{"id":567,"date":"2020-06-02T07:06:00","date_gmt":"2020-06-02T07:06:00","guid":{"rendered":"https:\/\/tensor.agenthub.uk\/?p=567"},"modified":"2024-05-16T07:54:48","modified_gmt":"2024-05-16T07:54:48","slug":"pytorch%e5%ae%9e%e7%8e%b0policy-gradient","status":"publish","type":"post","link":"https:\/\/tensorzen.blog\/?p=567","title":{"rendered":"PyTorch\u5b9e\u73b0Policy Gradient"},"content":{"rendered":"\n

\u5148\u6765\u56de\u5fc6\u4e00\u4e0b\u51e0\u4e2a\u53d8\u91cf\u7684\u5b9a\u4e49\uff0cPolicy Gradient\u7684\u5173\u952e\u662f\u901a\u8fc7Gradient\u6765\u66f4\u65b0Policy<\/p>\n\n\n\n

$$\\theta_{k+1} = \\theta_{k} + a \\nabla _{\\theta}J(\\pi_{\\theta})|_{\\theta_k}$$<\/p>\n\n\n\n

\u5176\u4e2d$\\pi_{\\theta}$\u662f\u53c2\u6570\u8bdd\u7684policy\uff0c$\\theta$\u662f\u5b83\u7684\u7cfb\u6570\uff0c$J(\\pi_{\\theta})$\u7528\u6765\u8861\u91cf\u5f53\u524dpolicy $\\pi_{\\theta}$\u7684\u6027\u80fd\uff0c\u54b1\u4eec\u8fd9\u91cc\u7528$\\pi_{\\theta}$\u7684\u671f\u671b\u6536\u76ca$E_{\\tau \\sim \\pi_{\\theta}}[R(\\tau)]$\u4f5c\u4e3apolicy\u7684\u6027\u80fd\uff0c$R(\\tau)$\u8868\u793a\u4e00\u5c40\u6e38\u620f\u7684\u6536\u76ca,$\\tau \\sim \\pi_{\\theta}$\u8868\u793a\u662f\u5728\u5f53\u524dpolicy $\\pi_{\\theta}$\u4e0b\u3002<\/p>\n\n\n\n

$\\nabla _{\\theta}J(\\pi_{\\theta})$\u7b49\u4e8e\u4e0b\u9762\u8fd9\u4e00\u4e32<\/p>\n\n\n\n

$$\\nabla _{\\theta}J(\\pi_{\\theta}) = E_{\\tau \\sim \\pi_{\\theta}} \\left [ \\sum_{t=0}^{T} \\nabla_{\\theta} \\log \\pi_{\\theta} (a_t | s_t)R(\\tau)\\right ]$$<\/p>\n\n\n\n

\u4e0a\u4e00\u7bc7\u6587\u7ae0\u4ecb\u7ecd\u4e86\u600e\u4e48\u5f97\u5230\u8fd9\u4e2a\u516c\u5f0f\u7684\uff0c\u5b9e\u9645\u4e0a\u5e76\u4e0d\u5f71\u54cd\u6211\u4eec\u7684\u5b9e\u73b0\u3002<\/p>\n\n\n\n

\u6e38\u620f\u73af\u5883\u6211\u4eec\u4f7f\u7528\u6700\u7ecf\u5178\u7684\u63a8\u5c0f\u8f66<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u8f66\u5b50\u4e0a\u7684\u6746\u513f\u4f1a\u56f4\u7ed5\u84dd\u8272\u70b9\u81ea\u7531\u8f6c\u52a8\uff0c\u5230\u8fbe\u4e00\u5b9a\u89d2\u5ea6\u6e38\u620f\u7ed3\u675f\uff0c\u6211\u4eec\u63a7\u5236\u8f66\u5b50\u5728\u6c34\u5e73\u7ebf\u4e0a\u5de6\u53f3\u79fb\u52a8\u6765\u4f7f\u6746\u513f\u5e73\u8861\uff0c\u5927\u6982\u5c31\u50cf\u6742\u6280\u6f14\u5458\u62ff\u4e00\u6839\u7af9\u7aff\u653e\u5728\u98df\u6307\u4e0a\u8ba9\u7af9\u7aff\u4e00\u76f4\u4e0d\u6389\u5730\u4e0a\u3002\u00a0<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u4ee3\u7801\u662f\u57fa\u4e8eOpenAI\u7684Spinning Up\u5b9e\u73b0\u7684\uff0c\u6240\u4ee5\u5982\u679c\u4f60\u6709\u5174\u8da3\u53ef\u4ee5\u76f4\u63a5\u770b\u539f\u6587\uff1ahttps:\/\/spinningup.openai.com\/en\/latest\/spinningup\/rl_intro3.html<\/p>\n\n\n\n

