ã²ãŒã ãã¢ã³ããºãã®åŸ©venãã®é²æ©ã¯ãäžæ £ããªç°å¢ã®ç 究ã«ãããææã®å矩èªãšããŠå€ãã®äººã«èããããŠããŸãã
奜å¥å¿ãéããŠç°å¢ãæ¢çŽ¢ãã匷ååŠç¿ãšãŒãžã§ã³ãã奚å±ããäºæž¬ããŒã¹ã®ã©ã³ãã ãããã¯ãŒã¯èžçïŒRNDïŒã¡ãœãããéçºããŸããã ãã®æ¹æ³ã¯ãã³ã³ãã¥ãŒã¿ãŒã²ãŒã ãã¢ã³ããºãã®åŸ©ven ãã§ã®äººéã®å¹³åçãªçµæãåããŠäžåããŸããïŒICLRã®å¿åã¢ããªã±ãŒã·ã§ã³ãé€ããçµæãç§ãã¡ã®ãã®ããæªãïŒã RNDã¯æå 端ã®å¹çæ§ãå®èšŒãã24宀ãã¹ãŠãå®æçã«æ€åºããäºåã®ãã¢ãã²ãŒã ã®åºæ¬çãªç¶æ ã«ã¢ã¯ã»ã¹ããããšãªã第1ã¬ãã«ãééããŸãã
RNDã¡ãœããã¯ãç¶æ ããŒã¿ã«ã©ã³ãã ã©ã³ãã ãã¥ãŒã©ã«ãããã¯ãŒã¯ãéãåãããçµæãäºæž¬ããè€éãã枬å®ããããšã«ããããšãŒãžã§ã³ãã®ãªãã¿ã®ãªãç¶æ ãžã®ç§»è¡ãåºæ¿ããŸãã æ¡ä»¶ããªãã¿ããªãå Žåãæçµçµæãäºæž¬ããããšã¯å°é£ã§ããããã¯ãå ±é ¬ãé«ãããšãæå³ããŸãã ãã®æ¹æ³ã¯ããããã匷ååŠç¿ã¢ã«ãŽãªãºã ã«é©çšã§ããå®è£ ãç°¡åã§ãã¹ã±ãŒãªã³ã°ã«å¹æçã§ãã 以äžã¯ãRNDã®å®è£ ãžã®ãªã³ã¯ã§ããRNDã¯ãèšäºã®çµæãåçŸããŠããŸãã
ç§åŠè«æã®ããã¹ã ã ã³ãŒã
ã¢ã³ãã¹ããªãã³ãžã®çµæ
ç®çã®ç®æšãéæããããã«ããšãŒãžã§ã³ãã¯æåã«ãç°å¢å ã§å¯èœãªã¢ã¯ã·ã§ã³ãšãç®æšã«åããé²æç¶æ³ãæ§æãããã®ã調æ»ããå¿ èŠããããŸãã ã²ãŒã ã®å€ãã®å ±é ¬ã·ã°ãã«ã¯ã«ãªãã¥ã©ã ãæäŸãããããåçŽãªç 究æŠç¥ã§ããç®æšãéæããã®ã«ååã§ãã DQNã®ãã¬ãŒã³ããŒã·ã§ã³ã«é¢ããæåã®äœæ¥ã§ã¯ ãMontezumaã®Revengeãã DQNãå¹³å人éã¹ã³ã¢ïŒ4700ïŒã®0ïŒ ã®çµæã瀺ããå¯äžã®ã²ãŒã ã§ããã åçŽãªã€ã³ããªãžã§ã³ã¹æŠç¥ã§ã¯ãå ±é ¬ãåéããããšã¯ã§ãããããã¬ãã«ã§æ°éšå±ããèŠã€ããããšãã§ããŸããã ãã以æ¥ãã¢ã³ããºãã®åŸ©venã®ã²ãŒã ã®é²æ©ã¯ãå€ãã®äººãäžæ £ããªç°å¢ã®ç 究ã®é²æ©ãšå矩ã§ãããšèŠãªããŠããŸãã
