{"id":10006281,"date":"2022-08-22T20:53:29","date_gmt":"2022-08-22T12:53:29","guid":{"rendered":"https:\/\/ityet.com\/?p=10006281"},"modified":"2022-08-22T20:53:29","modified_gmt":"2022-08-22T12:53:29","slug":"%e6%b7%b1%e5%ba%a6%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e6%8a%80%e6%9c%af%e6%a6%82%e8%bf%b0","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10006281","title":{"rendered":"\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6280\u672f\u6982\u8ff0"},"content":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u5728\u672c\u6587\u4e2d\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6280\u672f\uff0c\u5c06\u5f3a\u5316\u5b66\u4e60\u5206\u6210\u4e09\u5927\u7c7b\uff08value-based\u7b97\u6cd5\u3001policy-based\u7b97\u6cd5\u53ca\u7ed3\u5408\u4e24\u8005\u7684AC\u7b97\u6cd5\uff09\u6765\u8fdb\u884c\u4ecb\u7ecd\u3002\u9996\u5148\uff0c\u4ece\u6570\u5b66\u7406\u8bba\u89d2\u5ea6\u4ecb\u7ecd\u4e86\u5f3a\u5316\u5b66\u4e60\uff1b\u63a5\u7740\uff0c\u4ece\u4e0d\u540c\u9002\u7528\u65b9\u5411\u5bf9\u4e24\u7c7b\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u8fdb\u884c\u4ecb\u7ecd\uff1a\u57fa\u4e8e\u503c\u51fd\u6570\uff08Value-based\uff09\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5DQN\u548c\u57fa\u4e8e\u7b56\u7565\uff08Policy-based\uff09\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5PG\u3002\u6700\u540e\uff0c\u4ecb\u7ecd\u76ee\u524d\u5e94\u7528\u5e7f\u6cdb\u7684\u7ed3\u5408\u524d\u4e24\u4e2a\u7b97\u6cd5\u7684\u7ed3\u5408\u7269AC\uff08Actor-Critic\uff09\u7b97\u6cd5\u3002\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u4ecb\u7ecd\u5f3a\u5316\u5b66\u4e60\u4e3b\u8981\u7528\u6765\u5b66\u4e60\u4e00\u79cd\u6700\u5927\u5316\u667a\u80fd\u4f53\u4e0e\u73af\u5883\u4ea4\u4e92\u83b7\u5f97\u7684\u957f\u671f\u5956\u60e9\u503c\u7684\u7b56\u7565\uff0c\u5176\u5e38\u7528\u6765\u5904\u7406\u72b6\u6001\u7a7a\u95f4\u548c\u52a8\u4f5c\u7a7a\u95f4\u5c0f\u7684\u4efb\u52a1\uff0c\u5728\u5982\u4eca\u5927\u6570\u636e\u548c\u6df1\u5ea6\u5b66\u4e60\u5feb\u901f\u53d1\u5c55\u7684\u65f6\u4ee3\u4e0b\uff0c\u9488\u5bf9\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u65e0\u6cd5\u89e3\u51b3\u9ad8\u7ef4\u6570\u636e\u8f93\u5165\u7684\u95ee\u9898\uff0c2013\u5e74Mnih V\u7b49\u4eba\u9996\u6b21\u5c06\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff08Convolutional Neural Networks\uff0cCNN\uff091[3]\u5f15\u5165\u5f3a\u5316\u5b66\u4e60\u4e2d\uff0c\u63d0\u51fa\u4e86DQN\uff08Deep Q Learning Network\uff094\u7b97\u6cd5\uff0c\u81f3\u6b64\u56fd\u9645\u4e0a\u4fbf\u5f00\u59cb\u4e86\u5bf9\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\uff08Deep Reinforcement