DeepMind AI Masters Team Sewerani ndi Kuposa Anthu mu Quake III

Jambulani mbendera ndi njira yosavuta yopikisana yomwe imapezeka mwa owombera ambiri otchuka. Gulu lirilonse limakhala ndi cholembera m'munsi mwake, ndipo cholinga chake ndikujambula chikhomo cha gulu lomwe likulimbana nawo ndikuchibweretsa kwa icho chokha. Komabe, zomwe zimakhala zosavuta kwa anthu kuzimvetsetsa sizovuta kwa makina. Kuti agwire mbendera, anthu osakhala osewera (bots) amakonzedwa mwachizolowezi pogwiritsa ntchito ma heuristics ndi ma algorithms osavuta omwe amapereka ufulu wosankha komanso otsika kwambiri kwa anthu. Koma luntha lochita kupanga komanso kuphunzira pamakina zimalonjeza kusintha izi.

В nkhani, lofalitsidwa sabata ino m'magazini ya Science pafupifupi chaka chimodzi pambuyo pake chosindikizira, komanso blog yanu, ofufuza ochokera ku DeepMind, kampani ya ku London ya Zilembo, akufotokoza dongosolo lomwe silingangophunzira kusewera kulanda mbendera pamapu a id Software's Quake III Arena, komanso kupanga njira zatsopano zamagulu, osati zotsika kwa munthu.

"Palibe amene adauza AI momwe amasewera masewerawa, zidangokhala ndi zotsatira - kaya AI idamenya mdani wake kapena ayi. Ubwino wogwiritsa ntchito njirayi ndikuti sudziwa zomwe zingachitike mukamaphunzitsa othandizira," akutero Max Jaderberg, wasayansi wofufuza ku DeepMind yemwe m'mbuyomu adagwirapo ntchito pamakina ophunzirira makina a AlphaStar (posachedwa kwambiri. kupitirira gulu la anthu akatswiri mu Starcraft II). Anafotokozanso kuti njira yofunikira ya ntchito yawo yatsopano ndiyo, choyamba, kulimbikitsa kuphunzira, komwe kumagwiritsa ntchito mtundu wa mphotho kukankhira othandizira mapulogalamu kuti akwaniritse zolinga zawo, ndipo dongosolo la mphotho lidagwira ntchito mosasamala kanthu kuti gulu la AI lapambana kapena ayi. , koma chachiwiri, othandizira adaphunzitsidwa m'magulu, zomwe zidakakamiza AI kuti adziwe bwino kuyanjana kwamagulu kuyambira pachiyambi.

"Malinga ndi kafukufuku wofufuza, ichi ndi chachilendo pamachitidwe a algorithmic omwe ndi osangalatsa kwambiri," anawonjezera Max. "Momwe tidaphunzitsira AI yathu ikuwonetsa bwino momwe tingakulitsire ndikukhazikitsa malingaliro apamwamba achisinthiko."

Otchedwa For The Win (FTW), othandizira a DeepMind amaphunzira mwachindunji kuchokera ku pixel zowonekera pogwiritsa ntchito convolutional neural network, gulu la masamu (ma neurons) okonzedwa m'magawo otengera mawonekedwe amunthu. Deta yolandilidwa imatumizidwa ku maukonde awiri okhala ndi kukumbukira kwakanthawi kochepa (Chingerezi chanthawi yayitali kukumbukira - LSTM), yomwe imatha kuzindikira kudalira kwanthawi yayitali. Mmodzi wa iwo amayendetsa deta yogwira ntchito ndi liwiro lachangu, pamene winayo amagwira ntchito pang'onopang'ono kusanthula ndi kupanga njira. Onsewa amalumikizidwa ndi kukumbukira kosiyanasiyana, komwe amagwiritsa ntchito limodzi kulosera zakusintha kwamasewera ndikuchitapo kanthu kudzera mwa wowongolera masewera omwe amatsanzira.

Ponseponse, DeepMind idaphunzitsa othandizira 30, kuwapatsa osewera nawo angapo ndi omwe amawatsutsa kuti azisewera nawo, komanso makhadi osankhidwa mwachisawawa kuti AI isawakumbukire. Wothandizira aliyense anali ndi chizindikiro chake cha mphotho, kumulola kuti apange zolinga zake zamkati, monga kujambula mbendera. AI iliyonse payekha idasewera masewera pafupifupi 450 zikwizikwi olanda mbendera, zomwe ndi zofanana ndi zaka zinayi zamasewera.