2016幎ã«DQNãšã«ãŠã³ã¿ãŒã®ããŒãã¹ãçµã¿åãããããšã«ããã倧ããªé²æ©ãéæãããŸããããã®çµæããšãŒãžã§ã³ãã¯15éšå±ãèŠã€ããå¹³åã§çŽ3700ã§æé«ã®6600ãç²åŸããŸããããã以æ¥ãçµæã®å€§å¹ 㪠æ¹åã¯ã å°é家ãŸãã¯ãšãã¥ã¬ãŒã¿ã®åºæ¬ç¶æ ã«ã¢ã¯ã»ã¹ããã
1024人ã®åŸæ¥å¡ã察象ã«å€§èŠæš¡ãªRNDå®éšãå®æœãã 9åã®éå§ã§å¹³å10,000件 ã æé«ã®å¹³åçµæ14,500件ãåŸãŸãã ã ãããã®å ŽåãããšãŒãžã§ã³ãã¯20ã22ã®éšå±ãèŠã€ããŸããã ããã«ã1åã®å°ãããé·ãèµ·åïŒ10åã®ãã¡ïŒã§ã¯ã æ倧çµæã¯17,500ã§ãããã¯æåã®ã¬ãã«ãééããŠ24ã®éšå±ãã¹ãŠãèŠã€ããããšã«å¯Ÿå¿ããŸã ã 以äžã®ã°ã©ãã¯ãããã2ã€ã®å®éšãæ¯èŒããæŽæ°ãã©ã¡ãŒã¿ãŒã«å¿ããå¹³åå€ã瀺ããŠããŸãã
以äžã®èŠèŠåã¯ãå°èŠæš¡ã®å®éšã®é²æã瀺ããŠããŸãã ãšãŒãžã§ã³ãã¯ã奜å¥å¿ã®åœ±é¿ãåããŠãæ°ããéšå±ãéãããã€ã³ããç²åŸããæ¹æ³ãèŠã€ããŸããããã¬ãŒãã³ã°äžã«ããã®å€éšå ±é ¬ã«ãããåŸã§ãããã®éšå±ã«æ»ãããšã匷å¶ãããŸãã
ãšãŒãžã§ã³ããçºèŠããéšå±ãšãã¬ãŒãã³ã°äžã®å¹³åçµæã éšå±ã®éæ床ã¯ããšãŒãžã§ã³ãã®10ãã¹ã®ãã¡äœåæ€åºããããã«å¯Ÿå¿ããŸãã æ å
奜å¥å¿ã«åºã¥ãåºç¯ãªåŠç¿ç 究
RNDãéçºããåã«ãã«ãªãã©ã«ãã¢å€§åŠããŒã¯ã¬ãŒæ ¡ã®ã¹ã¿ãããšäžç·ã«ãç°å¢ãžã®å ±é ¬ãªãã§åŠç¿ãåŠã³ãŸããã 奜å¥å¿ã¯ãç¹å®ã®ã¿ã¹ã¯ã«å¯ŸããŠç¹å¥ã«èšèšãããå ±é ¬æ©èœã䜿çšããã®ã§ã¯ãªãããšãŒãžã§ã³ããããããç°å¢ãšå¯Ÿè©±ããããšãæããç°¡åãªæ¹æ³ãæäŸããŸããããã¯ãŸã åé¡ã®è§£æ±ºã«å¯Ÿå¿ããäºå®ã§ã¯ãããŸããã ALE ã Universe ã Malmo ã Gym ã Gym Retro ã Unity ã DeepMind Lab ã CommAIãªã©ã®ãããžã§ã¯ãã§ã¯ãæšæºåãããã€ã³ã¿ãŒãã§ãŒã¹ãä»ããŠãšãŒãžã§ã³ãçšã«å€æ°ã®ã·ãã¥ã¬ãŒããããç°å¢ãéãããŸãã ç¹å®ã®ç°å¢ã«åºæã§ã¯ãªãäžè¬åãããå ±é ¬æ©èœã䜿çšãããšãŒãžã§ã³ãã¯ãå¹ åºãç°å¢ã§åºæ¬çãªã¬ãã«ã®èœåãç²åŸã§ããŸãã ããã«ããã圌ã¯ç²Ÿå·§ãªå ±é ¬ããªããŠãæçšãªè¡åã決å®ããããšãã§ããŸãã
