ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¨å²ç¢
çµæ§åã«ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ãç¢ã®æèã¨ã³ã¸ã³ã«ä½¿ããã®ã§ã¯ãªããã¨ã»ãã®å°ãã ã試ã¿ã¦ã¿ãããã®æã¯ç¹ã«ææããªãçµãã£ãããåãããã«èãã¦ãã人ã¯ãã¦ãç§å¦ã¯é²ãã§ããã 以ä¸ã®è«æã¯ã深層ç³ã¿è¾¼ã¿ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¨ã¢ã³ãã«ã«ãæ¨æ¢ç´¢ãçµã¿åããã¦è¦ãã¨ããå ±åã§ããã
BETTER COMPUTER GO PLAYER WITH NEURAL NETWORK AND LONG-TERM PREDICTION
ããããç§ãªãéãã¢ããã¼ããåããªã¨æã£ããç§ãªãä¸è¨ã®ããã«ããã¨æãã
Tianã¨Zhuã®ããæ¹ã¯å±é¢ãå ¥åã¨ãã¦ä»¥éã®æãåºåã¨ãã¦å¦ç¿ããã¦ããã次ã®ä¸æã ãã ã¨ãã¤ã¢ã¹ããããããã¨3æå ã¾ã§ãåºåã¨ããã®ãç¹å¾´ã§ãããã¾ããå ¥åå¤ã¨ãã¦ã¯ç½ç³ã»é»ç³ã®é ç½®ãã³ã¦ã ãã§ãªããå¼å¸ç¹ãï¼ã®é£ã»ï¼ã®é£ã»ï¼ä»¥ä¸ã®é£ã®ä½ç½®ãç³ã®ç½®ããã¦ããæéã天å ããã®è·é¢ãç½®ãç³ã®æ°ã辺ã®ä½ç½®ãªã©ãå ¥åã¨ãã¦ããã19x19ã®ç»åãåè¨25æå ¥åã¨ãããããã¦ãå¦ç¿ããããã¥ã¼ã©ã«ãããã¯ã¼ã¯ãã¢ã³ãã«ã«ãæ¨æ¢ç´¢ã®ãã¬ã¤ã¢ã¦ãé¢æ°ã«ä½¿ã£ã¦ããããã ã
ãããããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¯æãåºåããã®ã§ã¯ãªãã¦ãè©ä¾¡é¢æ°ã«ä½¿ãã¹ãã ã¨ç§ã¯æããæ©æ¢°ã«ç¢ããããã¨ãã®é£ããã¯ã²ã¼ã æ¨ã®æã®æ°ãããã ããè©ä¾¡é¢æ°ãä½ãã®ãæ大ã®å°é£ã ããããã¥ã¼ã©ã«ãããã¯ã¼ã¯ãªããããå®ç¾ã§ããã¨äºæ³ãã¦ããã
åºåå¤ã¯ï¼éããèãããããããå±é¢ã§èªåãåªå¢ãå¦ãã®äºå¤ãè¿ãããããã¯ç¹ãçµå±æã«èªé£ï¼ä¸å½ã«ã¼ã«ã«ããã¦ï¼ã§ãã確çãåºåãããã©ã¡ããå ¥åå¤ã®å±é¢ã®å¾ã«ã²ã¼ã ãç¶ãã¦çµå±ã©ããªã£ããã¨ããã®ãæ師ãã¼ã¿ã¨ããã°ããã¨æãã®ã ã
åè ã§çµå±æã®å°ã®å·®ãåºåããªãã®ã¯ãã´ã¼ã«ã大差ãã¤ãããã¨ã§ã¯ãªãã¦åã¤ãã¨ã ããã§ãããã¢ã³ãã«ã«ãæ¨æ¢ç´¢ã§ããå½åã¯ãã¬ã¤ã¢ã¦ãæã®å°ã®å·®ã®æå¾ å¤ãé«ããããã«ãã¦ãããããããããã ã¨ã¢ã°ã¬ãã·ãã«ãªããããããããåç®å·®ã§åã¤ãããªæãçãããã®ãæå¹ã ã¨æãããã®ã§åæã®ã¿ãåºåããããã
å¾è ãåºåå¤ã¨ããããæå³ã¯ãç³ãç¤é¢ã«ããã¨ãç³ã®å¨ãã«éåå ´ã®ãããªãã®ãã§ããããã«æããã¨æãã®ã ããã©ããããæ°å¤åãããã¨ããåæ©ã§ããããããã§ãããå²ç¢ã¨ããã²ã¼ã ãçºå±ããã¯ãã ããã ãå¾è ã®å ´åã¯ä¸æ¼ãåã¡ã®å ´åã®ãã¼ã¿ã使ããªãã¨ããåé¡ãããããããªãã«è³¢ãæ¢åã®æèã¨ã³ã¸ã³ã«çµå±ã¾ã§æããããè¯ãã®ã ãããããã¡ãã®å ´åãè©ä¾¡é¢æ°ã¨ãã¦ä½¿ãã®ã¯ãå°ã®æå¾ å¤ã足ãåããã¦å·®ãè¦ãåæã ããããªã®ã§ãæèã¨ã³ã¸ã³ã¨ãã¦ã¯åè ã®æ¹ãç´æ¥çãªã®ã§åè ããããªãå¦ç¿ãããæ¹ãå¼·ããªããããããªãããåè ãå¾ãéç¨ã§ã©ã®éå¾è ãå¿ è¦ã¨ãªãã ããã¨ããä»®å®ãç½®ãã¦å¾è ãå¦ç¿ãããã¨ããããæ¹ãèããããã
