Mu Marichi 2017, OpenAI idapanga mafunde pagulu lophunzirira mwakuya ndi pepala "Evolution Strategies ngati Njira Yowonjezereka Yophunzitsira Kulimbikitsa" Ntchitoyi idafotokoza zotsatira zochititsa chidwi zokomera mfundo yakuti reinforcement learning (RL) siinasinthe, ndipo pophunzitsa ma neural network ovuta, ndikofunikira kuyesa njira zina. Kenako mkangano unayambika ponena za kufunikira kwa maphunziro olimbikitsira komanso momwe akuyenera kukhala ngati ukadaulo "woyenera kukhala nawo" pakuphunzitsa kuthetsa mavuto. Apa ndikufuna kunena kuti matekinoloje awiriwa sayenera kuonedwa ngati mpikisano, imodzi yomwe ili bwino kuposa ina; m'malo mwake, pamapeto pake amakwaniritsana. Zowonadi, ngati mukuganiza pang'ono pazomwe zimafunika kuti mupange General AI ndi machitidwe oterowo, omwe pakukhalapo kwawo angakhale okhoza kuphunzira, kuweruza ndi kukonzekera, ndiye kuti pafupifupi tidzafika pamapeto kuti izi kapena zophatikizana zidzafunika. Mwa njira, inali njira yophatikizira imeneyi yomwe chilengedwe chinabwera, chomwe chinapatsa zinyama ndi zinyama zina zapamwamba zanzeru zovuta panthawi ya chisinthiko.
Njira Zachisinthiko
Lingaliro lalikulu la pepala la OpenAI linali lakuti, m'malo mogwiritsa ntchito maphunziro olimbikitsana pamodzi ndi kubwereza kwachikhalidwe, adaphunzitsidwa bwino ndi neural network kuti athetse mavuto ovuta pogwiritsa ntchito zomwe adazitcha "njira yosinthika" (ES). Njira ya ES iyi imakhala ndi kusunga zolemera zolemera pa intaneti, kuphatikizapo othandizira angapo omwe amagwira ntchito mofanana ndi kugwiritsa ntchito magawo osankhidwa kuchokera kugawa uku. Wothandizira aliyense amagwira ntchito m'malo ake, ndipo akamaliza kuchuluka kwa magawo kapena magawo a gawo, algorithm imabweretsa mphotho yowonjezereka, yowonetsedwa ngati kulimba mtima. Poganizira za mtengo uwu, kugawidwa kwa magawo kungasinthidwe kwa othandizira opambana, kulepheretsa osapambana. Pobwereza ntchito yotereyi nthawi zambiri ndi kutenga nawo mbali kwa mazana a othandizira, ndizotheka kusuntha kugawa zolemera kumalo omwe angalole kuti ogwira ntchito apange ndondomeko yapamwamba yothetsera ntchito yomwe apatsidwa. Zowonadi, zotsatira zomwe zafotokozedwa m'nkhaniyi ndi zochititsa chidwi: zikuwonetsedwa kuti ngati muthamangitsa othandizira chikwi limodzi, ndiye kuti kuyenda kwa anthropomorphic pamiyendo iwiri kumatha kuphunziridwa pasanathe theka la ola (pamene ngakhale njira zapamwamba kwambiri za RL zimafuna kugwiritsa ntchito ndalama zambiri. kuposa ola limodzi pa izi). Kuti mumve zambiri, ndikupangira kuwerenga zabwino kwambiri positi kuchokera kwa olemba kuyesera, komanso nkhani yasayansi.
Njira zosiyanasiyana zophunzitsira kuyenda mowongoka kwa anthropomorphic, kuphunzira pogwiritsa ntchito njira ya ES kuchokera ku OpenAI.
Bokosi lakuda
Phindu lalikulu la njirayi ndikuti likhoza kufanana mosavuta. Ngakhale njira za RL, monga A3C, zimafuna kuti chidziwitso chisinthidwe pakati pa ulusi wogwira ntchito ndi seva ya parameter, ES imangofunika kuyerekezera kwa thupi ndi chidziwitso chogawa magawo. Ndi chifukwa cha kuphweka kumeneku kuti njirayi ili patsogolo kwambiri kuposa njira zamakono za RL ponena za luso lokulitsa. Komabe, zonsezi sizibwera pachabe: muyenera kukhathamiritsa maukonde molingana ndi mfundo ya black box. Pankhaniyi, "bokosi lakuda" limatanthawuza kuti panthawi yophunzitsira mawonekedwe amkati mwa netiweki amanyalanyazidwa kwathunthu, ndipo zotsatira zake zonse (mphoto ya gawo) zimagwiritsidwa ntchito, ndipo zimatengera ngati zolemera za netiweki inayake zidzakwaniritsidwa. kutengera kwa mibadwo yotsatira. M'mikhalidwe yomwe sitilandira ndemanga zambiri kuchokera ku chilengedwe-komanso m'mabvuto ambiri amtundu wa RL kutuluka kwa mphotho kumakhala kochepa kwambiri-vuto limachokera ku "bokosi lakuda" mpaka "bokosi lakuda kwathunthu." Pankhaniyi, mutha kuwonjezera zokolola zambiri, kotero, zowona, kunyengerera koteroko ndikoyenera. "Ndani amafunikira ma gradients ngati ali ndi phokoso lopanda chiyembekezo?" - ili ndilo lingaliro lonse.
Komabe, m'malo omwe mayankho amakhala achangu, zinthu zimayamba kusokonekera kwa ES. Gulu la OpenAI likufotokoza momwe maukonde osavuta a MNIST adaphunzitsidwa pogwiritsa ntchito ES, ndipo nthawi ino maphunzirowo anali ocheperako nthawi 1000. Chowonadi ndi chakuti chizindikiro cha gradient mumagulu azithunzi ndi chidziwitso kwambiri chokhudza momwe mungaphunzitsire gulu labwino la maukonde. Chifukwa chake, vuto limakhala locheperako ndi njira ya RL komanso zambiri zokhala ndi mphotho zochepa m'malo omwe amapanga phokoso laphokoso.