 8zTSP-K); 8zTSP, the full program, build 673813b (binary 42a3d624…, the same as the native build of 7f98b90); the laptop’s Intel i7-10875H (8 cores, 16 threads), one thread per real core on logical CPUs 0, 0–1, 0–3 and 0–6, nothing else running; seeds 1–10, at most 600 s, stopped at the optimum from outside the program; all 40 final tours checked again for this entry: every city once, length 96,132, the same input file.Pogoji: razvojni teki 15. novembra 2026, 11:55–12:14 (1, 2 in 4 jedra) in 14. novembra 2026 (7 jeder, preverjanje gradnje 673813b, 8zTSP-K); 8zTSP, polni program, gradnja 673813b (binarna datoteka 42a3d624…, enaka nativni gradnji 7f98b90); prenosnikov Intel i7-10875H (8 jeder, 16 niti), ena nit na pravo jedro na logičnih CPU 0, 0–1, 0–3 in 0–6, nič drugega ni teklo; semena 1–10, največ 600 s, ustavljeno ob optimumu od zunaj programa; vseh 40 končnih poti je bilo za ta vnos znova preverjenih: vsako mesto enkrat, dolžina 96.132, ista vhodna datoteka.

    ContentsKazalo

  12. · development runs of 15 Nov· razvojni teki 15. novembra

    In the browser the workers now share one search: nu3496 in about 13 secondsV brskalniku si delavci zdaj delijo eno iskanje: nu3496 v okoli 13 sekundah

    The browser version of 8zTSP now works like the native program: its 7 workers share one search instead of running separate tries. On our laptop it reached the nu3496 optimum in 10 of 10 runs, typically in about 13 seconds (the native program needs about 9 seconds with the same build; separate tries were estimated at about 38).Različica 8zTSP za brskalnik zdaj deluje kot nativni program: njenih 7 delavcev si deli eno iskanje, namesto da bi vsak izvajal svoj poskus. Na našem prenosniku je optimum nu3496 dosegla v 10 od 10 tekov, običajno v okoli 13 sekundah (nativni program z isto gradnjo potrebuje okoli 9 sekund; za ločene poskuse smo ocenili okoli 38).

    Try it in your browserPreizkusite v svojem brskalniku (password-protected; access on request): the workers share one search where your browser offers shared memory; by default the page uses all the cores your device reports, and a run on nu3496 stops by itself at the known optimum.(zaščiteno z geslom; dostop na prošnjo): delavci si delijo eno iskanje, kjer vaš brskalnik ponuja skupni pomnilnik; stran privzeto uporabi vsa jedra, ki jih javlja vaša naprava, tek na nu3496 pa se sam ustavi pri znanem optimumu.

    More on the shared search in the browserVeč o skupnem iskanju v brskalniku

    How it was measured.Kako smo merili. nu3496, seeds 1–10, 7 workers in one browser (headless Chromium 155) on logical CPUs 0–6 of our laptop, nothing else running, at most 600 s per run, a fresh browser profile for each seed. The test page watched the length shown and stopped the run when it reached 96,132, the known optimum; the solver itself is not told the optimum. The times are the solver’s, from the moment the 7 workers had started; with the start of the workers, the page’s clock shows under a tenth of a second more.nu3496, semena 1–10, 7 delavcev v enem brskalniku (Chromium 155 brez zaslona, headless) na logičnih CPU 0–6 našega prenosnika, nič drugega ni teklo, največ 600 s na tek, za vsako seme svež profil brskalnika. Testna stran je spremljala prikazano dolžino in ustavila tek, ko je dosegla 96.132, znani optimum; reševalnik sam optimuma ne pozna. Časi so reševalnikovi, od trenutka, ko je bilo 7 delavcev zagnanih; z zagonom delavcev vred ura strani pokaže manj kot desetinko sekunde več.

    nu3496: seconds to the optimum by seed, in the browser with 7 shared workers and natively with 7 threads (8zTSP-min, the run of 14 Nov below).nu3496: sekunde do optimuma po semenih, v brskalniku s 7 delavci s skupnim pomnilnikom in nativno s 7 nitmi (8zTSP-min, tek 14. novembra spodaj).

    SeedSemeBrowser, 7 workersBrskalnik, 7 delavcevNative, 7 threadsNativno, 7 niti
    112.612,69.19,1
    212.612,69.19,1
    338.038,027.427,4
    412.212,28.98,9
    512.012,08.68,6
    613.013,09.29,2
    789.889,864.364,3
    812.412,48.88,8
    937.337,326.426,4
    1013.013,09.49,4
    MedianMediana12.812,89.29,2
    MeanPovprečje25.325,318.118,1
    WorstNajslabši89.889,864.364,3

    Seed by seed.Seme za semenom. The browser took about 1.4 times as long as the native program on every seed, and the slow seeds are the same in both (3, 7 and 9). With a fixed number of steps instead of a time limit, the browser build with 7 shared workers ends with exactly the native build’s tour (nu3496 after 1,500 steps, DIMACS E100k.0 after one pass; 8zTSP-web’s checks): the same engine, in the browser more slowly.Brskalnik je pri vsakem semenu potreboval okoli 1,4-krat toliko časa kot nativni program, počasna semena pa so v obeh ista (3, 7 in 9). S fiksnim številom korakov namesto časovne omejitve gradnja za brskalnik s 7 delavci s skupnim pomnilnikom konča s točno isto potjo kot nativna gradnja (nu3496 po 1.500 korakih, DIMACS E100k.0 po enem prehodu; preverbe seje 8zTSP-web): isti motor, v brskalniku počasnejši.

    Written down before the run.Zapisano pred tekom. On 14 Nov we noted what would count: separate tries in the browser would need about 38 s (the 25 s estimate in the entry below, times the browser’s 1.5 from the pilot), and one shared search would count as a success at a median of up to 21.2 s, twice the 10.6 s native median we knew then (build 303eee6, 7 threads, 30 seeds). The median was 12.8 s.14. novembra smo zapisali, kaj bo štelo: ločeni poskusi v brskalniku bi potrebovali okoli 38 s (ocena 25 s iz vnosa spodaj, krat 1,5 za brskalnik iz pilota), eno skupno iskanje pa bo štelo za uspeh pri mediani do 21,2 s, dvakratniku nativne mediane 10,6 s, ki smo jo takrat poznali (gradnja 303eee6, 7 niti, 30 semen). Mediana je bila 12,8 s.

    The native program is unchanged.Nativni program je nespremenjen. The new code is only in the browser part (the threads compiled for WebAssembly); built natively, this version gives byte for byte the same programs as the one tested on 14 Nov, so all native results stand. Each of the 10 browser tours was checked again (by 8zTSP-K, and once more for this entry): every city once, length 96,132, the same input file.Nova koda je le v delu za brskalnik (niti, prevedene za WebAssembly); nativno prevedena ta različica da bajt za bajtom enaka programa kot tista, preizkušena 14. novembra, zato vsi nativni izidi veljajo. Vsako od 10 poti iz brskalnika je bilo znova preverjeno (8zTSP-K in še enkrat za ta vnos): vsako mesto enkrat, dolžina 96.132, ista vhodna datoteka.

    What it needs.Kaj potrebuje. The shared version needs a page that is cross-origin isolated, which takes two HTTP headers: only there can workers share memory (SharedArrayBuffer). The demo page now sends them, only on its own address and files, and opens with a password; where a browser does not offer shared memory, the demo falls back to separate tries and says so.Skupna različica potrebuje stran, ki je izolirana med izvori (cross-origin isolation), za kar sta potrebni dve glavi HTTP: le tam si delavci lahko delijo pomnilnik (SharedArrayBuffer). Predstavitvena stran ju zdaj pošilja, le na svojem naslovu in svojih datotekah, in se odpre z geslom; kjer brskalnik skupnega pomnilnika ne ponuja, preide na ločene poskuse in to napiše.

    Next.Naslednji korak. During the test the laptop used about 4 cores on average, not 7. Before we call the factor of about 1.4 the cost of WebAssembly alone, we measure how busy the browser workers and the native threads are on the same seeds: if the browser workers wait, we cut the waiting; if not, we look for the cost of WebAssembly on one core. Measured the same way: nu3496, 7 workers, 10 seeds, against the table above.Med preizkusom je prenosnik v povprečju uporabljal okoli 4 jedra, ne 7. Preden faktor okoli 1,4 pripišemo samemu WebAssemblyju, izmerimo, kako zaposleni so delavci v brskalniku in nativne niti na istih semenih: če delavci v brskalniku čakajo, čakanje skrajšamo; če ne, iščemo ceno WebAssemblyja na enem jedru. Merjeno enako: nu3496, 7 delavcev, 10 semen, proti tabeli zgoraj.

    Conditions: development runs of 15 Nov 2026, 01:18–01:23; kernel 7f98b90 (main 2a635dd), 8zTSP-min compiled to WebAssembly with threads (wasm-bindgen-rayon), tsp8z_web_bg.wasm sha256 6152b09c…, the same in all 10 runs; headless Chromium 155 on the laptop’s Intel i7-10875H, 7 workers on logical CPUs 0–6, nothing else running (other processes 0.05–0.12 cores); seeds 1–10, at most 600 s, stopped at the optimum from outside the solver. Native comparison: 8zTSP-min with 7 threads, kernel 673813b, the run of 14 Nov (entry below). Load: the whole laptop about 4 cores on average during the runs (30-second samples). Native builds of 7f98b90 with Rust 1.99.1: byte-identical to those of 673813b.Pogoji: razvojni teki 15. novembra 2026, 01:18–01:23; jedro 7f98b90 (main 2a635dd), 8zTSP-min, preveden v WebAssembly z nitmi (wasm-bindgen-rayon), tsp8z_web_bg.wasm sha256 6152b09c…, enak v vseh 10 tekih; Chromium 155 brez zaslona (headless) na prenosnikovem Intel i7-10875H, 7 delavcev na logičnih CPU 0–6, nič drugega ni teklo (drugi procesi 0,05–0,12 jedra); semena 1–10, največ 600 s, ustavljeno ob optimumu od zunaj reševalnika. Nativna primerjava: 8zTSP-min s 7 nitmi, jedro 673813b, tek 14. novembra (vnos spodaj). Obremenitev: celoten prenosnik med teki v povprečju okoli 4 jedra (vzorci po 30 sekund). Nativni gradnji 7f98b90 z Rust 1.99.1: bajtno enaki gradnjama 673813b.

    ContentsKazalo

  13. · development runs of 14 Nov· razvojni teki 14. novembra

    8zTSP-min: half the code, the same speed on nu3496, and the same engine in a browser8zTSP-min: pol kode, enaka hitrost na nu3496 in isti motor v brskalniku

    BD’s idea was to switch the parts of 8zTSP off one by one and keep only what carries the results: 8zTSP-min has the search for small maps (EAX with diversity-aware selection) and, for large ones, our own view of the map (BD’s idea) with our addition; it has about half the code of the full 8zTSP and about a quarter of LKH’s. With 7 threads it finds the nu3496 optimum just as fast as the full build (about 9 seconds in a typical run, 10 of 10 runs), and compiled to WebAssembly the same Rust code found it in a browser in about 40 seconds on one core.BD‑jeva zamisel je bila, da dele 8zTSP izklapljamo enega za drugim in obdržimo le tisto, kar nosi izide: 8zTSP-min ima za majhne zemljevide iskanje (EAX z izbiro z raznolikostjo), za velike pa naš lastni pogled na zemljevid (BD‑jeva zamisel) z našim dodatkom; ima okoli pol kode polnega 8zTSP in okoli četrtino kode LKH. S 7 nitmi najde optimum nu3496 enako hitro kot polna gradnja (okoli 9 sekund v običajnem teku, 10 od 10 tekov), prevedena v WebAssembly pa je ista koda v Rustu optimum v brskalniku našla v okoli 40 sekundah na enem jedru.

    Try it in your browserPreizkusite v svojem brskalniku: nu3496, random points or your own TSPLIB file; one worker is one core, and the speed depends on your computer.: nu3496, naključne točke ali vaša datoteka TSPLIB; en delavec je eno jedro, hitrost pa je odvisna od vašega računalnika.

    More on 8zTSP-minVeč o 8zTSP-min

    How it was built.Kako je nastal. With switches at build time (Rust features). The full build stays the default and gives the same results as before on 154 of the 155 cases of the golden test set (the remaining one differs only in the moment one log line is written; the tour is the same); the minimal build gives the same results on all 52 cases it supports. The kernel gains no dependencies: it still uses only Rust’s standard library. The browser version is the same Rust code compiled to WebAssembly with a small wrapper (wasm-bindgen), not a rewrite in JavaScript.S stikali ob gradnji (Rustove »features«). Polna gradnja ostaja privzeta in da na 154 od 155 primerov referenčnega nabora preizkusov enake izide kot prej (preostali se razlikuje le po trenutku, ko se zapiše ena vrstica dnevnika; tura je ista); minimalna gradnja da enake izide na vseh 52 primerih, ki jih podpira. Jedro ne dobi nobene odvisnosti: še vedno uporablja le standardno knjižnico Rusta. Različica za brskalnik je ista koda v Rustu, prevedena v WebAssembly z majhnim ovojem (wasm-bindgen), ne prepis v JavaScript.

    Code.Koda. Characters without spaces and comments, in the parts compiled into each build: 8zTSP-min 98,967; full 8zTSP 189,017; LKH 3.0.7 403,521; Concorde 2,217,976 plus the QSopt LP library. The 194,444 in the table above counts all of the kernel’s source files (src/*.rs), with their inline tests and every part that can be switched off.Znaki brez presledkov in komentarjev, v delih, ki se prevedejo v posamezno gradnjo: 8zTSP-min 98.967; polni 8zTSP 189.017; LKH 3.0.7 403.521; Concorde 2.217.976 in še knjižnica za linearno programiranje QSopt. Število 194.444 v tabeli zgoraj šteje vse izvorne datoteke jedra (src/*.rs), z vgrajenimi testi in vsemi deli, ki jih je mogoče izklopiti.

    Seven threads.Sedem niti. nu3496, seeds 1–10, nothing else running on the laptop: both builds found the optimum in 10 of 10 runs; median 9.2 s each, mean 18.1 s, the longest run about 64 s. Seed by seed the two take the same path.nu3496, semena 1–10, na prenosniku ni teklo nič drugega: obe gradnji sta optimum našli v 10 od 10 tekov; mediana po 9,2 s, povprečje 18,1 s, najdaljši tek okoli 64 s. Seme za semenom gresta obe po isti poti.

    nu3496 with 7 threads: seconds to the optimum by seed.nu3496 s 7 nitmi: sekunde do optimuma po semenih.

    SeedSemeFull 8zTSPPolni 8zTSP8zTSP-min
    19.19,19.19,1
    29.19,19.19,1
    327.427,427.427,4
    49.09,08.98,9
    58.68,68.68,6
    69.39,39.29,2
    764.064,064.364,3
    88.88,88.88,8
    926.626,626.426,4
    109.39,39.49,4
    MedianMediana9.29,29.29,2
    MeanPovprečje18.118,118.118,1
    WorstNajslabši64.064,064.364,3

    In the browserV brskalniku (a first pilot: one seed, one Web Worker, one core, Chromium): the nu3496 optimum after 37 s, natively 24 s with the same build, seed and core (the browser about 1.5 times slower); DIMACS E100k.0 after 2 minutes 0.87% above the Held–Karp bound in the browser, 0.80% natively. One seed each: a pilot, not a median.(prvi pilot: eno seme, en Web Worker, eno jedro, Chromium): optimum nu3496 po 37 s, nativno 24 s z isto gradnjo, semenom in jedrom (brskalnik okoli 1,5-krat počasnejši); DIMACS E100k.0 po 2 minutah 0,87 % nad mejo Held–Karp v brskalniku, 0,80 % nativno. Po eno seme: pilot, ne mediana.

    Next.Naslednji korak. At present several Web Workers run independent tries. On nu3496 that does not go much below about 25 s, even at native speed (the fastest of 7 independent tries, estimated from our 30 one-thread runs); the native threads share one search and need about 10 s. So the next version shares one search between the browser workers, measured with 7 workers and 10 seeds against the native 7 threads.Zdaj več Web Workerjev izvaja neodvisne poskuse. Na nu3496 to niti z nativno hitrostjo ne gre dosti pod okoli 25 s (najhitrejši od 7 neodvisnih poskusov, ocena iz naših 30 enonitnih tekov); nativne niti si delijo eno iskanje in potrebujejo okoli 10 s. Zato bo naslednja različica delila eno iskanje med delavci v brskalniku, merjena s 7 delavci in 10 semeni proti nativnim 7 nitim. (done 15 Nov, see above)(narejeno 15. nov., glej zgoraj)

    Conditions: development runs of 14 Nov 2026; kernel 673813b (main 9f7670a), the full build with the default switches, the minimal build with all of them off; the laptop’s Intel i7-10875H. Seven threads: logical CPUs 0–6, nothing else running, seeds 1–10, at most 600 s, stop at the optimum, the two builds alternating, 21:23–21:29; times from the lengths the kernel reported, final tours checked. On 7 of the 10 seeds the logged improvements of the two builds are identical; on the other 3 one row differs, because the log is written by the clock. Browser pilot: logical CPU 15, Chromium 155, Rust 1.99.1, seed 1; in the browser the time is the solver’s, from the moment the map has been read; natively it runs from the start of the program. Code: the same counting script as in the table above, over the parts each build compiles (8zTSP-web’s projection of the source, without macro expansion). The estimate for independent tries: the median of the fastest of 7, over all ways to pick 7 of the 30 one-thread runs of build 303eee6 (seeds 1–30, one run at a time on logical CPU 3): 24.7 s.Pogoji: razvojni teki 14. nov. 2026; jedro 673813b (main 9f7670a), polna gradnja s privzetimi stikali, minimalna z vsemi izklopljenimi; Intel i7-10875H v prenosniku. Sedem niti: logični CPU 0–6, ob tem ni teklo nič drugega, semena 1–10, največ 600 s, ustavitev ob optimumu, gradnji izmenično, 21:23–21:29; časi iz dolžin, ki jih je javilo jedro, končne ture preverjene. Na 7 od 10 semen sta zapisa izboljšav obeh gradenj enaka; na ostalih 3 se razlikuje ena vrstica, ker se dnevnik piše po uri. Pilot v brskalniku: logični CPU 15, Chromium 155, Rust 1.99.1, seme 1; v brskalniku je čas reševalnikov, od trenutka, ko je zemljevid prebran; nativno teče od zagona programa. Koda: isti skript za štetje kot v tabeli zgoraj, prek delov, ki jih posamezna gradnja prevede (projekcija izvorne kode seje 8zTSP-web, brez razširitve makrov). Ocena za neodvisne poskuse: mediana najhitrejšega od 7, prek vseh načinov izbire 7 od 30 enonitnih tekov gradnje 303eee6 (semena 1–30, en tek naenkrat na logičnem CPU 3): 24,7 s.

    ContentsKazalo

  14. · development runs of 14 Nov· razvojni teki 14. novembra

    Which parts carry the result: one switch at a timeKateri deli nosijo izid: eno stikalo naenkrat

    On nu3496 with 7 threads the default search had the shortest median time (9.6 s); each change lengthened the median (to 24–49 s), although some changes were faster on single seeds; our addition’s choice of partners mostly did not reach the optimum within 10 minutes. On 100,000 cities, three things clearly matter on all three seeds: our own order through the view (plain parts were worse), the quadrant neighbours, and BD’s MDL judge in our analysis of the view, which was worse at this run length. With more seeds (second round, 14 Nov evening): on 100,000 cities 3-opt in our analysis helps on 10 of 10 seeds and our addition on 8 of 10, each by about a hundredth of a percentage point or less; warm-up shows no clear effect. On nu3496, 3-opt was faster than the default on 20 of 30 seeds and has a slightly better mean, but a longer median: it trades the fast runs for shorter slow ones.Na nu3496 s 7 nitmi je imelo privzeto iskanje najkrajšo mediano časa (9,6 s); vsaka sprememba je mediano podaljšala (na 24–49 s), čeprav so bile nekatere spremembe hitrejše na posameznih semenih; z izbiro staršev našega dodatka optimum v 10 minutah večinoma ni bil dosežen. Pri 100.000 mestih so na vseh treh semenih jasno pomembne tri stvari: naš lastni vrstni red skozi pogled (goli deli so bili slabši), kvadrantni sosedje in BD‑jev sodnik MDL v naši analizi pogleda, ki je bil pri tej dolžini teka slabši. Z več semeni (drugi krog, 14. nov. zvečer): pri 100.000 mestih 3-opt v naši analizi pomaga na 10 od 10 semen in naš dodatek na 8 od 10, vsak za okoli stotinko odstotne točke ali manj; ogrevanje nima jasnega učinka. Na nu3496 je bil 3-opt hitrejši od privzetega na 20 od 30 semen in ima nekoliko boljše povprečje, a daljšo mediano: hitre teke zamenja za krajše počasne.

    More on which parts carry the resultVeč o tem, kateri deli nosijo izid
    Second round: nu3496, 7 threads, seeds 1–30, 600 s, alternatingDrugi krog: nu3496, 7 niti, semena 1–30, po 600 s, izmeničnoFound the optimumNašel optimumMedianMedianaMeanPovprečjeWorstNajslabšiFaster than the defaultHitrejši od privzetega
    Default (small maps)Privzeto (majhni zemljevidi)30 of 3030 od 309.4 s9,4 s19.3 s19,3 s64.0 s64,0 s—
    3-opt movesPoteze 3-opt30 of 3030 od 3021.3 s21,3 s18.3 s18,3 s64.5 s64,5 son 20 of 30 seedsna 20 od 30 semen
    Second round: DIMACS E100k.0, our analysis of the view, 7 threads, seeds 1–10, 120 sDrugi krog: DIMACS E100k.0, naša analiza pogleda, 7 niti, semena 1–10, 120 sAbove the Held–Karp bound (mean)Nad mejo Held–Karp (povprečje)Worse than the race settingsSlabše od dirkalnih nastavitevDifference (mean, percentage points)Razlika (povprečje, odstotne točke)
    Race settingsDirkalne nastavitve0.741%0,741 %——
    Our addition off (null control)Naš dodatek izklopljen (ničelna kontrola)0.747%0,747 %on 8 of 10 seedsna 8 od 10 semen+0.006+0,006
    Without warm-upBrez ogrevanja0.742%0,742 %on 7 of 10 seedsna 7 od 10 semen+0.001+0,001
    2-opt/Or-opt instead of 3-opt2-opt/Or-opt namesto 3-opt0.752%0,752 %on 10 of 10 seedsna 10 od 10 semen+0.011+0,011

    Conditions of the second round: build 303eee6, one run at a time on logical CPUs 0–6, 14 Nov 17:49–19:29; idle sessions sat on CPUs 7 and 15. nu3496: seeds 1–30, the default and 3-opt alternating seed by seed, 600 s each, time to the first tour of length 96,132. By seed, both settings show two typical times: the default about 9 s (17 of 30 runs) or about 27 s (10), 3-opt about 7 s (12) or about 22 s (15); three runs of each took over 30 s. DIMACS E100k.0: seeds 1–10, 120 s each, the gap of the final checked tour above the Held–Karp bound (224,330,692.1), averaged; »worse« counts the seeds on which the variant ended above the race settings with the same seed.Pogoji drugega kroga: gradnja 303eee6, en tek naenkrat na logičnih CPU 0–6, 14. novembra 17:49–19:29; na CPU 7 in 15 so mirovale seje. nu3496: semena 1–30, privzeto in 3-opt izmenično seme za semenom, po 600 s, čas do prve ture dolžine 96.132. Po semenih imata obe nastavitvi dva značilna časa: privzeta okoli 9 s (17 od 30 tekov) ali okoli 27 s (10), 3-opt okoli 7 s (12) ali okoli 22 s (15); po trije teki vsake so trajali več kot 30 s. DIMACS E100k.0: semena 1–10, po 120 s, odmik končne preverjene ture nad mejo Held–Karp (224.330.692,1), povprečen; »slabše« šteje semena, na katerih je različica končala nad dirkalnimi nastavitvami z istim semenom.

