InoteveraGaidhi rinotevera
Monte Carlo Tree Search
Tekinoroji
Nhungamiro yehunyanzvi
Markov cheni Monte Carlo (MCMC) inoshandisa inotsamira inodhirowa kubva kune yakanyatsogadzirwa Markov cheni kusvika kune tarisiro yekufungidzira pasi pechinangwa chekugovera, kazhinji mushure meBayesian.
Teerera kugwaro iriAI-yakarondedzerwa · ~3 min
Metropolis-Hastings neGibbs sampling inovaka maketani akadaro, asi fungidziro inobatsira inoda kutarisisa kusanganisa, saizi inoshanda yemuenzaniso uye convergence diagnostics.
Kugovera kwakawanda kwemashure hakugone kutorwa kubva zvakananga kushandisa yakapusa yakazvimirira-dhirowa nzira. MCMC inovaka cheni yeMarkov iyo yakagara yakamira yekugovera ndiyo inotarirwa. Imwe neimwe nyika itsva inotsamira pane yazvino mamiriro, uye pasi pemamiriro akakodzera, maavhareji epakati pemadhirowa anogona kuenzanisa zvinotarisirwa kumashure. Nekuti madhirowa anotsamira, huwandu hwemitsara yakachengetwa haina kufanana nehuwandu hweruzivo rwakazvimirira. MuMetropolis-Hastings, kugoverwa kwechikumbiro kunoratidza nyika inokwikwidza. Iyo algorithm inogamuchira mumiriri iyeye ane mukana zvichienderana nereshiyo yezvinotangwa densities uye, kana chikumbiro chiri asymmetric, gadziriso yekumberi nekudzosera kumashure zvingangoitika. Kana ikaramba, ketani inodzokorora mamiriro ekare. A symmetric random-walk proposal inorerutsa reshiyo, asi zano risina mwero rinogona kufamba kashoma kana kutora matanho asina kunyatsoshanda. Gibbs sampling inzira yakakosha yekumisikidza-yekugadzirisa: edza imwe neimwe shanduko kubva kune yakazara mamiriro ayo akapihwa maitiro azvino evamwe. Inoshanda kana izvo zvakagoverwa zviwanikwa zviripo uye ketani ine zvivakwa zvakakodzera. Fungidzira chinangwa chechikamu chimwe apo zvikumbiro zvinowanzorambwa nekuti saizi dzenhanho dzakakurisa. Iyo ketani inodzokorora nyika, ichiwedzera autocorrelation uye kudzikisa inoshanda saizi saizi. Matanho madiki anogona kugamuchira kazhinji asi ongorora zvishoma nezvishoma. Pane kutengeserana pakati pekugamuchirwa uye kufamba, uye kuongororwa kunofanirwa kuongorora kana maketani akasiyana achiongorora matunhu mamwe chete. Kurasa kudziya kwekutanga kunogona kuderedza kutsamira pakutanga, asi hapana nhamba yakatarwa yemadhirowa inovimbisa kusangana. Mhanya macheni akawanda kubva kunzvimbo dzakapararira dzekutanga, ongorora maronda ekutsvaga, uye ongorora R-heti uye inoshanda saizi yemuenzaniso wehuwandu hunokosha. Aya diagnostics anogona kutadza kuburitsa ese pathologies, kunyanya neakapatsanurwa modes kana zvisina kunyatsoongororwa miswe. Ongororawo sampler-chaiyo yambiro uye Monte Carlo kukanganisa. MCMC inofungidzira kusavimbika kwenhamba kubva kune inopera simulation; havabvisi kusava nechokwadi mumuenzaniso wenhamba kana data. Taura zvakatorwa, kuti zvakaonekwa sei uye kuti ndedzipi pfupiso dzekupedzisira dzakavimbika zvakakwana pasarudzo.
A hypothetical posterior ine maparameter maviri ane chimiro chakabatana. Metropolis-Hastings inokurudzira mumiriri ari padyo uye anoigamuchira zvinoenderana neinotarirwa-density reshiyo uye zano rekugadzirisa; zvakarambwa zvinosiya cheni payazvino.
