Umdlalo omuhle wakudala wokucasha ungaba uvivinyo oluhle lwe-artificial intelligence (AI) bots ukuze ubonise ukuthi zenza kanjani izinqumo nokusebenzelana kwazo nezinto ezihlukahlukene ezizungezile.
Kweyakhe , eshicilelwe abacwaningi abavela ku-OpenAI, inhlangano engenzi nzuzo yocwaningo lwezobunhloli esidume emdlalweni wekhompiyutha i-Dota 2, ososayensi bachaza ukuthi ama-agent alawulwa ubuhlakani bokwenziwa aqeqeshwa kanjani ukuze abe yinkimbinkimbi ekusesheni nasekucasheni komunye nomunye endaweni ebonakalayo. Imiphumela yocwaningo ibonise ukuthi iqembu lama-bots amabili lifunda ngempumelelo futhi ngokushesha kunanoma iyiphi i-ejenti eyodwa ngaphandle kombimbi.
Ososayensi baye basebenzisa indlela osekuyisikhathi eside yaduma ngayo , lapho ubuhlakani bokwenziwa bubekwe endaweni engaziwa, kuyilapho kunezindlela ezithile zokuxhumana nakho, kanye nesistimu yemivuzo nenhlawulo yomphumela owodwa noma omunye wezenzo zayo. Le ndlela iphumelela kakhulu ngenxa yekhono le-AI lokwenza izenzo ezihlukahlukene endaweni ebonakalayo ngesivinini esikhulu, izikhathi eziyizigidi ngokushesha kunalokho umuntu angakucabanga. Lokhu kuvumela ukuzama nephutha ukuthola amasu asebenza kahle kakhulu okuxazulula inkinga ethile. Kodwa le ndlela futhi inemikhawulo ethile, isibonelo, ukudala indawo kanye nokuqhuba imijikelezo eminingi yokuqeqesha kudinga izinsiza ezinkulu zekhompiyutha, futhi inqubo ngokwayo idinga uhlelo olunembile lokuqhathanisa imiphumela yezenzo ze-AI nomgomo wayo. Ngaphezu kwalokho, amakhono atholwe yi-ejenti ngale ndlela anqunyelwe emsebenzini ochaziwe futhi, uma i-AI isifundile ukubhekana nayo, ngeke kube khona ukuthuthuka okwengeziwe.
Ukuqeqesha i-AI ukuthi idlale umacashelana, ososayensi basebenzisa indlela ebizwa ngokuthi βUkuhlola okungaqondile,β okuyilapho ama-agent enenkululeko ephelele yokuthuthukisa ukuqonda kwawo umhlaba wegeyimu nokuthuthukisa amasu okuwina. Lokhu kufana nendlela yokufunda enama-ejenti amaningi abacwaningi bakwa-DeepMind abayisebenzisa lapho izinhlelo eziningi zobuhlakani bokwenziwa. . Njengalokhu, ama-agent e-AI awazange aqeqeshwe ngaphambili emithethweni yomdlalo, kodwa ngokuhamba kwesikhathi afunda amasu ayisisekelo futhi akwazi ngisho nokumangalisa abacwaningi ngezixazululo ezingezona ezincane.
Emdlalweni wokucasha, abenzeli abambalwa umsebenzi wabo bekuwukucasha bebedingeka ukuthi bagweme umugqa wokubuka wezimbangi zabo ngemuva kokuqala kancane ngesikhathi ithimba labasebenzeli bokusesha lisamile. Ngaphezu kwalokho, "umugqa wokubona" ββkulo mongo uyikhoni ye-135 degree phambi kwe-bot ngayinye. Ama-ejenti awakwazanga ukuya kude kakhulu ngaphandle kwendawo yokudlala futhi aphoqeleka ukuthi azulazule emakamelweni akhiwe ngokungahleliwe anekhono lokusebenzisa izinto zangaphandle (amabhokisi, izindonga ezigudlukayo, amarampu akhethekile) ahlakazekile angasetshenziswa kokubili ukudala ikhava nokungena kuwo. .
Ngenqubo ende yokufunda, ama-ejenti e-AI afunde amasu ayisithupha ahlukile, ngalinye libasize bathuthukela esigabeni esilandelayo somdlalo. Ekuqaleni, amaqembu afuna ukucasha avele abaleka futhi ajaha wodwa, kodwa ngemva kwemidlalo engaba yizigidi ezingu-25, iqembu labacashi lafunda ukuvimba amaphaseji ngamabhokisi futhi lakhe izindawo zokukhosela ngaphandle kwawo. Ngemuva kweminye imidlalo eyizigidi ezingama-75, iqembu lithole izindlela zokusebenzisa ama-ramp ukuze lingene kulezi zindawo zokucasha. Ngemva kweminye imizuliswano eyizigidi ezingu-10, abacashi bafunda ukuhudulela ama-ramps emaphethelweni enkundla yokudlala futhi bawavalele endaweni ukuze bavimbele izitha ukuba zingawasebenzisi.
Emuva kwemidlalo eyizigidi ezingu-388, abafunayo bafunde ukusebenzisa amarempu avinjiwe ukuze bagibele phezu kwamakhreyithi alethwe kubo, bese, beya kuwo ngqo, bangene ezindaweni ezicashe kuzo izitha ezidalwe ngezindonga eziphathwayo. Futhi ekugcineni, ngemva kwemidlalo eyizigidi ezingu-458, ithimba elicashile laphetha ngokuthi kwakudingeka livimbe zonke izinto bese lakha indawo yokukhosela, ngokusobala eyaholela ekunqobeni kwalo kokugcina.