Learning\uff0cDRL\uff09\u7684\u79d1\u7814\u5de5\u4f5c\u3002\u9664\u6b64\u4e4b\u5916\uff0c\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u9886\u57df\u4e2d\u4e00\u4e2a\u91cc\u7a0b\u724c\u4e8b\u4ef6\u662f2016\u5e74\u7684AlphaGo \u5bf9\u6218\u674e\u4e16\u77f3\u7684\u56f4\u68cb\u4e16\u7eaa\u5927\u62186\uff0c\u8c37\u6b4c\u65d7\u4e0b\u7684\u4eba\u5de5\u667a\u80fd\u56e2\u961fDeepMind \u57fa\u4e8e\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u5f00\u53d1\u51fa\u7684\u56f4\u68cb\u7a0b\u5e8f AlphaGo\u51fb\u8d25\u4e86\u4e16\u754c\u9876\u7ea7\u56f4\u68cb\u5927\u5e08\u674e\u4e16\u77f3\uff0c\u9707\u60ca\u4e86\u4e16\u754c\uff0c\u4e5f\u56e0\u6b64\u62c9\u5f00\u4e86\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u4ece\u5b66\u672f\u754c\u8d70\u5411\u5927\u4f17\u8ba4\u77e5\u7684\u5e37\u5e55\u3002\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7ed3\u5408\u4e86\u6df1\u5ea6\u5b66\u4e60[8]\uff08Deep Learning\uff0cDL\uff09\u7684\u7279\u5f81\u63d0\u53d6\u80fd\u529b\u548c\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning\uff0cRL\uff09\u7684\u51b3\u7b56\u80fd\u529b[9]\uff0c\u53ef\u4ee5\u76f4\u63a5\u6839\u636e\u8f93\u5165\u7684\u591a\u7ef4\u6570\u636e\u505a\u51fa\u6700\u4f18\u51b3\u7b56\u8f93\u51fa\uff0c\u662f\u4e00\u79cd\u7aef\u5bf9\u7aef\uff08end-to-end\uff09\u7684\u51b3\u7b56\u63a7\u5236\u7cfb\u7edf\uff0c\u5e7f\u6cdb\u5e94\u7528\u4e8e\u52a8\u6001\u51b3\u7b56\u3001\u5b9e\u65f6\u9884\u6d4b\u3001\u4eff\u771f\u6a21\u62df\u3001\u6e38\u620f\u535a\u5f08\u7b49\u9886\u57df\uff0c\u5176\u901a\u8fc7\u4e0e\u73af\u5883\u4e0d\u65ad\u5730\u8fdb\u884c\u5b9e\u65f6\u4ea4\u4e92\uff0c\u5c06\u73af\u5883\u4fe1\u606f\u4f5c\u4e3a\u8f93\u5165\u6765\u83b7\u53d6\u5931\u8d25\u6216\u6210\u529f\u7684\u7ecf\u9a8c\u6765\u66f4\u65b0\u51b3\u7b56\u7f51\u7edc\u7684\u53c2\u6570\uff0c\u4ece\u800c\u5b66\u4e60\u5230\u6700\u4f18\u51b3\u7b56\u3002\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u5982\u4e0b\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358688\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u4e0a\u56fe\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u4e2d\uff0c\u667a\u80fd\u4f53\u4e0e\u73af\u5883\u8fdb\u884c\u4ea4\u4e92\uff0c\u667a\u80fd\u4f53\u901a\u8fc7\u6df1\u5ea6\u5b66\u4e60\u5bf9\u73af\u5883\u72b6\u6001\u8fdb\u884c\u7279\u5f81\u63d0\u53d6\uff0c\u5c06\u7ed3\u679c\u4f20\u9012\u7ed9\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c\u51b3\u7b56\u5e76\u6267\u884c\u52a8\u4f5c\uff0c\u6267\u884c\u5b8c\u52a8\u4f5c\u540e\u5f97\u5230\u73af\u5883\u53cd\u9988\u7684\u65b0\u72b6\u6001\u548c\u5956\u60e9\u8fdb\u800c\u66f4\u65b0\u51b3\u7b56\u7b97\u6cd5\u3002\u6b64\u8fc7\u7a0b\u53cd\u590d\u8fed\u4ee3\uff0c\u6700\u7ec8\u4f7f\u667a\u80fd\u4f53\u5b66\u5230\u83b7\u5f97\u6700\u5927\u957f\u671f\u5956\u60e9\u503c\u7684\u7b56\u7565\u3002\u00a0\u2590 \u00a0\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7684\u6570\u5b66\u6a21\u578b\u00a0\u5f3a\u5316\u5b66\u4e60[10]\u662f\u4e00\u79cd\u51b3\u7b56\u7cfb\u7edf\uff0c\u5176\u57fa\u672c\u601d\u60f3\u662f\u901a\u8fc7\u4e0e\u73af\u5883\u8fdb\u884c\u5b9e\u65f6\u4ea4\u4e92\uff0c\u5728\u4e0d\u65ad\u5730\u5931\u8d25\u4e0e\u6210\u529f\u7684\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u7ecf\u9a8c\uff0c\u6700\u5927\u5316\u667a\u80fd\u4f53\uff08Agent\uff09\u4ece\u73af\u5883\u4e2d\u83b7\u5f97\u7684\u7d2f\u8ba1\u5956\u52b1\u503c\uff0c\u6700\u7ec8\u4f7f\u5f97\u667a\u80fd\u4f53\u5b66\u5230\u6700\u4f18\u7b56\u7565\uff08Policy\uff09\uff0c\u5176\u539f\u7406\u8fc7\u7a0b\u5982\u4e0b\u56fe\u6240\u793a\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358689\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u4e0a\u56fe\u5f3a\u5316\u5b66\u4e60\u57fa\u672c\u6a21\u578b\u4e2d\uff0c\u667a\u80fd\u4f53\uff08Agent\uff09\u662f\u5f3a\u5316\u5b66\u4e60\u7684\u52a8\u4f5c\u5b9e\u4f53\uff0c\u667a\u80fd\u4f53\u5728\u5f53\u524d\u72b6\u6001\u4e0b\u6839\u636e\u52a8\u4f5c\u9009\u62e9\u7b56\u7565\u6267\u884c\u52a8\u4f5c\uff0c\u6267\u884c\u8be5\u52a8\u4f5c\u540e\u5176\u5f97\u5230\u73af\u5883\u53cd\u9988\u5956\u60e9\u503c \u548c\u4e0b\u4e00\u72b6\u6001\uff0c\u5e76\u6839\u636e\u53cd\u9988\u4fe1\u606f\u66f4\u65b0\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u53c2\u6570\uff0c\u6b64\u8fc7\u7a0b\u4f1a\u53cd\u590d\u5faa\u73af\u4e0b\u53bb\uff0c\u6700\u7ec8\u667a\u80fd\u4f53\u5b66\u4e60\u5230\u5b8c\u6210\u76ee\u6807\u4efb\u52a1\u7684\u6700\u4f18\u7b56\u7565\u3002\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff08Markov Decision Process\uff0cMDP\uff09[11]\u662f\u5f3a\u5316\u5b66\u4e60\u7406\u8bba\u7684\u6570\u5b66\u63cf\u8ff0\uff0c\u5176\u53ef\u4ee5\u5c06\u5f3a\u5316\u5b66\u4e60\u95ee\u9898\u4ee5\u6982\u7387\u8bba\u7684\u5f62\u5f0f\u8868\u793a\u51fa\u6765\u3002\u5728MDP\u4e2d\u53ef\u5c06\u5f3a\u5316\u5b66\u4e60\u4ee5\u4e00\u4e2a\u5305\u542b\u56db\u4e2a\u5c5e\u6027\u7684\u5143\u7ec4\u8868\u793a\uff1a{S,A,P,R}\uff0c\u5176\u4e2d\uff1aS \u548c A \u5206\u522b\u8868\u793a\u667a\u80fd\u4f53\u7684\u73af\u5883\u72b6\u6001\u96c6\u548c\u667a\u80fd\u4f53\u53ef\u9009\u62e9\u7684\u52a8\u4f5c\u96c6\uff1b<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358690\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358691\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358692\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358693\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358694\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u2590 \u00a0\u57fa\u4e8e\u503c\u51fd\u6570\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u00a0\u57fa\u4e8e\u503c\u51fd\u6570\uff08Value