ãããããã«ïŒ
ç§ãã¡ã¯ãã³ãŒãããªããæããã«åŠè¡çãªæ§è³ªãæ〠XNUMX 幎åã®ããã¹ãã®ç¿»èš³ãããã«æçš¿ããããšã決å®ããããšã¯ããŸããããŸããããä»æ¥ã¯äŸå€ãèšããŸãã ãã®èšäºã®ã¿ã€ãã«ã§æèµ·ãããŠãããžã¬ã³ããå€ãã®èªè ãæ©ãŸããŠããããšãé¡ã£ãŠããŸãããŸããããªãã¯ãã®èšäºãè«ããŠããé²åæŠç¥ã«é¢ããåºæ¬çãªèäœãåæã§ãã§ã«èªãã§ããããããããèªãäºå®ã§ãã ç«ãããžããããïŒ
2017 幎 XNUMX æãOpenAI ã¯ãããšããè«æã§æ·±å±€åŠç¿ã³ãã¥ããã£ã«æ³¢çŽãåºããŸããã
é²åæŠç¥
OpenAI ã®è«æã®äž»ãªããŒãã¯ã匷ååŠç¿ãšåŸæ¥ã®éäŒæãçµã¿åãããŠäœ¿çšââãã代ããã«ããé²åæŠç¥ã(ES) ãšåŒã°ãããã®ã䜿çšããŠè€éãªåé¡ã解決ãããã¥ãŒã©ã« ãããã¯ãŒã¯ã®ãã¬ãŒãã³ã°ã«æåãããšããããšã§ããã ãã® ES ã¢ãããŒãã¯ãè€æ°ã®ãšãŒãžã§ã³ãã䞊è¡ããŠåäœãããã®ååžããéžæããããã©ã¡ãŒã¿ã䜿çšããããšã«ããããããã¯ãŒã¯å
šäœã®éã¿ã®ååžãç¶æããããšã§æ§æãããŸãã åãšãŒãžã§ã³ãã¯ç¬èªã®ç°å¢ã§åäœããæå®ãããæ°ã®ãšããœãŒããŸãã¯ãšããœãŒãã®ã¹ããŒãžãå®äºãããšãã¢ã«ãŽãªãºã ã¯ãã£ãããã¹ ã¹ã³ã¢ãšããŠè¡šçŸããã环ç©å ±é
¬ãè¿ããŸãã ãã®å€ãèæ
®ãããšããã©ã¡ãŒã¿ã®ååžãããæåãããšãŒãžã§ã³ãã«ã·ããããããŸãæåããŠããªããšãŒãžã§ã³ããé€å€ããããšãã§ããŸãã ãã®ãããªæäœãäœçŸäººãã®ãšãŒãžã§ã³ããåå ããŠäœçŸäžåãç¹°ãè¿ãããšã§ããšãŒãžã§ã³ããèªåã«å²ãåœãŠãããã¿ã¹ã¯ã解決ããããã®è³ªã®é«ãããªã·ãŒãçå®ã§ãã空éã«éã¿ã®é
åã移åããããšãã§ããŸãã å®éããã®èšäºã§çŽ¹ä»ãããŠããçµæã¯å°è±¡çã§ããXNUMX åã®ãšãŒãžã§ã³ãã䞊è¡ããŠå®è¡ãããšãæ¬äººåãããäºè¶³æ©è¡ã XNUMX å以å
ã«åŠç¿ã§ããããšã瀺ãããŠããŸã (ãã ããæãé«åºŠãª RL ææ³ã§ãã£ãŠããããå€ãã®è²»çšãå¿
èŠã§ã)ãããã«ã¯ XNUMX æé以äžïŒã ããã«è©³ããæ
å ±ã«ã€ããŠã¯ãåªããæžç±ãèªãããšããå§ãããŸãã