\u5148\u521b\u5efa\u73af\u5883\uff1a<\/p>\n\n\n\n

<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
import<\/span> gym<\/span><\/span>\nenv <\/span>=<\/span> gym.make(<\/span>'<\/span>CartPole-v1<\/span>'<\/span>)<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u6e38\u620f\u73af\u5883\u7684\u7684observation\u75314\u7ef4\u6570\u7ec4\u8868\u793a\uff0c\u81f3\u4e8e\u6bcf\u7ef4\u8868\u793a\u4ec0\u4e48\u610f\u601d\u5e76\u4e0d\u91cd\u8981\uff0c\u53ef\u4ee5\u91c7\u53d6\u7684\u52a8\u4f5c\u53ea\u6709\u5411\u5de6\u3001\u5411\u53f3\u6240\u4ee5\u52a8\u4f5c\u662f2\u7ef4\u7684\u3002policy\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u795e\u7ecf\u7f51\u7edc\uff0c\u628a\u5b83\u5b9a\u4e49\u6210\u5982\u4e0b\u5f62\u5f0f<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
import<\/span> torch<\/span><\/span>\nimport<\/span> torch.nn <\/span>as<\/span> nn<\/span><\/span>\nclass<\/span> <\/span>PolicyNet<\/span>(<\/span>nn<\/span>.<\/span>Module<\/span>):<\/span><\/span>\n    <\/span>def<\/span> <\/span>__init__<\/span>(<\/span>self<\/span>, <\/span>obs_space<\/span>, <\/span>act_space<\/span>):<\/span><\/span>\n        <\/span>super<\/span>().<\/span>__init__<\/span>()<\/span><\/span>\n        <\/span>self<\/span>.linear1 <\/span>=<\/span> nn.Linear(obs_space, <\/span>32<\/span>)<\/span><\/span>\n        <\/span>self<\/span>.activate <\/span>=<\/span> nn.Tanh()<\/span><\/span>\n        <\/span>self<\/span>.linear2 <\/span>=<\/span> nn.Linear(<\/span>32<\/span>, act_space)<\/span><\/span>\n    <\/span><\/span>\n    <\/span>def<\/span> <\/span>forward<\/span>(<\/span>self<\/span>, <\/span>x<\/span>):<\/span><\/span>\n        x <\/span>=<\/span> <\/span>self<\/span>.linear1(x)<\/span><\/span>\n        x <\/span>=<\/span> <\/span>self<\/span>.activate(x)<\/span><\/span>\n        x <\/span>=<\/span> <\/span>self<\/span>.linear2(x)<\/span><\/span>\n        <\/span>return<\/span> x<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u7f51\u7edc\u7684\u8f93\u51fa\u6ca1\u6709\u63a5softmax\uff0c\u540e\u9762\u5728\u6784\u9020\u5206\u5e03\u7684\u65f6\u5019\u8981\u6ce8\u610f\u4e0b\u3002\u5b9e\u4f8b\u5316\u65f6\u4f20\u5165\u4e24\u4e2a\u53c2\u6570\uff0c\u7b2c\u4e00\u4e2a\u53c2\u6570\u662fobservation\u7684\u7ef4\u5ea6\uff0c\u7b2c\u4e8c\u4e2a\u53c2\u6570\u662faction\u7684\u7ef4\u5ea6\uff0c\u5f53\u524d\u73af\u5883\u5206\u522b\u662f4\u548c2\uff1a<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
