I-Contra "i-Grandmaster-level Chess Ngaphandle Kokusesha" (2024): Kungani I-Pattern Recognition Alone Falls short

Iphepha le-Google DeepMind lango-2024 elithi grandmaster-level chess ngaphandle kwama-algorithms okusesha ajwayelekile lavusa ukungabaza okusheshayo nokunesisekelo esihle kuwo wonke umphakathi wocwaningo lwe-AI. Izimpikiswano eziphikisanayo zembula imikhawulo eyisisekelo ekushintsheni ukubonwa kwephethini eluhlaza ukuze kuhlaziywe ngendlela ehlelekile - izifundo ezidlulela ngale kwe-chess ibe yi-automation yebhizinisi, izinhlaka zokwenza izinqumo, kanye nendlela amapulatifomu afana ne-Mewayz ehlakaniphile esebenza ngayo kubasebenzisi abangaphezu kuka-138,000.

Iphepha Loqobo Lalifunani Ngempela?

Ucwaningo lwangempela, oluholwa ngu-Aram Ebrahimi kanye nozakwabo ku-Google DeepMind, luphakamise ukuthi imodeli ye-transformer enkulu ngokwanele eqeqeshwe ezindaweni ze-chess nokuhlola kwayo ingadlala ngamandla amakhulu ngaphandle kokusebenzisa ama-algorithms okusesha acacile njenge-minimax noma usesho lwesihlahla sase-Monte Carlo. Ngokungafani nezinjini ezifana ne-Stockfish noma i-AlphaZero, ezihlola izinkulungwane kuya ezigidini zezikhundla zesikhathi esizayo ngaphambi kokukhetha ukuthutha, le ndlela incike kunethiwekhi ye-neural eyenza izibikezelo ze-single-pass — empeleni "intuiting" umnyakazo ongcono kakhulu ukusuka ekubonweni kwephethini kuphela.

Isimangalo besinesibindi: uma imodeli ingathola ukuqonda okwanele kwesikhundla kusuka kudatha yokuqeqeshwa, ukubala kwe-brute-force kungase kungadingeki. Imiphumela yokuqala yebhentshimakhi ibonakale ithembisa, imodeli ifinyelela izilinganiso ze-Elo ebangeni le-grandmaster ngaphansi kwezimo ezithile zokuhlola.

Kungani Abagxeki Bephikisana Ukusesha Akukaze Kuqedwe Ngempela?

Impikiswano ephoqelela kakhulu ephikisanayo iqondise kusisekelo esimaphakathi sephepha. I-transformer yaqeqeshwa ezigidini zezikhundla ezihlolwe yi-Stockfish - injini ethembele kakhulu ekusesheni okujulile. Abagxeki baphikisa ngokuthi imodeli ayizange isuse ukusesha; yayihlakaza. Ukusesha kumane kulayishwe phambili kudatha yokuqeqeshwa kunokuba kwenziwe ngesikhathi sokunquma.

"Ukufuna imodeli kudlala i-chess 'ngaphandle kokusesha' ngenkathi uyiqeqesha emiphumeleni yenjini esekwe kufana nokuthi uxazulule i-maze ngaphandle kwemephu - ngemva kokubamba ngekhanda isisombululo omunye umuntu otholwe esebenzisa imephu."

Lo mehluko ubaluleke kakhulu. Imodeli ifunde izethulo ezicindezelwe zemiphumela yosesho, hhayi ukuqonda okuzimele kwesikhundla. Susa isignali yokuqeqesha esuselwe ekusesheni, futhi ukusebenza kuyagoqa. Lokhu kunokufana okuqondile kwezobuhlakani bebhizinisi: noma iliphi ithuluzi lesinqumo eliqhutshwa yi-AI lihle kuphela njengokuhlaziywa okuhlelekile okushumekwe emgqeni walo wokuqeqesha.

Kuphi Ukuqashelwa Kwephethini Okumsulwa Kwehle Phansi Ekusetshenzisweni?