    First round: nu3496, 7 threads, 10 seeds, 600 sPrvi krog: nu3496, 7 niti, 10 semen, 600 sFound the optimumNašel optimumMedian time to the optimumMediana časa do optimuma
    Default (small maps)Privzeto (majhni zemljevidi)10 of 1010 od 109.6 s9,6 s
    Without the diversity-aware selection (--no-entropy)Brez izbire z raznolikostjo (--no-entropy)10 of 1010 od 1025.0 s25,0 s
    BD’s multi-resolution start (--start t02:on)BD‑jev začetek v več ločljivostih (--start t02:on)10 of 1010 od 1026.8 s26,8 s
    Our addition’s choice of partnersIzbira staršev našega dodatka1 of 10 (498 s)1 od 10 (498 s)the others 0.006–0.095% above the optimum after 600 sostali po 600 s 0,006–0,095 % nad optimumom
    3-opt movesPoteze 3-opt10 of 1010 od 1023.9 s23,9 s
    LK movesPoteze LK10 of 1010 od 1049.3 s49,3 s
    First round: DIMACS E100k.0, our analysis of the view, 7 threads, 3 seeds, 120 sPrvi krog: DIMACS E100k.0, naša analiza pogleda, 7 niti, 3 semena, 120 sAbove the Held–Karp bound at 120 s (mean)Nad mejo Held–Karp po 120 s (povprečje)
    Race settingsDirkalne nastavitve0.749%0,749 %
    Our addition off (null control)Naš dodatek izklopljen (ničelna kontrola)0.754%0,754 %
    Without warm-upBrez ogrevanja0.757%0,757 %
    Plain parts of the view (without our addition)Goli deli pogleda (brez našega dodatka)0.767%0,767 %
    Without quadrant neighboursBrez kvadrantnih sosedov0.775%0,775 %
    2-opt/Or-opt instead of 3-opt2-opt/Or-opt namesto 3-opt0.756%0,756 %
    BD’s MDL judge of our analysis onBD‑jev sodnik MDL naše analize vklopljen0.791%0,791 %

    Conditions of the first round: build 303eee6, one run at a time on logical CPUs 0–6. nu3496: seeds 1–10, 600 s each, time to the first tour of length 96,132; the default row is the 7-thread series with seeds 1–10 (median 9.6 s; with 30 seeds 10.6 s, as in the box at the top). The first block ran 13:18–14:20 with LKH on CPU 7; the runs with a different start and with our addition’s partners were repeated 15:43–17:26, because the arena had dropped their option at the colon the first time (a fault of the tool, not of the kernel); during the repeat idle sessions sat on CPUs 7 and 15. DIMACS E100k.0 has a published Held–Karp bound (224,330,692.1); seeds 42, 142 and 242, the gap at 120 s averaged. Seed by seed against the race settings: the plain parts, no quadrant neighbours and the MDL judge were worse on 3 of 3 seeds; our addition off, no warm-up and 2-opt/Or-opt were worse on 2 of 3 seeds (on seed 242 slightly better).Pogoji prvega kroga: gradnja 303eee6, en tek naenkrat na logičnih CPU 0–6. nu3496: semena 1–10, po 600 s, čas do prve ture dolžine 96.132; vrstica »privzeto« je 7-nitna serija s semeni 1–10 (mediana 9,6 s; s 30 semeni 10,6 s, kot v okvirju na vrhu). Prvi blok je tekel 13:18–14:20, ob njem LKH na CPU 7; teka z drugačnim začetkom in s starši našega dodatka sta bila ponovljena 15:43–17:26, ker je arena prvič izgubila njuno opcijo pri dvopičju (napaka orodja, ne jedra); med ponovitvijo so na CPU 7 in 15 mirovale seje. DIMACS E100k.0 ima objavljeno mejo Held–Karp (224.330.692,1); semena 42, 142 in 242, odmik po 120 s povprečen. Po semenih v primerjavi z dirkalnimi nastavitvami: goli deli, brez kvadrantnih sosedov in sodnik MDL so bili slabši na 3 od 3 semen; naš dodatek izklopljen, brez ogrevanja in 2-opt/Or-opt so bili slabši na 2 od 3 semen (na semenu 242 nekoliko boljši).

    First round, nu3496 with 7 threads: seconds to the optimum by seed; in brackets, where a run did not find it within 600 s, the gap above it.Prvi krog, nu3496 s 7 nitmi: sekunde do optimuma po semenih; v oklepaju, kjer tek optimuma v 600 s ni našel, odmik nad njim.

    SeedSemeDefaultPrivzetoNo diversity selectionBrez izbire z raznolikostjoBD’s multi-resolution startBD‑jev začetek v več ločljivostih3-optLKPartners of our additionStarši našega dodatka
    19.49,412.712,79.29,224.624,649.749,7(0.010%)(0,010 %)
    29.29,240.640,627.227,224.924,917.017,0(0.037%)(0,037 %)
    328.728,731.531,59.19,17.77,7112.8112,8498.0498,0
    49.49,428.128,125.925,97.57,5265.6265,6(0.006%)(0,006 %)
    59.29,221.721,744.444,424.024,0108.8108,8(0.057%)(0,057 %)
    69.99,921.821,826.826,88.08,017.217,2(0.034%)(0,034 %)
    769.469,413.113,166.666,625.125,181.281,2(0.033%)(0,033 %)
    89.49,431.831,826.826,823.723,716.116,1(0.036%)(0,036 %)
    928.428,433.933,926.726,77.77,749.049,0(0.074%)(0,074 %)
    1010.110,14.34,39.29,224.424,416.716,7(0.095%)(0,095 %)
    MedianMediana9.69,625.025,026.826,823.923,949.349,3—
    MeanPovprečje19.319,323.923,927.227,217.817,873.473,4—
    WorstNajslabši69.469,440.640,666.666,625.125,1265.6265,6—

    ContentsKazalo

  15. · development runs of 14 Nov· razvojni teki 14. novembra

    Concorde at 100,000 cities: it starts and needs under 1 GB, but in 2 hours it does not get close to a proofConcorde pri 100.000 mestih: zažene se in porabi manj kot 1 GB, a v 2 urah ni niti blizu dokaza

    Concorde’s exact solver ran on DIMACS E100k.0 for 2 hours on one core with at most about 0.7 GB of memory. Its own best tour (228,109,622) and its proven lower bound after 16 rounds (225,686,848) were still about 1.1% apart, so a proof of the optimum was far away. The limit is time, not memory.Concordov natančni reševalec je na DIMACS E100k.0 tekel 2 uri na enem jedru z največ okoli 0,7 GB pomnilnika. Njegova najboljša tura (228.109.622) in dokazana spodnja meja po 16 krogih (225.686.848) sta bili še vedno okoli 1,1 % narazen, do dokaza optimuma je bilo torej daleč. Meja je čas, ne pomnilnik.

    More on Concorde at 100,000 citiesVeč o Concordu pri 100.000 mestih

    Conditions: Concorde 03.12.19, the 64-bit build with QSopt, default settings, seed 42; logical CPU 6, 14 Nov 11:02:56–13:02:56, stopped at the 2-hour limit; beside it the no-EAX test on CPUs 0–5 and LKH on CPU 7: a test of feasibility, not a race; memory sampled every 10 s (peak 663,516 kB as the kernel reports it). Its lower bound rose from round to round (round 16 after 6,589 s).Pogoji: Concorde 03.12.19, 64-bitni prevod s QSopt, privzete nastavitve, seme 42; logični CPU 6, 14. novembra 11:02:56–13:02:56, ustavljen ob meji 2 h; ob njem test brez EAX na CPU 0–5 in LKH na CPU 7: preizkus izvedljivosti, ne dirka; pomnilnik vzorčen vsakih 10 s (vrh 663.516 kB, kot ga poroča jedro sistema). Njegova spodnja meja je rasla iz kroga v krog (16. krog po 6.589 s).

    ContentsKazalo

  16. · development runs of 14 Nov· razvojni teki 14. novembra

    nu3496 without EAX: BD’s DCC and MDL judge make the kick search three times closernu3496 brez EAX: z BD‑jevim DCC in sodnikom MDL je iskanje s sunki trikrat bližje optimumu

    With EAX switched off (kick mode, one thread, 10 minutes), none of 40 runs found the optimum; the median gap was 1.14% with 2-opt/Or-opt, 0.39% with LK moves, 0.15% with BD’s DCC governor and 0.11% with BD’s MDL judge. With EAX the same build finds the optimum in about 30 s on one thread.Z izklopljenim EAX (način sunkov, ena nit, 10 minut) optimuma ni našel noben od 40 tekov; mediana odmika je bila 1,14 % z 2-opt/Or-opt, 0,39 % s potezami LK, 0,15 % z BD‑jevim krmilnikom DCC in 0,11 % z BD‑jevim sodnikom MDL. Z EAX ista gradnja na eni niti optimum najde v okoli 30 s.

    More on nu3496 without EAXVeč o nu3496 brez EAX
    nu3496, kick mode, one thread, 10 minnu3496, način sunkov, ena nit, 10 minMedian above the optimumMediana nad optimumomWorst of 10Najslabši od 10
    2-opt / Or-opt1.14%1,14 %1.53%1,53 %
    LK movesPoteze LK0.39%0,39 %0.90%0,90 %
    LK + BD’s DCC governorLK + BD‑jev krmilnik DCC0.15%0,15 %0.21%0,21 %
    LK + BD’s MDL judgeLK + BD‑jev sodnik MDL0.11%0,11 %0.16%0,16 %

    Conditions: development build 303eee6 in its kick mode (--mode ils) with four settings: 2-opt/Or-opt moves; LK moves (--moves lk); LK moves with BD’s DCC governor (--dcc on, limits 0.109375 and 0.125 from the hooks2 calibration); LK moves with BD’s MDL judge (--judge on). Seeds 1–10, 600 s each, six runs at a time on logical CPUs 0–5, 14 Nov 11:03–12:13; LKH ran on CPU 7 and Concorde on CPU 6 at the same time. For comparison: LKH with its default settings was 0.047% above the optimum after 10 min (11 Nov).Pogoji: razvojna gradnja 303eee6 v načinu sunkov (--mode ils) s štirimi nastavitvami: poteze 2-opt/Or-opt; poteze LK (--moves lk); poteze LK z BD‑jevim krmilnikom DCC (--dcc on, meji 0,109375 in 0,125 iz umerjanja hooks2); poteze LK z BD‑jevim sodnikom MDL (--judge on). Semena 1–10, po 600 s, šest tekov hkrati na logičnih CPU 0–5, 14. novembra 11:03–12:13; hkrati sta na CPU 7 tekel LKH in na CPU 6 Concorde. Za primerjavo: LKH s privzetimi nastavitvami je bil po 10 min 0,047 % nad optimumom (11. novembra).

    ContentsKazalo

  17. · nu3496, 14 Nov· nu3496, 14. novembra

    nu3496: the optimal tour in about 10 seconds with 7 threads; all 90 runs with 1, 7 and 14 threads found itnu3496: optimalna pot v približno 10 sekundah s 7 nitmi; našlo jo je vseh 90 tekov z 1, 7 in 14 nitmi

    Median 28.6 s on one core, 10.6 s with 7 threads and 9.8 s with 14 threads (both hyperthreads of 7 cores), 30 runs each. LKH did not find the optimum in 6 hours (a later run of 15–17 Nov: not after 16.8 hours either, best 96,152); Concorde proved it in a median of 70 min of processor time (three seeds: 55, 70 and 149 min).Mediana 28,6 s na enem jedru, 10,6 s s 7 nitmi in 9,8 s s 14 nitmi (po dve niti na vsakem od 7 jeder), po 30 tekov. LKH optimuma v 6 urah ni našel (poznejši tek 15.–17. nov.: tudi po 16,8 ure ne, najboljše 96.152); Concorde ga je dokazal v mediani 70 min procesorskega časa (tri semena: 55, 70 in 149 min).

    More on nu3496 with 1, 7 and 14 threadsVeč o nu3496 z 1, 7 in 14 nitmi

    The new build 303eee6 ran nu3496 (3,496 cities in Nicaragua, optimum 96,132) thirty times each with 1, 7 and 14 threads, one run at a time, and found the optimal tour in every run. With more threads the median came sooner: 28.6 s on one core, 10.6 s with 7 threads and 9.8 s with 14 threads (both hyperthreads of 7 cores; the laptop has 8 cores and 16 threads). For comparison, on one core of the same laptop: LKH 3.0.7 did not find the optimum in 6 hours (one run; 0.028% above it; a later run of 15–17 Nov had not reached it after 16.8 hours, best 96,152); Concorde proved it in a median of 70 min of processor time (three seeds: 55, 70 and 149 min; in the run with seed 42, on 11 Nov, the optimal tour appeared after 148 min by the clock and the proof after 169 min). On 11 Nov, after 30 s, LKH was 0.085% above the optimum (its first route came after 19 s) and Concorde 0.542%. 8zTSP finds the optimal tour; Concorde proves it is optimal. Our earlier arena 8Z-RP was still 0.688% above the optimum after a 16-hour run and stalled at 0.674%.Nova gradnja 303eee6 je nu3496 (3.496 mest v Nikaragvi, optimum 96.132) pognala po tridesetkrat z 1, 7 in 14 nitmi, en tek naenkrat, in optimalno pot našla v vsakem teku. Z več nitmi je mediana prišla prej: 28,6 s na enem jedru, 10,6 s s 7 nitmi in 9,8 s s 14 nitmi (po dve niti na vsakem od 7 jeder; prenosnik ima 8 jeder in 16 niti). Za primerjavo na enem jedru istega prenosnika: LKH 3.0.7 optimuma v 6 urah ni našel (en tek; 0,028 % nad njim; poznejši tek 15.–17. nov. ga tudi po 16,8 ure ni dosegel, najboljše 96.152); Concorde ga je dokazal v mediani 70 min procesorskega časa (tri semena: 55, 70 in 149 min; v teku s semenom 42, 11. novembra, se je optimalna pot pojavila po 148 min po uri, dokaz pa po 169 min). 11. novembra je bil LKH po 30 s 0,085 % nad optimumom (prvo pot je imel po 19 s), Concorde pa 0,542 %. 8zTSP optimalno pot najde, Concorde pa dokaže, da je optimalna. Naša prejšnja arena 8Z-RP je po 16-urnem teku ostala 0,688 % nad optimumom in se ustavila pri 0,674 %.

    nu3496, threadsnu3496, nitiRunsTekiFound the optimumNašel optimumTime to the optimum: min / median / maxČas do optimuma: najmanj / mediana / največMeanPovprečjeWithin 30 sV 30 s
    1303024.6 s / 28.6 s / 3.1 min24,6 s / 28,6 s / 3,1 min57.1 s57,1 s17 of 3017 od 30
    730309.2 s / 10.6 s / 69.4 s9,2 s / 10,6 s / 69,4 s21.1 s21,1 s23 of 3023 od 30
    1430308.5 s / 9.8 s / 63.5 s8,5 s / 9,8 s / 63,5 s19.6 s19,6 s27 of 3027 od 30

    Why it is hard.Zakaj je težko. A route through 3,496 cities can run in (3,495!)/2 different ways, about 10Pot skozi 3.496 mest lahko poteka na (3.495!)/2 različnih načinov, približno 1010,86810.868: a number with 10,869 digits. One thread of this laptop checks about 67 million tours per second (measured for 60 s on 14 Nov); checking all of them would take about 10: število z 10.869 števkami. Ena nit tega prenosnika preveri okoli 67 milijonov tur na sekundo (merjeno 60 s, 14. nov.); preizkus vseh bi trajal okoli 1010,85310.853 years, the age of the universe times 10let, kar je starost vesolja, pomnožena z 1010,84310.843. 8zTSP finds the optimal tour in about 10 seconds.. 8zTSP najde optimalno pot v približno 10 sekundah.

    CitiesMestaTime to check every tour on one threadČas za preizkus vseh poti na eni niti
    20about 29 yearsokoli 29 let
    24about 6 million yearsokoli 6 milijonov let
    26about 3.7 billion years (a quarter of the age of the universe)okoli 3,7 milijarde let (četrtina starosti vesolja)
    27about 96 billion years (about 7 times the age of the universe)okoli 96 milijard let (okoli 7-krat starost vesolja)
    29 (wi29, the smallest national instance)29 (wi29, najmanjši državni primer)about 5,000 times the age of the universe; 8zTSP: optimum in 0.00 sokoli 5.000-krat starost vesolja; 8zTSP: optimum v 0,00 s

    (n − 1)!/2 different tours for n cities; 6.67 × 10(n − 1)!/2 različnih poti za n mest; 6,67 × 107 tours per second on one thread (measured for 60 s on 14 Nov); the age of the universe 13.8 billion years.poti na sekundo na eni niti (merjeno 60 s, 14. nov.); starost vesolja 13,8 milijarde let.

    A little history.Malo zgodovine. In 1987 the largest TSP solved with a proof of optimality had 2,392 cities (pr2392). The next milestone came in 1994: 7,397 cities (pla7397; Applegate, Bixby, Chvátal and Cook). A map of 3,496 cities was beyond the record of 1987. Today, on one laptop, Concorde proves the optimum of nu3496 in about 55 minutes, and 8zTSP finds the optimal tour in about 10 seconds.Leta 1987 je imel največji TSP, rešen z dokazom optimalnosti, 2.392 mest (pr2392). Naslednji mejnik je prišel leta 1994: 7.397 mest (pla7397; Applegate, Bixby, Chvátal in Cook). Zemljevid s 3.496 mesti je bil nad rekordom iz leta 1987. Danes na enem prenosniku Concorde optimum nu3496 dokaže v približno 55 minutah, 8zTSP pa optimalno pot najde v približno 10 sekundah.

    Conditions: development build 303eee6 (sha256 of the binary 6ff6e90d…), default mode on nu3496 (the mode for small maps, used up to 12,000 cities), 14 Nov; times run from the start of the program and come from the lengths the kernel reported; the final tours were checked. Seeds 1–30 for each of 1, 7 and 14 threads, one run at a time, the one-thread runs on logical CPU 3; LKH ran on the eighth core (logical CPU 7) during all these runs: shared power and heat. First measurement, the morning of 14 Nov, before LKH’s run: one thread, seeds 1–10, six runs at a time on logical CPUs 0–5 (the procedure of the 0.5.0 study): the optimum in 10 of 10 runs, median 33 s. LKH 3.0.7: default settings, seed 42, one run on logical CPU 7 from 09:42 on 14 Nov (limit 6 h; its best 96,159 = 0.028% above the optimum after 20,758 s). Concorde 03.12.19 with QSopt, the same program in all three runs, on one core, from its own starting tour; processor time to the proof: 3,303 s (seed 7, 16 Nov), 4,209 s (seed 99, 16 Nov) and 8,938 s (seed 42, 11 Nov: a long side run with a limit of 4 h on a core shared with other jobs, where the optimal tour appeared after 148.1 min and the proof after 169.1 min by the clock). LKH and Concorde at 30 s: from their logs of 11 Nov (one run with one seed, at the same time as other programs, each on its own core). 8Z-RP: 96,793 at the end of its 16-hour run (1 September 2026), its best 96,780 found on 5 September 2026 (Pogoji: razvojna gradnja 303eee6 (sha256 binarne datoteke 6ff6e90d…), privzeti način na nu3496 (način za majhne zemljevide, ki se uporablja do 12.000 mest), 14. novembra; časi tečejo od zagona programa in izhajajo iz dolžin, ki jih je javilo jedro; končne poti so bile preverjene. Semena 1–30 za vsako od 1, 7 in 14 niti, en tek naenkrat, enonitni teki na logičnem CPU 3; LKH je med vsemi temi teki tekel na osmem jedru (logični CPU 7): skupna moč in toplota. Prva meritev, 14. novembra zjutraj, pred tekom LKH: ena nit, semena 1–10, šest tekov hkrati na logičnih CPU 0–5 (postopek študije 0.5.0): optimum v 10 od 10 tekov, mediana 33 s. LKH 3.0.7: privzete nastavitve, seme 42, en tek na logičnem CPU 7 od 09:42 14. novembra (meja 6 h; njegova najboljša 96.159 = 0,028 % nad optimumom po 20.758 s). Concorde 03.12.19 s QSopt, isti program v vseh treh tekih, na enem jedru, od svoje lastne začetne poti; procesorski čas do dokaza: 3.303 s (seme 7, 16. novembra), 4.209 s (seme 99, 16. novembra) in 8.938 s (seme 42, 11. novembra: dolg stranski tek z mejo 4 h na jedru, deljenem z drugimi posli, kjer se je optimalna pot pojavila po 148,1 min, dokaz pa po 169,1 min po uri). LKH in Concorde pri 30 s: iz njunih dnevnikov 11. novembra (en tek z enim semenom, hkrati z drugimi programi, vsak na svojem jedru). 8Z-RP: 96.793 ob koncu njegovega 16-urnega teka (1. septembra 2026), njegova najboljša pot 96.780 najdena 5. septembra 2026 (nu3496 statusstanje nu3496). Milestones: the TSP pages of the University of Waterloo, where nu3496 is one of the national instances with a known optimum. nu3496 has 3,496 cities but only 2,373 distinct locations; 1,840 cities share their location with one or more others.). Mejniki: strani TSP Univerze Waterloo, kjer je nu3496 eden od državnih primerov z znanim optimumom. nu3496 ima 3.496 mest, a le 2.373 različnih lokacij; 1.840 mest si lokacijo deli z vsaj enim drugim.

    ContentsKazalo

  18. · run of the morning of 14 Nov· tek 14. novembra dopoldne

    The new build with 7 threads, 3.16 million cities: at the 1% line after 4 min 43 sNova gradnja s 7 nitmi, 3,16 milijona mest: na črti 1 % po 4 min 43 s

    With 7 threads at 3.16M the new build reached the 1% line after 4 min 43 s, against 6 min 16 s for 0.6.0-z1 in the pilot; after 1 h it was 0.743% above the estimated reference.S 7 nitmi je nova gradnja pri 3,16M črto 1 % dosegla po 4 min 43 s, proti 6 min 16 s za 0.6.0-z1 v pilotu; po 1 h je bila 0,743 % nad ocenjeno referenco.

    More on 3.16 million cities with 7 threadsVeč o 3,16 milijona mest s 7 nitmi

    As at 10 million cities, we ran the new 8zTSP (development build 303eee6: our own view of the map with our addition, without LKH) with 7 threads, one per core, on the same map and in the same arena as the one-core test. It reached the 1% line after 4 min 43 s and ended 0.743% above the estimated reference after 1 h. The older 0.6.0-z1 with 7 threads needed 6 min 16 s in the pilot of 12 Nov; the new build on one core needed 19 min 14 s (0.825%). All times run from the start of the program and include reading the map and all preparation.Kot pri 10 milijonih mest smo novi 8zTSP (razvojna gradnja 303eee6: naš lastni pogled na zemljevid z našim dodatkom, brez LKH) pognali s 7 nitmi, po eno na jedro, na istem zemljevidu in v isti areni kot enojedrni preizkus. Črto 1 % je dosegel po 4 min 43 s in po 1 h končal 0,743 % nad ocenjeno referenco. Starejši 0.6.0-z1 s 7 nitmi je v pilotu 12. novembra potreboval 6 min 16 s, nova gradnja na enem jedru pa 19 min 14 s (0,825 %). Vsi časi tečejo od zagona programa in vključujejo branje zemljevida in vso pripravo.