ç§åŠè«æã®ããã¹ã ã ã³ãŒã
ååå¥ã®ã¿ã€ã ã¹ãããã§åŒ·åãããæšæºçãªãã¬ãŒãã³ã°èšå®ã§ã¯ããšãŒãžã§ã³ãã¯ã¢ã¯ã·ã§ã³ãç°å¢ã«éä¿¡ããåå¿ããŠããšãŒãžã§ã³ãã«æ°ãã芳å¯ã移è¡ã«å¯Ÿããå ±é ¬ãããã³ãšããœãŒãã®çµäºã®ã€ã³ãžã±ãŒã¿ãäžããŸãã åã®èšäºã§ã¯ã次ã®èŠ³å¯çµæã®ã¿ ãçæããããã«ç°å¢ãèšå®ããŸããã ããã§ããšãŒãžã§ã³ãã¯åœŒã®çµéšã«åºã¥ããŠæ¬¡ã®ç¶æ ã®äºæž¬ã¢ãã«ãç 究ããäºæž¬å ±é ¬ãå éšå ±é ¬ãšããŠäœ¿çšããŸãã ãã®çµæã圌ã¯äºæž¬äžèœã«æ¹ãããŠããŸãã ããšãã°ãã²ãŒã ã¢ã«ãŠã³ãã®å€æŽã¯ãã¢ã«ãŠã³ããç»é¢ã«è¡šç€ºãããå€æŽãäºæž¬ããã®ãå°é£ãªå Žåã«ã®ã¿å ±ãããŸãã éåžžããšãŒãžã§ã³ãã¯æ°ãããªããžã§ã¯ããšã®æçšãªçžäºäœçšãèŠã€ããŸãããã®ãããªçžäºäœçšã®çµæã¯ãéåžžãç°å¢ã®ä»ã®åŽé¢ãããäºæž¬ãé£ããããã§ãã
ä»ã® ç 究è ãšåæ§ã«ãç°å¢ã®ãã¹ãŠã®åŽé¢ãé¢é£ããŠãããã©ããã«é¢ä¿ãªããã¢ããªã³ã°ã®èŠ³å¯ã®ç¹åŸŽãéžæããããšãé¿ããŸããã é©ãã¹ãããšã«ãã©ã³ãã é¢æ°ã§ããããŸãæ©èœããããšãããããŸããã
奜å¥å¿agentsçãªãšãŒãžã§ã³ãã¯äœãããŸããïŒ
50以äžã®ç°ãªãç°å¢ã§ãšãŒãžã§ã³ãããã¹ãããäžèŠã©ã³ãã ãªåäœããç°å¢ãšã®æèçãªçžäºäœçšãŸã§ãããŸããŸãªèœåã芳å¯ããŸããã é©ããããšã«ãå Žåã«ãã£ãŠã¯ããšãŒãžã§ã³ãã¯ãªããšãã²ãŒã ãå®äºããããšãã§ããŸããããå€éšå ±é ¬ãéããŠç®æšãç¥ããããŠããŸããã§ããã
ãã¬ãŒãã³ã°éå§æã®å éšå ±é ¬
ã¬ãã«ã®æåã®ééã§ã®å éšå ±é ¬ã®ãžã£ã³ã
ãã¬ã€ã¯ã¢ãŠã -ãšãŒãžã§ã³ãããã¬ãŒãã³ã°ã®åæ段éã§ãããã¯ã®æ°ããæ§æãèŠããšããããã³æ°æéãã¬ãŒãã³ã°åŸåããŠã¬ãã«ãééãããšãã«ãå éšå ±é ¬ã«ãžã£ã³ãããŸãã
ãã³ -äž¡æ¹ã®ãã©ãããã©ãŒã ãåæã«å¶åŸ¡ããããã«ãšãŒãžã§ã³ããèšç·Žãã圌ã¯ã²ãŒã å ã§ããŒã«ãããŒãããããšãåŠã³ãŸããã ã²ãŒã å ã®AIã«å¯ŸããŠãã¬ãŒãã³ã°ãè¡ãå Žåã§ãããšãŒãžã§ã³ãã¯ã²ãŒã ãæ倧åããããšããŸããããåã€ããšã¯ã§ããŸããã§ããã