OpenAI ã® ES ã¡ãœããã䜿çšããŠç 究ããããæ¬äººåãããçŽç«æ©è¡ãæããããã®ããŸããŸãªæŠç¥ã
ãã©ãã¯ããã¯ã¹
ãã®æ¹æ³ã®å€§ããªå©ç¹ã¯ãç°¡åã«äžŠååã§ããããšã§ãã A3C ãªã©ã® RL ã¡ãœããã§ã¯ã¯ãŒã«ãŒ ã¹ã¬ãããšãã©ã¡ãŒã¿ ãµãŒããŒéã§æ å ±ã亀æããå¿ èŠããããŸãããES ã§ã¯é©å床æšå®ãšäžè¬åããããã©ã¡ãŒã¿ååžæ å ±ã®ã¿ãå¿ èŠã§ãã ãã®åçŽãã«ããããã®ææ³ã¯ã¹ã±ãŒãªã³ã°æ©èœã®ç¹ã§ææ°ã® RL ææ³ãããã¯ããã«åªããŠããŸãã ãã ããããããã¹ãŠãç¡é§ã«ãªãããã§ã¯ãããŸããããã©ã㯠ããã¯ã¹ã®ååã«åŸã£ãŠãããã¯ãŒã¯ãæé©åããå¿ èŠããããŸãã ãã®å Žåãããã©ã㯠ããã¯ã¹ããšã¯ããã¬ãŒãã³ã°äžã«ãããã¯ãŒã¯ã®å éšæ§é ãå®å šã«ç¡èŠãããå šäœçãªçµæ (ãšããœãŒãã®å ±é ¬) ã®ã¿ã䜿çšãããç¹å®ã®ãããã¯ãŒã¯ã®éã¿ãã©ã®ããã«å€åãããã¯ããã«äŸåããããšãæå³ããŸããåŸã®äžä»£ã«ãåãç¶ãããŸãã ç°å¢ããããŸããã£ãŒãããã¯ãåãåããªãç¶æ³ã§ã¯ããŸãå€ãã®åŸæ¥ã® RL åé¡ã§ã¯å ±é ¬ã®æµããéåžžã«åžèã§ãããããåé¡ã¯ãéšåçã«ãã©ã㯠ããã¯ã¹ããããå®å šã«ãã©ã㯠ããã¯ã¹ãã«ãªããŸãã ãã®å Žåãçç£æ§ãå€§å¹ ã«åäžã§ããããããã¡ããããã®ãããªåŠ¥åã¯æ£åœåãããŸãã ããšã«ãã絶æçã«ãã€ãºãå€ãå Žåã誰ãã°ã©ããŒã·ã§ã³ãå¿ èŠãšããã§ãããã?ã -ãããäžè¬çãªæèŠã§ãã
ãã ãããã£ãŒãããã¯ããã掻çºã«ãªãç¶æ³ã§ã¯ãES ã«ãšã£ãŠåé¡ãçºçãå§ããŸãã OpenAI ããŒã ã¯ãã·ã³ãã«ãª MNIST åé¡ãããã¯ãŒã¯ã ES ã䜿çšããŠãã¬ãŒãã³ã°ãããæ¹æ³ã«ã€ããŠèª¬æããŠããŸãããä»åã®ãã¬ãŒãã³ã°ã¯ 1000 åé ãã£ãã§ãã å®éã®ãšãããç»ååé¡ã«ãããåŸé ä¿¡å·ã¯ãããé©åãªåé¡ããããã¯ãŒã¯ã«æããæ¹æ³ã«é¢ããŠéåžžã«æçã§ãã ãããã£ãŠãåé¡ã¯ RL ææ³ã§ã¯å°ãªãããã€ãºã®å€ãåŸé ãçæããç°å¢ã§ã®å ±é ¬ããŸã°ããªå Žåã«åé¡ã倧ãããªããŸãã
èªç¶ã®è§£æ±ºç