    3,162,278 cities3.162.278 mestNew build, 7 threadsNova gradnja, 7 niti0.6.0-z1, 7 threads (pilot of 12 Nov)0.6.0-z1, 7 niti (pilot 12. novembra)New build, 1 thread (14 Nov)Nova gradnja, 1 nit (14. novembra)
    To the 1% lineDo črte 1 %4 min 43 s6 min 16 s19 min 14 s
    Best checked route in 1 h, above the estimated referenceNajboljša preverjena pot v 1 h, nad ocenjeno referenco0.743%0,743 %0.808%0,808 %0.825%0,825 %

    Conditions: DIMACS E3M.0, seed 42, one run of at most 1 h, 14 Nov; 7 threads on logical CPUs 0–6 (one per physical core) of the laptop’s Intel i7-10875H, on average at 2.91 GHz, the hottest zone at a median of 92 °C; the same memory limit as the one-core test (12 GB); no cooldown wait before the start. From about 35 to 37 min into the run, a text search of the page session used one thread of the eighth core (logical CPUs 7 and 15); CPUs 0–6 were not shared. The pilot run of 0.6.0-z1 (12 Nov) ran in the development arena, with development runs beside it on the laptop’s eighth core; the new build ran in the arena that saves and checks every counted route. The reference is 0.7080·√(N·A), an estimate of the Held–Karp bound, not a proven bound and not the optimum.Pogoji: DIMACS E3M.0, seme 42, en tek, največ 1 h, 14. novembra; 7 niti na logičnih CPU 0–6 (po ena na fizično jedro) procesorja Intel i7-10875H v prenosniku, v povprečju pri 2,91 GHz, najtoplejša cona z mediano 92 °C; enaka omejitev pomnilnika kot v enojedrnem preizkusu (12 GB); brez čakanja na ohladitev pred zagonom. Med približno 35. in 37. minuto teka je iskanje po besedilu seje strani zasedlo eno nit osmega jedra (logična CPU 7 in 15); CPU 0–6 niso bili deljeni. Pilotni tek 0.6.0-z1 (12. novembra) je tekel v razvojni areni, ob njem so na osmem jedru prenosnika tekli razvojni teki; nova gradnja je tekla v areni, ki shrani in preveri vsako šteto pot. Referenca je 0,7080·√(N·A), ocena meje Held–Karp, ne dokazana meja in ne optimum.

    ContentsKazalo

  19. · national instances, 14 Nov· državni primeri, 14. novembra

    Nine national instances: the optimal tour in 90 of 90 runsDevet državnih primerov: optimalna pot v 90 od 90 tekov

    The new build 303eee6 ran the study of release 0.5.0 again (ten seeds per instance, one core each, six runs at a time) and found the optimal tour in all 90 runs, with a shorter median time than 0.5.0 on uy734, zi929, lu980, rw1621, mu1979 and nu3496 (nu3496: 33 s against 43 s). The table is in the results (Nova gradnja 303eee6 je ponovila študijo izdaje 0.5.0 (deset semen na primer, vsak tek na enem jedru, šest tekov hkrati) in optimalno pot našla v vseh 90 tekih, s krajšim mediannim časom kot 0.5.0 na uy734, zi929, lu980, rw1621, mu1979 in nu3496 (nu3496: 33 s proti 43 s). Tabela je med rezultati (Nine national instances: the new buildDevet državnih primerov: nova gradnja); nu3496 with more threads has its own entry above.); nu3496 z več nitmi ima svoj vnos zgoraj.

    ContentsKazalo

  20. · fair one-core test of 14 Nov· pošteni enojedrni preizkus 14. novembra

    One core each, 1 million cities: 8zTSP at the 1% line after 5 min 33 s, LKH after 8 min 9 s, linkern after 5 min 51 sVsak na enem jedru, 1 milijon mest: 8zTSP je bil na črti 1 % po 5 min 33 s, LKH po 8 min 9 s, linkern po 5 min 51 s

    After 30 min on one core each: 8zTSP 0.772% above the Held–Karp bound, linkern 0.850%, LKH 0.868%. All three ran alone on the same core.Po 30 min, vsak na enem jedru: 8zTSP 0,772 % nad mejo Held–Karp, linkern 0,850 %, LKH 0,868 %. Vsi trije so tekli sami na istem jedru.

    More on 1 million cities, one core eachVeč o 1 milijonu mest, vsak na enem jedru

    As at 10 and 3.16 million cities, we ran 8zTSP, LKH and Concorde’s linkern one after another, each alone on the laptop, on the same single core, after the same cooldown, here for 30 minutes each, in the arena that saves and checks every route it counts. The new 8zTSP (development build 303eee6: our own view of the map with our addition, without LKH, one thread) reached the 1% line after 5 min 33 s and ended 0.772% above the Held–Karp bound. By its own log, LKH spent its first 6 minutes on preparation. LKH’s first route came after 8 min 9 s, already below the 1% line (0.950% above the Held–Karp bound); it improved it ten times in the 30 minutes and ended 0.868% above it. linkern reached the 1% line after 5 min 51 s and ended 0.850% above the Held–Karp bound. All times run from the start of each program and include reading the map and all preparation.Kot pri 10 in 3,16 milijona mest smo 8zTSP, LKH in Concordov linkern pognali enega za drugim, vsakega samega na prenosniku, na istem jedru, po enakem ohlajanju, tu po 30 minut, v areni, ki shrani in preveri vsako pot, ki jo šteje. Novi 8zTSP (razvojna gradnja 303eee6: naš lastni pogled na zemljevid z našim dodatkom, brez LKH, ena nit) je črto 1 % dosegel po 5 min 33 s in končal 0,772 % nad mejo Held–Karp. Po lastnem dnevniku je LKH prvih 6 minut porabil za pripravo. Prva pot LKH je prišla po 8 min 9 s, že pod črto 1 % (0,950 % nad mejo Held–Karp); v teh 30 minutah jo je desetkrat izboljšal in končal 0,868 % nad njo. linkern je črto 1 % dosegel po 5 min 51 s in končal 0,850 % nad mejo Held–Karp. Vsi časi tečejo od zagona programa in vključujejo branje zemljevida in vso pripravo.

    One core each, 1,000,000 citiesVsak na enem jedru, 1.000.000 mest8zTSPLKH 3.0.7linkern
    First route checked by the arenaPrva pot, ki jo je preverila arena16 s8 min 9 s22 s
    To the 1% lineDo črte 1 %5 min 33 s8 min 9 s5 min 51 s
    Best checked route in 30 min, above the Held–Karp boundNajboljša preverjena pot v 30 min, nad mejo Held–Karp0.772%0,772 %0.868%0,868 %0.850%0,850 %
    Pilot of 12 Nov, 8zTSP with 7 threads: to the 1% linePilot 12. novembra, 8zTSP s 7 nitmi: do črte 1 %2 min 3 s13 min 31 s9 min 56 s

    Conditions: DIMACS E1M.0, seed 42, one run per program, at most 30 min each, one after another on 14 Nov; logical CPU 3 of the laptop’s Intel i7-10875H, its twin thread left idle, on average at 4.43 (8zTSP), 4.53 (LKH) and 4.61 GHz (linkern), the hottest zone at a median of 100, 100 and 98 °C; before each start the hottest zone had to stay at or below 80 °C for 60 s (BD, 14 Nov); a memory limit of 12 GB per run. LKH 3.0.7 ran with the settings chosen for 1,000,000 cities in the pilot (POPMUSIC candidates, no subgradient ascent, POPMUSIC start tour). linkern ran in the race’s 64-bit build. The reference is the published Held–Karp bound of E1M.0, 708,703,512.9; the 1% line lies 1% above it. Routes are saved at set moments and whenever a route crosses 3, 2, 1.5 or 1%; the arena checks each one. LKH writes its route only when a trial ends; its last route of the 30 minutes, written after about 28 min, fell outside the arena’s last time band, so we checked that file after the run: a valid route, 0.868% above the Held–Karp bound. By their own logs, 8zTSP had its first route after 0.2 s and linkern after 5.8 s. Pilot row: 12 Nov, 8zTSP 0.6.0-z1 with 7 threads; LKH and linkern as the best of 4 one-core runs at the same time, with other work at the lowest priority on the eighth core. Earlier one-core runs at this size (11–12 Nov, with the older 8zTSP 0.5.0): 8zTSP 13.2 min, linkern 15.4 min, LKH 15.5 min.Pogoji: DIMACS E1M.0, seme 42, en tek na program, največ 30 min vsak, eden za drugim 14. novembra; logični CPU 3 procesorja Intel i7-10875H v prenosniku, druga nit istega jedra prosta, v povprečju pri 4,43 (8zTSP), 4,53 (LKH) in 4,61 GHz (linkern), najtoplejša cona z mediano 100, 100 in 98 °C; pred vsakim zagonom je morala najtoplejša cona 60 s ostati pri največ 80 °C (BD, 14. nov.); omejitev pomnilnika 12 GB na tek. LKH 3.0.7 je tekel z nastavitvami, izbranimi za 1.000.000 mest v pilotu (kandidati POPMUSIC, brez subgradientnega vzpona, začetna tura POPMUSIC). linkern je tekel v 64-bitni gradnji dirke. Referenca je objavljena meja Held–Karp za E1M.0, 708.703.512,9; črta 1 % leži 1 % nad njo. Poti se shranijo ob določenih trenutkih in vsakič, ko pot prečka 3, 2, 1,5 ali 1 %; arena vsako preveri. LKH zapiše pot le ob koncu poskusa; njegove zadnje poti v teh 30 minutah, zapisane po približno 28 min, arena v zadnjem časovnem pasu ni zajela, zato smo to datoteko preverili po teku: veljavna pot, 0,868 % nad mejo Held–Karp. Po lastnih dnevnikih je 8zTSP prvo pot imel po 0,2 s, linkern po 5,8 s. Vrstica pilota: 12. novembra, 8zTSP 0.6.0-z1 s 7 nitmi; LKH in linkern kot najboljša od 4 hkratnih enojedrnih tekov, ob njih je na osmem jedru z najnižjo prednostjo teklo drugo delo. Prejšnji enojedrni teki pri tej velikosti (11.–12. novembra, s starejšim 8zTSP 0.5.0): 8zTSP 13,2 min, linkern 15,4 min, LKH 15,5 min.

    ContentsKazalo

  21. · fair one-core test of 14 Nov· pošteni enojedrni preizkus 14. novembra

    One core each, 3.16 million cities: 8zTSP at the 1% line after 19 min, LKH after 31 min, linkern after 35 minVsak na enem jedru, 3,16 milijona mest: 8zTSP je bil na črti 1 % po 19 min, LKH po 31 min, linkern po 35 min

    After 1 h on one core each: 8zTSP 0.825% above the estimated reference, LKH 0.924%, linkern 0.941%. All three ran alone on the same core.Po 1 h, vsak na enem jedru: 8zTSP 0,825 % nad ocenjeno referenco, LKH 0,924 %, linkern 0,941 %. Vsi trije so tekli sami na istem jedru.

    More on 3.16 million cities, one core eachVeč o 3,16 milijona mest, vsak na enem jedru

    As at 10 million cities, we ran 8zTSP, LKH and Concorde’s linkern one after another, each alone on the laptop, on the same single core, after the same cooldown, here for 1 hour each, in the arena that saves and checks every route it counts. The new 8zTSP (development build 303eee6: our own view of the map with our addition, without LKH, one thread) reached the 1% line after 19 min 14 s and ended 0.825% above the estimated reference. By its own log, LKH spent its first 20 minutes on preparation. LKH’s first route came after 30 min 52 s, already below the 1% line (0.982% above the estimated reference); it improved it three times in the hour and ended 0.924% above it. linkern reached the 1% line after 35 min 4 s and ended 0.941% above the estimated reference. All times run from the start of each program and include reading the map and all preparation.Kot pri 10 milijonih mest smo 8zTSP, LKH in Concordov linkern pognali enega za drugim, vsakega samega na prenosniku, na istem jedru, po enakem ohlajanju, tu po 1 uro, v areni, ki shrani in preveri vsako pot, ki jo šteje. Novi 8zTSP (razvojna gradnja 303eee6: naš lastni pogled na zemljevid z našim dodatkom, brez LKH, ena nit) je črto 1 % dosegel po 19 min 14 s in končal 0,825 % nad ocenjeno referenco. Po lastnem dnevniku je LKH prvih 20 minut porabil za pripravo. Prva pot LKH je prišla po 30 min 52 s, že pod črto 1 % (0,982 % nad ocenjeno referenco); v tej uri jo je trikrat izboljšal in končal 0,924 % nad njo. linkern je črto 1 % dosegel po 35 min 4 s in končal 0,941 % nad ocenjeno referenco. Vsi časi tečejo od zagona programa in vključujejo branje zemljevida in vso pripravo.

    One core each, 3,162,278 citiesVsak na enem jedru, 3.162.278 mest8zTSPLKH 3.0.7linkern
    First route checked by the arenaPrva pot, ki jo je preverila arena33 s30 min 52 s97 s
    To the 1% lineDo črte 1 %19 min 14 s30 min 52 s35 min 4 s
    Best checked route in 1 h, above the estimated referenceNajboljša preverjena pot v 1 h, nad ocenjeno referenco0.825%0,825 %0.924%0,924 %0.941%0,941 %
    Pilot of 12 Nov, 8zTSP with 7 threads: to the 1% linePilot 12. novembra, 8zTSP s 7 nitmi: do črte 1 %6 min 16 s52 min 59 s49 min 33 s

    Conditions: DIMACS E3M.0, seed 42, one run per program, at most 1 h each, one after another on 14 Nov; logical CPU 3 of the laptop’s Intel i7-10875H, its twin thread left idle, on average at 4.37 (8zTSP), 4.54 (LKH) and 4.62 GHz (linkern), the hottest zone at a median of 100, 100 and 96 °C; before each start the hottest zone had to stay at or below 80 °C for 60 s (BD, 14 Nov); a memory limit of 12 GB per run. LKH 3.0.7 ran with the settings chosen for 1,000,000 cities in the pilot (POPMUSIC candidates, no subgradient ascent, POPMUSIC start tour). linkern ran in the race’s 64-bit build. The reference is 0.7080·√(N·A), an estimate of the Held–Karp bound, not a proven bound and not the optimum; the 1% line lies 1% above it. Routes are saved at set moments and whenever a route crosses 3, 2, 1.5 or 1%; the arena checks each one. LKH writes its route only when a trial ends; its last route of the hour, written after about 58 min, fell outside the arena’s last time band, so we checked that file after the run: a valid route, 0.924% above the estimated reference. 8zTSP’s first attempt stopped after 386 s because the laptop’s disk was full and it could not save its route; it was repeated from the start after space was freed (until then: 3 / 2 / 1.5% after 62.2 / 198.1 / 386.0 s). By their own logs, 8zTSP had its first route after 0.7 s and linkern after 22 s. Pilot row: 12 Nov, 8zTSP 0.6.0-z1 with 7 threads; LKH and linkern as the best of 4 one-core runs at the same time, with other work at the lowest priority on the eighth core. Earlier one-thread run of 8zTSP 0.5.0 at this size: 43 min 33 s (night of 11–12 Nov, sharing the laptop with other runs).Pogoji: DIMACS E3M.0, seme 42, en tek na program, največ 1 h vsak, eden za drugim 14. novembra; logični CPU 3 procesorja Intel i7-10875H v prenosniku, druga nit istega jedra prosta, v povprečju pri 4,37 (8zTSP), 4,54 (LKH) in 4,62 GHz (linkern), najtoplejša cona z mediano 100, 100 in 96 °C; pred vsakim zagonom je morala najtoplejša cona 60 s ostati pri največ 80 °C (BD, 14. nov.); omejitev pomnilnika 12 GB na tek. LKH 3.0.7 je tekel z nastavitvami, izbranimi za 1.000.000 mest v pilotu (kandidati POPMUSIC, brez subgradientnega vzpona, začetna tura POPMUSIC). linkern je tekel v 64-bitni gradnji dirke. Referenca je 0,7080·√(N·A), ocena meje Held–Karp, ne dokazana meja in ne optimum; črta 1 % leži 1 % nad njo. Poti se shranijo ob določenih trenutkih in vsakič, ko pot prečka 3, 2, 1,5 ali 1 %; arena vsako preveri. LKH zapiše pot le ob koncu poskusa; njegove zadnje poti v tej uri, zapisane po približno 58 min, arena v zadnjem časovnem pasu ni zajela, zato smo to datoteko preverili po teku: veljavna pot, 0,924 % nad ocenjeno referenco. Prvi poskus 8zTSP se je ustavil po 386 s, ker je bil disk prenosnika poln in poti ni mogel shraniti; po sprostitvi prostora je bil ponovljen od začetka (do takrat: 3 / 2 / 1,5 % po 62,2 / 198,1 / 386,0 s). Po lastnih dnevnikih je 8zTSP prvo pot imel po 0,7 s, linkern po 22 s. Vrstica pilota: 12. novembra, 8zTSP 0.6.0-z1 s 7 nitmi; LKH in linkern kot najboljša od 4 hkratnih enojedrnih tekov, ob njih je na osmem jedru z najnižjo prednostjo teklo drugo delo. Prejšnji enonitni tek 8zTSP 0.5.0 pri tej velikosti: 43 min 33 s (noč z 11. na 12. november, prenosnik si je delil z drugimi teki).

    ContentsKazalo

  22. · run of the night of 13–14 Nov· tek v noči s 13. na 14. november

    The new build with 7 threads, 10 million cities: at the 1% line after 15 minNova gradnja s 7 nitmi, 10 milijonov mest: na črti 1 % po 15 min

    With 7 threads at 10M the new build reached the 1% line after 15 min 20 s, against 17 min 10 s for 0.6.0-z1 in the pilot; after 2 h it was 0.745% above the estimated reference.S 7 nitmi je nova gradnja pri 10M črto 1 % dosegla po 15 min 20 s, proti 17 min 10 s za 0.6.0-z1 v pilotu; po 2 h je bila 0,745 % nad ocenjeno referenco.

    More on 10 million cities with 7 threadsVeč o 10 milijonih mest s 7 nitmi

    After the one-core test we ran the new 8zTSP (development build 303eee6: our own view of the map with our addition, without LKH) with 7 threads, one per core, alone on the laptop, on the same map and in the same arena. It reached the 1% line after 15 min 20 s and ended 0.745% above the estimated reference. It was below 3% by 55.3 s (its first checked route, 2.622%), below 2% after 171.2 s and below 1.5% after 312.1 s. At every point where both builds were measured with 7 threads, the new build was ahead of the older 0.6.0-z1 from the pilot of 12 Nov (0.872% against 0.899% at 30 min, 0.792% against 0.831% at 60 min, 0.745% against 0.786% at 2 h, and the 1% line after 15 min 20 s against 17 min 10 s), in one run per build under different conditions: the pilot run did not run alone, it went through the development arena, and its times there come from the lengths the solver reported. On one core the new build needed 56 min 4 s (0.857%). LKH and linkern did not change, so their pilot rows stay (as the best of 2 and of 4 one-core runs: LKH no route in 2 h, linkern 1.162% at best). All times run from the start of the program and include reading the map and all preparation.Po enojedrnem preizkusu smo novi 8zTSP (razvojna gradnja 303eee6: naš lastni pogled na zemljevid z našim dodatkom, brez LKH) pognali s 7 nitmi, po eno na jedro, samega na prenosniku, na istem zemljevidu in v isti areni. Črto 1 % je dosegel po 15 min 20 s in končal 0,745 % nad ocenjeno referenco. Pod 3 % je bil do 55,3 s (njegova prva preverjena pot, 2,622 %), pod 2 % po 171,2 s in pod 1,5 % po 312,1 s. V vsaki točki, kjer sta bili s 7 nitmi izmerjeni obe gradnji, je bila nova gradnja pred starejšim 0.6.0-z1 iz pilota 12. novembra (0,872 % proti 0,899 % po 30 min, 0,792 % proti 0,831 % po 60 min, 0,745 % proti 0,786 % po 2 h in črta 1 % po 15 min 20 s proti 17 min 10 s), pri enem teku na gradnjo in različnih pogojih: pilotni tek ni tekel sam, šel je skozi razvojno areno, njegovi časi pa so tam iz dolžin, ki jih je javil reševalnik. Na enem jedru je nova gradnja potrebovala 56 min 4 s (0,857 %). LKH in linkern se nista spremenila, zato njuni pilotni vrstici ostaneta (kot najboljši od 2 in od 4 enojedrnih tekov: LKH v 2 h brez poti, linkern najbolje 1,162 %). Vsi časi tečejo od zagona programa in vključujejo branje zemljevida in vso pripravo.

    10,000,000 cities10.000.000 mestNew build, 7 threadsNova gradnja, 7 niti0.6.0-z1, 7 threads (pilot of 12 Nov)0.6.0-z1, 7 niti (pilot 12. novembra)New build, 1 thread (13 Nov)Nova gradnja, 1 nit (13. novembra)
    To the 1% lineDo črte 1 %15 min 20 s17 min 10 s56 min 4 s
    Above the estimated reference at 30 minNad ocenjeno referenco po 30 min0.872%0,872 %0.899%0,899 %1.243%1,243 %
    Above the estimated reference at 60 minNad ocenjeno referenco po 60 min0.792%0,792 %0.831%0,831 %0.986%0,986 %
    Best checked route in 2 h, above the estimated referenceNajboljša preverjena pot v 2 h, nad ocenjeno referenco0.745%0,745 %0.786%0,786 %0.857%0,857 %
    ConditionsPogojialone on the laptop, CPUs 0–6; the arena checked every counted routesam na prenosniku, CPU 0–6; arena je preverila vsako šteto potdevelopment runs beside it on the eighth core; development arena: times from the lengths the solver reported, final route checkedob njem razvojni teki na osmem jedru; razvojna arena: časi iz dolžin, ki jih je javil reševalnik, končna pot preverjenaalone on the laptop, CPU 3; the arena checked every counted routesam na prenosniku, CPU 3; arena je preverila vsako šteto pot

    Conditions: DIMACS E10M.0, seed 42, one run of at most 2 h, 14 Nov; 7 threads on logical CPUs 0–6 (one per physical core) of the laptop’s Intel i7-10875H, nothing else running, on average at 2.84 GHz, the hottest zone at a median of 92 °C; the same memory limit as the one-core test (12 GB); no cooldown wait before the start (BD, 14 Nov). The pilot run of 0.6.0-z1 (12 Nov) ran in the development arena, with development runs beside it on the laptop’s eighth core; the new build ran in the arena that saves and checks every counted route. The reference is 0.7080·√(N·A), an estimate of the Held–Karp bound, not a proven bound and not the optimum.Pogoji: DIMACS E10M.0, seme 42, en tek, največ 2 h, 14. novembra; 7 niti na logičnih CPU 0–6 (po ena na fizično jedro) procesorja Intel i7-10875H v prenosniku, ob tem ni teklo nič drugega, v povprečju pri 2,84 GHz, najtoplejša cona z mediano 92 °C; enaka omejitev pomnilnika kot v enojedrnem preizkusu (12 GB); brez čakanja na ohladitev pred zagonom (BD, 14. nov.). Pilotni tek 0.6.0-z1 (12. novembra) je tekel v razvojni areni, ob njem so na osmem jedru prenosnika tekli razvojni teki; nova gradnja je tekla v areni, ki shrani in preveri vsako šteto pot. Referenca je 0,7080·√(N·A), ocena meje Held–Karp, ne dokazana meja in ne optimum.