ããŠãªã³ã° -ãšãŒãžã§ã³ãã¯ãå€éšå ±é ¬ãæ倧åããããã«çŽæ¥èšç·Žãããä»ã®ãšãŒãžã§ã³ããããã²ãŒã ãäžæã«ãã¬ã€ããããšãåŠã³ãŸããã ããã¯ãæçåŸã®ã¹ã³ã¢ããŒãã®äºæž¬ãå°é£ãªç¹æ» ã«ãšãŒãžã§ã³ããæ¹ãããããã ãšèããŠããŸãã
ããªãª -å éšå ±é ¬ã¯ãã²ãŒã ã®ç®æšã§ããã¬ãã«ã®é²è¡ãšç¹ã«ããäžèŽããŠããŸãã æ°ããèŠã€ãã£ããšãªã¢ã®è©³çŽ°ã¯äºæž¬ã§ããªãããããšãŒãžã§ã³ãã¯æ°ãããšãªã¢ã®æ€çŽ¢ã«å¯ŸããŠå ±é ¬ãåãåããŸãã ãã®çµæããšãŒãžã§ã³ãã¯11ã¬ãã«ãçºèŠããç§å¯ã®éšå±ãèŠã€ãããã¹ãåãããããŸããã
éšã ãããã¬ãã®åé¡
ã©ã³ãã ãªçµæã«æ¹ãããã¹ããããã·ã³ã®ã®ã£ã³ãã©ãŒãšããŠããšãŒãžã§ã³ãã¯ããã€ãºã®å€ããã¬ãã®åé¡ãã®çµæãšããŠã奜å¥å¿ã®trapã«é¥ãããšããããŸãã ãšãŒãžã§ã³ãã¯ç°å¢å ã®ã©ã³ãã æ§ã®åå ãèŠã€ããããã芳å¯ãç¶ãããã®ãããªé·ç§»ã«å¯ŸããŠåžžã«é«ãå éšå ±é ¬ãçµéšããŸãã ãã®ãããªãã©ããã®äŸã¯ãéçãã€ãºãçæãããã¬ããèŠãããšã§ãã ã©ã³ãã ãã£ã³ãã«ãåçããTVãåããUnityè¿·è·¯ã«ãšãŒãžã§ã³ããé 眮ããããšã§ãæåéããããå®èšŒããŸãã
éšã ãããã¬ããšè¿·è·¯ã®ãšãŒãžã§ã³ã
éšã ãããã¬ãã®ãªãè¿·è·¯ã®ãšãŒãžã§ã³ã
çè«çã«ã¯ããã€ãºã®å€ããã¬ãã®åé¡ã¯éåžžã«æ·±å»ã§ãããã¢ã³ãã¹ãã®åŸ©likeã®ãããªéåžžã«æ±ºå®çãªç°å¢ã§ã¯ã奜å¥å¿ã«ãã£ãŠãšãŒãžã§ã³ããéšå±ãèŠã€ããŠãªããžã§ã¯ããšããåãããããšãäºæ³ãããŸãã 奜å¥å¿ã«åºã¥ããŠæ¬¡ã®ç¶æ ãäºæž¬ããããã®ããã€ãã®ãªãã·ã§ã³ãè©ŠããŸããããªãµãŒãããŒãã¹ãšã²ãŒã ã¢ã«ãŠã³ããçµã¿åãããŸããã
ãããã®å®éšã§ã¯ããšãŒãžã§ã³ãã¯ãã€ãºã³ã³ãããŒã©ãŒãä»ããŠç°å¢ãå¶åŸ¡ããŸãããã€ãºã³ã³ãããŒã©ãŒã¯ãçŸåšã®ã¢ã¯ã·ã§ã³ã§ã¯ãªãæåŸã®ã¢ã¯ã·ã§ã³ãäœããã®ç¢ºçã§ç¹°ãè¿ããŸãã ç¹°ãè¿ãå¯èœãªãã¹ãã£ãããŒãã¢ã¯ã·ã§ã³ãåãããã®èšå®ã¯ãèšæ¶ãé²ãããã«ãAtariãªã©ã®å®å šã«æ±ºå®çãªã²ãŒã ã§ãšãŒãžã§ã³ãããã¬ãŒãã³ã°ããããã®ãã¹ããã©ã¯ãã£ã¹ãšããŠææ¡ãããŠããŸãã ãã¹ãã£ãããŒãã¢ã¯ã·ã§ã³ã¯ãéšå±ããéšå±ãžã®ç§»è¡ãäºæž¬äžèœã«ããŸãã
