I-Los Abaphenyi beGoogle bakhishwe izindaba mayelana nokwakhiwa kwalo kohlaka olusha olunweba ukuqeqeshwa kwamamodeli wezinhloli zokufakelwa ezinkulungwaneni zemishini. Umphumela ubizwa ngokuthi IMBEWU RL (ukufundwa okuqinisayo okujulile okuqinisayo).
Yilokho intuthuko ethembisayo ngoba kufanele nika amandla ama-algorithms obuhlakani bokufakelwa ukuthi aqeqeshwe ezigidini zezithombe ngomzuzwana futhi sinciphise izindleko zalokhu kuqeqeshwa ngo-80%, kusho uGoogle ephepheni locwaningo.
Lolu hlobo lokwehlisa umsebenzi lungasiza ukulinganisa inkundla yokudlala yeziqalo. ukuthi kuze kube manje abakwazanga ukuncintisana nabakhulu abanjengoGoogle emkhakheni we-AI. Izindleko zokuqeqesha amamodeli wokufunda wemishini esezingeni eliphezulu efwini ziphakeme ngokumangazayo. I-Google ikwenza kube semthethweni ukuvulwa kwekhodi ye-SEED RL, iphrojekthi ehlose ukukhulisa isilinganiso sezindleko / sokusebenza sokuqiniswa kokufunda.
Ukuqiniswa kokufunda kuyindlela ethile yokusebenziseka lapho amanxusa efunda ngemvelo yawo ngokuhlola nokwenza ngcono izenzo zawo ukuthola imivuzo eminingi.
Ku »SEED RL: I-Scalable and Efficient Deep-RL ene-Accelerated Central Inference", sethule umenzeli we-RL olinganisa izinkulungwane zemishini, evumela ukuqeqeshwa ezigidini zozimele ngomzuzwana futhi kuthuthukisa kakhulu ukusebenza kahle kwekhompyutha. Lokhu kutholakala ngokwakhiwa kwenoveli okusebenzisa ama-accelerators (i-GPU noma i-TPU) esikalini ngokufaka phakathi imodeli okufakwayo nokwethula ungqimba lwezokuxhumana olusheshayo.
Sikhombisa ukusebenza kwe-SEED RL kumabhentshi e-RL adumile afana ne-Google Research Football, i-Arcade Learning Environment, ne-DeepMind Lab, futhi sikhombisa ukuthi ngokusebenzisa amamodeli amakhulu, ukusebenza kwedatha kungakhuphuka. Ikhodi ivuliwe ku-Github kanye nezibonelo ezizosebenza ku-Google Cloud nge-GPU.
I-SEED RL isuselwa kuhlaka lweTensorFlow 2.0 y isebenza kusetshenziswa inhlanganisela yamayunithi wokucubungula ihluzo kanye namayunithi wokucubungula ama-tensor ukuqinisa ukuthathwa kwemodeli. Isilinganiso senziwa maphakathi kusetshenziswa ingxenye yokufunda eqeqesha imodeli.
Okuguqukayo nolwazi lwesimo semodeli eqondisiwe kugcinwa endaweni futhi ukubonwa kuzo kuthunyelwa kumfundi esigabeni ngasinye senqubo. I-SEED RL ibuye isebenzise umtapo wolwazi wenethiwekhi ngokususelwa kunhlaka yomthombo ovulekile we-RPC jikelele ukunciphisa ukubambezeleka.
I-Los Abaphenyi bakwaGoogle bathe ingxenye yokufunda by SGIYA ngoMASKANDI MEDIA inganwetshwa izinkulungwane zamakholi, Ngenkathi inani labalingisi elizophindaphindwa phakathi kokuthatha izilinganiso emvelweni nokwenza okunconyiwe kumodeli ukubikezela isenzo esilandelayo, kungezwa ezinkulungwaneni zemishini.
I-Google ihlole ukusebenza kwe-SEED RL ngokuyiqhathanisa nemvelo edumile yokufunda i-Arcade, imvelo ye-Google Research Football, nezindawo ezahlukahlukene ze-DeepMind Lab. Imiphumela ikhombisa ukuthi bakwazile ukuxazulula umsebenzi we-Google Research Football ngenkathi beqeqesha imodeli ezigidini ezingama-2,4 ozimele ngomzuzwana besebenzisa ama-chips angama-64 weyunithi yokucubungula ifu.
Ishesha cishe izikhathi ezingama-80 kunozimele bangaphambilini, kusho iGoogle.
"Lokhu kuhumushela ekusheshisweni kwesikhathi esibalulekile, njengoba ama-accelerator eshibhile kakhulu ekusebenzeni ngakunye kunama-CPU, izindleko zokuhlolwa zincishiswe kakhulu." Sikholelwa ukuthi i-SEED RL nemiphumela eveziwe ikhombisa ukuthi ukuqiniswa kokufunda sekubuye kwahlangana nakho konke okunye ukufunda okujulile maqondana nokusetshenziswa kwamafutha, "kubhala uLasse Espeholt, unjiniyela ocwaningweni kwaGoogle Research.
Ngokwakhiwa okulungiselelwe ukusetshenziswa kuma-accelerators anamuhla, kungokwemvelo ukukhulisa usayizi wemodeli ngomzamo wokukhulisa ukusebenza kwedatha.
AbakwaGoogle bathi ikhodi ye-SEED RL ibingumthombo ovulekile futhi iyatholakala eGithub, kanye nezibonelo ezibonisa ukuthi ungayisebenzisa kanjani ku-Google Cloud ngamayunithi wokucubungula ihluzo.
Ekugcineni, kulabo abanentshisekelo kulolu hlaka olusha, bangaya kusixhumanisi esilandelayo lapho bengathola khona imininingwane eminingi ngalo. Isixhumanisi yilokhu.
Umthombo: https://ai.googleblog.com/