Ɗaukar tuta hanya ce mai sauƙi mai sauƙi da aka samu a cikin shahararrun masu harbi da yawa. Kowace ƙungiya tana da alamar da ke a gindinta, kuma manufar ita ce ta kama alamar ƙungiyar da ke hamayya da kuma samun nasarar kawo ta ga kanta. Duk da haka, abin da ke da sauƙin fahimtar mutane ba shi da sauƙi ga inji. Don ɗaukar tuta, haruffan da ba na ɗan wasa ba (bots) ana tsara su ta al'ada ta hanyar amfani da ilimin lissafi da sauƙi na algorithms waɗanda ke ba da iyakacin yancin zaɓi kuma suna da ƙasa da ɗan adam. Amma basirar wucin gadi da koyan na'ura sun yi alkawarin canza wannan yanayin gaba ɗaya.
В , wanda aka buga wannan makon a cikin mujallar Kimiyya kimanin shekara guda bayan haka kazalika , Masu bincike daga DeepMind, wani reshen na Alphabet na London, sun bayyana tsarin da ba zai iya koyon buga tuta ba kawai a taswirar Quake III Arena na id Software, amma kuma yana haɓaka sabbin dabarun ƙungiya gaba ɗaya, ba ta wata hanya ƙasa da ɗan adam.
"Babu wanda ya gaya wa AI yadda za a buga wannan wasan, kawai yana da sakamako - ko AI ta doke abokin hamayyarta ko a'a. Kyakkyawan amfani da wannan hanyar ita ce ba za ku taɓa sanin wane hali zai fito ba lokacin da kuke horar da jami'ai, "in ji Max Jaderberg, masanin kimiyyar bincike a DeepMind wanda a baya ya yi aiki kan tsarin koyan na'ura AlphaStar (kwanan nan). ƙungiyar ƙwararrun mutane a cikin StarCraft II). Ya ci gaba da bayyana cewa, babbar hanyar sabon aikin nasu shi ne, na farko, ƙarfafa ilmantarwa, wanda ke amfani da wani nau'i na tsarin lada don tura wakilan software don cimma burin da aka tsara, kuma tsarin lada ya yi aiki ba tare da la'akari da ko kungiyar AI ta ci nasara ba ko a'a. , amma a cikin - na biyu, an horar da wakilai a kungiyoyi, wanda ya tilasta AI don kula da hulɗar ƙungiya daga farkon.
"Daga ra'ayi na bincike, wannan sabon abu ne don tsarin algorithmic wanda yake da ban sha'awa sosai," Max ya kara da cewa. "Hanyar da muka horar da AI ɗinmu yana nuna da kyau yadda ake ƙima da aiwatar da wasu ra'ayoyin juyin halitta na yau da kullun."
Mai tsokaci mai suna For The Win (FTW), wakilan DeepMind suna koyo kai tsaye daga pixels na allo ta amfani da hanyar sadarwa mai jujjuyawar jijiya, saitin ayyukan lissafi (neurons) wanda aka tsara a cikin yadudduka waɗanda aka kera bayan bawo na gani na ɗan adam. Ana watsa bayanan da aka karɓa zuwa cibiyoyin sadarwa guda biyu tare da ƙwaƙwalwar ɗan gajeren lokaci da yawa (Turanci dogon ƙwaƙwalwar ajiyar ɗan gajeren lokaci - LSTM), masu iya gane dogaro na dogon lokaci. Ɗayan su yana sarrafa bayanan aiki tare da saurin amsawa, yayin da ɗayan yana aiki a hankali don yin nazari da tsara dabarun. Dukansu suna da alaƙa da bambance-bambancen ƙwaƙwalwar ajiya, waɗanda suke amfani da su tare don hango canje-canje a cikin duniyar wasan da aiwatar da ayyuka ta hanyar mai sarrafa wasan kwaikwayo.
Gabaɗaya, DeepMind ya horar da wakilai 30, ya ba su ɗimbin abokan wasa da abokan hamayya don yin wasa tare, da kuma zaɓen katunan wasan ba da gangan ba don hana AI daga tunawa da su. Kowane wakili yana da nasa siginar lada, yana ba shi damar ƙirƙirar nasa manufofin ciki, kamar kama tuta. Kowane AI ya buga kusan wasanni dubu 450 na kama tuta, wanda yayi daidai da kusan shekaru huɗu na ƙwarewar wasan.