Okuhlaba umxhwele kakhulu ukuthi ngemuva kwemidlalo eyizigidi ezingama-22, ama-ejenti afunda ukuxhumanisa izenzo zabo kanye nokusebenza kahle kokubambisana kwabo kwanda esikhathini esizayo, isibonelo, ngamunye waletha ibhokisi lakhe noma udonga ukuze akhe indawo yokukhosela futhi wakhetha isabelo sakhe sezinto. block, ukwenza kube nzima umdlalo wobunzima kubaphikisi.
Ososayensi baphinde baphawula iphuzu elibalulekile elihlobene nomthelela wenani lezinto zokuqeqeshwa (inani ledatha elidluliswa kunethiwekhi ye-neural - "Usayizi weqoqo") ngesivinini sokufunda. Imodeli ezenzakalelayo idinga amameshi ayizigidi ezingu-132,3 emahoreni angu-34 okuqeqeshwa ukuze kufike lapho ithimba elicashile lifunde khona ukuvimba amarempu, kuyilapho idatha eyengeziwe ibangele ukuncipha okuphawulekayo kwesikhathi sokuqeqesha. Isibonelo, ukwandisa inani lamapharamitha (ingxenye yedatha etholwe phakathi nayo yonke inqubo yokuqeqesha) ukusuka ezigidini ezingu-0,5 ukuya ezigidini ezingu-5,8 kukhuphule ukusebenza kahle kwesampula ngezikhathi ezingu-2,2, nokwandisa usayizi wedatha yokufaka kusuka ku-64 KB kuya ku-128 KB ukunciphisa ukuqeqeshwa. isikhathi cishe isikhathi esisodwa nengxenye.
Ekupheleni komsebenzi wabo, abacwaningi banqume ukuhlola ukuthi kungakanani ukuqeqeshwa okungaphakathi kwegeyimu kungasiza abenzeli ukuthi babhekane nemisebenzi efanayo ngaphandle kwegeyimu. Kube nezivivinyo ezinhlanu sezizonke: ukuqwashisa ngenani lezinto (ukuqonda ukuthi into iyaqhubeka nokuba khona ngisho noma ingabonakali futhi ingasetshenziswa); "ukukhiya nokubuya" - ikhono lokukhumbula isikhundla somuntu sokuqala bese ubuyela kuso ngemva kokuqeda omunye umsebenzi owengeziwe; βukuvinjwa okulandelanayoβ - amabhokisi angu-4 ayebekwe ngokungahleliwe emakamelweni amathathu angenayo iminyango, kodwa anamarempu okungena ngaphakathi, ama-ejenti ayedinga ukuwathola futhi awavimbe wonke; ukubekwa kwamabhokisi ezindaweni ezinqunywe kusengaphambili; ukwakha isiphephelo ezungeze into ngendlela yesilinda.
Ngenxa yalokho, emisebenzini emithathu kwemihlanu, ama-bots ayethole ukuqeqeshwa kokuqala emdlalweni afunde ngokushesha futhi abonisa imiphumela engcono kune-AI eyayiqeqeshelwe ukuxazulula izinkinga kusukela ekuqaleni. Benza kangcono kancane ekuqedeni umsebenzi nasekubuyeleni endaweni yokuqala, bavimbe amabhokisi ngokulandelana emakamelweni avaliwe, futhi babeka amabhokisi ezindaweni ezinikeziwe, kodwa benza buthaka kancane ekuboneni inani lezinto nokudala ikhava ezungeze enye into.
Abacwaningi baveza imiphumela exubile yokuthi i-AI ifunda futhi ikhumbule kanjani amakhono athile. βSicabanga ukuthi imisebenzi lapho uqeqesho lwangaphambi komdlalo lwenze kahle kakhulu ibandakanya ukusebenzisa kabusha amakhono afundwe ngaphambilini ngendlela ejwayelekile, kuyilapho ukwenza imisebenzi esele kangcono kune-AI eqeqeshwe kusukela ekuqaleni kungadinga ukuwasebenzisa ngendlela ehlukile, okuyinto kunzima kakhulu,β kubhala ababhali ababambisene bomsebenzi. "Lo mphumela ugqamisa isidingo sokuthuthukisa izindlela zokuphinda usebenzise ngempumelelo amakhono atholwe ngokuqeqeshwa lapho edluliswa endaweni ethile eyiswa kwenye."
Umsebenzi owenziwayo umangalisa ngempela, ngoba ithemba lokusebenzisa le ndlela yokufundisa lingaphezu kwemingcele yanoma yimuphi umdlalo. Abacwaningi bathi umsebenzi wabo uyisinyathelo esibalulekile ekudaleni i-AI ngokuziphatha "okusekelwe ku-physics" kanye "nokuziphatha komuntu" okungahlola izifo, ukubikezela izakhiwo zama-molecule amaprotheni ayinkimbinkimbi futhi ahlaziye ama-CT scan.
Kuvidiyo engezansi ungabona ngokucacile ukuthi yonke inqubo yokufunda yenzeke kanjani, ukuthi i-AI yafunda kanjani ukusebenzisana, kanye namasu ayo aba ngobuqili futhi ayinkimbinkimbi.
Source: 3dnews.ru