I-DeepMind, eyaziwa ngophuhliso lwayo kwinkalo yobukrelekrele bokwenziwa kunye nokwakha uthungelwano lwe-neural olukwaziyo ukudlala imidlalo yekhompyuter kunye nebhodi kwinqanaba lomntu, elisanda kutyhilwa iprojekthi yeAlphaCode echaza ukuba njani inkqubo yokufunda ngomatshini wokuvelisa ikhowudi ukuba unokuthatha inxaxheba kukhuphiswano lwenkqubo kwiqonga Codeforces kwaye ubonise isiphumo avareji.
Kuxelwe ukuba iprojekthi isebenzisa i-"Transformer" ye-neural network architecture ngokudityaniswa nezinye iindlela zesampulu kunye nokucoca ukuvelisa iikhowudi ezahlukeneyo ezahlukeneyo ezihambelana nokubhaliweyo kolwimi lwendalo.
Indlela esebenza ngayo I-Alpha'ode isekwe kuhluzo, ngokwamaqela kunye nokuhlelwa, apho ke iqhubela phambili ukukhetha eyona khowudi yokusebenza ilungileyo kuluhlu oluveliswayo lweenketho, ethi ke ikhangelwe ukuqinisekisa ukuba iziphumo ezichanekileyo zifunyenwe (kumsebenzi ngamnye wokhuphiswano, umzekelo idatha yegalelo kunye nesiphumo esihambelanayo) kulo mzekelo, omele ufunyenwe emva kokuphunyezwa kweprogram).
Sichaza iAlphaCode, esebenzisa imifuziselo yolwimi esekwe kwisiguquli ukwenza ikhowudi kwizinga elingazange libonwe ngaphambili, emva koko ngobuchule ihluze iseti encinci yeenkqubo ezithembisayo.
Siqinisekisa ukusebenza kwethu sisebenzisa ukhuphiswano olubanjwe kwiCodeforces, iqonga elidumileyo elibamba ukhuphiswano oluqhelekileyo olutsala amashumi amawaka abangeneleli abavela kwihlabathi liphela abeze kuvavanya izakhono zabo zokubhala iikhowudi. Sikhethe ukhuphiswano lwamva nje lwe-10 ukuze luvavanywe, ngalinye lutsha kunedatha yethu yoqeqesho. I-AlphaCode yayimalunga nenqanaba kunye nomntu okhuphisana naye, ephawula okokuqala inkqubo yokuvelisa ikhowudi ye-AI ifikelele kwinqanaba lokhuphiswano lokusebenza kukhuphiswano lwenkqubo.
Kuqeqesho lwenkqubo eqikelelweyo ukufunda koomatshini, kuphawulwe ukuba ikhowudi yesiseko ekhoyo kwiindawo zokugcina zikawonke-wonke zeGitHub yasetyenziswa. Emva kokulungiselela imodeli yokuqala, isigaba sokuphucula senziwe ngokusekelwe kwingqokelela yekhowudi kunye nemizekelo yeengxaki kunye nezisombululo ezinikezelwa kubathathi-nxaxheba beCodeforces, CodeChef, HackerEarth, AtCoder kunye ne-Aizu.
Lilonke, kukwenziwa kweAlphaCode I-715 GB yekhowudi ye-GitHub esetyenzisiweyo kunye nemizekelo engaphezu kwesigidi yezisombululo kwiingxaki eziqhelekileyo zokhuphiswano. Ngaphambi kokuqhubeka nokuveliswa kwekhowudi, isicatshulwa somsebenzi sadlula kwisigaba esiqhelekileyo, apho yonke into engafanelekanga yayingabandakanywa kwaye kuphela iindawo ezibalulekileyo eziseleyo.
Ukuvavanya inkqubo, i-10 entsha ye-Codeforces ikhuphisana nabathathi-nxaxheba abangaphezu kwe-5.000 yakhethwa, ebanjwe emva kokugqiba uqeqesho lomzekelo wokufunda umatshini.
Ndingatsho ngokukhuselekileyo ukuba iziphumo zeAlphaCode zigqithise ulindelo lwam. Ndandithandabuza kuba nakwiingxaki ezilula zokukhuphisana, kudla ngokufuneka ukuba kungekuphela nje ukuphumeza i-algorithm, kodwa kwakhona (kwaye le yeyona nto inzima kakhulu) ukuyisungula. I-AlphaCode ikwazile ukwenza kwinqanaba lomntu okhuphisanayo omtsha othembisayo. Andikwazi ukulinda ukubona into ezayo!
UMIKE MIRZAYANOV
UMSEKI WEKHOWUDI
Iziphumo zezabelo zivumelekile yendlela yeAlphaCode yokungena phantse embindini wokufaneleka kolu khuphiswano (54,3%). I-AlphaCode iqikelelwe ukuba iphelele amanqaku ali-1238, iqinisekisa ukungena kwi-Top 28% kubo bonke abathathi-nxaxheba be-Codeforces abathathe inxaxheba kukhuphiswano kanye kwiinyanga ezi-6 ezidlulileyo.
Kufuneka kuqatshelwe ukuba kuqatshelwe ukuba iprojekthi ikwinqanaba lokuqala lophuhliso kwaye kwixesha elizayo licetywayo ukuphucula umgangatho wekhowudi eyenziwe, kunye nokuphuhlisa i-AlphaСode kwiinkqubo ezinceda ukubhala ikhowudi, okanye izixhobo zophuhliso lwesicelo abantu abangenazo izakhono zokucwangcisa banokusebenzisa.
Gqibela ukuba unomdla wokwazi okungakumbi ngayo, kufuneka ukwazi ukuba uphawu lophuhliso oluphambili lukwazi ukuvelisa ikhowudi kwiPython okanye i-C ++, ukuthatha njengegalelo lombhalo ingxelo yengxaki ngesiNgesi.
Ungajonga iinkcukacha Kule khonkco ilandelayo.