Wakilan FTW masu cikakken horarwa sun koyi yin amfani da dabarun gama gari ga kowane taswira, jerin gwano, da girman ƙungiyar. Sun koyi halayen ɗan adam kamar bin abokan wasansu, yin sansani a sansanin abokan gaba, da kuma kare tushensu daga maharan, kuma a hankali sun rasa tsarin da ba su da fa'ida kamar kallon abokan gaba sosai.
To wane sakamako aka samu? A cikin gasa ta mutum 40 inda mutane da wakilai suka yi wasa ba da gangan ba tare da juna, wakilan FTW sun yi fice sosai wajen samun nasarar 'yan wasan ɗan adam. Ƙimar AI ta Elo, wanda shine yuwuwar cin nasara, ya kasance 1600, idan aka kwatanta da 1300 don "ƙarfafa" 'yan wasan ɗan adam da 1050 don "matsakaicin" ɗan adam.
Wannan ba abin mamaki ba ne, tun da saurin amsawar AI ya fi girma fiye da na ɗan adam, wanda ya ba tsohon babban amfani a cikin gwaje-gwajen farko. Amma ko da lokacin da aka rage daidaiton wakilan kuma lokacin amsawa ya karu godiya ga ginanniyar 257 millisecond latency, AI har yanzu ya fi ɗan adam. Manyan 'yan wasa da na yau da kullun sun ci kashi 21% da 12% na jimlar wasannin, bi da bi.
Bugu da ƙari, bayan buga binciken, masana kimiyya sun yanke shawarar gwada wakilai akan taswirar Quake III Arena mai cikakken tsari tare da tsarin gine-gine masu rikitarwa da ƙarin abubuwa, kamar Crossings na gaba da Ironwood, inda AI ta fara samun nasarar kalubalantar mutane a wasan gwaji. . Lokacin da masu binciken suka kalli tsarin kunna hanyar sadarwa na jijiyoyi na wakilai, wato, ayyukan neurons da ke da alhakin tantance fitarwa bisa ga bayanan da ke shigowa, sun sami gungu masu wakiltar ɗakuna, yanayin tutoci, ganuwa na abokan aiki da abokan adawa, da kasantuwar ko rashin wakilai a sansanin abokan gaba ko na kungiya, da sauran muhimman abubuwan wasan. Wakilan da aka horar sun ma ƙunshi ƙwayoyin cuta waɗanda ke ɓoye takamaiman yanayi kai tsaye, kamar lokacin da wakili ya ɗauki tuta ko lokacin da abokin tarayya ke riƙe da ita.
"Ina tsammanin daya daga cikin abubuwan da za a duba shi ne cewa waɗannan ƙungiyoyin wakilai da yawa suna da ƙarfi sosai, kuma bincikenmu ya nuna hakan," in ji Jaderberg. "Wannan shine abin da muka koya don yin mafi kyau kuma mafi kyau a cikin 'yan shekarun da suka gabata-yadda za a magance matsalar ƙarfafa koyo." Kuma horarwar da aka inganta ta yi aiki sosai."
Thore Graepel, farfesa a fannin kimiyyar kwamfuta a Kwalejin Jami'ar London kuma masanin kimiyyar DeepMind, ya yi imanin cewa aikinsu yana nuna yiwuwar koyo na wakilai da yawa don makomar AI. Hakanan zai iya zama tushen bincike kan hulɗar ɗan adam da injina da tsarin da ke haɗa juna ko aiki tare.
“Sakamakonmu ya nuna cewa koyon ƙarfafawa wakilai da yawa na iya samun nasarar sarrafa wasa mai sarƙaƙiya har ƴan wasan ɗan adam ma sun yarda cewa ƴan wasan kwamfuta suna yin abokan wasa mafi kyau. Har ila yau, binciken ya ba da cikakken bincike mai zurfi game da yadda ƙwararrun wakilai ke nuna hali da aiki tare, in ji Grapel. "Abin da ya sa waɗannan sakamakon su kasance masu ban sha'awa shi ne cewa waɗannan jami'an sun fahimci yanayin su a cikin mutum na farko, (wato) kamar dan wasan ɗan adam. Don koyon yadda ake wasa da dabara da kuma ba da haɗin kai da abokan wasansu, dole ne waɗannan wakilai su dogara da martani daga sakamakon wasan, ba tare da wani malami ko koci ya nuna musu abin da za su yi ba."
source: 3dnews.ru