èªç¶ã®äŸããåŠã³ãAI ã®éçºæ¹æ³ãèããŠã¿ããšãå Žåã«ãã£ãŠã¯ AI ã¯æ¬¡ã®ããã«èããããšãã§ããŸãã
åºä¹³é¡ã®ç¥çè¡åã調ã¹ãŠã¿ããšãããã¯å¯æ¥ã«é¢é£ãã XNUMX ã€ã®ããã»ã¹ã®è€éãªçžäºåœ±é¿ã®çµæãšããŠåœ¢æãããŠããããšãããããŸãã ä»äººã®çµéšããåŠã¶ О å®è·µããåŠã¶ã åè ã¯èªç¶éžæã«ããé²åãšåäžèŠãããããšãå€ãã§ãããããã§ã¯ãšããžã§ããã£ã¯ã¹ããã€ã¯ããã€ãªãŒã ãããã³éºäŒçã«é¢é£ã®ãªãçç©éã§ã®çµéšã®å ±æãå¯èœã«ãããã®ä»ã®ã¡ã«ããºã ãèæ ®ããããã«ãããåºãçšèªã䜿çšããŠããŸãã XNUMX çªç®ã®ããã»ã¹ã§ããçµéšããã®åŠç¿ã¯ãåç©ãç涯ãéããŠåŠç¿ãããã¹ãŠã®æ å ±ã§ããããã®æ å ±ã¯ãã®åç©ãšå€çãšã®çžäºäœçšã«ãã£ãŠçŽæ¥æ±ºå®ãããŸãã ãã®ã«ããŽãªã«ã¯ãç©äœãèªèããåŠç¿ãããåŠç¿ããã»ã¹ã«åºæã®ã³ãã¥ãã±ãŒã·ã§ã³ã®ç¿åŸãŸã§ããããããã®ãå«ãŸããŸãã
倧ãŸãã«èšãã°ãèªç¶çã§çºçããããã 100 ã€ã®ããã»ã¹ã¯ããã¥ãŒã©ã« ãããã¯ãŒã¯ãæé©åããããã® XNUMX ã€ã®ãªãã·ã§ã³ã«ããšããããšãã§ããŸãã åŸé ã«é¢ããæ å ±ã䜿çšããŠçç©ã«é¢ããæ å ±ãæŽæ°ããé²åæŠç¥ã¯ãä»äººã®çµéšããåŠã¶ããšã«è¿ã¥ããŸãã åæ§ã«ãäœããã®çµéšãåŸãããšããšãŒãžã§ã³ãã®è¡åã«äœããã®å€åãããããåŸé æ³ã¯ãèªåèªèº«ã®çµéšããåŠç¿ããããšã«å¹æµããŸãã ããã XNUMX ã€ã®ã¢ãããŒãã®ãããããåç©ã«ãããŠã©ã®ãããªçš®é¡ã®ç¥çãªè¡åãèœåãçºéãããããèãããšããã®æ¯èŒã¯ããé¡èã«ãªããŸãã ã©ã¡ãã®å Žåãããé²åçæ¹æ³ãã¯ã人ãäžå®ã®é©å¿åºŠïŒçãç¶ããã®ã«ååãªïŒãçºéãããããšãå¯èœã«ããåå¿çè¡åã®ç 究ãä¿é²ããŸãã æ©ãããšã飌è²äžããéããããšãåŠã¶ããšã¯ãå€ãã®å Žåãå€ãã®åç©ã«éºäŒåã¬ãã«ã§ãçµã¿èŸŒãŸããŠããããããæ¬èœçãªãè¡åãšåçã§ãã ããã«ããã®äŸã¯ãå ±é ¬ã·ã°ãã«ãéåžžã«ãŸããªå Žå (ããšãã°ãèµ€ã¡ããã®è²æãæåãããšããäºå®) ã«é²åçææ³ãé©çšã§ããããšã確èªããŸãã