HejuruUbuyobozi bukurikira
Uburyo bwo Kumenya Hallucination
Tekiniki
Ubuyobozi bwa tekiniki
Uburyo bw'abakinnyi-banegura bahuza abiga babiri: 'umukinnyi' uhitamo ibikorwa na 'kunegura' ucira urubanza uko ibyo bikorwa byari byiza.
Umva iki gitaboAI-byavuzwe · ~2 min
This pairing makes reinforcement learning more stable and sample-efficient than using either approach alone.
Kwiga gushimangira bifite uburyo bubiri bwagutse: uburyo bushingiye kuri politiki bwiga neza icyo gukora, nuburyo bushingiye ku gaciro bwiga uko leta nziza ari nziza. Abakinnyi-banegura. Umukinnyi ni politiki isohora ibikorwa bishoboka; kunegura nigikorwa cyagaciro kigereranya kugaruka. Nyuma ya buri ntambwe, uwinegura abara ikosa-itandukaniro ryigihe gito ryerekana niba ibisubizo byari byiza cyangwa bibi kuruta uko byari byitezwe. Umukinnyi akoresha iri kosa kugirango asunike politiki yayo kubikorwa byatsinze ibiteganijwe kandi kure yibyo bidakora neza. Kuberako abanegura batanga ibipimo fatizo bito, igereranya ryabakinnyi ryikigereranyo ntirisakuza cyane kuruta muburyo bwa politiki-yerekana uburyo nka REINFORCE, mugihe ugikora umwanya uhoraho wibikorwa uburyo bwagaciro gusa nka Q-Kwiga bubona nabi.
Guhugura amaboko ya robo hamwe nabashinzwe kugenzura ibikorwa hamwe na torque zihoraho (urugero, ukoresheje PPO cyangwa SAC)
Guhuza imvugo nini yindimi ukoresheje RLHF, aho PPO (uburyo bwabakinnyi-banegura) itunganya ibisubizo kubihembo byigihembo
Kumenya ingamba zingamba zimikino nka StarCraft II na Dota 2
Data-center gukonjesha no kugenzura ingufu-biga biga guhinduka bikomeje
Umukinnyi-Kunegura ni inkingi ya RL igezweho. Algorithms nka A3C, A2C, PPO, SAC, na DDPG byose byubakiyeho, hiyongeraho amayeri nkintego zaciwe kugirango zigezweho zihamye, ibihembo bya entropiya kubushakashatsi, hamwe nabakinnyi babangikanye nibisohoka. Witege ko hazakomeza kwiyongera muri robo, imashini nini nini, hamwe na RL uhereye kubitekerezo byabantu kugirango bahuze imiterere yindimi, aho gutuza no gukora neza byingenzi.
Ibaruwa y'amakuru y'ubuntu
Inkuru eshatu zizewe za AI buri gitondo cy'iminsi y'akazi, zanditse mu rurimi rworoshye. Ubuntu iteka ryose, nta kwamamaza.
Imeri imwe buri munsi y'akazi. Kwiyimura mu buryo bumwe. Ntitwigera tugurisha cyangwa ngo dusangire aderesi yawe.
Isuzume
Ibisubizo ako kanya kuri buri kibazo, n'icyemezo gishobora gusangirwa gifite ID yagenzurwa umaze gutsinda isomo.
Fasha uburezi bwa AI bw'ubuntu. AI Understanding ni umuryango utegamiye kuri leta wa 501(c)(3) — nta kwamamaza, nta kwishyura, iteka ryose. Tanga impano
Uburyo bw'abakinnyi-banegura bahuza abiga babiri: 'umukinnyi' uhitamo ibikorwa na 'kunegura' ucira urubanza uko ibyo bikorwa byari byiza. Uku guhuza bituma kwiga gushimangira biga neza kandi byintangarugero kuruta gukoresha inzira yonyine.
Umunegura yiga imikorere yagaciro kandi atanga ibimenyetso byo gusuzuma (nkikosa rya TD) ubwira umukinnyi niba ibikorwa byayo byari byiza cyangwa bibi kuruta uko byari byitezwe.
Ibyiza bitandukanya uburyo ibikorwa byihariye biruta ibisubizo bisanzwe biva muri leta, bitanga ikimenyetso gisukuye kuruta kugaruka.
Gukuramo igereranyo cyagaciro kunegura nkibanze fatizo bigabanya itandukaniro ryikigereranyo cya gradient utongeyeho kubogama, kwiga byihuse.
Kuberako umukinnyi asobanura neza politiki, irashobora gusohora ibikorwa bikomeza (urugero, torque ihuriweho) idakeneye max hejuru yibikorwa byinshi bitagira akagero.
Umukinnyi ahindura politiki yayo mu cyerekezo cyatanzwe ninyungu zo kunegura / Ikimenyetso cya TD-ikosa, gishimangira ibikorwa byiza kuruta ibyateganijwe.
Komeza wige
Abandi bayobozi batoranijwe kuriyi ngingo
HejuruUbuyobozi bukurikira
Uburyo bwo Kumenya Hallucination
Tekiniki