policy_net <\/span>=<\/span> PolicyNet(obs_dim, n_act)<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u6709\u4e86\u8fd9\u4e2apolicy_net\u6211\u4eec\u5c31\u53ef\u4ee5\u6839\u636e\u5f53\u524d\u7684obs\u53d6\u6837\u4e00\u4e2a\u52a8\u4f5c\uff1a<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
def<\/span> <\/span>get_action<\/span>(<\/span>obs<\/span>):<\/span><\/span>\n    logits_output <\/span>=<\/span> policy_net(obs)<\/span><\/span>\n    act_distribution <\/span>=<\/span> Categorical(<\/span>logits<\/span>=<\/span>logits_output)<\/span><\/span>\n    <\/span>return<\/span> act_distribution.sample().item()<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u63a5\u4e0b\u6765\u770bgradient\u7684\u8ba1\u7b97\uff1a<\/p>\n\n\n\n
$$\\nabla _{\\theta}J(\\pi_{\\theta}) = E_{\\tau \\sim \\pi_{\\theta}} \\left [ \\sum_{t=0}^{T} \\nabla_{\\theta} \\log \\pi_{\\theta} (a_t | s_t)R(\\tau)\\right ]$$<\/p>\n\n\n\n
\u4e00\u79cd\u76f4\u89c9\u7684\u7406\u89e3\uff0c\u518d\u67d0\u4e2a\u72b6\u6001$s_t$\u4e0b\u6267\u884c\u67d0\u4e2a\u52a8\u4f5c$a_t$\uff0c\u5982\u679c\u8fd9\u4e2a\u52a8\u4f5c\u53ef\u4ee5\u5446\u4e86\u6bd4\u8f83\u9ad8\u7684\u6536\u76ca\u6211\u4eec\u5c31\u5e94\u8be5\u63d0\u5347\u5b83\uff08\u52a0\u5927\u8be5\u72b6\u6001\u6267\u884c\u8be5\u52a8\u4f5c\u7684\u6982\u7387\uff09\u8fd9\u6837\u4e0b\u6b21\u53d6\u6837\u5c31\u66f4\u6709\u53ef\u80fd\u53d6\u5230\u8fd9\u4e2a\u52a8\u4f5c\u4e86\u3002\u5982\u4f55\u6765\u8bc4\u4ef7$s_t$\u4e0b\u6267\u884c\u52a8\u4f5c$a_t$\u7684\u6536\u76ca\u597d\u574f\u5462\uff1f\u6211\u4eec\u53ef\u4ee5\u7528\u672c\u5c40\u6e38\u620f\u7684\u603b\u6536\u76ca\u6765\u8861\u91cf\uff0c\u53ef\u80fd\u4e0d\u597d\uff0c\u4f46\u662f\u80fd\u7528\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5982\u679c\u6211\u4eec\u8ba1\u7b97$s_0$\u5230$s_t$\u4e2d\u95f4\u6709\u591a\u5c11\u6b65\u6765\u4f5c\u4e3a\u8fd9\u5c40\u6e38\u620f\u7684\u6536\u76ca\uff0c\u4f7f\u7528$\\pi_{\\theta}$\u6267\u73a9\u4e86\u4e00\u5c40\u53ef\u80fd\u662f\u8fd9\u6837\u7684<\/p>\n\n\n\n
$$s_0, a_0^1, s_1^1,a_1^1,s_2^1,a_2^1,s_3^1,a_3^1,s_4^1,a_4^1,s_T$$<\/p>\n\n\n\n
\u4e8e\u662f\u8fd9\u4e00\u5c40\u7684\u6536\u76ca\u662f5,\u4e0a\u68071\u8868\u793a\u8fd9\u662f\u7b2c\u4e00\u5c40\u7684\u72b6\u6001\u8f6c\u79fb\u548c\u52a8\u4f5c\u9009\u62e9\uff0c\u8fd8\u662f\u7528\u8fd9\u4e2apolicy\u7b2c\u4e8c\u5c40\u662f\u8fd9\u6837\u7684<\/p>\n\n\n\n