ãã®ãããªå Žåããã®äºå®ãçºçããäœå¹Žãåã«å®è¡ãããå¯èœæ§ã®ããç¹å®ã®äžé£ã®è¡åãšå ±é ¬ãçžé¢ãããããšã¯äžå¯èœã§ãã äžæ¹ãES ã倱æããå Žåãã€ãŸãç»ååé¡ãèæ ®ãããšããã®çµæã¯ãXNUMX 幎以äžã«ããã£ãŠè¡ãããç¡æ°ã®è¡åå¿çåŠå®éšã§åŸãããåç©åŠç¿ã®çµæã«é©ãã»ã©å¹æµããŸãã
åç©ããåŠã¶
匷ååŠç¿ã§äœ¿çšãããææ³ã¯ãå€ãã®å Žåãå¿çåŠã®æç®ããçŽæ¥åŒçšãããŠããŸãã
çµéšããåŠç¿ããéã®äºæž¬ã®äžå¿çãªåœ¹å²ã¯ãäžèšã®ãã€ããã¯ã¹ã倧ããå€åãããŸãã 以åã¯éåžžã«ãŸã°ãã§ãããšèããããŠããã·ã°ãã« (ãšããœãŒãå ±é
¬) ã¯ãéåžžã«å¯ã§ããããšãå€æããŸããã çè«çã«ã¯ãç¶æ³ã¯æ¬¡ã®ãããªãã®ã§ããåºä¹³é¡ã®è³ã¯åžžã«ãæèŠåºæ¿ãšè¡åã®è€éãªæµãã«åºã¥ããŠçµæãèšç®ããŠããŸãããåç©ã¯åã«ãã®æµãã«æµžã£ãŠããã ãã§ãã ãã®å Žåãåç©ã®æçµçãªè¡åã¯ãäºæž¬ã®èª¿æŽãšè¡åã®çºå±ãå°ãããã«äœ¿çšããå¿
èŠããã匷ãä¿¡å·ãäžããŸãã è³ã¯ããããã¹ãŠã®ä¿¡å·ã䜿çšããŠãå°æ¥ã®äºæž¬ (ããã³ããã«å¿ããŠå®è¡ãããè¡åã®è³ª) ãæé©åããŸãã ãã®ã¢ãããŒãã®æŠèŠã¯ãåªããæžç±ã
ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®ããå å®ãããã¬ãŒãã³ã°
åžžã«äºæž¬ãè¡ãã®ã«å¿ããåºä¹³é¡ã®è³ã«åºæã®é«åºŠãªç¥çµæŽ»åã®åçã«åºã¥ããŠã匷ååŠç¿ãæè¿é²æ©ãããã®ãããªäºæž¬ã®éèŠæ§ãèæ ®ãããããã«ãªããŸããã 䌌ããããªäœåãããã« XNUMX ã€ãå§ãããŸãã
ãããã®è«æã®äž¡æ¹ã§ãèè ã¯ãå°æ¥ã®ç°å¢ã®ç¶æ ã«é¢ããäºæž¬çµæã䜿çšããŠããã¥ãŒã©ã« ãããã¯ãŒã¯ã®å žåçãªããã©ã«ã ããªã·ãŒãè£è¶³ããŠããŸãã æåã®èšäºã§ã¯ãããŸããŸãªæž¬å®å€æ°ã«äºæž¬ãé©çšããXNUMX çªç®ã®èšäºã§ã¯ãç°å¢ã®å€åããšãŒãžã§ã³ãã®åäœèªäœã«äºæž¬ãé©çšããŸãã ã©ã¡ãã®å Žåããæ£ã®åŒ·åã«é¢é£ãããŸã°ããªä¿¡å·ã¯ããè±å¯ã§æ å ±éãå€ããªããããè¿ éãªåŠç¿ãšããè€éãªåäœã®ç²åŸã®äž¡æ¹ãå¯èœã«ãªããŸãã ãã®ãããªæ¹åã¯ãåŸé ä¿¡å·ã䜿çšããæ¹æ³ã§ã®ã¿å©çšå¯èœã§ãããES ãªã©ã®ããã©ã㯠ããã¯ã¹ãåçã§åäœããæ¹æ³ã§ã¯å©çšã§ããŸããã
