{"id":2088,"date":"2024-09-20T10:54:48","date_gmt":"2024-09-20T02:54:48","guid":{"rendered":"https:\/\/www.gnn.club\/?p=2088"},"modified":"2024-09-20T10:54:48","modified_gmt":"2024-09-20T02:54:48","slug":"pong%e6%8c%91%e6%88%98%ef%bc%9a%e5%9f%ba%e4%ba%8e%e4%bb%b7%e5%80%bc%e3%80%81%e5%9f%ba%e4%ba%8e%e7%ad%96%e7%95%a5%e5%92%8c%e6%bc%94%e5%91%98-%e8%af%84%e8%ae%ba%e5%ae%b6%e5%bc%ba%e5%8c%96%e5%ad%a6","status":"publish","type":"post","link":"http:\/\/www.gnn.club\/?p=2088","title":{"rendered":"Pong\u6311\u6218\uff1a\u57fa\u4e8e\u4ef7\u503c\u3001\u57fa\u4e8e\u7b56\u7565\u548c\u6f14\u5458-\u8bc4\u8bba\u5bb6\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\u7684\u6bd4\u8f83"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u9879\u76ee\u80cc\u666f<\/h2>\n\n\n\n<p>\u968f\u7740\u4eba\u5de5\u667a\u80fd\u548c\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u8fc5\u901f\u53d1\u5c55\uff0c\u5f3a\u5316\u5b66\u4e60\u4f5c\u4e3a\u5176\u4e2d\u7684\u6838\u5fc3\u5206\u652f\uff0c\u5728\u89e3\u51b3\u590d\u6742\u51b3\u7b56\u95ee\u9898\u65b9\u9762\u663e\u793a\u51fa\u5de8\u5927\u6f5c\u529b\u3002\u5c24\u5176\u5728\u6e38\u620f\u9886\u57df\uff0c\u4e0d\u540c\u7c7b\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u5df2\u88ab\u8bc1\u5b9e\u80fd\u591f\u6709\u6548\u5730\u5904\u7406\u548c\u89e3\u51b3\u591a\u79cd\u6311\u6218\u3002\u5728\u672c\u9879\u76ee\u4e2d\uff0c\u6211\u4eec\u5c06\u63a2\u8ba8\u4e09\u79cd\u4e3b\u8981\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\u2014\u2014\u57fa\u4e8e\u4ef7\u503c\u7684Deep Q-Network\uff08DQN\uff09\u3001\u57fa\u4e8e\u7b56\u7565\u7684REINFORCE\u7b97\u6cd5\u4ee5\u53ca\u6f14\u5458-\u8bc4\u8bba\u5bb6\u65b9\u6cd5\u2014\u2014\u5728\u7ecf\u5178\u6e38\u620fPong\u4e2d\u7684\u5e94\u7528\u548c\u8868\u73b0\u3002Pong\u6e38\u620f\u4ee5\u5176\u7b80\u5355\u89c4\u5219\u548c\u76f4\u89c2\u73af\u5883\uff0c\u6210\u4e3a\u8bc4\u4f30\u548c\u6bd4\u8f83\u8fd9\u4e9b\u7b97\u6cd5\u7684\u7406\u60f3\u5e73\u53f0\u3002<br><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u9879\u76ee\u76ee\u6807<\/h2>\n\n\n\n<p>\u672c\u9879\u76ee\u65e8\u5728\u901a\u8fc7\u5b9e\u73b0\u5e76\u6bd4\u8f83\u4e09\u79cd\u4e0d\u540c\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\u6765\u63a2\u7d22Pong\u6e38\u620f\u7684AI\u5b9e\u73b0\u3002\u5177\u4f53\u76ee\u6807\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7efc\u5408\u7406\u89e3\u5f3a\u5316\u5b66\u4e60\u6a21\u578b<\/strong>\uff1a\u901a\u8fc7\u5b9e\u8df5\u6df1\u5165\u7406\u89e3DQN\u3001REINFORCE\u4ee5\u53ca\u6f14\u5458-\u8bc4\u8bba\u5bb6\u6a21\u578b\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u5e94\u7528\u548c\u7279\u70b9\u3002<\/li>\n\n\n\n<li><strong>\u591a\u6a21\u578b\u73af\u5883\u4ea4\u4e92<\/strong>\uff1a\u6784\u5efa\u5e76\u4f18\u5316Pong\u6e38\u620f\u73af\u5883\uff0c\u5b9e\u73b0\u8fd9\u4e09\u79cd\u4e0d\u540c\u7b97\u6cd5\u4e0e\u73af\u5883\u7684\u6709\u6548\u4ea4\u4e92\u3002<\/li>\n\n\n\n<li><strong>\u7f51\u7edc\u8bbe\u8ba1\u4e0e\u6027\u80fd\u6bd4\u8f83<\/strong>\uff1a\u4e3a\u6bcf\u79cd\u6a21\u578b\u8bbe\u8ba1\u9002\u5f53\u7684\u7f51\u7edc\u7ed3\u6784\uff0c\u5e76\u901a\u8fc7\u6027\u80fd\u6bd4\u8f83\u6765\u786e\u5b9a\u5404\u81ea\u7684\u4f18\u52a3\u3002<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u8bad\u7ec3\u4e0e\u6548\u679c\u8bc4\u4f30<\/strong>\uff1a\u8bad\u7ec3\u6bcf\u79cd\u6a21\u578b\u4ee5\u638c\u63e1Pong\u6e38\u620f\uff0c\u5e76\u5bf9\u5b83\u4eec\u5728\u4e0d\u540c\u60c5\u51b5\u4e0b\u7684\u8868\u73b0\u8fdb\u884c\u5168\u9762\u8bc4\u4f30\u3002<\/li>\n\n\n\n<li><strong>\u7b56\u7565\u5206\u6790\u4e0e\u4f18\u5316<\/strong>\uff1a\u5206\u6790\u5404\u6a21\u578b\u5728\u6e38\u620f\u4e2d\u7684\u7b56\u7565\uff0c\u5e76\u63a2\u7d22\u4f18\u5316\u65b9\u6cd5\u4ee5\u63d0\u5347\u6027\u80fd\u3002<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\">\u9879\u76ee\u5e94\u7528<\/h2>\n\n\n\n<p>\u901a\u8fc7\u672c\u9879\u76ee\uff0c\u6211\u4eec\u671f\u671b\u5b9e\u73b0\u4ee5\u4e0b\u5e94\u7528\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u591a\u7b97\u6cd5\u6280\u672f\u5c55\u793a<\/strong>\uff1a\u5c55\u793a\u548c\u6bd4\u8f83