    ContentsKazalo

  23. · fair one-core test of 13 Nov· pošten enojedrni preizkus 13. novembra

    One core each, 10 million cities: 8zTSP at the 1% line after 56 min, LKH after 1 h 49 min, linkern not within 2 h (best 1.060%)Vsak na enem jedru, 10 milijonov mest: 8zTSP je bil na črti 1 % po 56 min, LKH po 1 h 49 min, linkern v 2 h ne (najboljša 1,060 %)

    After 2 h on one core each: 8zTSP 0.857% above the estimated reference, LKH 0.960%, linkern 1.060%. All three ran alone on the same core.Po 2 h, vsak na enem jedru: 8zTSP 0,857 % nad ocenjeno referenco, LKH 0,960 %, linkern 1,060 %. Vsi trije so tekli sami na istem jedru.

    More on 10 million cities, one core eachVeč o 10 milijonih mest, vsak na enem jedru

    We ran 8zTSP, LKH and Concorde’s linkern one after another, each alone on the laptop, on the same single core, after the same cooldown, for 2 hours, in the arena we built for the full test: it saves and checks every route it counts. The new 8zTSP (development build 303eee6: our own view of the map with our addition, without LKH, one thread) reached the 1% line after 56 min 4 s and ended 0.857% above the reference. By its own log, LKH spent its first 65 minutes on preparation; its first route came after 1 h 48 min 39 s, already below the 1% line (0.960% above the reference), and stayed its best in the 2 hours. linkern did not reach the 1% line in the 2 hours; its best route was 1.060% above the reference. All times run from the start of each program and include reading the map and all preparation.8zTSP, LKH in Concordov linkern smo pognali enega za drugim, vsakega samega na prenosniku, na istem jedru, po enakem ohlajanju, za 2 uri, v areni, ki smo jo zgradili za polni preizkus: shrani in preveri vsako pot, ki jo šteje. Novi 8zTSP (razvojna gradnja 303eee6: naš lastni pogled na zemljevid z našim dodatkom, brez LKH, ena nit) je črto 1 % dosegel po 56 min 4 s in končal 0,857 % nad referenco. Po lastnem dnevniku je LKH prvih 65 minut porabil za pripravo; njegova prva pot je prišla po 1 h 48 min 39 s, že pod črto 1 % (0,960 % nad referenco), in ostala njegova najboljša v 2 urah. linkern v 2 urah ni dosegel črte 1 %; njegova najboljša pot je bila 1,060 % nad referenco. Vsi časi tečejo od zagona programa in vključujejo branje zemljevida in vso pripravo.

    One core each, 10,000,000 citiesVsak na enem jedru, 10.000.000 mest8zTSPLKH 3.0.7linkern
    First route checked by the arenaPrva pot, ki jo je preverila arena65 s1 h 48 min 39 s6 min 59 s
    To the 1% lineDo črte 1 %56 min 4 s1 h 48 min 39 snot in 2 hne v 2 h
    Best checked route in 2 h, above the referenceNajboljša preverjena pot v 2 h, nad referenco0.857%0,857 %0.960%0,960 %1.060%1,060 %
    Pilot of 12 Nov, 8zTSP with 7 threads: to the 1% linePilot 12. novembra, 8zTSP s 7 nitmi: do črte 1 %17 min 10 sno route in 2 hv 2 h brez potinot in 2 h (best 1.162%)ne v 2 h (najboljša 1,162 %)

    Conditions: DIMACS E10M.0, seed 42, one run per program, at most 2 h each, one after another on 13 Nov; logical CPU 3 of the laptop’s Intel i7-10875H, its twin thread left idle, on average at 4.39 (8zTSP), 4.52 (LKH) and 4.60 GHz (linkern), the hottest zone at a median of 100, 100 and 93 °C; before each start the hottest zone had to stay at or below 72 °C for 60 s. LKH 3.0.7 ran with the settings chosen for 1,000,000 cities in the pilot (POPMUSIC candidates, no subgradient ascent, POPMUSIC start tour); settings made for 10M might change its times. linkern ran in the race’s 64-bit build. The reference is 0.7080·√(N·A), an estimate of the Held–Karp bound, not a proven bound and not the optimum; the 1% line lies 1% above it. Routes are saved at set moments (the first at 65 s) and whenever a route crosses 3, 2, 1.5 or 1%; the arena checks each one. By their own logs, 8zTSP had its first route after 2.6 s and linkern after 76 s. Pilot row: 12 Nov, 8zTSP 0.6.0-z1 with 7 threads; LKH as the best of 2 and linkern as the best of 4 one-core runs at the same time.Pogoji: DIMACS E10M.0, seme 42, en tek na program, največ 2 h vsak, eden za drugim 13. novembra; logični CPU 3 procesorja Intel i7-10875H v prenosniku, druga nit istega jedra prosta, v povprečju pri 4,39 (8zTSP), 4,52 (LKH) in 4,60 GHz (linkern), najtoplejša cona z mediano 100, 100 in 93 °C; pred vsakim zagonom je morala najtoplejša cona 60 s ostati pri največ 72 °C. LKH 3.0.7 je tekel z nastavitvami, izbranimi za 1.000.000 mest v pilotu (kandidati POPMUSIC, brez subgradientnega vzpona, začetna tura POPMUSIC); nastavitve za 10M bi lahko spremenile njegove čase. linkern je tekel v 64-bitni gradnji dirke. Referenca je 0,7080·√(N·A), ocena meje Held–Karp, ne dokazana meja in ne optimum; črta 1 % leži 1 % nad njo. Poti se shranijo ob določenih trenutkih (prvič pri 65 s) in vsakič, ko pot prečka 3, 2, 1,5 ali 1 %; arena vsako preveri. Po lastnih dnevnikih je 8zTSP prvo pot imel po 2,6 s, linkern po 76 s. Vrstica pilota: 12. novembra, 8zTSP 0.6.0-z1 s 7 nitmi; LKH kot najboljši od 2 in linkern kot najboljši od 4 hkratnih enojedrnih tekov.

    ContentsKazalo

  24. · development runs of 13 Nov· razvojni teki 13. novembra

    LKH settings inside our analysis of the view: the default staysNastavitve LKH v naši analizi pogleda: privzeta ostane

    In the hybrid with our addition we tried 1 and 10 LKH trials per part of the view instead of 3, and LKH in our analysis for the whole run instead of only until a pass gains little (after that our own search takes over), paired by seed against the default. By the rule written down before the runs, a setting would replace the default only if it was shorter at 600 s in all three seeds on one map and in no seed longer than the noise on the other. None was, so the default stays: 3 trials, LKH only in the first passes.V hibridu z našim dodatkom smo namesto 3 poskusov LKH na del pogleda preizkusili 1 in 10, LKH pa smo v naši analizi pustili ves tek namesto le do prehoda z majhnim dobičkom (potem prevzame naše lastno iskanje); vse v parih po semenu proti privzeti nastavitvi. Po pravilu, zapisanem pred teki, bi nastavitev zamenjala privzeto le, če bi bila po 600 s krajša v vseh treh semenih na enem zemljevidu in na drugem v nobenem semenu daljša za več od šuma. Nobena ni bila, zato privzeta ostane: 3 poskusi, LKH le v prvih prehodih.

    More on LKH’s settingsVeč o nastavitvah LKH

    Clustered map, 1,000,000 citiesGručasti zemljevid, 1.000.000 mestour development instance (DIMACS clustered generator)naš razvojni primer (gručasti generator DIMACS)

    • 10 trials:10 poskusov: 0.196, 0.355 and 0.702% longer at 600 spo 600 s 0,196, 0,355 in 0,702 % daljša pot
    • LKH for the whole run:LKH ves tek: 1.181, 1.260 and 1.193% longer1,181, 1,260 in 1,193 % daljša
    • 1 trial:1 poskus: +0.017, −0.024 and −0.046% (minus = shorter), not shorter in all three seeds+0,017, −0,024 in −0,046 % (minus = krajša), ne krajša v vseh treh semenih

    Uniform map, 1,000,000 citiesEnakomerni zemljevid, 1.000.000 mestDIMACS E1M.0

    • 10 trials:10 poskusov: +0.008, −0.004 and −0.008% at 600 spo 600 s +0,008, −0,004 in −0,008 %
    • LKH for the whole run:LKH ves tek: +0.017, +0.007 and +0.006%+0,017, +0,007 in +0,006 %
    • 1 trial:1 poskus: +0.010, −0.007 and −0.006%+0,010, −0,007 in −0,006 %

    Conditions: development build 303eee6, the hybrid with our addition, 7 threads, 600 s per run, seeds 42, 142 and 242, one measured run at a time, 13 Nov (uniform map 08:56–13:56, clustered map 13:55–17:08). Part of the runs had other work beside them on the laptop’s eighth core: on the uniform map 7 of the 12 runs (on average 0.4–1.5 of its two threads busy), on the clustered map 5 of the 12 runs (all four of seed 242 and one of seed 142, beside a one-core run of 8zTSP 0.6.0-z1).Pogoji: razvojna gradnja 303eee6, hibrid z našim dodatkom, 7 niti, 600 s na tek, semena 42, 142 in 242, en merjeni tek naenkrat, 13. novembra (enakomerni zemljevid 08:56–13:56, gručasti 13:55–17:08). Del tekov je imel ob sebi drugo delo na osmem jedru prenosnika: na enakomernem zemljevidu 7 od 12 tekov (v povprečju zasedenih 0,4–1,5 od njegovih dveh niti), na gručastem 5 od 12 tekov (vsi štirje semena 242 in eden semena 142, ob enojedrnem teku 8zTSP 0.6.0-z1).

    ContentsKazalo

  25. · development runs of 13 Nov· razvojni teki 13. novembra

    Renumbering the cities stays a switchPreštevilčenje mest ostane stikalo

    Renumbering the cities so that cities close on the map also sit close in the program’s memory, saves time at 10,000,000 cities: the same work (a warm-up and two passes over the map) took 13.2–13.6% less time, at the same length (−0.0043, +0.0003 and −0.0037%). At 1,000,000 cities, at 600 s, it was level on the uniform map (+0.002, −0.004 and −0.008%), but on the clustered map one seed was 0.030% longer (+0.030, +0.019 and +0.004%), more than the largest difference we measured between two runs of the same command (0.026%). By the rule written down before the runs it is therefore not on by default; it stays a switch.Preštevilčenje mest, tako da mesta, ki so si blizu na zemljevidu, ležijo blizu tudi v pomnilniku programa, pri 10.000.000 mestih prihrani čas: isto delo (ogrevanje in dva prehoda čez zemljevid) je trajalo 13,2–13,6 % manj časa, pri enaki dolžini (−0,0043, +0,0003 in −0,0037 %). Pri 1.000.000 mestih je bilo po 600 s na enakomernem zemljevidu izenačeno (+0,002, −0,004 in −0,008 %), na gručastem pa je bilo eno seme 0,030 % daljše (+0,030, +0,019 in +0,004 %), kar je več od največje razlike, ki smo jo izmerili med dvema tekoma istega ukaza (0,026 %). Po pravilu, zapisanem pred teki, zato ni privzeto vklopljeno; ostane stikalo.

    More on renumbering the citiesVeč o preštevilčenju mest

    Conditions: 10,000,000 cities: DIMACS E10M.0, the hybrid without our addition, counted runs with the same work in each (20,000 kicks of the warm-up, then two passes of 10,000 parts), 7 threads, 3 seeds, 13 Nov 06:08–07:49, build ea3671a (303eee6 plus a switch not used in these runs). 1,000,000 cities: the same blocks as in the entry above.Pogoji: 10.000.000 mest: DIMACS E10M.0, hibrid brez našega dodatka, šteti teki z enakim delom v vsakem (20.000 sunkov ogrevanja, nato dva prehoda po 10.000 delov), 7 niti, 3 semena, 13. novembra 06:08–07:49, gradnja ea3671a (303eee6 in stikalo, ki v teh tekih ni bilo uporabljeno). 1.000.000 mest: isti bloki kot v vnosu zgoraj.

    ContentsKazalo

  26. · development runs of 12–13 Nov· razvojni teki 12. in 13. novembra

    Our addition also works without LKHNaš dodatek deluje tudi brez LKH

    We ran 8zTSP’s own search, without LKH inside our analysis of the view, with and without our addition, paired by seed. Our addition shortened its route at 600 s about as much as it shortens the hybrid’s (LKH inside our analysis), and with our addition our own search without LKH ended level with the hybrid with our addition on the uniform map and level or slightly shorter on the clustered map. On the clustered map it got there sooner; on the uniform map it did not.Pognali smo lastno iskanje 8zTSP, brez LKH v naši analizi pogleda, z našim dodatkom in brez njega, v parih po semenu. Naš dodatek mu je pot po 600 s skrajšal približno toliko kot hibridu (LKH v naši analizi), naše lastno iskanje brez LKH pa je z našim dodatkom končalo izenačeno s hibridom z našim dodatkom na enakomernem zemljevidu, na gručastem pa izenačeno ali nekoliko krajše. Na gručastem zemljevidu je tja prišlo prej, na enakomernem ne.

    More on our addition without LKHVeč o našem dodatku brez LKH

    Clustered map, 1,000,000 citiesGručasti zemljevid, 1.000.000 mestour development instance (DIMACS clustered generator)naš razvojni primer (gručasti generator DIMACS)

    • The gain of our addition without LKH:Dobiček našega dodatka brez LKH: 0.412, 0.408 and 0.400% shorter at 600 s (in the hybrid, box on our addition above: 0.408, 0.397 and 0.353%)po 600 s 0,412, 0,408 in 0,400 % krajša pot (v hibridu, okvir o našem dodatku zgoraj: 0,408, 0,397 in 0,353 %)
    • Against the hybrid with our additionProti hibridu z našim dodatkom at 600 s: −0.046, −0.005 (across a reboot) and −0.121% (minus = shorter)po 600 s: −0,046, −0,005 (čez ponovni zagon) in −0,121 % (minus = krajša)
    • Sooner:Prej: the length the hybrid without our addition had at 600 s was reached after 117, 174 and 124 s; by the hybrid with our addition after 373, 361 and 406 sdolžino, ki jo je imel hibrid brez našega dodatka po 600 s, je naše iskanje doseglo po 117, 174 in 124 s, hibrid z našim dodatkom pa po 373, 361 in 406 s

    Uniform map, 1,000,000 citiesEnakomerni zemljevid, 1.000.000 mestDIMACS E1M.0

    • The gain of our addition without LKH:Dobiček našega dodatka brez LKH: 0.102, 0.107 and 0.106% shorter at 600 s (in the hybrid, box on our addition above: 0.109, 0.105 and 0.104%)po 600 s 0,102, 0,107 in 0,106 % krajša pot (v hibridu, okvir o našem dodatku zgoraj: 0,109, 0,105 in 0,104 %)
    • Against the hybrid with our additionProti hibridu z našim dodatkom at 600 s: +0.009, −0.004 and −0.009%po 600 s: +0,009, −0,004 in −0,009 %
    • Not sooner:Ne prej: 1% above the Held–Karp bound after 74–91 s; the hybrid with our addition, in the same block, after 75–85 s1 % nad mejo Held–Karp po 74–91 s; hibrid z našim dodatkom v istem bloku po 75–85 s

    Conditions: development build 303eee6; 3 seeds, 600 s per run, 7 threads, one program at a time; paired comparisons inside 8zTSP only, LKH and linkern were not run with this build. Clustered runs on the night of 12–13 Nov: a reboot split one seed (its later runs are compared with a base run made after the reboot), for part of the night a game or a video ran on the laptop, and in parts of three runs other short programs ran on the measuring cores. Uniform runs on 13 Nov 08:56–13:56, part of them beside light load on the laptop’s other cores. The clocks of the measuring cores differed between runs (2.2–3.0 GHz); this moves the times more than the lengths at 600 s.Pogoji: razvojna gradnja 303eee6; 3 semena, 600 s na tek, 7 niti, en program naenkrat; primerjave v parih samo znotraj 8zTSP, LKH in linkern s to gradnjo nista tekla. Gručasti teki v noči z 12. na 13. november: ponovni zagon je razdelil eno seme (njegovi poznejši teki so primerjani z osnovnim tekom po ponovnem zagonu), del noči je na prenosniku tekla igra ali video, v delih treh tekov pa so na merilnih jedrih tekli drugi kratki programi. Enakomerni teki 13. novembra 08:56–13:56, del njih ob rahli obremenitvi drugih jeder prenosnika. Frekvence merilnih jeder so se med teki razlikovale (2,2–3,0 GHz); to bolj premakne čase kot dolžine po 600 s.

    ContentsKazalo

Downloads: our tours, to check for yourselfPrenosi: naše poti, za lastno preverjanje

Each tour is a TSPLIB tour file: the cities in the order of the round trip, numbered as in the map’s file. To check one, downloadVsaka pot je datoteka poti v obliki TSPLIB: mesta v vrstnem redu kroga, oštevilčena kot v datoteki zemljevida. Za preverjanje prenesi verify_tour.py (Python 3, no other packages) and the map, and run one line:(Python 3, brez drugih paketov) in zemljevid ter poženi eno vrstico:

python3 verify_tour.py nu3496.tsp nu3496_96132.tour

It answersOdgovori valid tour: 3496 cities, each once; length 96132: every city exactly once, and the length with the TSPLIB rounding that Concorde and LKH also use. It reads gzipped files too. The SHA-256 of every file is in: vsako mesto natanko enkrat in dolžina z zaokroževanjem TSPLIB, ki ga uporabljata tudi Concorde in LKH. Bere tudi stisnjene datoteke (.gz). SHA-256 vsake datoteke je v sha256sums (sha256sum -c sha256sums).

nu3496 · Nicaragua · 3,496 citiesnu3496 · Nikaragva · 3.496 mest

Length 96,132, the known optimum. Found by 8zTSP build 563f8fa with 7 threads, seed 1, after 5.6 s (Dolžina 96.132, znani optimum. Našel jo je 8zTSP, gradnja 563f8fa, s 7 nitmi, seme 1, po 5,6 s (15 Nov 202615. nov. 2026).

nu3496_96132.tour · 17 kB
1d310c77d3da7ab411f387f676983c1b5adaf992c7de9aac9a91c8d048190c4f

The map:Zemljevid: nu3496.tsp among the National TSPsnu3496.tsp med nacionalnimi TSP of the University of Waterloo.Univerze Waterloo.

si6035 · Slovenia · 6,035 settlementssi6035 · Slovenija · 6.035 naselij

Best known tour: 8,728,190 m, reached by 8zTSP on 18 Nov at 05:35 by fusing five of its own earlier tours, and by LKH 3.0.14 from scratch the same morning; the two tours are identical. At 08:55 on 18 Nov, Concorde’s lower bound was 8,727,732 m; the proof is running. The file below is our tour ofNajboljša znana pot: 8.728.190 m; 8zTSP jo je dosegel 18. novembra ob 05:35 z zlitjem petih svojih prejšnjih poti, LKH 3.0.14 pa isto jutro od začetka; poti sta enaki. Ob 08:55 18. novembra je bila Concordova spodnja meja 8.727.732 m; dokaz teče. V spodnji datoteki je naša pot s 16 Nov 202616. nov. 2026, 8,728,254 m, found by combining three 8zTSP tours with the fusion method from BD’s arena (BD × GPT)., 8.728.254 m, najdena z združitvijo treh poti 8zTSP z metodo zlivanja iz BD-jeve arene (BD × GPT).

si6035_gurs_20261115_8728254.tour · 29 kB
8418babce634d44a037958e68909a5645e89100033fe42ade584eb28de0c5b6d

si6035_gurs_20261115.tsp (the map) · 111 kB(zemljevid) · 111 kB
5f8e7fffff3b51a97f21d99331205e01d46bf40ccda84127e6ce3848eb9de3b8

si6035_gurs_20261115_ids.csv (the names of the settlements, in the map’s order) · 447 kB(imena naselij v vrstnem redu zemljevida) · 447 kB
165c88cd4fd086d3f6adeec4ad21610527928228e8698508934f0b0d2cc24417

Source of the map:Vir zemljevida: Surveying and Mapping Authority of the Republic of Slovenia (GURS), RPE settlementsGeodetska uprava Republike Slovenije (GURS), RPE Naselja, taken on 15 Nov 2026,, prevzem 15. novembra 2026, CC BY 4.0. Our change: one point per settlement, the area-weighted centre of its official area, rounded to metres (D96/TM); straight lines, not roads.. Naša sprememba: ena točka na naselje, težišče njegovega uradnega območja, zaokroženo na metre (D96/TM); zračne črte, ne ceste.

DIMACS E1M.0 · 1,000,000 citiesDIMACS E1M.0 · 1.000.000 mest

Length 714,171,458, 0.772% above the Held–Karp bound (708,703,512.9). Found by 8zTSP build 303eee6 on one core, seed 42, in a run of 30 min (Dolžina 714.171.458, 0,772 % nad mejo Held–Karp (708.703.512,9). Našel jo je 8zTSP, gradnja 303eee6, na enem jedru, seme 42, v teku 30 min (14 Nov 202614. nov. 2026).

e1m.0_714171458.tour.gz · 3.3 MB· 3,3 MB
3f07c865e2ef0b98c73b6f3609dde6901ad2fbb2623d200cb72cc2b6421cc850

The map: make it with the generator of theZemljevid: naredi ga z generatorjem DIMACS TSP Challenge (portgen.c in its codes.tar):(portgen.c v njegovem codes.tar): portgen 1000000 1000000 > E1M.0.

BD × AI Lab. BD, a self-described layman, is the human architect and seed giver; the kernel and the measurements were built with frontier LLMs (Claude, GPT). Credits: our own view of the map, BD’s idea, first built by Claude (C) on 22 March 2026, then implemented by GPT · an addition to it, Claude C1, made from BD’s idea in a session BD prepared free of old TSP knowledge; BD calls it the first Claude invention · Rank, Don’t Eliminate, BD × GPT · MDL × DCC: BD (DCC from his CFH theory, first demo by Gemini; MDL as the judge that chooses the governor) · EAX: Nagata and Kobayashi. Reference solvers: LKH 3.0.7 by Keld Helsgaun; Concorde and its linkern by Applegate, Bixby, Chvátal and Cook.BD, po lastnih besedah laik, je človeški arhitekt in dajalec semen; jedro in meritve so nastali s frontier LLM-ji (Claude, GPT). Zasluge: naš lastni pogled na zemljevid, BD‑jeva zamisel, ki jo je 22. marca 2026 prvi izvedel Claude (C), nato pa GPT · dodatek k njemu, Claude C1, iz BD‑jeve zamisli v seji, ki jo je BD odtujil od okov starega znanja o TSP; BD ga imenuje prvi Claudov izum · Rank, Don’t Eliminate, BD × GPT · MDL × DCC: BD (DCC iz njegove teorije CFH, prvi demo Gemini; MDL kot sodnik, ki izbere krmilnik) · EAX: Nagata in Kobayashi. Referenčna reševalca: LKH 3.0.7, avtor Keld Helsgaun; Concorde in njegov linkern, avtorji Applegate, Bixby, Chvátal in Cook.