$$s_0, a_0^2, s_1^2,a_1^2,s_2^2,a_2^2,s_3^2,a_3^2,s_4^2,a_4^2,s_5^2,a_5^2,s_6^2,a_6^2,s_7^2,a_7^2,s_T$$<\/p>\n\n\n\n
\u90a3\u4e48\u7b2c\u4e8c\u5c40\u7684\u6536\u76ca\u662f8 ,\u6240\u4ee5\u7b2c\u4e8c\u5c40\u6253\u7684\u6bd4\u8f83\u597d\uff0c\u5e94\u8be5\u5f3a\u5316\u7b2c\u4e8c\u5c40\u4e2d\u6bcf\u4e2a\u72b6\u6001\u4e0b\u6267\u884c\u7684\u90a3\u4e2a\u52a8\u4f5c\u7684\u6982\u7387\uff0c\u8ba1\u7b97\u68af\u5ea6\u7684\u65f6\u5019\u7b2c\u4e8c\u5c40\u5199\u6210<\/p>\n\n\n\n
$$\\nabla _{\\theta} \\log \\pi_{\\theta}(s_0|a_0^2) \\times 8 + \\nabla _{\\theta} \\log \\pi_{\\theta}(s_1|a_1^2) \\times 8 + \\nabla _{\\theta} \\log \\pi_{\\theta}(s_2|a_2^2) \\times 8 …$$<\/p>\n\n\n\n
\u800c\u7b2c\u4e00\u5c40\u662f<\/p>\n\n\n\n
$$\\nabla _{\\theta} \\log \\pi_{\\theta}(s_0|a_0^1) \\times 5 + \\nabla _{\\theta} \\log \\pi_{\\theta}(s_1|a_1^1) \\times 5 + \\nabla _{\\theta} \\log \\pi_{\\theta}(s_2|a_2^1) \\times 5 …$$<\/p>\n\n\n\n
\u663e\u7136\u7b2c\u4e8c\u5c40\u4e58\u4ee58\u7684\u90a3\u90e8\u5206\u5728\u68af\u5ea6\u66f4\u65b0\u4e2d\u66f4\u6709\u4f18\u52bf\uff0c\u4e8e\u662f\u7b2c\u4e8c\u5c40\u6267\u884c\u7684\u90a3\u4e9b\u52a8\u4f5c\u90fd\u5f97\u5230\u4e86\u5f3a\u5316\uff0c\u5f53\u7136\u8fd9\u4e2a\u53ea\u662f\u80fd\u7528\u800c\u5df2\u8fd8\u4e0d\u591f\u597d\uff0c\u6700\u7406\u60f3\u7684\u80af\u5b9a\u662f$Q_{\\pi}(s,a)$\u4f46\u662f\u8fd9\u4e2a\u4e0d\u597d\u7b97\u554a\uff0c\u5148\u5c06\u5c31\u7740\u7528\u5427\uff5e\u65e2\u7136\u6211\u4eec\u8fbe\u6210\u5171\u8bc6\uff0c\u5c31\u770b\u5728pytorch\u4e2d\u600e\u4e48\u5b9e\u73b0\u4e0a\u8ff0\u903b\u8f91\u4e86\u3002<\/p>\n\n\n\n
\u516c\u5f0f\u7ed9\u51fa\u7684gradient\u6211\u4eec\u4e0d\u7528\u5b83\uff0c\u56e0\u4e3aPyTorch\u8fd9\u4e00\u7c7b\u7684\u8ba1\u7b97\u56fe\u6846\u67b6\u7684\u7f16\u7a0b\u903b\u8f91\u662f\u6211\u4eec\u63d0\u4f9bobjective function\u5b83\u7ed9\u81ea\u52a8\u7b97\u68af\u5ea6\uff0c\u4e8e\u662f\u6211\u4eec\u4e2agradient\u8be5\u5199\u6210objective\u5f62\u5f0f<\/p>\n\n\n\n
$$J(\\pi_{\\theta}) = E_{\\tau \\sim \\pi_{\\theta}} [\\log \\pi_{\\theta} (a_t|s_t) R(\\tau)]$$<\/p>\n\n\n\n
\u54b1\u4e0d\u5173\u5fc3\u8fd9\u4e2aobjective\u5230\u5e95\u662f\u5565\u610f\u601d\uff0c\u5b83\u53ea\u662f\u7528\u6765\u5728PyTorch\u4e2d\u8ba1\u7b97\u68af\u5ea6\u6765\u53cd\u5411\u4f20\u64ad\u7528\u7684<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