DQN\u3001REINFORCE\u548c\u6f14\u5458-\u8bc4\u8bba\u5bb6\u6a21\u578b\u5728\u6e38\u620fAI\u9886\u57df\u7684\u5e94\u7528\u6548\u679c\uff0c\u5c24\u5176\u5728\u590d\u6742\u51b3\u7b56\u548c\u63a7\u5236\u65b9\u9762\u3002<\/li>\n\n\n\n<li><strong>\u6559\u80b2\u4e0e\u7814\u7a76\u4ef7\u503c<\/strong>\uff1a\u4e3a\u7814\u7a76\u4e0d\u540c\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u7684\u5b66\u751f\u548c\u7814\u7a76\u4eba\u5458\u63d0\u4f9b\u4e00\u4e2a\u7efc\u5408\u6027\u7684\u6848\u4f8b\u7814\u7a76\u3002<\/li>\n\n\n\n<li><strong>\u6df1\u5165\u7b97\u6cd5\u5206\u6790\u4e0e\u6539\u8fdb<\/strong>\uff1a\u901a\u8fc7\u5bf9\u6bd4\u5206\u6790\u5728Pong\u6e38\u620f\u4e2d\u7684\u5e94\u7528\u6548\u679c\uff0c\u63a2\u8ba8\u5404\u7b97\u6cd5\u7684\u4f18\u52bf\u3001\u5c40\u9650\u6027\u53ca\u6539\u8fdb\u65b9\u5411\u3002<\/li>\n<\/ol>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"wznav_3\">\u6570\u636e\u96c6\u63cf\u8ff0<\/h2>\n\n\n\n<p>\u672c\u9879\u76ee\u4f7f\u7528\u7684\u6570\u636e\u96c6\u57fa\u4e8e\u7ecf\u5178\u7684Pong\u6e38\u620f\uff0c\u5305\u62ec\u4ee5\u4e0b\u5173\u952e\u7279\u5f81\u548c\u5143\u7d20\uff0c\u8fd9\u4e9b\u5c06\u7528\u4e8e\u8bad\u7ec3\u548c\u8bc4\u4f30DQN\u6a21\u578b\uff0c\u5176\u4ed6\u4e24\u79cd\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u4e5f\u5927\u591a\u91c7\u7528\u4ee5\u4e0b\u7279\u5f81\uff0c\u4e0d\u540c\u7684\u90e8\u5206\u4f1a\u5728\u4ee3\u7801\u90e8\u5206\u8bb2\u89e3\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>action_placeholder<\/strong>\uff1a\u4ee3\u8868AI\u5728\u6e38\u620f\u4e2d\u53ef\u80fd\u91c7\u53d6\u7684\u52a8\u4f5c\uff0c\u4f8b\u5982\u4e0a\u79fb\u3001\u4e0b\u79fb\u548c\u4e0d\u52a8\u3002<\/li>\n\n\n\n<li><strong>target_placeholder<\/strong>\uff1a\u7528\u4e8e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u76ee\u6807\u503c\uff0c\u5373\u9884\u671f\u7684\u8f93\u51fa\u7ed3\u679c\u3002<\/li>\n\n\n\n<li><strong>loss<\/strong>\uff1a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u635f\u5931\u51fd\u6570\uff0c\u7528\u4e8e\u8bc4\u4f30\u6a21\u578b\u7684\u8868\u73b0\u3002<\/li>\n\n\n\n<li><strong>checkpoint<\/strong>\uff1a\u6a21\u578b\u7684\u4fdd\u5b58\u548c\u52a0\u8f7d\u70b9\uff0c\u7528\u4e8e\u8bb0\u5f55\u8bad\u7ec3\u8fdb\u5ea6\u3002<\/li>\n\n\n\n<li><strong>game_agent<\/strong>\uff1a\u6e38\u620f\u4ee3\u7406\uff0c\u7528\u4e8e\u63a7\u5236\u6e38\u620f\u73af\u5883\u548c\u751f\u6210\u6e38\u620f\u6570\u636e\u3002<\/li>\n\n\n\n<li><strong>data_deque<\/strong>\uff1a\u7528\u4e8e\u5b58\u50a8\u6e38\u620f\u8fc7\u7a0b\u4e2d\u7684\u6570\u636e\uff0c\u4f8b\u5982\u5e27\u3001\u5956\u52b1\u548c\u52a8\u4f5c\u3002<\/li>\n\n\n\n<li><strong>reward<\/strong>\uff1a\u6e38\u620f\u4e2d\u7ed9\u4e88AI\u7684\u5373\u65f6\u53cd\u9988\uff0c\u6839\u636e\u5176\u52a8\u4f5c\u548c\u6e38\u620f\u7ed3\u679c\u7ed9\u51fa\u3002<\/li>\n\n\n\n<li><strong>action<\/strong>\uff1aAI\u5728\u6bcf\u4e00\u5e27\u4e2d\u91c7\u53d6\u7684\u5b9e\u9645\u52a8\u4f5c\u3002<\/li>\n\n\n\n<li><strong>paddle_1_score<\/strong>&nbsp;\u548c&nbsp;<strong>paddle_2_score<\/strong>\uff1a\u6e38\u620f\u53cc\u65b9\u7684\u5f97\u5206\u60c5\u51b5\uff0c\u7528\u4e8e\u8bc4\u4f30AI\u7684\u8868\u73b0\u3002<\/li>\n\n\n\n<li><strong>action_pred<\/strong>\uff1a\u7531\u795e\u7ecf\u7f51\u7edc\u9884\u6d4b\u7684\u52a8\u4f5c\u3002<\/li>\n\n\n\n<li><strong>num_frames<\/strong>\uff1a\u8bb0\u5f55\u5904\u7406\u7684\u6e38\u620f\u5e27\u6570\u3002<\/li>\n\n\n\n<li><strong>num_games<\/strong>\uff1a\u8bb0\u5f55\u73a9\u8fc7\u7684\u6e38\u620f\u6b21\u6570\u3002<\/li>\n\n\n\n<li><strong>num_win_games<\/strong>\uff1a\u8bb0\u5f55AI\u8d62\u5f97\u7684\u6e38\u620f\u6b21\u6570\u3002<\/li>\n\n\n\n<li><strong>minibatch<\/strong>\uff1a\u7528\u4e8e\u8bad\u7ec3\u7684\u5c0f\u6279\u91cf\u6570\u636e\u3002<\/li>\n\n\n\n<li><strong>reward_batch<\/strong>\uff1a\u5c0f\u6279\u91cf\u6570\u636e\u4e2d\u7684\u5956\u52b1\u503c\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"wznav_4\">\u9879\u76ee\u5b9e\u73b0<\/h2>\n\n\n\n<p>\u672c\u9879\u76ee\u901a\u8fc7\u4ee5\u4e0b\u5173\u952e\u6b65\u9aa4\u5b9e\u73b0DQN\u7b97\u6cd5\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