ã©ã³ãã ãããã¯ãŒã¯èžç
次ã®ç¶æ ãäºæž¬ããããšã¯ããã€ãºã®å€ããã¬ãã®åé¡ãæ¬è³ªçã«åãããããããäºæž¬ãšã©ãŒã®æ¬¡ã®é¢é£ããåå ãç¹å®ããŸããã
- èŠå 1 äºæž¬åã以åã«æ€èšããäŸããäžè¬åã«å€±æããå Žåãäºæž¬ãšã©ãŒã¯å€§ãããªããŸãã æ°ããçµéšã¯ãé«ãäºæž¬ãšã©ãŒã«å¯Ÿå¿ããŸãã
- ãã¡ã¯ã¿ãŒ2 ã 確ççäºæž¬ç®æšã®ããã«ãäºæž¬ãšã©ãŒã¯é«ããªããŸãã
- èŠå 3 ã äºæž¬ã«å¿ èŠãªæ å ±ãäžè¶³ããŠããããããŸãã¯äºæž¬ã¢ãã«ã®ã¯ã©ã¹ãç®çé¢æ°ã®è€éãã«é©åããã«ã¯å¶éã倧ãããããããäºæž¬ãšã©ãŒã倧ãããªããŸãã
èŠå 1ãšèŠå 2ã¯ãã€ãºã®å€ããã¬ãã®åé¡ã«ã€ãªããã®ã«å¯ŸããèŠå 1ã¯çµéšã®æ°èŠæ§ãå®éåããããããšã©ãŒã®æçšãªåå ã§ãããšå€æããŸããã èŠå 2ããã³3ãåé¿ããããã«ã次ã®ç¶æ èªäœãèæ ®ããŠã次ã®ç¶æ ã§ã®ã©ã³ãã ã§åæåãããäžå®ã®ãã¥ãŒã©ã«ãããã¯ãŒã¯ã®çºè¡ã®äºæž¬ã«åºã¥ãæ°ããç 究ããŒãã¹ã§ããRNDãéçºããŸãã ã
çŽèŠ³ã¯ãäºæž¬ã¢ãã«ã¯åœŒå¥³ãèšç·Žãããç¶æ ãäºæž¬ããéã®èª€å·®ãå°ããããšã瀺åããŠããŸãã ç¹ã«ãã©ã³ãã ã«åæåããããã¥ãŒã©ã«ãããã¯ãŒã¯ã®çºè¡ã«é¢ãããšãŒãžã§ã³ãã®äºæž¬ã¯ããšãŒãžã§ã³ãã以åã«é »ç¹ã«äŒã£ãç¶æ ãããæ°ããç¶æ ã®æ¹ãæ£ç¢ºã§ã¯ãããŸããã åæäºæž¬åé¡ã䜿çšããå©ç¹ã¯ã決å®è«çïŒãã€ãã¹ãã¡ã¯ã¿ãŒ2ïŒã§ãããé¢æ°ã®ã¯ã©ã¹å ã§ãäºæž¬åãã¿ãŒã²ãããããã¯ãŒã¯ãšåãã¢ãŒããã¯ãã£ã®äºæž¬åïŒãã€ãã¹ãã¡ã¯ã¿ãŒ3ïŒãéžæã§ããããšã§ãã ããã«ããããã€ãºã®å€ããã¬ãã®RNDåé¡ã解æ¶ãããŸãã