The evidence, section by sectionPodatki in zapisi, razdelek za razdelkom

What this page isKaj je ta stran

8zTSP is a new TSP kernel. Release 0.5.0 was declared on 11 November 2026. Its comparisons with LKH are reported below; the separately labelled comparison with Concorde used the early release 0.2.0. All measurements were made on the same laptop. This page reports those measurements, with their limits. The results are preliminary.8zTSP je novo jedro za TSP. Izdaja 0.5.0 je bila razglašena 11. novembra 2026. Njene primerjave z LKH so spodaj; posebej označena primerjava s Concordom je bila narejena z zgodnjo izdajo 0.2.0. Vse meritve so z istega prenosnika. Stran poroča o teh meritvah in njihovih omejitvah. Rezultati so predhodni.

More on this pageVeč o tej strani

Where it comes from.Od kod prihaja. 8zTSP grows out of the BD × AI Lab TSP work, including TSPTS, TSPES and the8zTSP raste iz dela BD × AI Lab na TSP, med drugim iz TSPTS, TSPES in 8Z-RP researchraziskav 8Z-RP (see also the(glej tudi research portfolioportfelj raziskav). It carries forward our own view of the map, BD's idea, as that research implemented it. The other ideas of that research, among them the addition to it made by Claude in a session BD prepared, from BD's idea, are candidates for controlled experiments in the new kernel (see). Naprej nosi naš lastni pogled na zemljevid, BD-jevo zamisel, kakor so ga te raziskave izvedle. Druge zamisli teh raziskav, med njimi dodatek k njemu, ki ga je Claude naredil v seji, ki jo je pripravil BD, iz BD-jeve zamisli, so kandidatke za nadzorovane poskuse v novem jedru (glej NextNaprej). 8zTSP is a new kernel, distinct from the 8zTSP-R research lines of that portfolio.). 8zTSP je novo jedro in ni ena od raziskovalnih vej 8zTSP-R iz tega portfelja.

GoalCilj

A human layman and frontier LLMs (BD × AI Lab) first developed our own ideas without the existing solvers (our own view of the map, our addition to it, MDL × DCC). We are now combining them with the best solutions in the world (LKH as the local search inside our analysis of the view, EAX and other methods from the literature) into the hybrid 8zTSP. The goal is to show that on an ordinary laptop, on uniformly random instances with 1, 3.16 and 10 million cities, the hybrid reaches a good tour several times faster than LKH 3.0.7 and Concorde's linkern on the same computer. This is a goal, not a measured result: a factor will be stated only after the full test under a locked protocol (several seeds, reserved confirmation instances, an independent review).Človeški laik in frontier LLM-ji (BD × AI Lab) smo najprej brez obstoječih rešitev razvili lastne zamisli (naš lastni pogled na zemljevid, naš dodatek k njemu, MDL × DCC). Zdaj jih združujemo z najboljšimi rešitvami sveta (LKH kot lokalno iskanje v naši analizi pogleda, EAX in druge metode iz literature) v hibrid 8zTSP. Cilj je pokazati, da na navadnem prenosniku pri enakomerno naključnih primerih z 1, 3,16 in 10 milijoni mest hibrid pride do dobre ture večkrat hitreje kot LKH 3.0.7 in Concordov linkern na istem računalniku. To je cilj, ne izmerjen izid: faktor bomo navedli šele po polnem preizkusu po zaklenjenem protokolu (več semen, rezervirani potrditveni primeri, neodvisen pregled).

More on the goalVeč o cilju

Current development indicatorTrenutni razvojni kazalnik

Development indicator:Razvojni kazalnik: on E1M.0 (1,000,000 cities, one seed, development) the time to 1% above the Held–Karp bound was 122.8 s for 8zTSP 0.6.0-z1 on 7 threads (the program's own report; the final tour is verified), 810.7 s for LKH 3.0.7 with POPMUSIC candidates and without subgradient ascent and 596.5 s for Concorde's linkern (both as the best of 4 independent runs, one core each; the program's own report). These are the runs of stage 1 of the pilot, one program at a time. In stage 0, which chose the number of independent runs, the same settings took 727.8 s (LKH) and 583.7 s (linkern). The pilot ended on 12 November: its times to 1% at all three sizes are at the top of this page (na E1M.0 (1.000.000 mest, eno seme, razvoj) je bil čas do 1 % nad mejo Held–Karp 122,8 s pri 8zTSP 0.6.0-z1 na 7 nitih (poročilo programa samega; končna tura je preverjena), 810,7 s pri LKH 3.0.7 s kandidati POPMUSIC in brez subgradientnega vzpona in 596,5 s pri Concordovem linkernu (oba kot najboljši od 4 neodvisnih tekov, vsak na enem jedru; poročilo programa samega). To so teki stopnje 1 pilota, en program naenkrat. V stopnji 0, ki je izbrala število neodvisnih tekov, sta isti nastavitvi potrebovali 727,8 s (LKH) in 583,7 s (linkern). Pilot se je končal 12. novembra: njegovi časi do 1 % pri vseh treh velikostih so na vrhu te strani (Same laptop, same finish lineIsti prenosnik, ista ciljna črta); the records of its first stage are under); zapisi njegove prve stopnje so pod Pilot of the goal testPilot ciljnega preizkusa.

How it was measuredKako je bilo merjeno

8zTSP 0.5.0

8zTSP is a kernel written in Rust that uses only the standard library. Release 0.5.0 prints "0.5.0-rc3": the last candidate was declared without recompiling. Up to 12,000 cities the default mode is the search for small maps, in which tours are merged; above that the default is our own recursive analysis of the view with sequential 3-opt. The large-instance measurements below used the mode for large maps.8zTSP je jedro, napisano v Rustu, ki uporablja samo standardno knjižnico. Izdaja 0.5.0 izpiše »0.5.0-rc3«: zadnji kandidat je bil razglašen brez ponovnega prevajanja. Do 12.000 mest je privzeti način iskanje za majhne zemljevide, v katerem se ture zlivajo; nad tem je privzeta naša lastna rekurzivna analiza pogleda z zaporednim 3-opt. Meritve velikih primerov spodaj so uporabile način za velike zemljevide.

More on 8zTSP 0.5.0Več o 8zTSP 0.5.0

Parts of the kernel and where each comes from.Deli jedra in od kod kateri prihaja.

PartDelWhat it does hereKaj dela tukajOriginIzvor
Our own view of the mapNaš lastni pogled na zemljevidFinds neighbours and builds start tours without an n × n table. The running time depends on the geometry: many cities close together can make the search quadratic. Candidates in the mode for large maps: 4 nearest + 1 per quadrant (8 candidates).Poišče sosede in zgradi začetne ture brez tabele n × n. Čas je odvisen od geometrije: veliko mest blizu skupaj lahko naredi iskanje kvadratično. Kandidati v načinu za velike zemljevide: 4 najbližji + 1 na kvadrant (8 kandidatov).
BD × AI LabOur own view of the map of the earlier research: BD's idea, first built by Claude (C) on 22 March 2026, then implemented by GPT.Naš lastni pogled na zemljevid iz prejšnjih raziskav: BD-jeva zamisel, ki jo je 22. marca 2026 prvi izvedel Claude (C), nato pa GPT.
StandardStandardnoNeighbour lists with quadrants (Bentley; Johnson and McGeoch).Seznami sosedov s kvadranti (Bentley; Johnson in McGeoch).
Our own recursive analysis of the viewNaša lastna rekurzivna analiza pogledaThe search works on parts of about a thousand cities at once, in parallel, and keeps refining them; how the parts are formed is our own. A development variant has also been tried; it is not part of release 0.5.0 (seeIskanje dela na delih s približno tisoč mesti hkrati, vzporedno, in jih sproti izboljšuje; kako deli nastanejo, je naše lastno. Preizkušena je bila tudi razvojna različica; ta ni del izdaje 0.5.0 (glej NextNaprej).
BD's ideaBD-jeva zamiselOur own view of the map: BD’s idea. In 8zTSP it works as our own recursive analysis of the view, on parts of about 1,000 cities solved in parallel.Naš lastni pogled na zemljevid: BD‑jeva zamisel. V 8zTSP deluje kot naša lastna rekurzivna analiza pogleda, na delih po približno 1.000 mest, ki se rešujejo vzporedno.
StandardStandardnoReplacing the outside part of the tour by a fixed edge is known from the subproblem partitioning of LKH.Nadomestitev zunanjega dela ture s fiksnim robom je znana iz delitve na podprobleme v LKH.
2-opt and Or-opt with a work queue2-opt in Or-opt z delovno vrstoLocal search: 2-opt reverses a segment, Or-opt moves a short segment.Lokalno iskanje: 2-opt obrne odsek, Or-opt prestavi kratek odsek.
StandardStandardnoCroes (2-opt); Or (Or-opt).
Sequential 3-optZaporedni 3-optThe default deeper local move in our analysis of the view; available as an explicit option in the kick mode.Privzeta globlja lokalna poteza v naši analizi pogleda; v načinu sunkov na voljo kot izrecna možnost.
StandardStandardnoLin (3-opt).
Search with tour mergingIskanje z zlivanjem turA set of tours is improved by merging two tours through their difference cycles (AB-cycles): the edge assembly crossover (EAX).Množica tur se izboljšuje z zlivanjem dveh tur prek ciklov njunih razlik (AB-cikli): križanje s sestavljanjem robov (EAX).
StandardStandardnoNagata and Kobayashi.Nagata in Kobayashi.
Related to an idea of the earlier research (seeSorodno zamisli iz prejšnjih raziskav (glej NextNaprej); it is not our addition.); ni naš dodatek.
Diversity-aware selectionIzbira z upoštevanjem raznolikostiUses the change in edge diversity when scoring which child enters the set of tours.Pri ocenjevanju, kateri otrok vstopi v množico tur, upošteva spremembo raznolikosti robov.
StandardStandardnoInspired by the diversity-aware selection of the EAX research (Nagata and Kobayashi); the rule is adapted in this implementation.Po zgledu izbire z upoštevanjem raznolikosti iz raziskav EAX (Nagata in Kobayashi); pravilo je v tej izvedbi prilagojeno.

Not in release 0.5.0.Ni v izdaji 0.5.0. Release 0.5.0 does not include plug-ins for our addition, DCC or Rank, Don't Eliminate, the other ideas of the BD × AI Lab research; whether they add anything will be measured in the idea arena (seeIzdaja 0.5.0 ne vsebuje vtičnikov za naš dodatek, DCC ali Rank, Don't Eliminate, druge zamisli raziskav BD × AI Lab; ali kaj prispevajo, bo izmerjeno v areni zamisli (glej NextNaprej). The mechanics are known methods: the merging is EAX, and 2-opt, Or-opt, 3-opt and neighbour lists with quadrants are from the literature. Our own view of the map is BD's idea, first implemented in the earlier BD × AI Lab research. What is specific to 8zTSP is how these parts are put together; no claim of priority is made.). Mehanike so znane metode: zlivanje je EAX, 2-opt, Or-opt, 3-opt in seznami sosedov s kvadranti pa so iz literature. Naš lastni pogled na zemljevid je BD-jeva zamisel, prvič izvedena v prejšnjih raziskavah BD × AI Lab. Posebnost 8zTSP je v tem, kako so ti deli sestavljeni; prvenstva ne trdimo.

Results of 8zTSPRezultati 8zTSP

Preliminary. One laptop, 11 November 2026; conditions are stated with every table.Predhodno. En prenosnik, 11. november 2026; pogoji so navedeni ob vsaki tabeli.

Nine national instances: the new build, ten seeds, one coreDevet državnih primerov: nova gradnja, deset semen, eno jedro

The new build 303eee6 (14 Nov 2026) in the same study as release 0.5.0: the optimum was reached in 90 of 90 runs; its median time to the optimum was shorter than 0.5.0’s on six of the nine instances and equal on the other three.Nova gradnja 303eee6 (14. nov. 2026) v isti študiji kot izdaja 0.5.0: optimum je dosegla v 90 od 90 tekov; njena mediana časa do optimuma je bila krajša kot pri 0.5.0 na šestih od devetih primerov in enaka na ostalih treh.

More on the results of 8zTSPVeč o rezultatih 8zTSP

The new build 303eee6, one core, ten seeds on each of the nine national instancesNova gradnja 303eee6, eno jedro, deset semen na vsakem od devetih državnih primerov (14 Nov 2026, the procedure of the 0.5.0 study, six runs at a time on logical CPUs 0–5). Percentages are over the known optimum.(14. nov. 2026, postopek študije 0.5.0, šest tekov hkrati na logičnih CPU 0–5). Odstotki so nad znanim optimumom.

InstancePrimerTime limit per runČasovna meja na tekSeedsSemenReached the optimumDoseglo optimumTime to the optimum: min / median / maxČas do optimuma: najmanj / mediana / največAfter 10 s: median (worst) above the optimumPo 10 s: mediana (najslabše) nad optimumomRelease 0.5.0: median timeIzdaja 0.5.0: mediana časa
wi2920 s10100.00 s / 0.00 s / 0.01 s0,00 s / 0,00 s / 0,01 s0.000% (0.000%)0,000 % (0,000 %)0.00 s0,00 s
dj3820 s10100.00 s / 0.00 s / 0.01 s0,00 s / 0,00 s / 0,01 s0.000% (0.000%)0,000 % (0,000 %)0.00 s0,00 s
qa19420 s10100.06 s / 0.08 s / 0.09 s0,06 s / 0,08 s / 0,09 s0.000% (0.000%)0,000 % (0,000 %)0.08 s0,08 s
uy73460 s10101.00 s / 1.21 s / 5.87 s1,00 s / 1,21 s / 5,87 s0.000% (0.000%)0,000 % (0,000 %)1.49 s1,49 s
zi92960 s10101.90 s / 2.20 s / 5.54 s1,90 s / 2,20 s / 5,54 s0.000% (0.000%)0,000 % (0,000 %)2.69 s2,69 s
lu980120 s10101.17 s / 5.83 s / 15 s1,17 s / 5,83 s / 15 s0.000% (0.009%)0,000 % (0,009 %)6.87 s6,87 s
rw1621300 s10102.67 s / 15 s / 26 s2,67 s / 15 s / 26 s0.008% (0.012%)0,008 % (0,012 %)19 s
mu197960 s10105.82 s / 12 s / 54 s5,82 s / 12 s / 54 s0.001% (0.001%)0,001 % (0,001 %)14 s
nu3496600 s101029 s / 33 s / 3.4 min29 s / 33 s / 3,4 min2.416% (2.620%)2,416 % (2,620 %)43 s

Nine national instances, ten seeds, one coreDevet državnih primerov, deset semen, eno jedro

Release 0.5.0, mode for small maps: the optimum was reached in 90 of 90 runs. The median time to the optimum is 0.00 s on wi29 and dj38 and 43 s on nu3496; the longest single run took 3.6 min (nu3496). In the earlier same-machine comparison with release 0.2.0, LKH 3.0.7 with default settings reached the optimum on 7 of the 9 instances within 10 minutes, 8zTSP on all 9.Izdaja 0.5.0, način za majhne zemljevide: optimum je bil dosežen v 90 od 90 tekov. Mediana časa do optimuma je 0,00 s na wi29 in dj38 ter 43 s na nu3496; najdaljši posamezni tek je trajal 3,6 min (nu3496). V prejšnji primerjavi z izdajo 0.2.0 na istem računalniku je LKH 3.0.7 s privzetimi nastavitvami v 10 minutah dosegel optimum na 7 od 9 primerov, 8zTSP na vseh 9.

Release 0.5.0, one core, ten seeds on each of the nine national instancesIzdaja 0.5.0, eno jedro, deset semen na vsakem od devetih državnih primerov (the study that was the condition for declaring the release). Percentages are over the known optimum.(študija, ki je bila pogoj za razglasitev izdaje). Odstotki so nad znanim optimumom.

InstancePrimerTime limit per runČasovna meja na tekSeedsSemenReached the optimumDoseglo optimumTime to the optimum: min / median / maxČas do optimuma: najmanj / mediana / največAfter 10 s: median (worst) above the optimumPo 10 s: mediana (najslabše) nad optimumom
wi2920 s10100.00 s / 0.00 s / 0.01 s0,00 s / 0,00 s / 0,01 s0.000% (0.000%)0,000 % (0,000 %)
dj3820 s10100.00 s / 0.00 s / 0.01 s0,00 s / 0,00 s / 0,01 s0.000% (0.000%)0,000 % (0,000 %)
qa19420 s10100.07 s / 0.08 s / 0.10 s0,07 s / 0,08 s / 0,10 s0.000% (0.000%)0,000 % (0,000 %)
uy73460 s10101.10 s / 1.49 s / 6.71 s1,10 s / 1,49 s / 6,71 s0.000% (0.000%)0,000 % (0,000 %)
zi92960 s10102.29 s / 2.69 s / 6.64 s2,29 s / 2,69 s / 6,64 s0.000% (0.000%)0,000 % (0,000 %)
lu980120 s10101.45 s / 6.87 s / 17 s1,45 s / 6,87 s / 17 s0.000% (0.009%)0,000 % (0,009 %)
rw1621300 s10103.56 s / 19 s / 31 s3,56 s / 19 s / 31 s0.008% (0.012%)0,008 % (0,012 %)
mu197960 s10107.49 s / 14 s / 58 s7,49 s / 14 s / 58 s0.001% (0.001%)0,001 % (0,001 %)
nu3496600 s101037 s / 43 s / 3.6 min37 s / 43 s / 3,6 min2.924% (3.276%)2,924 % (3,276 %)

Nine national instances: an early release next to LKH and ConcordeDevet državnih primerov: zgodnja izdaja ob LKH in Concordu

This is an early release of the kernel,To je zgodnja izdaja jedra, 0.2.0, not 0.5.0: one core each, all at the same time, 600 s., ne 0.5.0: vsak na enem jedru, vsi hkrati, 600 s.

8zTSP 0.2.0 next to LKH (default settings) and Concorde.8zTSP 0.2.0 ob LKH (privzete nastavitve) in Concordu. In each cell: values after 10 s / 60 s / 10 min, or the time at which the optimum was reached. Percentages are over the known optimum.V vsaki celici: vrednosti po 10 s / 60 s / 10 min ali čas, ob katerem je bil dosežen optimum. Odstotki so nad znanim optimumom.

InstancePrimer8zTSP 0.2.0LKH 3.0.7, default settingsLKH 3.0.7, privzete nastavitveConcorde
wi29optimum after 0.00 spo 0,00 soptimum after 0.01 spo 0,01 soptimum after 0.02 s; proof after 0.04 spo 0,02 s; dokaz po 0,04 s
dj38optimum after 0.01 spo 0,01 soptimum after 0.01 spo 0,01 soptimum after 0.02 s; proof after 0.04 spo 0,02 s; dokaz po 0,04 s
qa194optimum after 0.06 spo 0,06 soptimum after 0.07 spo 0,07 soptimum after 0.19 s; proof after 0.67 spo 0,19 s; dokaz po 0,67 s
uy734optimum after 14 spo 14 s10 s / 60 s / 10 min: 0.004% /10 s / 60 s / 10 min: 0,004 % / optimum / optimumoptimum after 5.04 spo 5,04 soptimum after 3.8 min; proof after 4.9 minpo 3,8 min; dokaz po 4,9 min10 s / 60 s / 10 min: 0.190% / 0.003% /10 s / 60 s / 10 min: 0,190 % / 0,003 % / optimum
zi929optimum after 1.30 spo 1,30 soptimum after 3.70 spo 3,70 soptimum after 2.8 min; proof after 2.8 minpo 2,8 min; dokaz po 2,8 min10 s / 60 s / 10 min: 0.261% / 0.051% /10 s / 60 s / 10 min: 0,261 % / 0,051 % / optimum
lu980optimum after 4.53 spo 4,53 soptimum after 4.8 minpo 4,8 min10 s / 60 s / 10 min: 0.168% / 0.062% /10 s / 60 s / 10 min: 0,168 % / 0,062 % / optimumoptimum after 4.9 minpo 4,9 min*10 s / 60 s / 10 min: 0.150% / 0.150% /10 s / 60 s / 10 min: 0,150 % / 0,150 % / optimum
rw1621optimum after 22 spo 22 s10 s / 60 s / 10 min: 0.008% /10 s / 60 s / 10 min: 0,008 % / optimum / optimum0.461% / 0.223% / 0.104%0,461 % / 0,223 % / 0,104 %0.549% / 0.549% / 0.023%0,549 % / 0,549 % / 0,023 %*
mu1979optimum after 12 spo 12 s10 s / 60 s / 10 min: 0.003% /10 s / 60 s / 10 min: 0,003 % / optimum / optimumoptimum after 31 spo 31 s10 s / 60 s / 10 min: 0.124% /10 s / 60 s / 10 min: 0,124 % / optimum / optimum0.241% / 0.241% / 0.002%0,241 % / 0,241 % / 0,002 %*
nu3496optimum after 5.9 minpo 5,9 min10 s / 60 s / 10 min: 1.803% / 0.008% /10 s / 60 s / 10 min: 1,803 % / 0,008 % / optimumno tourni ture / 0.073% / 0.047%/ 0,073 % / 0,047 %0.542% / 0.542% / 0.051%0,542 % / 0,542 % / 0,051 %*

* Concorde does not write its tour as it goes, so when it is stopped at 600 s there is no tour on disk; its intermediate lengths are from its log.* Concorde ture ne zapisuje sproti, zato ob ustavitvi pri 600 s ture ni na disku; njegove vmesne dolžine so iz dnevnika.

  • 8zTSP 0.2.0 reached the optimum on all nine instances within 10 minutes; LKH with default settings on seven of the nine (on uy734 LKH was sooner: 5.04 s against 14 s).8zTSP 0.2.0 je v 10 minutah dosegel optimum na vseh devetih primerih; LKH s privzetimi nastavitvami na sedmih od devetih (na uy734 je bil LKH hitrejši: 5,04 s proti 14 s).
  • Release 0.2.0 reports a new best tour only at the end of a generation, so its measured time to the optimum is somewhat longer than the time at which the tour was found; this works against 8zTSP.Izdaja 0.2.0 novo najboljšo turo objavi šele ob koncu generacije, zato je njen izmerjeni čas do optimuma nekoliko daljši od časa, ko je bila tura najdena; to je v škodo 8zTSP.
  • The cores were not rotated between the solvers.Jedra se med reševalci niso menjala.

Concorde: long side runsConcorde: dolgi teki ob strani (limit 4 h, on a core shared with other jobs): time to the optimal tour and to the proof of optimality.(meja 4 h, na jedru, deljenem z drugimi posli): čas do optimalne ture in do dokaza optimalnosti.

InstancePrimerBest tourNajboljša turaProof of optimalityDokaz optimalnosti
rw1621optimum after 10.1 minpo 10,1 minafter 128.2 minpo 128,2 min
mu1979optimum after 12.4 minpo 12,4 minafter 72.9 minpo 72,9 min
nu3496optimum after 148.1 minpo 148,1 minafter 169.1 minpo 169,1 min

Large instances next to LKHVeliki primeri ob LKH

8zTSP 0.5.0 in the mode for large maps next to LKH 3.0.7 with the geometric settings (see8zTSP 0.5.0 v načinu za velike zemljevide ob LKH 3.0.7 z geometrijskimi nastavitvami (glej MethodMetoda). Six-core entries: 8zTSP with six threads in one run against six independent runs of LKH, of which the best counts. One-core entries: 8zTSP and LKH ran at the same time on two fixed cores. In the tables,). Vnosi s šestimi jedri: 8zTSP s šestimi nitmi v enem teku proti šestim neodvisnim tekom LKH, od katerih šteje najboljši. Vnosi z enim jedrom: 8zTSP in LKH sta tekla hkrati na dveh stalnih jedrih. V tabelah no tourni ture means that the solver had not yet written a tour at that time. Percentages on the DIMACS instances are over the published Held–Karp bound; on sw24978 over the optimum.pomeni, da reševalec ob tem času še ni zapisal ture. Odstotki pri primerih DIMACS so nad objavljeno mejo Held–Karp; pri sw24978 nad optimumom.