7f51\u7edc\u6784\u5efa<\/strong>\uff1a\u901a\u8fc7<code>createNetwork<\/code>\u51fd\u6570\u6784\u5efa\u6df1\u5ea6\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\uff0c\u4f7f\u7528<code>initWeightVariable<\/code>\u548c<code>initBiasVariable<\/code>\u521d\u59cb\u5316\u7f51\u7edc\u6743\u91cd\u548c\u504f\u7f6e\u3002<\/li>\n\n\n\n<li><strong>\u8bad\u7ec3\u6d41\u7a0b<\/strong>\uff1a\u5229\u7528<code>train<\/code>\u65b9\u6cd5\u5bf9\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\uff0c\u5305\u62ec\u6570\u636e\u7684\u6536\u96c6\u3001\u7f51\u7edc\u7684\u8bad\u7ec3\u548c\u6a21\u578b\u7684\u4fdd\u5b58\u3002<\/li>\n\n\n\n<li><strong>\u52a8\u4f5c\u51b3\u7b56<\/strong>\uff1a\u5728\u6bcf\u4e00\u6e38\u620f\u5e27\u4e2d\uff0cAI\u6839\u636e\u7f51\u7edc\u7684\u9884\u6d4b(<code>action_pred<\/code>)\u6765\u51b3\u5b9a\u5176\u52a8\u4f5c\u3002<\/li>\n\n\n\n<li><strong>\u6027\u80fd\u8bc4\u4f30<\/strong>\uff1a\u901a\u8fc7\u7edf\u8ba1<code>num_games<\/code>\u548c<code>num_win_games<\/code>\u6765\u8bc4\u4f30AI\u7684\u6e38\u620f\u8868\u73b0\u3002<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u4f18\u5316<\/strong>\uff1a\u901a\u8fc7\u5206\u6790<code>loss<\/code>\u548c<code>reward_batch<\/code>\u6765\u8c03\u6574\u548c\u4f18\u5316\u6a21\u578b\u3002<\/li>\n<\/ol>\n\n\n\n<p>REINFORCE\u7684\u533a\u522b\u5728\u4e8e\uff0c\u5728\u52a8\u4f5c\u9009\u62e9\u7684\u65f6\u5019\uff0cREINFORCE\u57fa\u4e8e\u7b56\u7565\u7f51\u7edc\u751f\u6210\u7684<code>action_probabilities<\/code> \uff08\u5728\u5f53\u524d\u72b6\u6001\u4e0b\u91c7\u53d6\u6bcf\u4e2a\u52a8\u4f5c\u7684\u6982\u7387\uff09\u91c7\u53d6\u52a8\u4f5c\u3002<\/p>\n\n\n\n<p>\u800cAC\u91c7\u53d6\u52a8\u4f5c\u65f6\u57fa\u4e8e\u6f14\u5458\u7f51\u7edc\u751f\u6210\u7684<code>action_probabilities<\/code> \u9009\u62e9\u52a8\u4f5c<code>action_index<\/code>\uff1b\u8bc4\u8bba\u5bb6\u7f51\u7edc\u5219\u8d1f\u8d23\u8bc4\u4f30\u5f53\u524d\u72b6\u6001<code>state<\/code>\u7684\u4ef7\u503c\uff0c\u5373<code>state_val<\/code>\u3002\u8fd9\u4e2a\u4ef7\u503c\u7528\u4e8e\u8bc4\u4f30\u5f53\u524d\u7b56\u7565\u7684\u597d\u574f\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"wznav_5\">\u6a21\u578b\u9009\u62e9<\/h2>\n\n\n\n<p>\u5728\u672c\u9879\u76ee\u4e2d\uff0c\u6211\u4eec\u5c06\u4f7f\u7528\u4e09\u79cd\u4e0d\u540c\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\uff1aDeep Q-Network\uff08DQN\uff09\u3001REINFORCE\u4ee5\u53ca\u6f14\u5458-\u8bc4\u8bba\u5bb6\u6a21\u578b\u3002\u4ee5\u4e0b\u662f\u5bf9\u6bcf\u79cd\u6a21\u578b\u7684\u8be6\u7ec6\u4ecb\u7ecd\u548c\u6240\u9700\u7684\u4f9d\u8d56\u5e93\u3002<\/p>\n\n\n\n<p><strong>Deep Q-Network\uff08DQN\uff09\uff08\u57fa\u4e8e\u4ef7\u503c\u7684\u5f3a\u5316\u5b66\u4e60\uff09\uff1a<\/strong><\/p>\n\n\n\n<p>DQN\u662f\u4e00\u79cd\u7ed3\u5408\u4e86\u6df1\u5ea6\u5b66\u4e60\u548c\u5f3a\u5316\u5b66\u4e60\u7684\u65b9\u6cd5\uff0c\u7528\u4e8e\u5904\u7406\u5177\u6709\u9ad8\u7ef4\u5ea6\u8f93\u5165\u7a7a\u95f4\u7684\u590d\u6742\u4efb\u52a1\u3002\u5728\u672c\u9879\u76ee\u4e2d\uff0cDQN\u901a\u8fc7\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u6765\u5904\u7406\u6e38\u620f\u7684\u89c6\u89c9\u8f93\u5165\uff0c\u5e76\u5b66\u4e60\u5728\u6e38\u620f\u4e2d\u505a\u51fa\u6700\u4f18\u7684\u51b3\u7b56\u3002\u4e0e\u4f20\u7edf\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u76f8\u6bd4\uff0cDQN\u80fd\u591f\u66f4\u6709\u6548\u5730\u5904\u7406\u5177\u6709\u5927\u89c4\u6a21\u72b6\u6001\u7a7a\u95f4\u7684\u95ee\u9898\uff0c\u5982\u89c6\u9891\u6e38\u620f\u3002<\/p>\n\n\n\n<p><strong>REINFORCE\uff08\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\uff09\uff1a<\/strong><\/p>\n\n\n\n<p>REINFORCE\u662f\u4e00\u79cd\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff0c\u5b83\u76f4\u63a5\u4f18\u5316\u4e86\u7b56\u7565\u672c\u8eab\uff0c\u800c\u4e0d\u662f\u50cfDQN\u90a3\u6837\u95f4\u63a5\u4f18\u5316\u503c\u51fd\u6570\u3002\u8fd9\u79cd\u65b9\u6cd5\u901a\u8fc7\u68af\u5ea6\u4e0a\u5347\u6765\u4f18\u5316\u7b56\u7565\uff0c\u4ee5\u6700\u5927\u5316\u7d2f\u79ef\u5956\u52b1\u3002REINFORCE\u9002\u7528\u4e8e\u89e3\u51b3\u5177\