based\uff09\u7684\u5b66\u4e60\u65b9\u6cd5\u662f\u4e00\u79cd\u6c42\u89e3\u6700\u4f18\u503c\u51fd\u6570\u4ece\u800c\u83b7\u53d6\u6700\u4f18\u7b56\u7565\u7684\u65b9\u6cd5\u3002\u503c\u51fd\u6570\u8f93\u51fa\u7684\u6700\u4f18\u52a8\u4f5c\u6700\u7ec8\u4f1a\u6536\u655b\u5230\u786e\u5b9a\u7684\u52a8\u4f5c\uff0c\u4ece\u800c\u5b66\u4e60\u5230\u4e00\u79cd\u786e\u5b9a\u6027\u7b56\u7565\u3002\u5f53\u5728\u52a8\u4f5c\u7a7a\u95f4\u8fde\u7eed\u7684\u60c5\u51b5\u4e0b\u8be5\u65b9\u6cd5\u5b58\u5728\u7ef4\u5ea6\u707e\u96be\u3001\u8ba1\u7b97\u4ee3\u4ef7\u5927\u95ee\u9898\uff0c\u867d\u7136\u53ef\u4ee5\u5c06\u52a8\u4f5c\u7a7a\u95f4\u79bb\u6563\u5316\u5904\u7406\uff0c\u4f46\u79bb\u6563\u95f4\u8ddd\u4e0d\u6613\u786e\u5b9a\uff0c\u8fc7\u5927\u4f1a\u5bfc\u81f4\u7b97\u6cd5\u53d6\u4e0d\u5230\u6700\u4f18\uff0c\u8fc7\u5c0f\u4f1a\u4f7f\u5f97\u52a8\u4f5c\u7a7a\u95f4\u8fc7\u5927\uff0c\u5f71\u54cd\u7b97\u6cd5\u901f\u5ea6\u3002\u56e0\u6b64\u8be5\u65b9\u6cd5\u5e38\u88ab\u7528\u4e8e\u79bb\u6563\u52a8\u4f5c\u7a7a\u95f4\u4e0b\u52a8\u4f5c\u7b56\u7565\u4e3a\u786e\u5b9a\u6027\u7684\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u3002\u00a0\u57fa\u4e8e\u503c\u51fd\u6570\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u5229\u7528CNN\u6765\u903c\u8fd1\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u7684\u52a8\u4f5c\u503c\u51fd\u6570\uff0c\u4ee3\u8868\u7b97\u6cd5\u5c31\u662fDQN\u7b97\u6cd5\uff0c DQN\u7b97\u6cd5\u6846\u67b6\u5982\u4e0b\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358695\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358696\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358697\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u2590 \u00a0\u57fa\u4e8e\u7b56\u7565\u68af\u5ea6\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u00a0\u57fa\u4e8e\u7b56\u7565\u68af\u5ea6\uff08Policy based\uff09\u7684\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u662f\u4e00\u79cd\u6700\u5927\u5316\u7b56\u7565\u76ee\u6807\u51fd\u6570\u4ece\u800c\u83b7\u53d6\u6700\u4f18\u7b56\u7565\u7684\u65b9\u6cd5\uff0c\u5176\u76f8\u6bd4Value based\u65b9\u6cd5\u533a\u522b\u5728\u4e8e\uff1a\u53ef\u4ee5\u5b66\u4e60\u5230\u4e00\u79cd\u6700\u4f18\u5316\u968f\u673a\u7b56\u7565\u3002Policy based\u65b9\u6cd5\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u76f4\u63a5\u5b66\u4e60\u7b56\u7565\u51fd\u6570\uff0c\u968f\u7740\u7b56\u7565\u68af\u5ea6\u65b9\u5411\u4f18\u5316\u7b56\u7565\u51fd\u6570\u53c2\u6570\uff0c\u4f7f\u5f97\u7b56\u7565\u76ee\u6807\u51fd\u6570\u6700\u5927\u5316\uff0c\u6700\u7ec8\u7b56\u7565\u8f93\u51fa\u6700\u4f18\u52a8\u4f5c\u5206\u5e03\u3002\u7b56\u7565\u4ee5\u4e00\u5b9a\u6982\u7387\u8f93\u51fa\u52a8\u4f5c\uff0c\u6bcf\u6b21\u7ed3\u679c\u5f88\u53ef\u80fd\u4e0d\u4e00\u6837\uff0c\u6545\u4e0d\u9002\u5408\u5e94\u7528\u4e8e\u50cfCR\u9891\u8c31\u534f\u540c\u8fd9\u7c7b\u52a8\u4f5c\u7b56\u7565\u4e3a\u786e\u5b9a\u6027\u7684\u95ee\u9898\u4e2d\u3002Policy