Ukuhlolwa kokuhlola okwenziwa abacwaningi abazimele kudalule izindlela zokuhluleka ezibucayi ezisithwe ama-benchmarks asekuqaleni:

• Izindawo zamaqhinga ajulile: Imodeli ibilokhu igeja izinhlanganisela ezidinga ukubalwa okudlula ukunyakaza okungu-4-5, lapho izinjini zomdabu zihamba phambili ngezihlahla zokusesha ezisobala.
• Izimo ze-novel endgame: Izikhundla ezingaphandle kokusatshalaliswa kokuqeqeshwa zidalule ukungakwazi kwemodeli ukucabanga ngezimiso zokuqala, okuholela kumaphutha ayisisekelo okungekho umkhulu ongumuntu ongawenza.
• Ukuqina kokumelana: Lapho abaphikisi beqondisa ngamabomu imidlalo ezindaweni ezingajwayelekile, i-Elo yemodeli yehla kakhulu — iphakamisa ukubamba ngekhanda kunokuqonda kwangempela.
• Ukungaguquguquki ngaphansi kwengcindezi: Nakuba ukusebenza okumaphakathi kubonakala izinga le-grandmaster-level, ukuhluka bekuphakeme kakhulu kunogogo abangabantu noma izinjini ezisekelwe ekusesheni, okunamaphutha okuyinhlekelele okwenzeka ngamanani angahambisani nokudlala kwe-grandmaster kwangempela.
• Isimo sokukala inkimbinkimbi: Njengoba ubunkimbinkimbi bebhodi bukhula, igebe phakathi kwemodeli yamahhala yosesho nezinjini ezisekelwe ekusesheni landa kakhulu kunokuba libe ngomugqa.

Isho Ukuthini Le mpikiswano Ngezinhlelo Zebhizinisi Eziqhutshwa yi-AI?

Impikiswano ye-chess-without-search ikhanyisa ukungezwani enhliziyweni yokusetshenziswa kwesimanje kwe-AI. Ukubonwa kwephethini nokuhlaziya okuhlelekile akushintsheki - kuyahambisana. Amasistimu asebenza kahle kakhulu ahlanganisa izimpendulo ezisheshayo nokucabanga okuhlelekile lapho ukubaluleka kuphezulu.

Lokhu kuyisakhiwo esingemuva kwesistimu yokusebenza yebhizinisi yamamojula angama-207 ka-Mewayz. Kunokuthembela kuphela ku-heuristics yokufanisa iphethini noma ingqondo esekelwe emthethweni kuphela, inkundla ihlanganisa zombili izindlela kuwo wonke umsebenzi wayo wokuzenzakalela, i-CRM, ukuphathwa kwephrojekthi, namamojula ezezimali. Iziphakamiso ezisheshayo ezisuselwe kwiphethini zisingatha izinqumo ezijwayelekile, kuyilapho izinhlaka zokuhlaziya ezihlelekile zisebenzisana nezimo eziyinkimbinkimbi - ukubonisa ukuthi izinjini ze-chess eziqine kakhulu zibhanqa kanjani ukuhlolwa kwenethiwekhi ye-neural nosesho oluqondisiwe.

Isifundo esivela ku-contra analysis sicacile: amasistimu athi asusa ukucabanga okuhlelekile ngokuvuna umuzwa omsulwa ashaye usilingi wokusebenza. Kungakhathaliseki ukuthi uphatha indawo ye-chess noma umbhobho webhizinisi, inhlanganisela yokubonwa kwephethini esheshayo nokuhlaziya ngamabomu ihlala idlula noma iyiphi indlela yokuzihlukanisa.

Kufanele Sizihlole Kanjani Izimangalo "ze-AI" eziqhubekayo?