u6709\u9ad8\u7ef4\u5ea6\u52a8\u4f5c\u7a7a\u95f4\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p><strong>AC\uff08\u6f14\u5458-\u8bc4\u8bba\u5bb6\u6a21\u578b\uff09\uff1a<\/strong><\/p>\n\n\n\n<p>\u6f14\u5458-\u8bc4\u8bba\u5bb6\u65b9\u6cd5\u7ed3\u5408\u4e86\u57fa\u4e8e\u503c\u7684\u548c\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u7684\u4f18\u70b9\u3002\u5728\u8fd9\u79cd\u65b9\u6cd5\u4e2d\uff0c\"\u6f14\u5458\"\u8d1f\u8d23\u751f\u6210\u52a8\u4f5c\uff0c\u800c\"\u8bc4\u8bba\u5bb6\"\u8bc4\u4f30\u8fd9\u4e9b\u52a8\u4f5c\u5e76\u5f15\u5bfc\u6f14\u5458\u7684\u5b66\u4e60\u3002\u8fd9\u79cd\u53cc\u91cd\u673a\u5236\u4f7f\u5f97\u6f14\u5458-\u8bc4\u8bba\u5bb6\u65b9\u6cd5\u5728\u7a33\u5b9a\u6027\u548c\u6548\u7387\u65b9\u9762\u90fd\u6709\u6240\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u5148\u6765\u901a\u8fc7\u4e00\u6bb5\u8fd9\u4e09\u4e2a\u7b97\u6cd5\u5b9e\u73b0\u7ecf\u5178\u7684Pong\u6e38\u620f\u7684\u89c6\u9891\uff0c\u5c55\u793a\u4e00\u4e0b\u6548\u679c\u4ee5\u53caPong\u6e38\u620f\u7684\u754c\u9762\u3002Pong\u6e38\u620f\u7b80\u5355\u6765\u8bf4\u5c31\u662f\u63a7\u5236\u62cd\u5b50\u53bb\u5b8c\u6210\u51fb\u7403\uff0c\u51fb\u7403\u4e0d\u5f97\u5206\uff0c\u4f46\u662f\u6ca1\u6709\u51fb\u4e2d\u7403\uff0c\u5bf9\u624b\u52a0\u5206\uff0c\u6e38\u620f\u4e2d\u8bbe\u7f6e\u83b7\u5f9720\u5206\u5373\u8d62\u5f97\u4e00\u5c40\u3002\u753b\u9762\u5de6\u8fb9\u7684\u7eff\u8272\u62cd\u5b50\u662f\u7531\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u63a7\u5236\u7684\uff0c\u89c6\u9891\u5982\u4e0b\u6240\u793a\uff0c\u70b9\u51fb\u5373\u53ef\u89c2\u770b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-video\"><video controls src=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240920104657230.mp4\"><\/video><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u4f9d\u8d56\u5e93<\/h2>\n\n\n\n<p><strong>1.TensorFlow\uff1a<\/strong><\/p>\n\n\n\n<p>TensorFlow\u662f\u4e00\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u652f\u6301\u591a\u79cd\u64cd\u4f5c\u7cfb\u7edf\u548c\u5e73\u53f0\u3002\u5728\u672c\u9879\u76ee\u4e2d\uff0c\u5b83\u88ab\u7528\u6765\u6784\u5efa\u548c\u8bad\u7ec3DQN\u6a21\u578b\u3002TensorFlow\u7684\u7075\u6d3b\u6027\u548c\u5f3a\u5927\u7684\u8ba1\u7b97\u80fd\u529b\u4f7f\u5176\u6210\u4e3a\u5b9e\u73b0\u590d\u6742\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u7684\u7406\u60f3\u9009\u62e9\u3002<\/p>\n\n\n\n<p><strong>2.NumPy\uff1a<\/strong><\/p>\n\n\n\n<p>NumPy\u662fPython\u4e2d\u7528\u4e8e\u79d1\u5b66\u8ba1\u7b97\u7684\u57fa\u7840\u5305\uff0c\u63d0\u4f9b\u4e86\u9ad8\u6027\u80fd\u7684\u591a\u7ef4\u6570\u7ec4\u5bf9\u8c61\u53ca\u76f8\u5173\u5de5\u5177\u3002\u5728\u672c\u9879\u76ee\u4e2d\uff0cNumPy\u7528\u4e8e\u6570\u636e\u5904\u7406\u548c\u6570\u503c\u8ba1\u7b97\uff0c\u5982\u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5904\u7406\u6e38\u620f\u5e27\u548c\u5956\u52b1\u6570\u636e\u3002<\/p>\n\n\n\n<p><strong>3.OpenCV\uff1a<\/strong><\/p>\n\n\n\n<p>OpenCV\uff08Open Source Computer Vision Library\uff09\u662f\u4e00\u4e2a\u5f00\u6e90\u7684\u8ba1\u7b97\u673a\u89c6\u89c9\u548c\u673a\u5668\u5b66\u4e60\u8f6f\u4ef6\u5e93\u3002\u5728\u672c\u9879\u76ee\u4e2d\uff0cOpenCV\u7528\u4e8e\u5904\u7406\u6e38\u620f\u7684\u56fe\u50cf\u6570\u636e\uff0c\u4f8b\u5982\u901a\u8fc7\u8c03\u6574\u56fe\u50cf\u5927\u5c0f\u6765\u51c6\u5907\u8f93\u5165\u6570\u636e\u3002<\/p>\n\n\n\n<p>\u6ce8\u610f\uff1a\u8981\u60f3\u5728\u8ba1\u7b97\u673a\u4e0a\u8fd0\u884c\u76f8\u5173\u4ee3\u7801\uff0c\u9996\u5148\u9700\u8981\u4e0b\u8f7d\u76f8\u5173\u5e93\uff0c\u4ee3\u7801\u53ef\u4ee5\u5728tensorflow2\u4e0a\u8fd0\u884c\uff0c\u5982\u53ea\u6709tensorflow1\u53ef\u4ee5\u6839\u636e\u4ee3\u7801\u4e2d\u7684\u63d0\u793a\u5185\u5bb9\u53bb\u66f4\u6539\u4e00\u4e9b\u5185\u5bb9\u5373\u53ef\u3002\u5e76\u4e14\u8bf7\u786e\u4fdd\u4fdd\u5b89\u88c5\u7684\u5e93\u7248\u672c\u76f8\u4e92\u517c\u5bb9\uff0c\u7279\u522b\u662f\u8bf7\u786e\u4fddPython\u7248\u672c\u4e0eTensorFlow 2\u517c\u5bb9\uff0c\u5efa\u8bae\u5728Python\u865a\u62df\u73af\u5883\u4e2d\u5b89\u88c5TensorFlow\uff0c\u4ee5\u907f\u514d\u7248\u672c\u51b2\u7a81\u3002\u4f7f\u7528<code>pip install tensorflow<\/code>\u5373\u53ef\u914d\u7f6e\u73af\u5883\uff0c\u5982\u679c\u5b58\u5728\u7f51\u7edc\u9650\u5236\uff0c\u53ef\u4ee5\u4f7f\u7528\u955c\u50cf\u6e90\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"wznav_7\">DQN\u4ee3\u7801\u5b9e\u73b0<\/h2>\n\n\n\n<p><strong>1.