def<\/span> <\/span>compute_loss<\/span>(<\/span>obs<\/span>, <\/span>act<\/span>, <\/span>weights<\/span>):<\/span><\/span>\n    logits_output <\/span>=<\/span> policy_net(obs)<\/span><\/span>\n    actions_dist <\/span>=<\/span> Categorical(<\/span>logits<\/span>=<\/span>logits_output)<\/span><\/span>\n    logp <\/span>=<\/span> actions_dist.log_prob(act)<\/span><\/span>\n    <\/span>return<\/span> <\/span>-<\/span>1<\/span> <\/span>*<\/span> (logp <\/span>*<\/span> weights).mean()<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u63a5\u6536\u4e09\u4e2a\u53c2\u6570\uff0c\u7b2c\u4e00\u4e2a\u662f\u5f53\u524d$\\pi_{\\theta}$\u73a9\u7684\u90a3\u597d\u591a\u597d\u591a\u5c40\u6e38\u620f\u7684\u6240\u6709observation\uff0c\u7b2c\u4e8c\u4e2a\u662f\u8fd9\u4e9bobservation\u51fa\u73b0\u65f6\u6240\u91c7\u53d6\u7684\u52a8\u4f5c\uff0c\u7b2c\u4e09\u4e2a\u662f\u5bf9\u8fd9\u4e2a\u52a8\u4f5c\u7684\u8bc4\u4ef7\uff0c\u4e5f\u5c31\u662f\u6240\u5728\u7684\u90a3\u5c40\u6e38\u620f\u7684\u6536\u76ca\u3002\u8fd8\u662f\u4e0a\u9762\u7684\u4f8b\u5b50\uff0c\u8f93\u5165\u7684\u5185\u5bb9\u5c31\u662f\uff1a<\/p>\n\n\n\n
$$obs = s_0,s_1^1,s_2^1,s_3^1,s_4^1,s_0,s_1^2,s_2^2,s_3^2,s_4^2,s_5^2,s_6^2,s_7^2$$<\/p>\n\n\n\n
$$act=a_0^1,a_1^1,a_2^1,a_3^1,a_4^1,a_0^2,a_1^2,a_2^2,a_3^2,a_4^2,a_5^2,a_6^2,a_7^2$$<\/p>\n\n\n\n
$$weights=5,5,5,5,5,8,8,8,8,8,8,8,8$$<\/p>\n\n\n\n
\u7ed3\u675f\u72b6\u6001$s_T$\u7528\u4e0d\u4e0a\uff0c\u9010\u884c\u89e3\u91ca\u4e0b<\/p>\n\n\n\n
    \n
  1. \u9996\u5148\u4f7f\u7528policy_net\u83b7\u53d6\u6bcf\u4e2aobservation\u4e0b\u4e24\u4e2a\u52a8\u4f5c\u7684logits(\u7ecf\u8fc7softmax\u540e\u5f97\u5230\u6982\u7387)<\/li>\n\n\n\n
  2. \u5b9a\u4e49Categorical\u5206\u5e03\uff0c\u8fd9\u91cc\u8981\u6307\u5b9a\u53c2\u6570logits,\u9ed8\u8ba4\u53c2\u6570\u63a5\u6536\u7684\u662f\u6982\u7387<\/li>\n\n\n\n
  3. \u83b7\u5f97\u5f53\u65f6\u6e38\u620f\u65f6\u52a8\u4f5c\u7684log probability<\/li>\n\n\n\n
  4. \u6700\u540e\u8ba1\u7b97\u5747\u503c\u7684\u65f6\u5019\u7ed9\u52a0\u4e86\u4e00\u4e2a\u8d1f\u53f7\uff0c\u8fd9\u662f\u56e0\u4e3apytorch\u7684\u4f18\u5316\u5668\u9ed8\u8ba4\u662f\u7528\u6765\u505agradient descent\u7684\u3002<\/li>\n<\/ol>\n\n\n\n
    \u63a5\u4e0b\u6765\u7684\u5de5\u4f5c\u5c31\u662f\u8ba9\u5de5\u7a0b\u5faa\u73af\u8d77\u6765\uff0c\u8fed\u4ee3\u66f4\u65b0policy\u3002\u5927\u81f4\u8fc7\u7a0b\u5c31\u662f\uff0c\u4f7f\u7528\u5f53\u524d\u7684policy\u00a0\uff0c\u73a9\u5f88\u591a\u5f88\u591a\u5c40\u6e38\u620f\uff0c\u8fd9\u91cc\u662f\u901a\u8fc7step\u6765\u9650\u5236\u7684\u73a9\u591a\u4e45\uff0c\u6bd4\u5982\u73a95000\u6b65\uff0c\u6e38\u620f\u8fc7\u7a0b\u4e2d\u91c7\u7528\u7684act\u7531$\\pi_{\\theta}$\u51b3\u5b9a\uff0c\u8bb0\u5f55\u4e0b5000\u6