2ã€ã®å ±é ¬ãããŒã«2ã€ã®å€ã䜿çšãããæãè¿ãããªã·ãŒã®æé©åã®åœ¢åŒã§ãããããã·ãŒããªã·ãŒæé©å ïŒ PPO ïŒãéããŠãç 究ããŒãã¹ãšå€éšå ±é ¬ãçµã¿åãããŸããã ããã«ãããããŸããŸãªå ±é ¬ã«ããŸããŸãªå²åŒãé©çšãããããšããœãŒããšéãšããœãŒãã®å ±é ¬ãçµã¿åããããããããšãã§ããŸãã ãã®ãããªè¿œå ã®æè»æ§ã«ãããç§ãã¡ã®æé«ã®ãšãŒãžã§ã³ãã¯ãã¢ã³ãã¹ãã®åŸ©venã®æåã®ã¬ãã«ã§24宀äž22宀ãèŠã€ããããšãå€ããæ®ãã®2宀ãèŠã€ããåŸã«æåã®ã¬ãã«ãééããããšããããŸãã åãæ¹æ³ã¯ãVentureããã³Gravitarã²ãŒã ã§èšé²çãªããã©ãŒãã³ã¹ã瀺ããŠããŸãã
以äžã®èŠèŠåã¯ããšãŒãžã§ã³ããæåã«ããŒããèŠã€ããã¢ã³ããºããªãã³ãžãšããœãŒãã®å éšå ±é ¬ã®ã°ã©ãã瀺ããŠããŸãã

æèœãªå®è£ ãéèŠã§ã
é©åãªã¢ã«ãŽãªãºã ãéžæããã«ã¯ããã€ãºã®å€ããã¬ãã®åé¡ã«å¯Ÿããæåæ§ãªã©ãäžè¬çãªèæ ®äºé ãèæ ®ããããšãéèŠã§ãã ããããåçŽãªã¢ã«ãŽãªãºã ã®éåžžã«å°ããªå€æŽãããã®æå¹æ§ã«åŒ·ã圱é¿ããããšãããããŸãããæåã®éšå±ããåºãããªããšãŒãžã§ã³ãããæåã®ã¬ãã«ãééãããšãŒãžã§ã³ããŸã§ã ãã¬ãŒãã³ã°ã«å®å®æ§ãè¿œå ããããã«ãç¹æ§ã®é£œåãåé¿ããäºæž¬å¯èœãªç¯å²ã«å éšå ±é ¬ããããããŸããã ãŸãããã°ãèŠã€ããŠä¿®æ£ãããã³ã«RNDã®æå¹æ§ãå€§å¹ ã«æ¹åãããŠããããšã«æ°ä»ããŸãã ïŒç§ãã¡ã®ãæ°ã«å ¥ãã«ã¯ãé åã®ã©ã³ãã ãªãŒãåãå«ãŸããŠãããããå€éšå ±é ¬ãéãšããœãŒããšèŠãªããããšããäºå®ã«ã€ãªãããŸãããããã¯ãå€éšäŸ¡å€é¢æ°ã«ã€ããŠèããåŸã«åããŠå®çŸããŸãããçãããåšææ§ã«èŠããŸããïŒã ãããã®è©³çŽ°ãä¿®æ£ããããšã¯ã以åã®äœæ¥ãšæŠå¿µçã«é¡äŒŒããã¢ã«ãŽãªãºã ã䜿çšããå Žåã§ããé«ãããã©ãŒãã³ã¹ãéæããããã®éèŠãªéšåã«ãªããŸããã ããããå¯èœãªéãåçŽãªã¢ã«ãŽãªãºã ãéžæããããšãæåã§ããçç±ã®1ã€ã§ãã
ä»åŸã®ä»äº
次ã®åéã®ãããªãç 究ãæäŸããŸãã
- ããŸããŸãªç 究æ¹æ³ã®å©ç¹ã®åæãšãããããçµã¿åãããæ°ããæ¹æ³ã®æ¢çŽ¢ã
- å ±é ¬ãªãã§å€ãã®ç°ãªãç°å¢ã§å¥œå¥å¿agentçãªãšãŒãžã§ã³ãããã¬ãŒãã³ã°ããå ±é ¬ã®ããã¿ãŒã²ããç°å¢ãžã®ç§»è¡ãåŠç¿ããŸãã
- é·æã«ãããå調ãœãªã¥ãŒã·ã§ã³ãå«ãã°ããŒãã«ã€ã³ããªãžã§ã³ã¹ã