è©ä¾¡é¢æ°ãä½ããããç¢ã§ãã¢ã«ãã¡ãã¼ã¿æ³ã復権ããã¨æãã®ã§ããã¢ã³ãã«ã«ãæ¨æ¢ç´¢ã¯çè«çæ ¹æ ãæªãããããªãã¨ãªã好ãã§ãªãã®ã ãããããã¢ã«ãã¡ãã¼ã¿æ³ã§ãã©ã®æãåãåããã¨ãã©ã®æã¯æ·±ãã¾ã§ä¼¸ã°ããã¨ãã£ãåé¡ãä»ã®è¨ç®æ©ã®ãã¯ã¼ã§ã¯ããããçããã®ã§çµå±ã¢ã³ãã«ã«ãæ¨æ¢ç´¢ã®ããã«ãªã£ã¦ãã¾ãã®ã ãããï¼
çµå±æã«èªé£ã§ãã確çã使ãå ´åã«ã¯ç¢ºåº¦ã¨ããã®ãã©ã®æã伸ã°ããã¨ããã®ã«ã¯ä½¿ãããããããªããåæã®äºå¤ãåºåããå ´åã¯[0, 1]ã®åºéå ã®å¤ãè¿ããã¨ã«ãªãã®ã§ã0.5ã«è¿ãå¤ãè¿ããããªå ´åã¯ç¢ºåº¦ãä½ãã®ã§ãã£ã¨æ¨ã伸ã°ãã¨ããããã«ãããç¹ãèªåã®å°ãã©ããã®ç¢ºçãè¿ãå ´åã«ã¯ãç¤å ¨ä½ã«å¯¾ãã¦0.5ããã®å·®ã®ï¼ä¹å¹³åå¤ã確度ã¨ããã¨ããããã«ãªãã®ã ããã
æãåãã®ã¯é£ããããããããããåã£ã¦ãã¾ã£ãæã«æåæããããããããªãããã ãå°æ£ã§ãæè¿ã¯æåãã¯æµè¡ããªãããããç¢ã«ã¯æ¬å½ã«ä½ã®æå³ããªãæã¨ããã®ãããã®ã§ãèªæãªã®ããå°éã«åã£ã¦ãããããªãã®ã ãããã
Tianã¨Zhuã®ããæ¹ã¯å ¥åå¤ãå¤ãããããã«æããããã§ãå è¡ç 究ããã¯æ¸ããã¦ãããããããã®è¾ºã¯è²ã 試ãã¦å¹æããªãã®ã¯åã£ã¦ãããããªãã®ããªã
æ¸ããã®ã§ãªãã¦å ãããªããç§ã¯ãã¥ã¼ã©ã«ãããã¯ã¼ã¯ã¿ãããªè¤éãªææ³ã§ãªãã¦ç¢ºå®ã«åæ¢ãããããªåå§çãªã¢ã«ã´ãªãºã ã§å¾ãæ»æ´»å¤å®ã®çµæãå ¥ãããããã¨æããã¢ã«ã´ãªãºããã¯ãªè¨ç®ã¯ã³ã³ãã¥ã¼ã¿ã¯å¾æãªã®ã§ãããã¯æãªããã®ææ³ã§ãããã¦ãåãè¦ãããããªãã¨ã ããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã«ãããããããªè¨è¨ã«ç§ãªããããã¾ããåºç¤ã»ä¸ç¤ã»çµç¤ã ã¨æããã«ä½¿ã£ã¦ããè³ãéãã®ã§ãåããã¥ã¼ã©ã«ãããã¯ã¼ã¯ã使ããããªãã¦ãã¨ã¯ããªãããªãã¤ã¾ããææ°ãå¤æ°ã«å ããã¨ãããã¨ã«ãªããç¹ã«åºç¤ã¯ã©ã¡ãã®ç³ãæã¤ãã§æ¦ç¥ãå¤ããã®ã§ãããå¤æ°ã«å ¥ãã®ãï¼ã£ã¦ãã£ã¦ããã¨å¤æ°ãç¡éã«å¢ãã¦ãã£ã¦ãã¾ãã¡ã ãªã
ãã¡ãããæ¬å½ã«è¯ãã¯ç¥ããªãã試ããããããããè³¢ã人ãæè¦ã交æãåããªããå ¨åã§æ½¤æ²¢ãªäºç®ã®ä¸ã«ãã£ã¦ãããã®ã«å人ã§åæ¦ãããã«ãå俵ã«ããç«ã¦ãªããã¡ããããå®å ¨ä½ã«ãªãããããã°ã誰ãç§ãéãå ¥ãã¦ããããããã¦ãããªãã ããããå®ç¸¾ãªãããé£ããããªãéã«ããªããããä½ããããã«ã¯å®ç¸¾ãå¿ è¦ã ããå®ç¸¾ãä½ãã«ã¯ãã®ä½ããããªããã°ãªããªãã¨ããåé¡ã¯è§£æ±ºã§ããªããã®ã ãããããããå®ç¾ã§ããã人é¡ã¯æ大å¤æ°ã®æ大幸ç¦ãæã«å ¥ããããæ°ãããã®ã ã
追è¨ã2016-03-10
GoogleãNatureã«è«æåºãåã«ãã®è¨äºæ¸ãã¦ãã¦è¯ãã£ãâ¦â¦åããããªãã¨ãã¦ããâ¦â¦