ããã«ãçµéšãšåŸé
æ³ããåŠã¶æ¹ãã¯ããã«å¹æçã§ãã ES ææ³ã䜿çšãããšã匷ååŠç¿ã䜿çšãããããéãç¹å®ã®åé¡ãç 究ã§ããå Žåã§ããES æŠç¥ã«ã¯ RL ãããäœåãå€ãã®ããŒã¿ãå«ãŸãããšããäºå®ã«ãããå©ç¹ãéæãããŸããã ãã®å Žåãåç©ã®åŠç¿åçãåæ ãããšãä»ã®èª°ãã®äŸããåŠç¿ããçµæã¯äœäžä»£ãçµã£ãŠããçŸããŸãããåç©ããã®æèšãæ°žé ã«åŠã¶ã«ã¯ãåç¬ã§çµéšããåäžã®åºæ¥äºã§ååãªå ŽåããããŸãã ã¿ãããªæãã§
ã§ã¯ãããããçµã¿åãããŠã¿ãŠã¯ãããã§ãããã?
ãã®èšäºã®å€§éšåã¯ãç§ã RL ææ³ãæå±ããŠãããšããå°è±¡ãäžããå¯èœæ§ããããŸãã ããããå®éã«ã¯ãé·æçã«ã¯ãäž¡æ¹ã®æ¹æ³ãçµã¿åãããŠããããããæé©ãªç¶æ³ã§äœ¿çšãããããšãæåã®è§£æ±ºçã§ãããšèããŠããŸãã æããã«ãå€ãã®äºåŸå¯Ÿå¿ããªã·ãŒã®å ŽåããŸãã¯æ£ã®åŒ·åã®ä¿¡å·ãéåžžã«ãŸã°ããªç¶æ³ã§ã¯ãç¹ã«å€§èŠæš¡ãªäžŠåãã¬ãŒãã³ã°ãå®è¡ã§ããèªç±ã«äœ¿ããã³ã³ãã¥ãŒãã£ã³ã°èœåãããå Žåã¯ãES ãåã¡ãŸãã äžæ¹ã匷ååŠç¿ãŸãã¯æåž«ããåŠç¿ã䜿çšããåŸé æ³ã¯ãåºç¯ãªãã£ãŒãããã¯ã«ã¢ã¯ã»ã¹ã§ããããå°ãªãããŒã¿ã§åé¡ãè¿ éã«è§£æ±ºããæ¹æ³ãåŠã¶å¿ èŠãããå Žåã«åœ¹ç«ã¡ãŸãã
èªç¶ã«ç®ãåãããšãæåã®æ¹æ³ãæ¬è³ªçã« XNUMX çªç®ã®æ¹æ³ã®åºç€ãç¯ãããšãããããŸãã ããããåºä¹³é¡ãé²åã®éçšã§ãç°å¢ããæ¥ãè€éãªä¿¡å·ããéåžžã«å¹ççã«åŠç¿ã§ããè³ãçºéãããŠããçç±ã§ãã ãããã£ãŠã質åã¯æªè§£æ±ºã®ãŸãŸã§ãã ããããé²åæŠç¥ã¯ãåŸé
åŠç¿æ³ã«ã圹ç«ã€å¹æçãªåŠç¿ã¢ãŒããã¯ãã£ãçºæããã®ã«åœ¹ç«ã€ã§ãããã çµå±ã®ãšãããèªç¶ãèŠã€ãã解決çã¯å®éã«éåžžã«æåããŠããŸãã
åºæïŒ habr.com