based\u65b9\u6cd5\u5bb9\u6613\u6536\u655b\u5230\u5c40\u90e8\u6781\u503c\uff0c\u800cValue based\u63a2\u7d22\u80fd\u529b\u5f3a\u53ef\u4ee5\u627e\u5230\u6700\u4f73\u503c\u51fd\u6570\u3002\u00a0\u4e0a\u8ff0\u4e2d\u7684\u7b56\u7565\u76ee\u6807\u51fd\u6570\u7528\u6765\u8861\u91cf\u7b56\u7565\u7684\u6027\u80fd\u3002\u5b9a\u4e49\u5982\u4e0b\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358698\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><\/p>\n<p><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358699\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>\u2590 \u00a0AC\u7b97\u6cd5\u00a0R=\/Sigma^{T-1}_{t=1}{r_t}\u00a0AC\u7b97\u6cd5\u6846\u67b6\u88ab\u5e7f\u6cdb\u5e94\u7528\u4e8e\u5b9e\u9645\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u4e2d\uff0c\u8be5\u6846\u67b6\u96c6\u6210\u4e86\u503c\u51fd\u6570\u4f30\u8ba1\u7b97\u6cd5\u548c\u7b56\u7565\u641c\u7d22\u7b97\u6cd5\uff0c\u662f\u89e3\u51b3\u5b9e\u9645\u95ee\u9898\u65f6\u6700\u5e38\u8003\u8651\u7684\u6846\u67b6\u3002\u4f17\u6240\u5468\u77e5\u7684alphago\u4fbf\u7528\u4e86AC\u6846\u67b6\u3002\u800c\u4e14\u5728\u5f3a\u5316\u5b66\u4e60\u9886\u57df\u6700\u53d7\u6b22\u8fce\u7684A3C\u7b97\u6cd5\uff0cDDPG\u7b97\u6cd5\uff0cPPO\u7b97\u6cd5\u7b49\u90fd\u662fAC\u6846\u67b6\u3002Actor-Critic\u7b97\u6cd5\u5206\u4e3a\u4e24\u90e8\u5206\uff0cActor\u7684\u524d\u8eab\u662fpolicy gradient\uff0cpolicy gradient\u53ef\u4ee5\u8f7b\u677e\u5730\u5728\u8fde\u7eed\u52a8\u4f5c\u7a7a\u95f4\u5185\u9009\u62e9\u5408\u9002\u7684\u52a8\u4f5c\uff08value-based\u7684Q-learning\u53ea\u80fd\u89e3\u51b3\u79bb\u6563\u52a8\u4f5c\u7a7a\u95f4\u7684\u95ee\u9898\uff09\u3002\u4f46\u662f\u53c8\u56e0\u4e3aActor\u662f\u57fa\u4e8e\u4e00\u4e2aepisode\u7684return\u6765\u8fdb\u884c\u66f4\u65b0\u7684\uff0c\u6240\u4ee5\u5b66\u4e60\u6548\u7387\u6bd4\u8f83\u6162\u3002\u8fd9\u65f6\u5019\u6211\u4eec\u53d1\u73b0\u4f7f\u7528\u4e00\u4e2avalue-based\u7684\u7b97\u6cd5\u4f5c\u4e3aCritic\u5c31\u53ef\u4ee5\u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u65b9\u6cd5\u5b9e\u73b0\u5355\u6b65\u66f4\u65b0\uff0c\u8fd9\u5176\u5b9e\u53ef\u4ee5\u770b\u505a\u662f\u62ff\u504f\u5dee\u6362\u65b9\u5dee\uff0c\u4f7f\u5f97\u65b9\u5dee\u53d8\u5c0f\u3002\u8fd9\u6837\u4e24\u79cd\u7b97\u6cd5\u76f8\u4e92\u8865\u5145\u7ed3\u5408\u5c31\u5f62\u6210\u4e86Actor-Critic\u7b97\u6cd5\u3002\u6846\u67b6\u5982\u4e0b\uff1a<br \/><img decoding=\"async\" referrerpolicy=\"no-referrer\" src=\"https:\/\/segmentfault.com\/img\/remote\/1460000042358700\" alt=\"\u56fe\u7247\" title=\"\u56fe\u7247\" loading=\"lazy\"><br \/>Actor \u57fa\u4e8e\u6982\u7387\u5206\u5e03\u9009\u62e9\u884c\u4e3a, Critic \u57fa\u4e8e Actor \u751f\u6210\u7684\u884c\u4e3a\u8bc4\u5224\u5f97\u5206, Actor \u518d\u6839\u636e Critic \u7684\u8bc4\u5206\u4fee\u6539\u9009\u884c\u4e3a\u7684\u6982\u7387\u3002\u4e0b\u9762\u5206\u6790\u4f18\u7f3a\u70b9\uff1a\u00a0\u4f18\u70b9\uff1a\u53ef\u4ee5\u8fdb\u884c\u5355\u6b65\u66f4\u65b0\uff0c\u4e0d\u9700\u8981\u8dd1\u5b8c\u4e00\u4e2aepisode\u518d\u66f4\u65b0\u7f51\u7edc\u53c2\u6570\uff0c\u76f8\u8f83\u4e8e\u4f20\u7edf\u7684PG\u66f4\u65b0\u66f4\u5feb\u3002\u4f20\u7edfPG\u5bf9\u4ef7\u503c\u7684\u4f30\u8ba1\u867d\u7136\u662f\u65e0\u504f\u7684\uff0c\u4f46\u65b9\u5dee\u8f83\u5927\uff0cAC\u65b9\u6cd5\u727a\u7272\u4e86\u4e00\u70b9\u504f\u5dee\uff0c\u4f46\u80fd\u591f\u6709\u6548\u964d\u4f4e\u65b9\u5dee\uff1b\u00a0\u7f3a\u70b9\uff1aActor\u7684\u884c\u4e3a\u53d6\u51b3\u4e8e Critic \u7684Value\uff0c\u4f46\u662f\u56e0\u4e3a Critic\u672c\u8eab\u5c31\u5f88\u96be\u6536\u655b\u548cactor\u4e00\u8d77\u66f4\u65b0\u7684\u8bdd\u5c31\u66f4\u96be\u6536\u655b\u4e86\u3002\uff08\u4e3a\u4e86\u89e3\u51b3\u6536\u655b\u95ee\u9898\uff0c Deepmind \u63d0\u51fa\u4e86 Actor Critic \u5347\u7ea7\u7248\u00a0Deep Deterministic Policy Gradient\uff0c\u540e\u8005\u878d\u5408\u4e86 DQN \u7684\u4e00\u4e9b trick, \u89e3\u51b3\u4e86\u6536\u655b\u96be\u7684\u95ee\u9898\uff09\u3002\u00a0\u53c2\u8003\u6587\u732eKetkar N . Convolutional Neural Networks[J]. 2017.Aghdam H H , Heravi E J . Convolutional Neural Networks[M]\/\/ Guide to Convolutional Neural Networks. Springer International Publishing, 2017.Gu, Jiuxiang, Wang, et al. Recent advances in convolutional neural networks[J]. PATTERN RECOGNITION, 2018.MINH V, KAVUKCUOGLU K, SILVER D, et al.Playing atari with deep reinforcement learning[J].Computer Science, 2013, 1-9.MNIH V,KAVUKCUOGLU K,SILVER D,et al.Human-level control through deep reinforcement learning[J].Nature,2015,518(7540):529-533.\u66f9\u8a89\u680a. \u4eceAlphaGO\u6218\u80dc\u674e\u4e16\u77f3\u7aa5\u63a2\u4eba\u5de5\u667a\u80fd\u53d1\u5c55\u65b9\u5411[J]. \u7535\u8111\u8ff7, 2018, 115(12):188.\u5218\u7ecd\u6850. \u4eceAlphaGo\u5b8c\u80dc\u674e\u4e16\u77f3\u770b\u4eba\u5de5\u667a\u80fd\u4e0e\u4eba\u7c7b\u53d1\u5c55[J]. \u79d1\u5b66\u5bb6, 2016(16).Lecun Y, Bengio Y, Hinton G. Deep learning.[J]. 2015, 521(7553):436.\u8d75\u51ac\u658c,\u90b5\u5764,\u6731\u5706\u6052,\u674e\u680b,\u9648\u4e9a\u5189,\u738b\u6d77\u6d9b,\u5218\u5fb7\u8363,\u5468\u5f64,\u738b\u6210\u7ea2.\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u7efc\u8ff0:\u517c\u8bba\u8ba1\u7b97\u673a\u56f4\u68cb\u7684\u53d1\u5c55[J].