100,000 cities (DIMACS E100k.0)100.000 mest (DIMACS E100k.0)

Percent over the Held–Karp bound 224,330,692.1, after 60 s / 2 min / 5 min. LKH at seed s used the seeds s to s+5; its run with seed 42 comes from the earlier large-instance series.Odstotek nad mejo Held–Karp 224.330.692,1, po 60 s / 2 min / 5 min. LKH je pri semenu s uporabil semena od s do s+5; njegov tek s semenom 42 je iz prejšnje serije velikih primerov.

Six cores, three seeds.Šest jeder, tri semena. 8zTSP 0.5.0 with six threads against LKH with six independent runs.8zTSP 0.5.0 s šestimi nitmi proti LKH s šestimi neodvisnimi teki.

SeedSeme8zTSP 0.5.0, six threads8zTSP 0.5.0, šest nitiLKH, six independent runsLKH, šest neodvisnih tekov
60 s2 min5 min60 s2 min5 min
420.784%0,784 %0.740%0,740 %0.714%0,714 %no tourni ture0.792%0,792 %0.754%0,754 %
1420.820%0,820 %0.767%0,767 %0.714%0,714 %no tourni ture0.804%0,804 %0.772%0,772 %
2420.796%0,796 %0.770%0,770 %0.718%0,718 %no tourni ture0.797%0,797 %0.762%0,762 %

One core, three seeds.Eno jedro, tri semena. 8zTSP 0.5.0 with one thread against one LKH run.8zTSP 0.5.0 z eno nitjo proti enemu teku LKH.

SeedSeme8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nitLKH, one runLKH, en tek
60 s2 min5 min60 s2 min5 min
420.953%0,953 %0.858%0,858 %0.784%0,784 %no tourni ture0.816%0,816 %0.786%0,786 %
1421.004%1,004 %0.886%0,886 %0.814%0,814 %no tourni ture0.822%0,822 %0.785%0,785 %
2420.952%0,952 %0.865%0,865 %0.788%0,788 %0.829%0,829 %0.807%0,807 %0.785%0,785 %
Run details, seed 42: first tour, best tour, checkPodrobnosti tekov, seme 42: prva tura, najboljša tura, preverba
  • 8zTSP 0.5.0: first tour after 0.02 s; best tour 225,932,479 (0.714%) after 4.2 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,02 s; najboljša tura 225.932.479 (0,714 %) po 4,2 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, six independent runs: first tour after 64 s; best tour 226,021,810 (0.754%) after 5.0 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, šest neodvisnih tekov: prva tura po 64 s; najboljša tura 226.021.810 (0,754 %) po 5,0 min; končna tura ponovno prebrana z diska in preverjena: da.
  • 8zTSP 0.5.0: first tour after 0.02 s; best tour 226,088,996 (0.784%) after 4.9 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,02 s; najboljša tura 226.088.996 (0,784 %) po 4,9 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, one run: first tour after 68 s; best tour 226,093,678 (0.786%) after 4.8 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, en tek: prva tura po 68 s; najboljša tura 226.093.678 (0,786 %) po 4,8 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, repeat run: first tour after 61 s; best tour 226,064,416 (0.773%) after 4.9 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, ponovljen tek: prva tura po 61 s; najboljša tura 226.064.416 (0,773 %) po 4,9 min; končna tura ponovno prebrana z diska in preverjena: da.

First tour (six cores): 8zTSP within the first second, LKH after 64 s / 63 s / 63 s. Two LKH runs with the same seed differ by about 0.01 percentage points after 5 minutes (one core, seed 42: 0.786% and, in the repeat run, 0.773%).Prva tura (šest jeder): 8zTSP v prvi sekundi, LKH po 64 s / 63 s / 63 s. Dva teka LKH z istim semenom se po 5 minutah razlikujeta za približno 0,01 odstotne točke (eno jedro, seme 42: 0,786 % in v ponovljenem teku 0,773 %).

1,000,000 cities (DIMACS E1M.0)1.000.000 mest (DIMACS E1M.0)

One seed.Eno seme. Percent over the Held–Karp bound 708,703,512.9.Odstotek nad mejo Held–Karp 708.703.512,9.

TimeČasSix coresŠest jederOne coreEno jedro
8zTSP 0.5.0, six threads8zTSP 0.5.0, šest nitiLKH, six independent runsLKH, šest neodvisnih tekov8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nitLKH, one runLKH, en tek
10 s5.811%5,811 %no tourni ture7.243%7,243 %no tourni ture
60 s2.258%2,258 %no tourni ture2.779%2,779 %no tourni ture
5.0 min5,0 min0.923%0,923 %no tourni ture1.360%1,360 %no tourni ture
10.0 min10,0 min0.876%0,876 %no tourni ture1.055%1,055 %no tourni ture
20.0 min20,0 min0.833%0,833 %0.825%0,825 %0.942%0,942 %0.833%0,833 %
30.0 min30,0 min0.815%0,815 %0.810%0,810 %0.906%0,906 %0.816%0,816 %
Run details: first tour, best tour, checkPodrobnosti tekov: prva tura, najboljša tura, preverba
  • 8zTSP 0.5.0: first tour after 0.17 s; best tour 714,480,494 (0.815%) after 30.0 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,17 s; najboljša tura 714.480.494 (0,815 %) po 30,0 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, six independent runs: first tour after 15.8 min; best tour 714,444,823 (0.810%) after 29.4 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, šest neodvisnih tekov: prva tura po 15,8 min; najboljša tura 714.444.823 (0,810 %) po 29,4 min; končna tura ponovno prebrana z diska in preverjena: da.
  • 8zTSP 0.5.0: first tour after 0.27 s; best tour 715,121,800 (0.906%) after 30.0 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,27 s; najboljša tura 715.121.800 (0,906 %) po 30,0 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, one run: first tour after 18.6 min; best tour 714,486,658 (0.816%) after 27.2 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, en tek: prva tura po 18,6 min; najboljša tura 714.486.658 (0,816 %) po 27,2 min; končna tura ponovno prebrana z diska in preverjena: da.

Time to the first recorded tour.Čas do prve zapisane ture. On E1M.0, 8zTSP has a tour after 0.17 s (six threads) and 0.27 s (one thread); LKH needs 15.8 min (six runs) and 18.6 min (one run) for its first tour. After 5 minutes 8zTSP with six threads is at 0.923% over the bound (one seed). The clocks start at different points (seeNa E1M.0 ima 8zTSP turo po 0,17 s (šest niti) in 0,27 s (ena nit); LKH za prvo turo potrebuje 15,8 min (šest tekov) in 18,6 min (en tek). Po 5 minutah je 8zTSP s šestimi nitmi pri 0,923 % nad mejo (eno seme). Uri začneta teči ob različnih trenutkih (glej MethodMetoda).

10,000 cities (DIMACS E10k.0)10.000 mest (DIMACS E10k.0)

One seed.Eno seme. Percent over the Held–Karp bound 71,362,276.4. 8zTSP in the mode for large maps, which was requested explicitly (the default up to and including 12,000 cities is the mode for small maps).Odstotek nad mejo Held–Karp 71.362.276,4. 8zTSP v načinu za velike zemljevide, ki je bil zahtevan izrecno (do vključno 12.000 mest je privzet način za majhne zemljevide).

TimeČasSix coresŠest jederOne coreEno jedro
8zTSP 0.5.0, six threads8zTSP 0.5.0, šest nitiLKH, six independent runsLKH, šest neodvisnih tekov8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nitLKH, one runLKH, en tekLKH, repeat runLKH, ponovljen tek
1.00 s1,00 s1.277%1,277 %no tourni ture1.779%1,779 %no tourni tureno tourni ture
5.00 s5,00 s0.839%0,839 %0.818%0,818 %1.042%1,042 %0.847%0,847 %0.776%0,776 %
10 s0.826%0,826 %0.741%0,741 %0.882%0,882 %0.757%0,757 %0.744%0,744 %
30 s0.778%0,778 %0.720%0,720 %0.827%0,827 %0.724%0,724 %0.720%0,720 %
60 s0.748%0,748 %0.713%0,713 %0.808%0,808 %0.720%0,720 %0.720%0,720 %
2.0 min2,0 min0.720%0,720 %0.711%0,711 %0.778%0,778 %0.718%0,718 %0.718%0,718 %
Run details: first tour, best tour, checkPodrobnosti tekov: prva tura, najboljša tura, preverba
  • 8zTSP 0.5.0: first tour after 0.00 s; best tour 71,876,300 (0.720%) after 117 s; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,00 s; najboljša tura 71.876.300 (0,720 %) po 117 s; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, six independent runs: first tour after 4.65 s; best tour 71,869,997 (0.711%) after 76 s; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, šest neodvisnih tekov: prva tura po 4,65 s; najboljša tura 71.869.997 (0,711 %) po 76 s; končna tura ponovno prebrana z diska in preverjena: da.
  • 8zTSP 0.5.0: first tour after 0.00 s; best tour 71,917,582 (0.778%) after 102 s; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,00 s; najboljša tura 71.917.582 (0,778 %) po 102 s; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, one run: first tour after 4.97 s; best tour 71,874,687 (0.718%) after 117 s; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, en tek: prva tura po 4,97 s; najboljša tura 71.874.687 (0,718 %) po 117 s; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, repeat run: first tour after 2.94 s; best tour 71,874,687 (0.718%) after 83 s; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, ponovljen tek: prva tura po 2,94 s; najboljša tura 71.874.687 (0,718 %) po 83 s; končna tura ponovno prebrana z diska in preverjena: da.

24,978 cities (sw24978, Sweden)24.978 mest (sw24978, Švedska)

One seed.Eno seme. Percent over the optimum 855,597. The candidate 0.5.0-rc1 is shown next to the release for comparison.Odstotek nad optimumom 855.597. Kandidat 0.5.0-rc1 je za primerjavo prikazan ob izdaji.

TimeČasSix coresŠest jederOne coreEno jedro
8zTSP 0.5.0-rc1 (candidate), six threads8zTSP 0.5.0-rc1 (kandidat), šest niti8zTSP 0.5.0, six threads8zTSP 0.5.0, šest nitiLKH, geometric settings, six independent runsLKH, geometrijske nastavitve, šest neodvisnih tekovLKH, default settings, six independent runsLKH, privzete nastavitve, šest neodvisnih tekov8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nitLKH, one runLKH, en tekLKH, repeat runLKH, ponovljen tek
10 s1.378%1,378 %0.178%0,178 %no tourni tureno tourni ture0.654%0,654 %no tourni ture0.173%0,173 %
30 s1.271%1,271 %0.103%0,103 %0.097%0,097 %no tourni ture0.195%0,195 %0.114%0,114 %0.065%0,065 %
60 s0.114%0,114 %0.076%0,076 %0.054%0,054 %no tourni ture0.156%0,156 %0.058%0,058 %0.045%0,045 %
5.0 min5,0 min0.012%0,012 %0.034%0,034 %0.022%0,022 %no tourni ture0.076%0,076 %0.026%0,026 %0.025%0,025 %
10.0 min10,0 min0.010%0,010 %0.029%0,029 %0.010%0,010 %no tourni ture0.045%0,045 %0.025%0,025 %0.022%0,022 %
20.0 min20,0 min0.010%0,010 %–0.003%0,003 %no tourni ture–––
Run details: first tour, best tour, checkPodrobnosti tekov: prva tura, najboljša tura, preverba
  • 8zTSP 0.5.0-rc1 (candidate): first tour after 0.00 s; best tour 855,681 (0.010%) after 8.2 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0-rc1 (kandidat): prva tura po 0,00 s; najboljša tura 855.681 (0,010 %) po 8,2 min; končna tura ponovno prebrana z diska in preverjena: da.
  • 8zTSP 0.5.0: first tour after 0.00 s; best tour 855,847 (0.029%) after 8.2 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,00 s; najboljša tura 855.847 (0,029 %) po 8,2 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, six independent runs: first tour after 13 s; best tour 855,620 (0.003%) after 18.8 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, šest neodvisnih tekov: prva tura po 13 s; najboljša tura 855.620 (0,003 %) po 18,8 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, default settings, six independent runs: no tour in 20.0 min.LKH, privzete nastavitve, šest neodvisnih tekov: v 20,0 min nobene ture.
  • 8zTSP 0.5.0: first tour after 0.01 s; best tour 855,986 (0.045%) after 9.8 min; final tour re-read from disk and checked: yes.8zTSP 0.5.0: prva tura po 0,01 s; najboljša tura 855.986 (0,045 %) po 9,8 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, one run: first tour after 14 s; best tour 855,808 (0.025%) after 9.8 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, en tek: prva tura po 14 s; najboljša tura 855.808 (0,025 %) po 9,8 min; končna tura ponovno prebrana z diska in preverjena: da.
  • LKH, geometric settings, repeat run: first tour after 8.49 s; best tour 855,783 (0.022%) after 9.2 min; final tour re-read from disk and checked: yes.LKH, geometrijske nastavitve, ponovljen tek: prva tura po 8,49 s; najboljša tura 855.783 (0,022 %) po 9,2 min; končna tura ponovno prebrana z diska in preverjena: da.

The reading, with its conditionsKako to brati, s pogoji

A limited advantage on E100k.0Omejena prednost na E100k.0

  • 100,000 cities (DIMACS E100k.0), six threads of 8zTSP 0.5.0 against six independent runs of LKH 3.0.7 with geometric settings, on the same laptop:100.000 mest (DIMACS E100k.0), šest niti 8zTSP 0.5.0 proti šestim neodvisnim tekom LKH 3.0.7 z geometrijskimi nastavitvami, na istem prenosniku: 8zTSP is ahead after 2 minutes (0.740% / 0.767% / 0.770% against 0.792% / 0.804% / 0.797%) and after 5 minutes (0.714% / 0.714% / 0.718% against 0.754% / 0.772% / 0.762%), in three of three seeds.8zTSP je spredaj po 2 minutah (0,740 % / 0,767 % / 0,770 % proti 0,792 % / 0,804 % / 0,797 %) in po 5 minutah (0,714 % / 0,714 % / 0,718 % proti 0,754 % / 0,772 % / 0,762 %), pri treh od treh semen.

LKH level or aheadLKH izenačen ali spredaj

  • One core, 100,000 cities:Eno jedro, 100.000 mest: LKH is ahead after 2 minutes in three of three seeds (0.816% / 0.822% / 0.807% against 0.858% / 0.886% / 0.865%) and after 5 minutes LKH is shorter in two of the three seeds, by 0.029 and 0.003 percentage points, and 8zTSP in one, by 0.002 (0.786% / 0.785% / 0.785% against 0.784% / 0.814% / 0.788%).LKH je po 2 minutah spredaj pri treh od treh semen (0,816 % / 0,822 % / 0,807 % proti 0,858 % / 0,886 % / 0,865 %), po 5 minutah pa je LKH krajši pri dveh od treh semen, za 0,029 in 0,003 odstotne točke, 8zTSP pa pri enem, za 0,002 (0,786 % / 0,785 % / 0,785 % proti 0,784 % / 0,814 % / 0,788 %).
  • One core, 1,000,000 cities:Eno jedro, 1.000.000 mest: LKH is ahead once it has its first tour (after 20 minutes 0.833% against 0.942%; after 30 minutes 0.816% against 0.906%). LKH’s first tour came after 18.6 min; 8zTSP 0.5.0 was within 1% of the bound after 13.2 min.LKH je spredaj, ko ima svojo prvo turo (po 20 minutah 0,833 % proti 0,942 %; po 30 minutah 0,816 % proti 0,906 %). LKH je prvo turo dobil po 18,6 min; 8zTSP 0.5.0 je bil po 13,2 min manj kot 1 % nad mejo.
  • Six threads, 1,000,000 cities:Šest niti, 1.000.000 mest: after 20 minutes LKH 0.825% and 8zTSP 0.833%; after 30 minutes LKH 0.810% and 8zTSP 0.815%. The difference after 30 minutes is 0.005 percentage points. For scale, two same-seed LKH runs on E100k.0 differed by about 0.01 percentage points after five minutes; more repeated runs are needed to interpret this difference.po 20 minutah LKH 0,825 % in 8zTSP 0,833 %; po 30 minutah LKH 0,810 % in 8zTSP 0,815 %. Razlika po 30 minutah je 0,005 odstotne točke. Za merilo: dva teka LKH z istim semenom na E100k.0 sta se po petih minutah razlikovala za približno 0,01 odstotne točke; za razlago te razlike je potrebnih več ponovljenih tekov.
  • 10,000 cities (E10k.0):10.000 mest (E10k.0): LKH is ahead at every measured time from its first tour (after 2 minutes 0.711% against 0.720% with six threads, and 0.718% against 0.778% with one thread). At 1 s it has no tour yet.LKH je od svoje prve ture spredaj pri vsakem merjenem času (po 2 minutah 0,711 % proti 0,720 % s šestimi nitmi in 0,718 % proti 0,778 % z eno nitjo). Pri 1 s še nima ture.
  • 24,978 cities (sw24978), six cores:24.978 mest (sw24978), šest jeder: in this seed LKH had shorter tours than release 0.5.0 at every measured time at which both had a tour. The earlier candidate 0.5.0-rc1 was ahead at five minutes and level at ten minutes; a comparison across candidates does not replace a comparison over several seeds. After 20 minutes LKH is at 0.003%; 8zTSP 0.5.0 has no value at that time and its candidate 0.5.0-rc1 is at 0.010%. Late results of a single seed differ more among our own candidates than between us and LKH. On one core LKH is ahead at every time from its first tour.pri tem semenu je imel LKH krajše ture kot izdaja 0.5.0 pri vsakem merjenem času, ob katerem sta oba imela turo. Prejšnji kandidat 0.5.0-rc1 je bil po petih minutah spredaj in po desetih izenačen; primerjava med kandidati ne nadomesti primerjave na več semenih. Po 20 minutah je LKH pri 0,003 %; 8zTSP 0.5.0 za ta čas nima vrednosti, njegov kandidat 0.5.0-rc1 pa je pri 0,010 %. Pozni izidi enega semena se med našimi kandidati razlikujejo bolj kot med nami in LKH. Na enem jedru je LKH od svoje prve ture spredaj pri vsakem času.

Development results, 11–12 November 2026Razvojni rezultati, 11. in 12. november 2026

Development runs, one seed per entry, one laptop; conditions are stated with every table.Razvojni teki, eno seme na vnos, en prenosnik; pogoji so navedeni ob vsaki tabeli.

These are results of development runs, not of the locked test described underTo so rezultati razvojnih tekov, ne zaklenjenega preizkusa, opisanega v razdelku GoalCilj: one seed, development instances, one laptop. They are partial and can change. The conditions come first, then the tables.: eno seme, razvojni primeri, en prenosnik. So delni in se lahko spremenijo. Najprej so pogoji, nato tabele.

More on the development resultsVeč o razvojnih rezultatih
  • Programs.Programi. 8zTSP 0.5.0 in the mode for large maps, on six threads and on one thread. The other programs ran on one core each, all at the same time next to each other.8zTSP 0.5.0 v načinu za velike zemljevide, na šestih nitih in na eni niti. Drugi programi so tekli vsak na enem jedru, vsi hkrati drug ob drugem.
  • LKH variants.Različice LKH. "LKH, geometric settings, six independent runs" is the entry of the earlier tables. "LKH POPMUSIC" is LKH 3.0.7 with POPMUSIC candidates and an initial period of 100 (the settings of the POPMUSIC report that comes with LKH). "LKH POPMUSIC without subgradient ascent" is the same without the subgradient ascent and with the best POPMUSIC tour as the starting tour. "LKH, one POPMUSIC tour" is a single strong POPMUSIC tour with no Lin–Kernighan trials after it (settings from the same report); it ends by itself.»LKH, geometrijske nastavitve, šest neodvisnih tekov« je vnos iz prejšnjih tabel. »LKH POPMUSIC« je LKH 3.0.7 s kandidati POPMUSIC in začetnim obdobjem 100 (nastavitve iz poročila o POPMUSIC, ki je priloženo LKH). »LKH POPMUSIC brez subgradientnega vzpona« je isto brez subgradientnega vzpona in z najboljšo turo POPMUSIC kot začetno. »LKH, ena tura POPMUSIC« je ena sama močna tura POPMUSIC brez poskusov Lin–Kernighan po njej (nastavitve iz istega poročila); konča sama.
  • linkern. Concorde's chained Lin–Kernighan program with a given time limit; "linkern, limit 300 s" has a limit of 300 s.Concordov program chained Lin–Kernighan z dano časovno mejo; »linkern, meja 300 s« ima mejo 300 s.
  • Percentages.Odstotki. On E1M.0 the percentages are over the published Held–Karp bound (seePri E1M.0 so odstotki nad objavljeno mejo Held–Karp (glej MethodMetoda). E3M.0 and E10M.0 were made with the official generator portgen and have no published bound; their reference is the estimate 0.7080·√(N·A), rounded up, where N is the number of cities and A the area. Their percentages are therefore above an estimate, not above the optimum.). E3M.0 in E10M.0 sta narejena z uradnim generatorjem portgen in nimata objavljene meje; njuna referenca je ocena 0,7080·√(N·A), zaokrožena navzgor, kjer je N število mest in A površina. Njuni odstotki so zato nad oceno, ne nad optimumom.
  • Processor.Procesor. Under full load the processor runs at 92–99 °C and lowers its clock. The column "processor during the run" gives the average clock of the cores of that run and the share of ten-second samples at 95 °C or more.Pri polni obremenitvi je procesor pri 92 do 99 °C in zniža takt. Stolpec »procesor med tekom« navaja povprečni takt jeder tega teka in delež desetsekundnih vzorcev z najmanj 95 °C.
  • Clocks.Ure. 8zTSP 0.5.0 measures time from the end of reading the input file (reading took 0.25 s at E1M.0, 1.0 s at E3M.0 and 2.7 s at E10M.0); LKH and linkern measure from the start of the process (see8zTSP 0.5.0 meri čas od konca branja vhodne datoteke (branje je trajalo 0,25 s pri E1M.0, 1,0 s pri E3M.0 in 2,7 s pri E10M.0); LKH in linkern merita od zagona procesa (glej MethodMetoda).
  • Reading the tables.Branje tabel. In the tables,V tabelah no tourni ture means that the program had not yet written a tour at that time, andpomeni, da program ob tem času še ni zapisal ture, not reachedni doseženo that the entry never had a tour that good. A dash (pa da vnos nikoli ni imel tako dobre ture. Črtica (–) means that there is no value: the time limit of that entry was already over at that time, or the value was not recorded. The entries have different time limits, so compare them at equal times.) pomeni, da vrednosti ni: časovna meja tega vnosa je bila ob tem času že pretekla ali vrednost ni bila zapisana. Vnosi imajo različne časovne meje, zato jih primerjajte pri enakih časih.

Ladder to very large instancesLestev do zelo velikih primerov

1,000,000 cities (DIMACS E1M.0), percent over the Held–Karp bound1.000.000 mest (DIMACS E1M.0), odstotki nad mejo Held–Karp

Quality over time.Kakovost skozi čas. One seed per entry. The first three columns repeat the values of the E1M.0 table above at the times that both tables share; the other columns are new.Eno seme na vnos. Prvi trije stolpci ponavljajo vrednosti iz zgornje tabele za E1M.0 ob časih, ki jih imata tabeli skupne; drugi stolpci so novi.