Sampler yeGibbs inogadziridza parameter nekudhirowa kubva pakugovera kwayo kwakapihwa iwo azvino maitiro emamwe ese ma paramita. Kudzokorora izvi zvigadziriso zvekugadzirisa zvinogadzira cheni yakabatana kana mamiriro eiyo sampler asangana.
Chikwata chinomhanyisa cheni dzakati wandei kubva kune akapararira ekutanga kukosha uye inoenzanisa yekuteedzera marongero, nhanho-yakajairwa R-heti uye yakawanda / muswe inoshanda saizi saizi. Nzira dzakafanana chete hadzibatirwe sehumbowo hwekuti cheni yega yega yakaongorora chinangwa.
Zviuru zviviri zvine chekuita nekudhirowa zvinogona kuve neruzivo rushoma pane nhamba yakafanana yedhirowa yakazvimirira. Saizi yemuenzaniso inoshanda inofungidzira kurasikirwa kweruzivo kune huwandu hwechido.
Ongororo yeBayesian inogona kutaurirana kuvimbika kwekufananidza nekutaura cheni kuverenga, warmup, R-hat, yakawanda uye muswe ESS, Monte Carlo kukanganisa uye sampler yambiro ine posterior pfupiso. Zvikwata zvinofanirwa kusarudza diagnostics inosungirirwa kune yavo chaiyo fungidziro, kusanganisira nemuswe mukana kana sarudzo dzichienderana nezviitiko zvisingawanzo. Kana maketani asingabvumirani kana kuongorora zvishoma nezvishoma, wedzera sampling chete mushure mekunzwisisa geometry uye sampler maitiro. Zvirinani otomatiki diagnostics inogona kutarisana nezvinetso, asi vanoongorora vachiri kuda kuongorora chimiro chemuenzaniso uye kuteedzera maitiro. Izvi zvinoita kuti magumo ekuverenga aonekwe pachinzvimbo chekupa fungidziro yekumashure sekunge yaive maverengero chaiwo.
Tsamba yemahara
Nhatu dzakasimbiswa dzeAI nyaya mangwanani ega ega pakati pevhiki, dzakanyorwa neChirungu chakajeka. Yemahara zvachose, hapana ads.
Imwe email zuva rega rega revhiki. Bvisa kunyoresa kamwe chete. Hatimbotengese kana kugovera kero yako.
Zviedze iwe pachako
Mhinduro yepakarepo pamhinduro yega yega, uye chitupa chinogovaniswa chine ID inobatika kana wapedza kosi.
Tsigira dzidzo yemahara yeAI. AI Understanding ndeye 501(c)(3) isingabatsiri - hapana ads, hapana paywall, zvachose. Ita chipo
Markov cheni Monte Carlo (MCMC) inoshandisa inotsamira inodhirowa kubva kune yakanyatsogadzirwa Markov cheni kusvika kune tarisiro yekufungidzira pasi pechinangwa chekugovera, kazhinji mushure meBayesian. Metropolis-Hastings neGibbs sampling inovaka maketani akadaro, asi fungidziro inobatsira inoda kutarisisa kusanganisa, saizi inoshanda yemuenzaniso uye convergence diagnostics.
Iyo Markov cheni inobatanidza inoteedzana inodhirowa, saka autocorrelation inoderedza iyo yakazvimirira ruzivo inomiririrwa nesaizi yemuenzaniso.
Chiyero chekugamuchira chinosanganisira q(ikozvino|mumiriri)/q(mumiriri|zvino) kugadzirisa chikumbiro asymmetry.
Yega yega Gibbs nhanho inokwevera kubva kune imwe yakazara mamiriro emamiriro pane zvazvino kukosha kwezvakasara zvinosiyana.
Nyika dzekutanga dzakasiyana dzinobatsira kuongorora kana maketani achisangana uye kusanganiswa mukati meiyo yakajairika chinangwa kugovera.
ESS inopfupikisa kuti ingani ruzivo rwunotsamira runodhirowa inopa huwandu kana ichienzaniswa neakazvimirira masampuli.
Ramba uchidzidza
Mamwe madhairekitori akasarudzirwa nyaya iyi
InoteveraGaidhi rinotevera
Monte Carlo Tree Search
Tekinoroji