å²ç¢ã®åè² ã§ã¯äººå·¥ç¥è½ï¼AIï¼ãåã£ããããã¼ã«ã¼ç¨ã®AIéçºã«ã¯ãå²ç¢ã«ã¯ãªããã¾ãã¾ãªèª²é¡ãåå¨ããã人éã®åããäºæ¸¬ä¸å¯è½ã ã¨ããç¹ããã®ã²ã¨ã¤ã ã
ãããããã¼ã«ã¼ãå·§ã¿ã«ãã¬ã¤ã§ããAIã®ç 究ã¯åå°ã§è¡ããã¦ãããããã¦ãã®ã»ã©ãããã®äººéã®ããã©ã¼ãã³ã¹ã«ããªããè¿ã¥ãããã¨ããã·ã¹ãã ã«é¢ããç 究è«æãçºè¡¨ãããã
å²ç¢ã§ä¸çãã£ã³ããªã³ã«åå©ããDeepMindï¼ãã£ã¼ããã¤ã³ãï¼ã®éçºã«ãåå ãã¦ãããã¤ã´ã£ããã»ã·ã«ã´ã¡ã¼ãå«ããã¦ãã´ã¡ã¼ã·ãã£ã»ã«ã¬ãã¸ã»ãã³ãã³ã®ç 究ãã¼ã ã¯ããä¸å®å ¨æ å ±ã²ã¼ã ã«ãããã»ã«ããã¬ã¤ããã®æ·±å±¤å¼·åå¦ç¿ï¼Deep Reinforcement Learning from Self-Play in Imperfect-Information Gamesï¼ãã¨ããè«æï¼PDFï¼ãå ¬éããã
ãã®ç 究ã§ã¯ãããããµã¹ã»ãã¼ã«ãã ãï¼ãã¼ã«ã¼ã®ä¸ç¨®ã§ãç±³å½ã®ã«ã¸ãã§ã¯ä¸è¬çï¼ã¨ãåç´åãããã¼ã«ã¼ãã«ããã¯ï¼Leducï¼ãããã¬ã¤ã§ããä¸é£ã®å¼·åã¢ã«ã´ãªãºã ãä½æãããã
ç 究ãã¼ã ã«ããã¨ããã®AIã¯æ¦ç¥ã«é¢ããäºåç¥èããªãã¦ãã²ã¼ã ãå¦ç¿ãããã¨ãã§ããã²ã¨ãã§æ¶ç©ºã®è©¦åãè¡ããã¨ã§ç¬å¦ãã¦ããã¨ããã
è«æã«ããã¨ãä½æãããããã¥ã¼ã©ã«ã»ãã£ã¯ãã£ã·ã£ã¹ã»ã»ã«ããã¬ã¤ï¼Neural Fictitious Self-Playï¼ãæ³ã¯ã深層強åå¦ç¿ã使ç¨ãã¦ããã²ã¼ã ã§ã®å¯¾æ¦çµé¨ããç´æ¥å¦ã¶ãã¨èª¬æãã¦ããããã¥ã¼ã©ã«ãããã¯ã¼ã¯ãæ´»ç¨ãã¤ã¤ãééãããå¦ç¿ãã¦ã²ã¼ã ã«åã¤æ¹æ³ãç·¨ã¿åºãã®ã ã
ç 究è ãã¡ã«ããã¨ãä½æããã¢ãã«ã¯ãã«ããã¯ã§ã¯ããã·ã¥åè¡¡ï¼ã»ãã®ãã¬ã¤ã¤ã¼ã®æ¦ç¥ãæä¸ã¨ããå ´åãã©ã®ãã¬ã¤ã¤ã¼ãèªåã®æ¦ç¥ãå¤æ´ãããã¨ã«ãã£ã¦ããé«ãå©å¾ãå¾ããã¨ãã§ããªãæ¦ç¥ï¼ãã·ãã¥ã¬ã¼ãã§ããããããµã¹ã»ãã¼ã«ãã ã§ããããã«è¿ãç¶æ ãå®ç¾ã§ããã¨ããã
è«æçè ã§ããç 究çã®ãã¤ã³ãªããæ°ã¯ãã¬ã¼ãã£ã¢ã³ãç´ã®è¨äºã§ãããã®ææ³ã¯ãæ¦ç¥ãæ±ããããå®ä¸çã®åé¡ã«ãé©ç¨ã§ããã¨èãã¦ãã¾ããã¨èªã£ã¦ããã
ãªãã2015å¹´4æã«ã¯ãã«ã¼ãã®ã¼ã¡ãã³å¤§å¦ãéçºããAIããåãã¦ãããµã¹ã»ãã¼ã«ãã ã®è©¦åã§äººéã¨å¯¾æ¦ãã¦ããï¼æ¥æ¬èªçè¨äºï¼ã
14æ¥éãããã¦è¡ããããã®è©¦åã§ã¯ã人éã73ä¸2,713ãã«åã£ã¦çµãã£ããã¡ãªã¿ã«äººéå´ã¨AIå´ãæããéé¡ã¯åè¨ã§çè«ä¸1å7,000ä¸ãã«ã«ä¸ã£ã¦ããã
AIãè¦æ¦ããã®ã¯ãæãéãä¸ãã¦ãã人éã¸ã®å¯¾å¿æ¹æ³ã ã£ãã人éã«ããè³ããäºæ¸¬ã§ããªãã¨ãããã¨ã¯ãAIãã²ã¼ã ãææ¡ã§ãã¦ããªãã¨ãããã¨ã¨åãã ããã ã
ã¾ãAIããæã®ãªãã«ããã«ã¼ãããªãã²ã¼ã ã«å½±é¿ãä¸ããããç解ã§ããªãã¨ããç¹ãã人éã«ã¨ã£ã¦ã®ã¢ãã´ã¡ã³ãã¼ã¸ã¨ãªã£ããã¤ã¾ã人éã«ã¨ã£ã¦ãã³ã³ãã¥ã¼ã¿ã¼ããã°ã©ã ãå¼±ãæã§ãã©ããããã¦ãããã©ãããè¦åããã®ã¯ç°¡åã ã£ãã®ã ã
TEXT BY MATT BURGESS
TRANSLATION BY MIHO AMANO, HIROKO GOHARA/GALILEO