TimeČas8zTSP 0.5.0, six threads8zTSP 0.5.0, šest niti8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nitLKH, geometric settings, six independent runsLKH, geometrijske nastavitve, šest neodvisnih tekovLKH POPMUSICLKH POPMUSIC without subgradient ascentLKH POPMUSIC brez subgradientnega vzponaLKH, one POPMUSIC tourLKH, ena tura POPMUSIClinkern, limit 60 slinkern, meja 60 slinkern, limit 300 slinkern, meja 300 slinkern, limit 900 slinkern, meja 900 slinkern, limit 1800 slinkern, meja 1800 s
60 s2.258%2,258 %2.779%2,779 %no tourni tureno tourni tureno tourni tureno tourni tureno tourni tureno tourni tureno tourni tureno tourni ture
2.0 min2,0 min1.160%1,160 %2.294%2,294 %no tourni tureno tourni tureno tourni tureno tourni ture2.045%2,045 %no tourni tureno tourni tureno tourni ture
5.0 min5,0 min0.923%0,923 %1.360%1,360 %no tourni tureno tourni tureno tourni tureno tourni ture2.045%2,045 %no tourni tureno tourni tureno tourni ture
10.0 min10,0 min0.876%0,876 %1.055%1,055 %no tourni tureno tourni tureno tourni ture2.657%2,657 %2.045%2,045 %1.238%1,238 %no tourni tureno tourni ture
15.0 min15,0 min0.848%0,848 %0.980%0,980 %no tourni tureno tourni tureno tourni ture2.657%2,657 %–1.238%1,238 %no tourni tureno tourni ture
20.0 min20,0 min0.833%0,833 %0.942%0,942 %0.825%0,825 %no tourni ture0.950%0,950 %2.657%2,657 %––0.988%0,988 %no tourni ture
30.0 min30,0 min0.815%0,815 %0.906%0,906 %0.810%0,810 %0.826%0,826 %0.897%0,897 %2.657%2,657 %–––no tourni ture
45.0 min45,0 min–––0.813%0,813 %0.870%0,870 %2.657%2,657 %–––0.904%0,904 %
60.0 min60,0 min–––0.810%0,810 %0.867%0,867 %2.657%2,657 %––––
90.0 min90,0 min–––0.804%0,804 %0.857%0,857 %–––––

Time to reach a gap, best tour, peak memory and processor.Čas do odmika, najboljša tura, največja poraba pomnilnika in procesor. The columns "to 3.0%" to "to 0.9%" give the time at which a tour at or below that percentage over the bound was first recorded. When several columns show the same time, the first recorded tour was already at or below the lowest of those percentages. "Best tour" is the best value and the time at which it was reached. A dash means that no value was recorded.Stolpci »do 3,0 %« do »do 0,9 %« navajajo čas, ob katerem je bila prvič zapisana tura z odstotkom nad mejo, ki je enak ali manjši od navedenega. Če je v več stolpcih isti čas, je bila prva zapisana tura že pri najnižjem od teh odstotkov ali pod njim. »Najboljša tura« je najboljša vrednost in čas, ob katerem je bila dosežena. Črtica pomeni, da vrednost ni bila zapisana.

EntryVnosto 3.0%do 3,0 %to 2.0%do 2,0 %to 1.5%do 1,5 %to 1.0%do 1,0 %to 0.9%do 0,9 %Best tourNajboljša turaPeak memoryNajvečja poraba pomnilnikaProcessor during the runProcesor med tekom
8zTSP 0.5.0, six threads8zTSP 0.5.0, šest niti29 s67 s86 s3.1 min3,1 min6.7 min6,7 min0.815% after 30.0 min0,815 % po 30,0 min––
8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nit47 s2.6 min2,6 min4.2 min4,2 min13.2 min13,2 minnot reachedni doseženo0.906% after 30.0 min0,906 % po 30,0 min––
LKH, geometric settings, six independent runsLKH, geometrijske nastavitve, šest neodvisnih tekov15.8 min15,8 min15.8 min15,8 min15.8 min15,8 min15.8 min15,8 min15.8 min15,8 min0.810% after 29.4 min0,810 % po 29,4 min––
LKH POPMUSIC26.8 min26,8 min26.8 min26,8 min26.8 min26,8 min26.8 min26,8 min26.8 min26,8 min0.804% after 89.5 min0,804 % po 89,5 min694 MB2,945 MHz, 1% of samples ≥ 95 °C2.945 MHz, 1 % vzorcev ≥ 95 °C
LKH POPMUSIC without subgradient ascentLKH POPMUSIC brez subgradientnega vzpona15.5 min15,5 min15.5 min15,5 min15.5 min15,5 min15.5 min15,5 min23.7 min23,7 min0.857% after 88.6 min0,857 % po 88,6 min696 MB2,943 MHz, 1% of samples ≥ 95 °C2.943 MHz, 1 % vzorcev ≥ 95 °C
LKH, one POPMUSIC tourLKH, ena tura POPMUSIC5.8 min5,8 minnot reachedni doseženonot reachedni doseženonot reachedni doseženonot reachedni doseženo2.657% after 5.8 min2,657 % po 5,8 min585 MB2,688 MHz, 0% of samples ≥ 95 °C2.688 MHz, 0 % vzorcev ≥ 95 °C
linkern, limit 60 slinkern, meja 60 s76 snot reachedni doseženonot reachedni doseženonot reachedni doseženonot reachedni doseženo2.045% after 76 s2,045 % po 76 s226 MB2,700 MHz, 0% of samples ≥ 95 °C2.700 MHz, 0 % vzorcev ≥ 95 °C
linkern, limit 300 slinkern, meja 300 s5.3 min5,3 min5.3 min5,3 min5.3 min5,3 minnot reachedni doseženonot reachedni doseženo1.238% after 5.3 min1,238 % po 5,3 min226 MB2,690 MHz, 0% of samples ≥ 95 °C2.690 MHz, 0 % vzorcev ≥ 95 °C
linkern, limit 900 slinkern, meja 900 s15.4 min15,4 min15.4 min15,4 min15.4 min15,4 min15.4 min15,4 minnot reachedni doseženo0.988% after 15.4 min0,988 % po 15,4 min226 MB2,725 MHz, 0% of samples ≥ 95 °C2.725 MHz, 0 % vzorcev ≥ 95 °C
linkern, limit 1800 slinkern, meja 1800 s30.5 min30,5 min30.5 min30,5 min30.5 min30,5 min30.5 min30,5 minnot reachedni doseženo0.904% after 30.5 min0,904 % po 30,5 min226 MB2,926 MHz, 1% of samples ≥ 95 °C2.926 MHz, 1 % vzorcev ≥ 95 °C

3,162,278 cities (DIMACS E3M.0), percent over the estimated reference3.162.278 mest (DIMACS E3M.0), odstotki nad ocenjeno referenco

Quality over time.Kakovost skozi čas. One seed per entry; percent over the estimated reference, not over the optimum.Eno seme na vnos; odstotki nad ocenjeno referenco, ne nad optimumom.

TimeČas8zTSP 0.5.0, six threads8zTSP 0.5.0, šest niti8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nitLKH, one POPMUSIC tourLKH, ena tura POPMUSIClinkern, limit 300 slinkern, meja 300 slinkern, limit 900 slinkern, meja 900 s
60 s4.682%4,682 %5.403%5,403 %no tourni tureno tourni tureno tourni ture
2.0 min2,0 min3.121%3,121 %3.714%3,714 %no tourni tureno tourni tureno tourni ture
5.0 min5,0 min1.314%1,314 %2.851%2,851 %no tourni tureno tourni tureno tourni ture
10.0 min10,0 min0.992%0,992 %2.022%2,022 %no tourni ture1.960%1,960 %no tourni ture
15.0 min15,0 min0.921%0,921 %1.576%1,576 %no tourni ture1.960%1,960 %no tourni ture
20.0 min20,0 min0.892%0,892 %1.334%1,334 %2.758%2,758 %1.960%1,960 %1.359%1,359 %
30.0 min30,0 min0.866%0,866 %1.115%1,115 %2.758%2,758 %–1.359%1,359 %
45.0 min45,0 min0.842%0,842 %0.992%0,992 %2.758%2,758 %––
60.0 min60,0 min–0.942%0,942 %2.758%2,758 %––
90.0 min90,0 min–0.893%0,893 %2.758%2,758 %––

Time to reach a gap, best tour, peak memory and processor.Čas do odmika, najboljša tura, največja poraba pomnilnika in procesor. Columns as for E1M.0, over the estimated reference.Stolpci kot pri E1M.0, nad ocenjeno referenco.

EntryVnosto 3.0%do 3,0 %to 2.0%do 2,0 %to 1.5%do 1,5 %to 1.0%do 1,0 %to 0.9%do 0,9 %Best tourNajboljša turaPeak memoryNajvečja poraba pomnilnikaProcessor during the runProcesor med tekom
8zTSP 0.5.0, six threads8zTSP 0.5.0, šest niti2.0 min2,0 min3.0 min3,0 min4.1 min4,1 min9.6 min9,6 min18.5 min18,5 min0.842% after 45.0 min0,842 % po 45,0 min541 MB2,430 MHz, 43% of samples ≥ 95 °C2.430 MHz, 43 % vzorcev ≥ 95 °C
8zTSP 0.5.0, one thread8zTSP 0.5.0, ena nit4.2 min4,2 min10.3 min10,3 min16.0 min16,0 min43.5 min43,5 min83.1 min83,1 min0.893% after 89.9 min0,893 % po 89,9 min532 MB2,942 MHz, 1% of samples ≥ 95 °C2.942 MHz, 1 % vzorcev ≥ 95 °C
LKH, one POPMUSIC tourLKH, ena tura POPMUSIC19.8 min19,8 minnot reachedni doseženonot reachedni doseženonot reachedni doseženonot reachedni doseženo2.758% after 19.8 min2,758 % po 19,8 min1,867 MB1.867 MB2,722 MHz, 0% of samples ≥ 95 °C2.722 MHz, 0 % vzorcev ≥ 95 °C
linkern, limit 300 slinkern, meja 300 s5.9 min5,9 min5.9 min5,9 minnot reachedni doseženonot reachedni doseženonot reachedni doseženo1.960% after 5.9 min1,960 % po 5,9 min740 MB2,905 MHz, 0% of samples ≥ 95 °C2.905 MHz, 0 % vzorcev ≥ 95 °C
linkern, limit 900 slinkern, meja 900 s15.9 min15,9 min15.9 min15,9 min15.9 min15,9 minnot reachedni doseženonot reachedni doseženo1.359% after 15.9 min1,359 % po 15,9 min740 MB3,019 MHz, 0% of samples ≥ 95 °C3.019 MHz, 0 % vzorcev ≥ 95 °C

10,000,000 cities (DIMACS E10M.0), percent over the estimated reference10.000.000 mest (DIMACS E10M.0), odstotki nad ocenjeno referenco

Quality over time.Kakovost skozi čas. One seed per entry; percent over the estimated reference, not over the optimum.Eno seme na vnos; odstotki nad ocenjeno referenco, ne nad optimumom.

TimeČas8zTSP 0.5.0, six threads8zTSP 0.5.0, šest nitilinkern, limit 1800 slinkern, meja 1800 s
60 s38.229%38,229 %no tourni ture
2.0 min2,0 min7.090%7,090 %no tourni ture
5.0 min5,0 min4.952%4,952 %no tourni ture
10.0 min10,0 min4.108%4,108 %no tourni ture
15.0 min15,0 min3.361%3,361 %no tourni ture
20.0 min20,0 min2.558%2,558 %no tourni ture
30.0 min30,0 min1.957%1,957 %no tourni ture
45.0 min45,0 min1.391%1,391 %1.891%1,891 %
60.0 min60,0 min1.138%1,138 %1.891%1,891 %
90.0 min90,0 min0.954%0,954 %–
120.0 min120,0 min0.875%0,875 %–

Time to reach a gap, best tour, peak memory and processor.Čas do odmika, najboljša tura, največja poraba pomnilnika in procesor. Columns as for E1M.0, over the estimated reference.Stolpci kot pri E1M.0, nad ocenjeno referenco.

EntryVnosto 3.0%do 3,0 %to 2.0%do 2,0 %to 1.5%do 1,5 %to 1.0%do 1,0 %to 0.9%do 0,9 %Best tourNajboljša turaPeak memoryNajvečja poraba pomnilnikaProcessor during the runProcesor med tekom
8zTSP 0.5.0, six threads8zTSP 0.5.0, šest niti17.2 min17,2 min29.0 min29,0 min40.1 min40,1 min78.4 min78,4 min110.1 min110,1 min0.875% after 119.9 min0,875 % po 119,9 min1,655 MB1.655 MB2,633 MHz, 47% of samples ≥ 95 °C2.633 MHz, 47 % vzorcev ≥ 95 °C
linkern, limit 1800 slinkern, meja 1800 s33.8 min33,8 min33.8 min33,8 minnot reachedni doseženonot reachedni doseženonot reachedni doseženo1.891% after 33.8 min1,891 % po 33,8 min2,452 MB2.452 MB2,746 MHz, 0% of samples ≥ 95 °C2.746 MHz, 0 % vzorcev ≥ 95 °C

A finding of the night: 8zTSP 0.5.0 at E10M.0 ran almost on one threadUgotovitev noči: 8zTSP 0.5.0 je pri E10M.0 tekel skoraj na eni niti

The thread that returns the results of the parts into the tour was 80 to 94% busy, the six worker threads about 1%: every returned result rewrote the whole array of ten million cities. The candidate 0.6.0-z1 returns the results of a group of parts in chunks and writes the array once per group (the same tours, checked). It is measured below.Nit, ki rezultate delov vrača v turo, je bila zasedena 80 do 94 %, šest delovnih niti okoli 1 %: vsak vrnjen rezultat je na novo zapisal celo polje desetih milijonov mest. Kandidat 0.6.0-z1 rezultate skupine delov vrača v kosih in polje zapiše enkrat na skupino (iste ture, preverjeno). Izmerjen je spodaj.

Candidate 0.6.0-z1Kandidat 0.6.0-z1

The candidate 0.6.0-z1 returns the results of a group of parts in chunks (from 200,000 cities upward; the same tours as before), and its warm-up by sections of the tour on all threads is reworked. It is given the logical CPUs of seven of the eight physical cores (both logical CPUs of each, 14 in all); the eighth physical core stays with the system and other work. The number of threads for the large instances was chosen by a rule written down before the runs: the fastest to 1% above the bound on E1M.0, otherwise the smallest gap after 600 s. These are development measurements with one seed. The goal test next to LKH and linkern on equal resources follows under a locked protocol (seeKandidat 0.6.0-z1 rezultate skupine delov vrača v kosih (od 200.000 mest naprej; iste ture kot prej), njegovo ogrevanje po odsekih ture na vseh nitih pa je predelano. Na voljo ima logične CPU sedmih od osmih fizičnih jeder (oba logična CPU-ja vsakega, skupaj 14); osmo fizično jedro ostane sistemu in drugemu delu. Število niti za velike primere je izbrano po pravilu, zapisanem pred teki: najhitreje do 1 % nad mejo na E1M.0, sicer najmanjši odmik po 600 s. To so razvojne meritve z enim semenom. Ciljni preizkus ob LKH in linkernu na enakih virih sledi po zaklenjenem protokolu (glej GoalCilj), starting with a development pilot on 12 November.), najprej z razvojnim pilotom 12. novembra.

Scaling on E1M.0: 14, 7 and 6 threadsSkaliranje na E1M.0: 14, 7 in 6 niti

Scaling.Skaliranje. DIMACS E1M.0, 0.6.0-z1 with warm-up by sections, 600 s; percent over the Held–Karp bound.DIMACS E1M.0, 0.6.0-z1 z ogrevanjem po odsekih, 600 s; odstotki nad mejo Held–Karp.

TimeČas14 threads (both logical CPUs of 7 physical cores)14 niti (oba logična CPU-ja 7 fizičnih jeder)7 threads (7 physical cores)7 niti (7 fizičnih jeder)6 threads (6 physical cores)6 niti (6 fizičnih jeder)
30 s1.621%1,621 %1.621%1,621 %1.654%1,654 %
60 s1.236%1,236 %1.215%1,215 %1.239%1,239 %
2.0 min2,0 min1.016%1,016 %1.007%1,007 %1.023%1,023 %
5.0 min5,0 min0.909%0,909 %0.907%0,907 %0.910%0,910 %
10.0 min10,0 min0.862%0,862 %0.866%0,866 %0.868%0,868 %

Time to reach a gap, best tour and peak memory.Čas do odmika, najboljša tura in največja poraba pomnilnika. Columns as in the tables above.Stolpci kot v zgornjih tabelah.

EntryVnosto 3.0%do 3,0 %to 2.0%do 2,0 %to 1.5%do 1,5 %to 1.0%do 1,0 %to 0.9%do 0,9 %Best tourNajboljša turaPeak memoryNajvečja poraba pomnilnika
14 threads (both logical CPUs of 7 physical cores)14 niti (oba logična CPU-ja 7 fizičnih jeder)7.60 s7,60 s20 s39 s2.2 min2,2 min5.7 min5,7 min0.862% after 9.8 min0,862 % po 9,8 min253 MB
7 threads (7 physical cores)7 niti (7 fizičnih jeder)6.95 s6,95 s20 s38 s2.1 min2,1 min5.6 min5,6 min0.866% after 9.9 min0,866 % po 9,9 min214 MB
6 threads (6 physical cores)6 niti (6 fizičnih jeder)7.59 s7,59 s21 s40 s2.3 min2,3 min5.8 min5,8 min0.868% after 9.9 min0,868 % po 9,9 min207 MB

By this rule 7 threads were chosen. With 14 threads (both logical CPUs of each core) the kernel is not faster than with seven threads on seven physical cores; the large instances are therefore measured with seven threads, and the 14-thread run stays shown.Po tem pravilu je izbranih 7 niti. Pri 14 nitih (oba logična CPU-ja vsakega jedra) jedro ni hitrejše kot pri sedmih nitih na sedmih fizičnih jedrih; zato velike primere meri s sedmimi nitmi, 14-nitni tek pa ostane prikazan.

E10M.0: 0.6.0-z1 on seven threads next to 0.5.0 on six threadsE10M.0: 0.6.0-z1 s sedmimi nitmi ob 0.5.0 s šestimi

E10M.0, percent over the estimated reference.E10M.0, odstotki nad ocenjeno referenco. 0.6.0-z1 on 7 threads with warm-up by sections next to 0.5.0 on 6 threads. The 0.5.0 column repeats the values of the E10M.0 ladder above at the times that both tables share.0.6.0-z1 na 7 nitih z ogrevanjem po odsekih ob 0.5.0 na 6 nitih. Stolpec 0.5.0 ponavlja vrednosti zgornje lestve za E10M.0 ob časih, ki jih imata tabeli skupne.

TimeČas8zTSP 0.6.0-z1, 7 threads, warm-up by sections8zTSP 0.6.0-z1, 7 niti, ogrevanje po odsekih8zTSP 0.5.0, 6 threads8zTSP 0.5.0, 6 niti
30 s7.691%7,691 %38.229%38,229 %
60 s3.131%3,131 %38.229%38,229 %
2.0 min2,0 min2.301%2,301 %7.090%7,090 %
5.0 min5,0 min1.717%1,717 %4.952%4,952 %
10.0 min10,0 min1.233%1,233 %4.108%4,108 %
15.0 min15,0 min1.055%1,055 %3.361%3,361 %
20.0 min20,0 min0.982%0,982 %2.558%2,558 %
30.0 min30,0 min0.905%0,905 %1.957%1,957 %
45.0 min45,0 min0.854%0,854 %1.391%1,391 %
60.0 min60,0 min0.831%0,831 %1.138%1,138 %

Time to reach a gap, best tour and peak memory.Čas do odmika, najboljša tura in največja poraba pomnilnika. Columns as in the tables above.Stolpci kot v zgornjih tabelah.

EntryVnosto 3.0%do 3,0 %to 2.0%do 2,0 %to 1.5%do 1,5 %to 1.0%do 1,0 %to 0.9%do 0,9 %Best tourNajboljša turaPeak memoryNajvečja poraba pomnilnika
8zTSP 0.6.0-z1, 7 threads, warm-up by sections8zTSP 0.6.0-z1, 7 niti, ogrevanje po odsekih96 s3.6 min3,6 min6.4 min6,4 min18.4 min18,4 min31.2 min31,2 min0.831% after 59.9 min0,831 % po 59,9 min1,671 MB1.671 MB
8zTSP 0.5.0, 6 threads8zTSP 0.5.0, 6 niti17.2 min17,2 min29.0 min29,0 min40.1 min40,1 min78.4 min78,4 min110.1 min110,1 min0.875% after 119.9 min0,875 % po 119,9 min1,655 MB1.655 MB

At every time shown, 0.6.0-z1 is at a lower percentage than 0.5.0; note that it ran on 7 threads and 0.5.0 on 6.Ob vsakem prikazanem času je 0.6.0-z1 pri nižjem odstotku kot 0.5.0; pri tem je tekel na 7 nitih, 0.5.0 pa na 6.

Does a numbering that keeps neighbours close speed up the kernel?Ali oštevilčenje, ki ohrani bližino, pospeši jedro?

E1M.0, 0.6.0-z1, one thread, 900 s, seed 42.E1M.0, 0.6.0-z1, ena nit, 900 s, seme 42. The instance is the same (cities and coordinates); only the city numbers differ. Percent over the Held–Karp bound; time by the kernel's clock.Primer je isti (mesta in koordinate), razlikujejo se samo številke mest. Odstotki nad mejo Held–Karp; čas po uri jedra.

TimeČasOwn numberingLastno oštevilčenjeNumbering that keeps neighbours closeOštevilčenje, ki ohrani bližino
30 s3.609%3,609 %2.532%2,532 %
60 s3.175%3,175 %2.283%2,283 %
2.0 min2,0 min2.308%2,308 %1.843%1,843 %
5.0 min5,0 min1.301%1,301 %1.210%1,210 %
10.0 min10,0 min1.026%1,026 %1.007%1,007 %
15.0 min15,0 min0.957%0,957 %0.952%0,952 %

At every time shown the instance with that numbering is at a lower percentage: the difference is large early (2.532% against 3.609% at 30 s) and small at the end (0.952% against 0.957% at 15.0 min).Ob vsakem prikazanem času je primer s takim oštevilčenjem pri nižjem odstotku: razlika je zgodaj velika (2,532 % proti 3,609 % po 30 s) in na koncu majhna (0,952 % proti 0,957 % po 15,0 min).

Hybrid 0.6.0-hyb1 (LKH as the local search in our analysis of the view) on 100,000 and 316,000 citiesHibrid 0.6.0-hyb1 (LKH kot lokalno iskanje v naši analizi pogleda) na 100.000 in 316.000 mestih

The hybrid uses LKH 3.0.7 as the local search inside our kernel’s analysis of the view. All values below are development values: one seed, a development version of 12 November, and processors that were not reserved. Only the rows of one table are compared with each other, not rows of different tables or the tables of the other sections. Percentages are over the published Held–Karp bound of the instance.Hibrid uporablja LKH 3.0.7 kot lokalno iskanje v analizi pogleda našega jedra. Vse spodnje vrednosti so razvojne: eno seme, razvojna različica z dne 12. novembra in procesorji, ki niso bili rezervirani. Med seboj se primerjajo samo vrstice iste tabele, ne vrstice različnih tabel ali tabele drugih razdelkov. Odstotki so nad objavljeno mejo Held–Karp primera.