\u5bfc\u5165\u5fc5\u8981\u7684\u5e93<\/strong><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import os\n\nimport cv2\n\nimport random\n\nimport numpy as np\n\nimport tensorflow as tf\n\nimport tensorflow.compat.v1 as tf\n\ntf.disable_v2_behavior()# \u7981\u7528 TensorFlow 2 \u884c\u4e3a\uff0c\u4ee5\u4fbf\u5728 TensorFlow 1 \u73af\u5883\u4e2d\u8fd0\u884c\uff0c\u5982\u679c\u5728TensorFlow 1 \u73af\u5883\u4e2d\u8fd0\u884c\u5220\u9664\u8fd9\u884c\u5373\u53ef\n\nfrom collections import deque\n\nfrom gameAPI.agent import gameAgent<\/code><\/pre>\n\n\n\n<p><strong>2.\u521b\u5efaDQN\u6a21\u578b\u7c7b<\/strong><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>class DQN():\n\ndef __init__(self, config, **kwargs):\n\nself.config = config # \u521d\u59cb\u5316DQN\u6a21\u578b\u914d\u7f6e<\/code><\/pre>\n\n\n\n<p><strong>3.DQN\u6a21\u578b\u8bad\u7ec3\u4ee3\u7801<\/strong><\/p>\n\n\n\n<p>\u8fd9\u90e8\u5206<code>action_pred=np.argmax(action_pred)<\/code>\u4f53\u73b0\u4e86DQN\u57fa\u4e8e\u4ef7\u503c\u7684\u7279\u70b9\uff0c\u5728\u9009\u53d6\u52a8\u4f5c\u65f6\uff0cDQN\u4f1a\u9009\u62e9\u5f97\u5206\u6700\u9ad8\u7684\u52a8\u4f5c<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def train(self, session):\n\n# \u521b\u5efa\u795e\u7ecf\u7f51\u7edc\n\nnet_in, net_out = self.__createNetwork() # \u795e\u7ecf\u7f51\u7edc\u7684\u8f93\u5165\u548c\u8f93\u51fa\n\n# \u5b9a\u4e49\u52a8\u4f5c\u548c\u76ee\u6807\u7684\u5360\u4f4d\u7b26\n\naction_placeholder = tf.placeholder('float', &#91;None, 3]) # \u8868\u793a\u6e38\u620f\u4e2d\u7684\u52a8\u4f5c\n\ntarget_placeholder = tf.placeholder('float', &#91;None]) # \u8868\u793a\u9884\u671f\u76ee\u6807\u503c\n\n# \u5b9a\u4e49\u635f\u5931\u51fd\u6570\n\nloss = tf.reduce_mean(tf.square(target_placeholder - tf.reduce_sum(tf.multiply(net_out, action_placeholder), reduction_indices=1)))\n\n# \u5b9a\u4e49\u8bad\u7ec3\u6b65\u9aa4\n\ntrain_step = tf.train.AdamOptimizer(self.config.lr).minimize(loss)\n\n# \u6e38\u620f\u4ee3\u7406\u548c\u6570\u636e\u961f\u5217\n\ngame_agent = gameAgent() # \u6e38\u620f\u4ee3\u7406\uff0c\u7528\u4e8e\u4e0e\u6e38\u620f\u73af\u5883\u4ea4\u4e92\n\ndata_deque = deque() # \u6570\u636e\u961f\u5217\uff0c\u7528\u4e8e\u5b58\u50a8\u6e38\u620f\u7ecf\u9a8c\n\n# \u4e3b\u5faa\u73af\uff1a\u6e38\u620f\u73a9\u6cd5\u548c\u7f51\u7edc\u8bad\u7ec3\n\nwhile True:\n\t\t\t# random decide#\u4f53\u73b0\u4e86\u63a2\u7d22\n\t\t\tif random.random() &lt;= prob or frame_pre is None:\n\t\t\t\tframe_now, action, reward, terminal, paddle_1_score, paddle_2_score = game_agent.nextFrame(action=None)\n\t\t\t\tframe_now = cv2.resize(frame_now, self.config.frame_size)\n\t\t\t# decide by network\n\t\t\telse:#\u4f53\u73b0\u4e86\u5229\u7528\n\t\t\t\taction_pred = net_out.eval(feed_dict={net_in: &#91;frame_pre]})\n\t\t\t\taction_pred = np.argmax(action_pred)#\u8fd9\u91cc\u8bf4\u660eDQN\u9009\u53d6\u52a8\u4f5c\u8003\u8651\u7684\u662f\u52a8\u4f5c\u5f97\u5206\u7684\u9ad8\u4f4e\uff0c\u5373\u57fa\u4e8e\u4ef7\u503c\n\t\t\t\tif action_pred == 0:\n\t\t\t\t\taction_pred = &#91;1, 0, 0]\n\t\t\t\telif action_pred == 1:\n\t\t\t\t\taction_pred = &#91;0, 1, 0]\n\t\t\t\telif action_pred == 2:\n\t\t\t\t\taction_pred = &#91;0, 0, 1]\n\t\t\t\telse:\n\t\t\t\t\traise RuntimeError('Hhhhh, your code in net_out for action should be wrong I think...')\n\t\t\t\tframe_now, action, reward, terminal, paddle_1_score, paddle_2_score = game_agent.nextFrame(action=action_pred)\n\t\t\t\tframe_now = cv2.resize(frame_now, self.config.frame_size) \/ 255.\n\npass # \u7701\u7565\u5177\u4f53\u7684\u6e38\u620f\u903b\u8f91\u548c\u8bad\u7ec3\u8fc7\u7a0b<\/code><\/pre>\n\n\n\n<p><strong>4.