Ama-agumenti okuphikisa asungula uhlaka oluwusizo lokuhlola ngokujulile izimangalo zocwaningo lwe-AI. Okokuqala, hlola ukuthi ingabe ikhono elifuniwe lifinyelelwe ngempela noma laphinde lasatshalaliswa kabusha - ingabe isistimu isuse ukusesha, noma yakufihla kunqubo yokuqeqesha? Okwesibili, ukuhlola ukusebenza kokufakwayo okuphikisayo nokungaphandle kokusabalalisa, hhayi nje amabhentshimakhi avumayo. Okwesithathu, linganisa ukungaguquguquki kanye nokusebenza okubi kakhulu ngokuhambisana nezilinganiso, njengoba isistimu edlala ngokumangalisayo u-90% wesikhathi kodwa ephutha ngendlela eyinhlekelele U-10% wesikhathi awulona izinga lika- grandmaster nganoma yimuphi umqondo onengqondo.

Lezi zimiso zokuhlola zisebenza ngokulinganayo lapho amabhizinisi ehlola amathuluzi axhaswe yi-AI okusebenza kwawo. Izilinganiso ezisezingeni eliphezulu zingasitha ubuthakathaka obubucayi obuvela ngaphansi kwezimo zomhlaba wangempela - iqiniso elazisa indlela kaMewayz yokwakha ukwethembeka kuwo wonke amamojula ayo e-ecosystem.

Imibuzo Evame Ukubuzwa

Ingabe imodeli ye-chess-without-search ifinyelele izinga le-grandmaster?

Ngaphansi kwezimo zebhentshimakhi ezilawulwayo, imodeli ithole izilinganiso ze-Elo ebangeni le-grandmaster. Kodwa-ke, ukuhlola okuzimele kwembule ukungahambisani, ubungozi bokuphikisana, nezindawo eziyimpumputhe zamaqhinga ezibukela phansi ukuhlukaniswa kwe-grandmaster. Umdlalo we-Grandmaster wangempela udinga ukwethembeka nokujula imodeli engazange ibonise njalo, okwenza isimangalo sibe mncane ngokobuchwepheshe kunokuba sivumeleke kabanzi.

Ingabe ucwaningo lwe-AI chess olungaseshiwe lusebalulekile naphezu kwalokhu kugxeka?

Nakanjani. Ucwaningo lubonise ukuthi izakhiwo ze-transformer zingacindezela amanani amakhulu olwazi lwe-chess ekuhlolweni okusheshayo kwe-single-pass. Lokhu kunezinhlelo zokusebenza ezisebenzayo zokuhlola cishe okusheshayo, usizo lokuqeqesha, nezinhlelo ezixubile. Izimpikiswano eziphikisanayo azilwenzi ucwaningo - zenza kahle imikhawulo yalo bese zibekela inselele isiphetho esishiwo ngokweqile.

Ingabe le mpikiswano ihlobana kanjani nokukhetha amathuluzi ezishintshayo zebhizinisi?

Isifundo esibalulekile ukuthi ukuzenzakalela okusebenzayo kudinga ukufanisa indlela yokucabanga efanele kuhlobo lomsebenzi ngamunye. Izinqumo ezilula, eziphindaphindayo zizuza ekubonweni kwephethini okusheshayo. Izinqumo eziyinkimbinkimbi, eziphakeme zidinga ukuhlaziywa okuhlelekile. Izinkundla ezihamba phambili - njenge-OS yebhizinisi edidiyelwe ye-Mewayz - zihlanganisa zombili, ziqinisekise ukuthi akukho ndlela eyodwa eba umgoqo noma iphuzu lokwehluleka kuyo yonke imisebenzi yakho.

Usulungele ukuqhuba ibhizinisi lakho kusistimu eyakhelwe kokubili isivinini nokujula? I-Mewayz ihlanganisa amamojula ahlanganisiwe angu-207 nokuzenzakalela okuhlakaniphile okudizayinelwe inkimbinkimbi yomhlaba wangempela — hhayi amatiyetha ebhentshimakhi. Izinhlelo ziqala ku-$19/mo emaqenjini afuna ukwethembeka kuwo wonke amazinga. Qala isilingo sakho samahhala ku-app.mewayz.com futhi uzizwele ukuthi isistimu yokusebenza yebhizinisi yangempela injani.