  • Our ladder (without LKH)Naša lestev (brez LKH) is the work of the kernel itself on the parts of the view, with its fixed ladder of sizes.je delo jedra samega na delih pogleda z njegovo stalno lestvijo velikosti.
  • Hybrid: LKH first, then the ladder.Hibrid: LKH najprej, nato lestev. LKH works in every part of the view until a pass gains less than a set level; then the ladder takes over.LKH dela v vsakem delu pogleda, dokler prehod ne pridobi manj od določene ravni; nato prevzame lestev.
  • Hybrid: only LKH.Hibrid: samo LKH. LKH works in every part of the view.LKH dela v vsakem delu pogleda.
  • Hybrid: LKH in the ladder.Hibrid: LKH v lestvi. LKH works in every part of the view, and the ladder raises the number of LKH trials before it widens the parts.LKH dela v vsakem delu pogleda, lestev pa poveča število poskusov LKH, preden dele razširi.

The values at the times shown are the lengths that the kernel reported on its own clock; they are not verified tour versions. The final tour is the tour at the end of the run, with its length recomputed from the tour file and checked as a permutation; it can be slightly shorter than the last value shown.Vrednosti ob prikazanih časih so dolžine, ki jih je jedro sporočilo s svojo uro; niso preverjene verzije tur. Končna tura je tura ob koncu teka, njena dolžina je znova izračunana iz datoteke ture in preverjena kot permutacija; lahko je malo krajša od zadnje prikazane vrednosti.

E100k.0, 120 s, two threadsE100k.0, 120 s, dve niti

E100k.0, 120 s, two threads.E100k.0, 120 s, dve niti. Seed 42. The two threads ran on the two logical CPUs of one physical core at low priority, shared with other work.Seme 42. Niti sta tekli na dveh logičnih CPU enega fizičnega jedra z nizko prioriteto, ki ju je delilo drugo delo.

EntryVnos30 s60 s90 s120 sFinal tourKončna tura
Our ladder (without LKH)Naša lestev (brez LKH)1.706%1,706 %1.111%1,111 %0.963%0,963 %0.892%0,892 %0.892%0,892 %
Hybrid: LKH first, then the ladderHibrid: LKH najprej, nato lestev1.256%1,256 %0.990%0,990 %0.911%0,911 %0.878%0,878 %0.878%0,878 %
Hybrid: only LKHHibrid: samo LKH1.370%1,370 %1.022%1,022 %0.936%0,936 %0.910%0,910 %0.910%0,910 %
Hybrid: LKH in the ladderHibrid: LKH v lestvi1.380%1,380 %1.028%1,028 %0.944%0,944 %0.897%0,897 %0.896%0,896 %

At every time shown, "LKH first, then the ladder" is at a lower percentage than our ladder (0.878% against 0.892% at 120 s). With LKH in every part ("only LKH" and "LKH in the ladder") the percentage is lower until 90 s and higher at 120 s (0.910% and 0.897% against 0.892%).Ob vsakem prikazanem času je »LKH najprej, nato lestev« pri nižjem odstotku kot naša lestev (0,878 % proti 0,892 % po 120 s). Pri LKH v vsakem delu (»samo LKH« in »LKH v lestvi«) je odstotek do 90 s nižji, po 120 s pa višji (0,910 % in 0,897 % proti 0,892 %).

E100k.0, 300 s, two threadsE100k.0, 300 s, dve niti

E100k.0, 300 s, two threads.E100k.0, 300 s, dve niti. Seed 42. The same two logical CPUs at low priority, while the pilot ran on the other cores.Seme 42. Ista dva logična CPU z nizko prioriteto, medtem ko je na drugih jedrih tekel pilot.

EntryVnos30 s60 s120 s180 s240 s300 sFinal tourKončna tura
Our ladder (without LKH)Naša lestev (brez LKH)1.153%1,153 %0.912%0,912 %0.819%0,819 %0.796%0,796 %0.778%0,778 %0.764%0,764 %0.764%0,764 %
Hybrid: LKH first, then the ladderHibrid: LKH najprej, nato lestev0.991%0,991 %0.878%0,878 %0.824%0,824 %0.799%0,799 %0.783%0,783 %0.771%0,771 %0.771%0,771 %

At 30 s and 60 s "LKH first, then the ladder" is lower than our ladder (0.991% against 1.153%; 0.878% against 0.912%); from 120 s on it is a little higher (0.824% against 0.819% at 120 s; 0.771% against 0.764% at 300 s).Po 30 s in po 60 s je »LKH najprej, nato lestev« nižje od naše lestve (0,991 % proti 1,153 %; 0,878 % proti 0,912 %); od 120 s naprej je malo višje (0,824 % proti 0,819 % po 120 s; 0,771 % proti 0,764 % po 300 s).

E316k.0, 300 s, two threadsE316k.0, 300 s, dve niti

E316k.0, 300 s, two threads.E316k.0, 300 s, dve niti. Seed 42. The processors were not reserved.Seme 42. Procesorji niso bili rezervirani.

EntryVnos30 s60 s90 s120 s180 s240 s300 sFinal tourKončna tura
Our ladder (without LKH)Naša lestev (brez LKH)3.523%3,523 %2.599%2,599 %1.999%1,999 %1.554%1,554 %1.307%1,307 %1.168%1,168 %1.097%1,097 %1.097%1,097 %
Hybrid: LKH first, then the ladderHibrid: LKH najprej, nato lestev3.767%3,767 %2.568%2,568 %1.708%1,708 %1.368%1,368 %1.159%1,159 %1.094%1,094 %1.044%1,044 %1.044%1,044 %

At 30 s the hybrid is higher than our ladder (3.767% against 3.523%); from 60 s on it is lower (2.568% against 2.599% at 60 s; 1.044% against 1.097% at 300 s).Po 30 s je hibrid višji od naše lestve (3,767 % proti 3,523 %); od 60 s naprej je nižji (2,568 % proti 2,599 % po 60 s; 1,044 % proti 1,097 % po 300 s).

Fewer LKH trials and larger parts: E100k.0, 300 s, one threadManj poskusov LKH in večji deli: E100k.0, 300 s, ena nit

Hybrid with fewer LKH trials and larger parts, E100k.0, 300 s, one thread.Hibrid z manj poskusi LKH in večjimi deli, E100k.0, 300 s, ena nit. One seed; one core at low priority. The time to 1% and the value at 300 s are from verified tour versions.Eno seme; eno jedro z nizko prioriteto. Čas do 1 % in vrednost ob 300 s sta iz preverjenih verzij tur.

EntryVnosTo 1% (verified)Do 1 % (preverjeno)At 300 s (verified)Ob 300 s (preverjeno)
3 trials, parts of 1,000–8,000 cities (hyb1 setting)3 poskusi, deli po 1.000–8.000 mest (nastavitev hyb1)71 s0.806%0,806 %
1 trial, parts of 1,000–8,000 cities1 poskus, deli po 1.000–8.000 mest97 s0.839%0,839 %
1 trial, parts of 3,000–24,000 cities1 poskus, deli po 3.000–24.000 mest2.6 min2,6 min0.876%0,876 %
1 trial, parts of 10,000–80,000 cities1 poskus, deli po 10.000–80.000 mest4.6 min4,6 min0.998%0,998 %

LKH by subproblems with one trial per piece (a finding in favour of the competitor)LKH po podproblemih z enim poskusom na kos (ugotovitev v prid tekmecu)

LKH 3.0.7 by subproblems starts from a POPMUSIC tour and then works in rounds of subproblems: pieces of the stated size, with the stated number of trials per piece. This setting of LKH is not in the tables above. It is shown here because, with one trial per piece, it reaches 1% early.LKH 3.0.7 s podproblemi začne z turo POPMUSIC, nato dela v krogih podproblemov: kosov navedene velikosti z navedenim številom poskusov na kos. Te nastavitve LKH ni v zgornjih tabelah. Prikazana je, ker pri enem poskusu na kos zgodaj pride do 1 %.

E100k.0, 600 s, one coreE100k.0, 600 s, eno jedro

E100k.0, 600 s, one core.E100k.0, 600 s, eno jedro. One seed; one core at low priority, shared with other work. "At 300 s" and "At the end" are the lengths that LKH printed; "To 1%" is from verified tour versions. For the rows with 100 and 30 trials only the record of the older arena exists, without verified versions. A run with 3 trials that shared its core with a test suite was set aside and repeated; the repeat is shown.Eno seme; eno jedro z nizko prioriteto, deljeno z drugim delom. »Ob 300 s« in »Ob koncu« sta dolžini, ki ju je izpisal LKH; »Do 1 %« je iz preverjenih verzij tur. Pri vrstah s 100 in 30 poskusi obstaja samo zapis starejše arene, brez preverjenih verzij. Tek s 3 poskusi, ki je delil jedro z zbirko testov, je odložen in ponovljen; prikazan je ponovljeni tek.

Trials per piecePoskusov na kosPiece size (cities)Velikost kosa (mest)At 300 s (LKH output)Ob 300 s (izpis LKH)At the end (LKH output)Ob koncu (izpis LKH)To 1% (verified)Do 1 % (preverjeno)
10010,00010.0002.312%2,312 %1.697%1,697 %no verified record (older arena)ni preverjenega zapisa (starejša arena)
3010,00010.0001.163%1,163 %1.021%1,021 %no verified record (older arena)ni preverjenega zapisa (starejša arena)
1010,00010.0000.969%0,969 %0.944%0,944 %4.1 min4,1 min
310,00010.0000.920%0,920 %0.872%0,872 %101 s
110,00010.0000.835%0,835 %0.823%0,823 %58 s
13,0003.0000.823%0,823 %0.788%0,788 %96 s
130,00030.0000.897%0,897 %0.897%0,897 %73 s

With one trial per piece LKH reaches 1% fastest in this table (58 s with pieces of 10,000 cities). With 3 and 10 trials it needs 101 s and 4.1 min; with 30 and 100 trials its output ends at 1.021% and 1.697%, so there is no verified time to 1%. With one trial and pieces of 3,000 or 30,000 cities it reaches 1% after 96 s and 73 s and ends at 0.788% and 0.897%.Pri enem poskusu na kos LKH v tej tabeli najhitreje pride do 1 % (58 s s kosi po 10.000 mest). Pri 3 in 10 poskusih potrebuje 101 s in 4,1 min; pri 30 in 100 poskusih se njegov izpis konča pri 1,021 % in 1,697 %, zato preverjenega časa do 1 % ni. Pri enem poskusu in kosih po 3.000 ali 30.000 mest pride do 1 % po 96 s in 73 s in se konča pri 0,788 % in 0,897 %.

On E1M.0 the same LKH (1 trial, pieces of 10,000 cities) needed, in 3,600 s on one core (shared, at low priority), 9.0 min / 11.4 min / 13.3 min / 23.3 min to reach 3% / 2% / 1.5% / 1% above the bound, and was at 0.894% after 1 h (verified versions). After the run had passed 1%, a short test on the second logical CPU of its core disturbed the late part of the curve a little.Pri E1M.0 je isti LKH (1 poskus, kosi po 10.000 mest) v 3.600 s na enem jedru (deljenem, z nizko prioriteto) do 3 % / 2 % / 1,5 % / 1 % nad mejo potreboval 9,0 min / 11,4 min / 13,3 min / 23,3 min, po 1 h pa je bil pri 0,894 % (preverjene verzije). Potem ko je tek prešel 1 %, je kratek preizkus na drugem logičnem CPU istega jedra njegov pozni del malo motil.

Pilot of the goal test on 1,000,000 citiesPilot ciljnega preizkusa na 1.000.000 mest

The pilot is a development run of the goal test on E1M.0: one seed (42), one program at a time, on the same laptop. 8zTSP 0.6.0-z1 runs on 7 threads. LKH 3.0.7 (POPMUSIC candidates, without subgradient ascent) and Concorde's linkern run as several independent runs on one core each, and the best of them counts at each moment. Stage 0 chose the number of independent runs; stage 1 runs the chosen settings.Pilot je razvojni tek ciljnega preizkusa na E1M.0: eno seme (42), en program naenkrat, na istem prenosniku. 8zTSP 0.6.0-z1 teče na 7 nitih. LKH 3.0.7 (kandidati POPMUSIC, brez subgradientnega vzpona) in Concordov linkern tečeta kot več neodvisnih tekov, vsak na enem jedru, šteje pa najboljši v vsakem trenutku. Stopnja 0 je izbrala število neodvisnih tekov; stopnja 1 požene izbrane nastavitve.

Values.Vrednosti. Final tour: recomputed from the tour file and checked (verified). Times to 1% and 0.9%: from the program's own report, not from verified tour versions. Peak memory: the sum over all processes of the run, under a cap of 15 GB. Percentages are over the published Held–Karp bound; for E10M.0 over the estimated reference.Končna tura: znova izračunana iz datoteke ture in preverjena (preverjeno). Časi do 1 % in 0,9 %: iz poročila programa samega, ne iz preverjenih verzij tur. Največja poraba pomnilnika: vsota vseh procesov teka, pod mejo 15 GB. Odstotki so nad objavljeno mejo Held–Karp; pri E10M.0 nad ocenjeno referenco.

Stage 1, E1M.0, 1,800 s.Stopnja 1, E1M.0, 1.800 s. The pilot runs with the chosen settings.Pilotski teki z izbranimi nastavitvami.

EntryVnosTo 1%Do 1 %To 0.9%Do 0,9 %Final tour (verified)Končna tura (preverjeno)Above the referenceNad referencoPeak memoryNajvečja poraba pomnilnika
8zTSP 0.6.0-z1, 7 threads, warm-up by sections8zTSP 0.6.0-z1, 7 niti, ogrevanje po odsekih122.8 s122,8 s319.9 s319,9 s714,463,887714.463.8870.813%0,813 %316 MB
LKH 3.0.7, POPMUSIC without subgradient ascent, 4 independent runsLKH 3.0.7, POPMUSIC brez subgradientnega vzpona, 4 neodvisni teki810.7 s810,7 s1,318.3 s1.318,3 s714,973,318714.973.3180.885%0,885 %2,781 MB2.781 MB
linkern, 4 independent runslinkern, 4 neodvisni teki596.5 s596,5 s1,431.9 s1.431,9 s714,966,793714.966.7930.884%0,884 %1,035 MB1.035 MB

Stage 0, E1M.0.Stopnja 0, E1M.0. The runs behind the choice of the number of independent runs: LKH 1,800 s, linkern 900 s. For both, 4 independent runs were chosen.Teki, na katerih temelji izbira števila neodvisnih tekov: LKH 1.800 s, linkern 900 s. Pri obeh so izbrani 4 neodvisni teki.

EntryVnosTo 1%Do 1 %To 0.9%Do 0,9 %Final tour (verified)Končna tura (preverjeno)Above the referenceNad referencoPeak memoryNajvečja poraba pomnilnika
LKH 3.0.7, POPMUSIC without subgradient ascent, 4 independent runsLKH 3.0.7, POPMUSIC brez subgradientnega vzpona, 4 neodvisni teki727.8 s727,8 s1,131.2 s1.131,2 s714,946,616714.946.6160.881%0,881 %2,774 MB2.774 MB
LKH 3.0.7, POPMUSIC without subgradient ascent, 7 independent runsLKH 3.0.7, POPMUSIC brez subgradientnega vzpona, 7 neodvisnih tekov905.4 s905,4 s1,356.7 s1.356,7 s714,976,996714.976.9960.885%0,885 %4,857 MB4.857 MB
LKH 3.0.7, POPMUSIC without subgradient ascent, 14 independent runsLKH 3.0.7, POPMUSIC brez subgradientnega vzpona, 14 neodvisnih tekov1,516.9 s1.516,9 snot reachedni doseženo715,420,773715.420.7730.948%0,948 %9,706 MB9.706 MB
linkern, 4 independent runslinkern, 4 neodvisni teki583.7 s583,7 snot reachedni doseženo715,441,732715.441.7320.951%0,951 %1,035 MB1.035 MB
linkern, 7 independent runslinkern, 7 neodvisnih tekov718.7 s718,7 snot reachedni doseženo715,599,466715.599.4660.973%0,973 %1,811 MB1.811 MB
linkern, 14 independent runslinkern, 14 neodvisnih tekovnot reachedni doseženonot reachedni doseženonot verifiedni preverjeno–3,622 MB3.622 MB

8zTSP 0.6.0-z1, 7 threads, with the default warm-up8zTSP 0.6.0-z1, 7 niti, s privzetim ogrevanjem (without the warm-up by sections used in stage 1), on E1M.0 and on E10M.0; one seed.(brez ogrevanja po odsekih, uporabljenega v stopnji 1), na E1M.0 in na E10M.0; eno seme.

EntryVnosTo 1%Do 1 %To 0.9%Do 0,9 %Final tour (verified)Končna tura (preverjeno)Above the referenceNad referencoPeak memoryNajvečja poraba pomnilnika
8zTSP 0.6.0-z1, 7 threads, default warm-up, E1M.0 (600 s)8zTSP 0.6.0-z1, 7 niti, privzeto ogrevanje, E1M.0 (600 s)161.6 s161,6 s350.1 s350,1 s714,826,289714.826.2890.864%0,864 %210 MB
8zTSP 0.6.0-z1, 7 threads, default warm-up, E10M.0 (3,600 s)8zTSP 0.6.0-z1, 7 niti, privzeto ogrevanje, E10M.0 (3.600 s)1,141.2 s1.141,2 s1,783.3 s1.783,3 s2,257,393,3892.257.393.3890.826%0,826 %1,619 MB1.619 MB

The pilot ended on 12 November; its comparison with LKH and linkern at 1,000,000, 3,162,278 and 10,000,000 cities is at the top of this page (Pilot se je končal 12. novembra; njegova primerjava z LKH in linkernom pri 1.000.000, 3.162.278 in 10.000.000 mestih je na vrhu te strani (Same laptop, same finish lineIsti prenosnik, ista ciljna črta).

How it was built and checkedKako je nastalo in bilo preverjeno

LimitsOmejitve

Claim discipline.Disciplina trditev. What these results do and do not support:Kaj ti rezultati podpirajo in česa ne:

  • Preliminary.Predhodno. All instances are development instances: the kernel and its settings were chosen on them (the nine national instances, E10k.0, E100k.0, E1M.0 and sw24978). The reserved confirmation set of instances has not been used.Vsi primeri so razvojni: jedro in njegove nastavitve so bili izbrani na njih (devet državnih primerov, E10k.0, E100k.0, E1M.0 in sw24978). Rezervirani potrditveni nabor primerov ni bil uporabljen.
  • Seeds.Semena. Ten on each national instance; three on E100k.0 (seeds 42, 142 and 242); one on every other large instance.Deset na vsakem državnem primeru; tri na E100k.0 (semena 42, 142 in 242); eno na vsakem drugem velikem primeru.
  • Noise.Šum. Two same-seed LKH runs on E100k.0 differed by about 0.01 percentage points after five minutes. Small differences have not been tested for robustness across repeated runs.Dva teka LKH z istim semenom na E100k.0 sta se po petih minutah razlikovala za približno 0,01 odstotne točke. Majhne razlike niso bile preizkušene na ponovljenih tekih.
  • One laptop.En prenosnik. The cores were not rotated between solvers; the two one-core entries ran on two fixed cores.Jedra se med reševalci niso menjala; vnosa z enim jedrom sta tekla na dveh stalnih jedrih.
  • LKH. Default settings on the national instances; on the large instances one set of geometric settings, not tuned per instance, run as six independent runs when six cores were used. With default settings LKH wrote no tour within 20 minutes at 24,978 cities, so there is no comparison with default settings at the large sizes.Na državnih primerih privzete nastavitve; na velikih primerih en nabor geometrijskih nastavitev, ne uglašen za posamezen primer, pri šestih jedrih kot šest neodvisnih tekov. S privzetimi nastavitvami LKH pri 24.978 mestih v 20 minutah ni zapisal ture, zato pri velikih velikostih ni primerjave s privzetimi nastavitvami.
  • Clocks.Ure. The clock of 8zTSP starts after the input file has been read; the clock of LKH starts at process start.Ura 8zTSP začne teči po branju vhodne datoteke; ura LKH ob zagonu procesa.
  • Bounds.Meje. Percentages on E10k.0, E100k.0 and E1M.0 are over a lower bound, not over the optimum.Odstotki pri E10k.0, E100k.0 in E1M.0 so nad spodnjo mejo, ne nad optimumom.
  • No proof.Brez dokaza. 8zTSP gives no proof of optimality; Concorde does.8zTSP ne da dokaza optimalnosti; Concorde ga da.
  • Concorde values marked *Vrednosti Concorda z zvezdico * are from its log; its tour was not on disk when it was stopped.so iz njegovega dnevnika; ob ustavitvi njegove ture ni bilo na disku.
  • Reconstructed records.Obnovljena zapisa. Two arena records from earlier series were reconstructed from the solver's own complete logs after the arena's follower stopped on a half-written line; the original records were kept and the final tours were rechecked. No number on this page comes from those two records.Dva zapisa arene iz prejšnjih serij sta bila obnovljena iz reševalčevih lastnih popolnih zapisov, potem ko se je sledilnik arene ustavil na napol zapisani vrstici; izvirna zapisa sta ohranjena, končni turi sta bili ponovno preverjeni. Nobena številka na tej strani ni iz teh dveh zapisov.
  • Known limitations of release 0.5.0.Znane omejitve izdaje 0.5.0. The deadline is not checked inside the construction of a single start tour or a single crossover; our view of the map is not linear if very many cities fall close together; the memory option is an estimate, not a hard limit.Rok se ne preverja znotraj gradnje posamezne začetne ture ali posameznega križanja; naš pogled na zemljevid ni linearen, če zelo veliko mest pade blizu skupaj; možnost za pomnilnik je ocena, ne trda meja.
  • Code.Koda. The 8zTSP code is not public, and no licence has been chosen yet.Koda 8zTSP ni javna in licenca še ni izbrana.

NextNaprej

State on 17 November 2026. Next: the build is frozen (from 18 November) and recorded on this laptop against the established solvers; then a three-day full test on the same laptop; then 100 million and 1 billion cities, and the World TSP. No confirmatory result is claimed here yet. The idea arena for the other ideas of the research (Rank, Don’t Eliminate, DCC as the governor of the search, and the rest) comes after the full test.Stanje 17. novembra 2026. Naslednji korak: gradnja bo zamrznjena (od 18. novembra) in posneta na tem prenosniku ob uveljavljenih reševalnikih; nato tridnevni polni preizkus na istem prenosniku; nato 100 milijonov in 1 milijarda mest ter svetovni TSP. Potrditvenega izida tu še ne navajamo. Arena idej za druge zamisli raziskave (Rank, Don’t Eliminate, DCC kot krmilnik iskanja in druge) pride po polnem preizkusu.

8zTSP-Q (planned): we assess that the 8zTSP architecture also suits quantum computers, as co-processors for parts of a problem.8zTSP-Q (v načrtu): ocenjujemo, da je arhitektura 8zTSP primerna tudi za kvantne računalnike, kot sopomočnike za dele problema.

CreditsZasluge

No programs, sources or mirrors of third-party software are offered on this page. The 8zTSP code is not public and has no licence yet.Na tej strani ni programov, izvorne kode ali zrcal tuje programske opreme. Koda 8zTSP ni javna in še nima licence.