\u795e\u7ecf\u7f51\u7edc\u6784\u5efa\u51fd\u6570<\/strong><\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def __createNetwork(self):#\u6784\u5efaDQN\u6a21\u578b\u7684\u795e\u7ecf\u7f51\u7edc\u67b6\u6784\uff0c\u5305\u62ec\u5377\u79ef\u5c42\u548c\u5168\u8fde\u63a5\u5c42\uff0c\u7528\u4e8e\u7279\u5f81\u63d0\u53d6\u548c\u52a8\u4f5c\u51b3\u7b56\u3002\n\nw_conv1 = self.__initWeightVariable(&#91;8, 8, 3, 32])\n\nb_conv1 = self.__initBiasVariable(&#91;32])\n\n# ... \u795e\u7ecf\u7f51\u7edc\u7684\u5177\u4f53\u6784\u5efa\u8fc7\u7a0b ...\n\npass # \u7701\u7565\u5177\u4f53\u7684\u7f51\u7edc\u6784\u5efa\u4ee3\u7801<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">REINFORCE\u4ee3\u7801\u5b9e\u73b0<\/h2>\n\n\n\n<p>\u7531\u4e8e\u8fd9\u90e8\u5206\u4ee3\u7801\u5927\u4f53\u4e0a\u548cDQN\u4ee3\u7801\u5dee\u522b\u4e0d\u5927\uff0c\u6240\u4ee5\u53ea\u5c55\u793a\u5b83\u4eec\u4e0d\u540c\u7684\u90e8\u5206\uff0c\u5373\uff1a<\/p>\n\n\n\n<p><strong>1.REINFORCE\u6a21\u578b\u8bad\u7ec3\u4ee3\u7801<\/strong><\/p>\n\n\n\n<p>\u5728\u8fd9\u4e2a\u4ee3\u7801\u6bb5\u4e2d\uff0c<code>action_probabilities<\/code> \u662f\u7531\u7b56\u7565\u7f51\u7edc\u751f\u6210\u7684\uff0c\u8868\u793a\u5728\u5f53\u524d\u72b6\u6001\u4e0b\u91c7\u53d6\u6bcf\u4e2a\u52a8\u4f5c\u7684\u6982\u7387\u3002\u7136\u540e\uff0c<code>np.random.choice<\/code> \u51fd\u6570\u57fa\u4e8e\u8fd9\u4e9b\u6982\u7387\u6765\u968f\u673a\u9009\u62e9\u4e00\u4e2a\u52a8\u4f5c\u6267\u884c\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u65b9\u6cd5\u4e0e\u57fa\u4e8e\u4ef7\u503c\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08\u5982DQN\uff09\u4e0d\u540c\u3002\u5728\u57fa\u4e8e\u4ef7\u503c\u7684\u65b9\u6cd5\u4e2d\uff0c\u52a8\u4f5c\u7684\u9009\u62e9\u901a\u5e38\u662f\u57fa\u4e8e\u9884\u6d4b\u7684\u6700\u5927\u503c\u51fd\u6570\uff08\u6216Q\u503c\uff09\u6765\u9009\u62e9\u7684\uff0c\u5373\u9009\u62e9\u80fd\u5e26\u6765\u6700\u5927\u9884\u671f\u56de\u62a5\u7684\u52a8\u4f5c\u3002\u800c\u5728\u57fa\u4e8e\u7b56\u7565\u7684\u65b9\u6cd5\uff08\u5982REINFORCE\uff09\u4e2d\uff0c\u52a8\u4f5c\u7684\u9009\u62e9\u662f\u76f4\u63a5\u7531\u7b56\u7565\u7f51\u7edc\u8f93\u51fa\u7684\u6982\u7387\u5206\u5e03\u51b3\u5b9a\u7684\u3002\u8fd9\u5c31\u610f\u5473\u7740\u5373\u4f7f\u67d0\u4e2a\u52a8\u4f5c\u7684\u9884\u671f\u56de\u62a5\u4e0d\u662f\u6700\u9ad8\u7684\uff0c\u53ea\u8981\u7b56\u7565\u7f51\u7edc\u7ed9\u51fa\u4e86\u4e00\u5b9a\u7684\u6982\u7387\uff0c\u8fd9\u4e2a\u52a8\u4f5c\u4ecd\u7136\u6709\u53ef\u80fd\u88ab\u9009\u62e9\u3002\u8fd9\u6837\u7684\u7b56\u7565\u5141\u8bb8\u7b97\u6cd5\u63a2\u7d22\u975e\u6700\u4f18\u52a8\u4f5c\uff0c\u6709\u52a9\u4e8e\u627e\u5230\u5168\u5c40\u6700\u4f18\u89e3\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def train(self, session):\n        # create network\n        net_in, action_probs = self._create_network()  # \u7b56\u7565\u7f51\u7edc\u8f93\u51fa\u52a8\u4f5c\u6982\u7387\uff0c\u8fd9\u91cc\u4f53\u73b0REINFORCE\u57fa\u4e8e\u7b56\u7565\u7684\u7279\u70b9\uff1a\u6839\u636e\u52a8\u4f5c\u52a8\u4f5c\u6982\u7387\u751f\u6210\u7b56\u7565\uff0c\u6700\u540e\u9009\u62e9\u7684\u662f\u67d0\u4e2a\u7b56\u7565\n        actions_placeholder = tf.placeholder(tf.int32, &#91;None])\n        rewards_placeholder = tf.placeholder(tf.float32, &#91;None])\n\n        # \u83b7\u53d6\u9009\u4e2d\u52a8\u4f5c\u7684\u6982\u7387\n        action_masks = tf.one_hot(actions_placeholder, 3)  # \u5047\u8bbe\u67093\u4e2a\u52a8\u4f5c\n        picked_action_probs = tf.reduce_sum(action_probs * action_masks, axis=1)\n\n        # \u635f\u5931\u51fd\u6570\u548c\u8bad\u7ec3\u6b65\u9aa4\n        loss = -tf.reduce_mean(tf.log(picked_action_probs) * rewards_placeholder)\n        train_step = tf.train.AdamOptimizer(self.config.lr).minimize(loss)\n\n        # game agent\n        game_agent = gameAgent()\n\n        # \u8bad\u7ec3\u5faa\u73af\n        for episode in range(self.config.num_episodes):\n            #--\u5ffd\u7565\u90e8\u5206\u4ee3\u7801\n\n            while not done:\n                # \u83b7\u53d6\u52a8\u4f5c\u6982\u7387\n                action_probabilities = session.run(action_probs, feed_dict={net_in: &#91;state]})\n                action_index = np.random.choice(range(3), p=action_probabilities&#91;0])#\u4f53\u73b0\u4e86REINFORCE\u8fd9\u79cd\u57fa\u4e8e\u7b56\u7565\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u7684\u7279\u70b9\n                frame_info = game_agent.nextFrame(&#91;1 if i == action_index else 0 for i in range(3)])\n                next_state = frame_info&#91;0]  # \u63d0\u53d6 next frame          \n                reward = frame_info&#91;2]  # \u63d0\u53d6 reward\n                done = frame_info&#91;3]  # \u63d0\u53d6 terminal\n                episode_states.append(state)\n                episode_actions.append(action_index)\n                episode_rewards.append(reward)\n\n                state = next_state \n\n            # \u8ba1\u7b97\u7d2f\u79ef\u56de\u62a5\n            cumulative_rewards = self._compute_cumulative_rewards(episode_rewards)\n<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">AC\u4ee3\u7801\u5b9e\u73b0\uff1a<\/h2>\n\n\n\n<p><strong>1.\u795e\u7ecf\u7f51\u7edc\u6784\u5efa\u51fd\u6570<\/strong><\/p>\n\n\n\n<p>\u4e0d\u540c\u4e8eDQN\u548cREINFORCE\uff0cAC\u8981\u6784\u5efa\u56db\u4e2a\u795e\u7ecf\u7f51\u7edc\uff1a2\u4e2a\u6f14\u5458\u7f51\u7edc\uff0c2\u4e2a\u8bc4\u8bba\u5bb6\u7f51\u7edc\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def _create_networks(self):\n\t\t\t# \u8f93\u5165\u5c42\n\t\t\tx = tf.placeholder(tf.float32, &#91;None, *self.config.frame_size, 3])\n\n\t\t\t# \u5171\u4eab\u7f51\u7edc\u5c42\n\t\t\tconv1 = tf.layers.conv2d(inputs=x, filters=32, kernel_size=&#91;8, 8], strides=4, activation=tf.nn.relu)\n\t\t\tconv2 = tf.layers.conv2d(inputs=conv1, filters=64, kernel_size=&#91;4, 4], strides=2, activation=tf.nn.relu)\n\t\t\tconv3 = tf.layers.conv2d(inputs=conv2, filters=64, kernel_size=&#91;3, 3], strides=1, activation=tf.nn.relu)\n\t\t\tflattened = tf.layers.flatten(conv3)\n\n\t\t\t# \u6f14\u5458\u7f51\u7edc\n\t\t\tfc_actor = tf.layers.dense(flattened, 512, activation=tf.nn.relu)\n\t\t\taction_probs = tf.layers.dense(fc_actor, self.config.num_actions, activation=tf.nn.softmax)\n\n\t\t\t# \u8bc4\u8bba\u5bb6\u7f51\u7edc\n\t\t\tfc_critic = tf.layers.dense(flattened, 512, activation=tf.nn.relu)\n\t\t\tstate_value = tf.layers.dense(fc_critic, 1)\n\n\t\t\treturn (x, action_probs), (x, state_value)<\/code><\/pre>\n\n\n\n<p><strong>2.AC\u6a21\u578b\u8bad\u7ec3\u4ee3\u7801<\/strong><\/p>\n\n\n\n<p>\u5728\u6f14\u5458-\u8bc4\u8bba\u5bb6\uff08Actor-Critic\uff09\u6a21\u578b\u4ee3\u7801\u4e2d\uff0c\u52a8\u4f5c\u7684\u9009\u62e9\u548c\u8bc4\u4ef7\u5206\u522b\u7531\u6f14\u5458\uff08Actor\uff09\u7f51\u7edc\u548c\u8bc4\u8bba\u5bb6\uff08Critic\uff09\u7f51\u7edc\u6765\u5b9e\u73b0\u3002<code>action_probabilities<\/code> \u7531\u6f14\u5458\u7f51\u7edc\u6839\u636e\u5f53\u524d\u72b6\u6001<code>state<\/code>\u8ba1\u7b97\u4e86\u91c7\u53d6\u6bcf\u4e2a\u53ef\u80fd\u52a8\u4f5c\u7684\u6982\u7387\uff0c\u968f\u540e\u6839\u636e\u6f14\u5458\u7f51\u7edc\u8f93\u51fa\u7684\u52a8\u4f5c\u6982\u7387\uff0c\u4ee3\u7801\u968f\u673a\u9009\u62e9\u4e86\u4e00\u4e2a\u52a8\u4f5c<code>action_index<\/code>\uff1b\u540c\u6837\u5728\u8fd9\u4e00\u6b65\u4e2d\uff0c\u8bc4\u8bba\u5bb6\u7f51\u7edc\u8bc4\u4f30\u4e86\u5f53\u524d\u72b6\u6001<code>state<\/code>\u7684\u4ef7\u503c\uff0c\u5373<code>state_val<\/code>\u3002\u8fd9\u4e2a\u4ef7\u503c\u7528\u4e8e\u8bc4\u4f30\u5f53\u524d\u7b56\u7565\u7684\u597d\u574f\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def train(self, session):\n\t\t\t# \u521b\u5efa\u7f51\u7edc\n\t\t\t(actor_in, action_probs), (critic_in, state_value) = self._create_networks()\n\n\t\t\t# \u8bbe\u7f6e\u5360\u4f4d\u7b26\n\t\t\tactions_placeholder = tf.placeholder(tf.int32, &#91;None])\n\t\t\trewards_placeholder = tf.placeholder(tf.float32, &#91;None])\n\t\t\tadvantages_placeholder = tf.placeholder(tf.float32, &#91;None])\n\n\t\t\t# \u8bc4\u8bba\u5bb6\u7f51\u7edc\u7684\u635f\u5931\u51fd\u6570\n\t\t\tcritic_loss = tf.reduce_mean(tf.square(rewards_placeholder - state_value))\n\n\t\t\t# \u6f14\u5458\u7f51\u7edc\u7684\u635f\u5931\u51fd\u6570\n\t\t\taction_masks = tf.one_hot(actions_placeholder, self.config.num_actions)\n\t\t\tpicked_action_probs = tf.reduce_sum(action_probs * action_masks, axis=1)\n\t\t\tactor_loss = -tf.reduce_mean(tf.log(picked_action_probs) * advantages_placeholder)\n\n\t\t\t# \u4f18\u5316\u5668\n\t\t\tactor_optimizer = tf.train.AdamOptimizer(self.config.lr).minimize(actor_loss)\n\t\t\tcritic_optimizer = tf.train.AdamOptimizer(self.config.lr).minimize(critic_loss)\n\n\t\t\tsaver = tf.train.Saver()\n\t\t\tsession.run(tf.global_variables_initializer())\n\n\t\t\t# game agent\n\t\t\tgame_agent = gameAgent()\n\n\t\t\t# \u8bad\u7ec3\u5faa\u73af\n\t\t\tfor episode in range(self.config.num_episodes):\n\t\t\t\tframe_info = game_agent.nextFrame(None)\n\t\t\t\tstate = frame_info&#91;0]  # \u63d0\u53d6 frame\n\t\t\t\tstate = cv2.resize(state, (96, 96))  # \u8c03\u6574 state \u5927\u5c0f\n\t\t\t\tepisode_rewards = &#91;]\n\t\t\t\tepisode_actions = &#91;]\n\t\t\t\tepisode_states = &#91;]\n\t\t\t\tepisode_values = &#91;]\n\t\t\t\tdone = False\n\t\t\t\twhile not done:\n\t\t\t\t\t# \u83b7\u53d6\u52a8\u4f5c\u6982\u7387\u548c\u72b6\u6001\u4ef7\u503c\n\t\t\t\t\taction_probabilities, state_val = session.run(&#91;action_probs, state_value], feed_dict={actor_in: &#91;state], critic_in: &#91;state]})#\u6f14\u5458\u9009\u62e9\u52a8\u4f5c\uff0c\u51b3\u5b9a\u4e86action_probabilities\uff1b\u8bc4\u8bba\u5bb6\u8bc4\u4ef7\u52a8\u4f5c\uff0c\u51b3\u5b9a\u4e86state_val\n\t\t\t\t\taction_index = np.random.choice(range(3), p=action_probabilities&#91;0])\n\t\t\t\t\taction = &#91;0, 0, 0]\n\t\t\t\t\taction&#91;action_index] = 1\n\n\t\t\t\t\tframe_info = game_agent.nextFrame(action)\n\t\t\t\t\tnext_state = frame_info&#91;0]\n\t\t\t\t\tnext_state = cv2.resize(next_state, (96, 96))\n\t\t\t\t\treward = frame_info&#91;2]\n\t\t\t\t\tdone = frame_info&#91;3]\n\n\t\t\t\t\tepisode_states.append(state)\n\t\t\t\t\tepisode_actions.append(action_index)\n\t\t\t\t\tepisode_rewards.append(reward)\n\t\t\t\t\tepisode_values.append(state_val&#91;0]&#91;0])\n\t\t\t\t\tstate = next_state \n\n\t\t\t\t# \u8ba1\u7b97\u7d2f\u79ef\u56de\u62a5\u548c\u4f18\u52bf\n\t\t\t\tcumulative_rewards = self._compute_cumulative_rewards(episode_rewards)\n\t\t\t\tadvantages = cumulative_rewards - np.array(episode_values)<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u9879\u76ee\u4ee3\u7801\u4e0b\u8f7d<\/h2>\n\n\n\n<div class=\"wp-block-file\"><a id=\"wp-block-file--media-45da5b8f-246a-49b4-89f8-2b44dc864054\" href=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240920105054494.zip\">20240223183120879<\/a><a href=\"https:\/\/gnnclub-1311496010.cos.ap-beijing.myqcloud.com\/wp-content\/uploads\/2024\/09\/20240920105054494.zip\" class=\"wp-block-file__button wp-element-button\" download aria-describedby=\"wp-block-file--media-45da5b8f-246a-49b4-89f8-2b44dc864054\">\u4e0b\u8f7d<\/a><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u9879\u76ee\u80cc\u666f \u968f\u7740\u4eba\u5de5\u667a\u80fd\u548c\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u8fc5\u901f\u53d1\u5c55\uff0c\u5f3a\u5316\u5b66\u4e60\u4f5c\u4e3a\u5176\u4e2d\u7684\u6838\u5fc3\u5206\u652f\uff0c\u5728\u89e3\u51b3\u590d\u6742\u51b3\u7b56\u95ee\u9898\u65b9\u9762\u663e\u793a\u51fa\u5de8\u5927\u6f5c [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":2091,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2088","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/2088","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2088"}],"version-history":[{"count":1,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/2088\/revisions"}],"predecessor-version":[{"id":2092,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/posts\/2088\/revisions\/2092"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=\/wp\/v2\/media\/2091"}],"wp:attachment":[{"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2088"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2088"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.gnn.club\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2088"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}