\u63a7\u5236\u7406\u8bba\u4e0e\u5e94\u7528,2016,33(06):701-717.9\u9648\u5b66\u677e, \u6768\u5b9c\u6c11. \u5f3a\u5316\u5b66\u4e60\u7814\u7a76\u7efc\u8ff0[J]. \u8ba1\u7b97\u673a\u5e94\u7528\u7814\u7a76(8):40-44+50.L. C. Thomas. Markov Decision Processes[J]. European Journal of Operational Research, 1995, 46(6):792-793.Bardi M, Bardi M. Optimal control and viscosity solutions of Hamilton-Jacobi-Bellman equations \/[M]\/\/ Optimal control and viscosity solutions of Hamilton-Jacobi-Bellman equations. 1997.Ruder S . An overview of gradient descent optimization algorithms[J]. 2016.R. Johnson, T. Zhang. Accelerating stochastic gradient descent using predictive variance reduction[J]. News in physiological sciences, 2013, 1(3):315-323.Sutton, Richard&#038; Mcallester, David&#038; Singh, Satinder&#038; Mansour, Yishay. (2000). Policy Gradient Methods for Reinforcement Learning with Function Approximation. Adv. Neural Inf. Process. Syst. 12.Silver, David&#038; Lever, Guy&#038; Heess, Nicolas&#038; Degris, Thomas&#038; Wierstra, Daan&#038; Riedmiller, Martin. (2014). Deterministic Policy Gradient Algorithms. 31st International Conference on Machine Learning, ICML 2014. 1.Lillicrap, Timothy&#038; Hunt, Jonathan&#038; Pritzel, Alexander&#038; Heess, Nicolas&#038; Erez, Tom&#038; Tassa, Yuval&#038; Silver, David&#038; Wierstra, Daan. (2015). Continuous control with deep reinforcement learning. CoRR.\u539f\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/link.segmentfault.com\/?enc=AQa6t%2B03RQnDKRKKouQoDw%3D%3D.n99MbPu3Nt93iqdFRF21vpysXbWkd%2FO%2B1Ey%2FtdSVVQ06YKv0crFY9mSlybdxMuaD\" rel=\"nofollow\">https:\/\/click.aliyun.com\/m\/10&#8230;<\/a>\u3001\u672c\u6587\u4e3a\u963f\u91cc\u4e91\u539f\u521b\u5185\u5bb9\uff0c\u672a\u7ecf\u5141\u8bb8\u4e0d\u5f97\u8f6c\u8f7d\u3002<\/p>\n<p><a href=\"https:\/\/segmentfault.com\/a\/1190000042358686\">https:\/\/segmentfault.com\/a\/1190000042358686<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7b80\u4ecb\uff1a\u00a0\u5728\u672c\u6587\u4e2d\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u6280\u672f\uff0c\u5c06\u5f3a\u5316\u5b66\u4e60\u5206\u6210\u4e09\u5927\u7c7b\uff08value-based\u7b97\u6cd5\u3001policy-b [&hellip;]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[],"tags":[],"class_list":["post-10006281","post","type-post","status-publish","format-standard","hentry"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006281","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10006281"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006281\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10006281"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10006281"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10006281"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}