Othandizira a FTW ophunzitsidwa bwino aphunzira kugwiritsa ntchito njira zomwe zimafanana ndi mapu aliwonse, mndandanda wamagulu, ndi kukula kwa timu. Anaphunzira makhalidwe a anthu monga kutsatira anzawo a m’timu, kumanga msasa pamalo a adani, ndi kuteteza maziko awo kwa omwe akuukira, ndipo pang’onopang’ono anataya njira zosapindulitsa kwenikweni monga kuyang’anira wothandizana naye kwambiri.

Ndiye zotsatira zake zinali zotani? Pampikisano wa anthu 40 momwe anthu ndi othandizira adasewera limodzi komanso motsutsana wina ndi mnzake, othandizira a FTW adapambana kwambiri kupambana kwa osewera. Chiwerengero cha Elo cha AI, chomwe ndi mwayi wopambana, chinali 1600, poyerekeza ndi 1300 kwa osewera "amphamvu" aumunthu ndi 1050 kwa "average" player player.

Izi sizosadabwitsa, chifukwa liwiro la AI ndilokwera kwambiri kuposa la munthu, zomwe zidapatsa mwayi wakale pakuyesa koyambirira. Koma ngakhale kulondola kwa othandizira kudachepetsedwa ndipo nthawi yoyankha idakwera chifukwa cha 257 millisecond latency, AI idapambanabe anthu. Osewera apamwamba komanso wamba adapambana 21% ndi 12% yamasewera onse, motsatana.

Komanso, phunzirolo litatha, asayansi adaganiza zoyesa oyimira pamapu a Quake III Arena omwe ali ndi zomangamanga zovuta komanso zina zowonjezera, monga Future Crossings ndi Ironwood, kumene AI inayamba kutsutsa anthu bwino pamasewera oyesera. . Ofufuzawo atayang'ana mawonekedwe a neural network activation ya othandizira, ndiye kuti, ntchito za ma neuron omwe amawunikira zomwe zimachokera kuzomwe zikubwera, adapeza magulu oyimira zipinda, momwe mbendera, mawonekedwe a osewera nawo ndi otsutsa, ndi kukhalapo kapena kusapezeka kwa othandizira pamalo a adani, kapena kutengera timu, ndi mbali zina zazikulu zamasewera. Othandizira ophunzitsidwawo anali ndi ma neuron omwe amalemba zochitika zenizeni mwachindunji, monga pamene mbendera inatengedwa ndi wothandizira kapena pamene wothandizana naye anali kuigwira.

"Ndikuganiza kuti chimodzi mwazinthu zoyenera kuyang'ana ndikuti magulu a othandizira ambiri ndi amphamvu kwambiri, ndipo kafukufuku wathu akuwonetsa izi," akutero Jaderberg. "Izi ndi zomwe takhala tikuphunzira kuchita bwino komanso bwino zaka zingapo zapitazi-momwe tingathetsere vuto la maphunziro olimbikitsa." Ndipo maphunziro owonjezerawa adagwira ntchito bwino kwambiri. ”

Thore Graepel, pulofesa wa sayansi ya makompyuta ku University College London komanso wasayansi wa DeepMind, akukhulupirira kuti ntchito yawo ikuwonetsa kuthekera kwa kuphunzira kwa ma agent ambiri mtsogolo mwa AI. Zitha kukhalanso maziko opangira kafukufuku wokhudzana ndi makina a anthu ndi machitidwe omwe amathandizirana kapena kugwirira ntchito limodzi.

"Zotsatira zathu zikuwonetsa kuti kuphunzira kolimbitsa ma agent ambiri kumatha kukhala bwino pamasewera ovuta mpaka pomwe osewera amafika pokhulupirira kuti osewera apakompyuta amapanga anzawo abwinoko. Kafukufukuyu amaperekanso chidwi chozama mozama momwe othandizira ophunzitsidwa amachitira ndi kugwirira ntchito limodzi, akutero Grapel. "Chomwe chimapangitsa izi kukhala zosangalatsa kwambiri ndikuti othandizirawa amawona malo omwe amakhala ngati munthu woyamba, [ndiko] ngati wosewera munthu. Kuti aphunzire kusewera mwaluso komanso kugwirizana ndi anzawo, othandizirawa amayenera kudalira mayankho azotsatira zamasewera, popanda mphunzitsi kapena mphunzitsi aliyense kuwawonetsa zoyenera kuchita. "

Source: 3dnews.ru