b65\u7684[observation, action, weight] \u4ea4\u7ed9compute_loss()\u51fd\u6570\u8ba1\u7b97\u51faloss\uff0c\u4f7f\u7528pytorch\u63d0\u4f9b\u7684\u4f18\u5316\u5668\u6765\u66f4\u65b0policy\u3002\u6bcf\u66f4\u65b0\u4e00\u6b21policy\u5c31\u5f97\u91cd\u65b0\u53d6\u68375000\u6b65\uff0c\u8fd9\u4e2a\u5927\u6982\u662f\u8981\u5f3a\u8c03\u4e00\u4e0b\u7684\u3002\u8fd9\u4e2a\u53d6\u6837\u7684\u6b65\u6570\u4e0d\u80fd\u592a\u5c11\uff0c\u8981\u4e0d\u7136\u5f88\u591aobservation\u548caction\u90fd\u6ca1\u6709\u53d6\u6837\u5230\u3002\u6211\u628a\u4ee3\u7801\u8d34\u51fa\u6765\uff0cSpinning Up\u5df2\u7ecf\u7ed9\u4ee3\u7801\u589e\u52a0\u4e86\u6bd4\u8f83\u8be6\u7ec6\u7684\u6ce8\u91ca\uff0c\u65b0\u7248\u672c\u7684gym\u4f1a\u6709\u9002\u914d\u95ee\u9898\uff0c\u8fd9\u91cc\u4fee\u6539\u4e86\u4e0b\u3002<\/p>\n\n\n\n
    <\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
    # for training policy<\/span><\/span>\ndef<\/span> <\/span>train_one_epoch<\/span>():<\/span><\/span>\n    <\/span># make some empty lists for logging.<\/span><\/span>\n    batch_obs <\/span>=<\/span> []          <\/span># for observations<\/span><\/span>\n    batch_acts <\/span>=<\/span> []         <\/span># for actions<\/span><\/span>\n    batch_weights <\/span>=<\/span> []      <\/span># for R(tau) weighting in policy gradient<\/span><\/span>\n    batch_rets <\/span>=<\/span> []         <\/span># for measuring episode returns<\/span><\/span>\n    batch_lens <\/span>=<\/span> []         <\/span># for measuring episode lengths<\/span><\/span>\n<\/span>\n    <\/span># reset episode-specific variables<\/span><\/span>\n    obs, info <\/span>=<\/span> env.reset()       <\/span># first obs comes from starting distribution<\/span><\/span>\n    done <\/span>=<\/span> <\/span>False<\/span>            <\/span># signal from environment that episode is over<\/span><\/span>\n    <\/span># \u7528\u6765\u88c5\u586b\u6bcf\u4e00\u6b65\u7684reward\uff0c\u4e00\u5c40\u6e38\u620f\u6253\u5b8c\u4e4b\u540esum\u4e00\u5c31\u662f\u8fd9\u5c40\u6e38\u620f\u7684\u6536\u76ca(\u4e0a\u9762\u7684\u4f8b\u5b505\u548c8)<\/span><\/span>\n    ep_rews <\/span>=<\/span> []            <\/span># list for rewards accrued throughout ep <\/span><\/span>\n<\/span>\n    <\/span># render first episode of each epoch<\/span><\/span>\n    finished_rendering_this_epoch <\/span>=<\/span> <\/span>False<\/span><\/span>\n<\/span>\n    <\/span># collect experience by acting in the environment with current policy<\/span><\/span>\n    <\/span># \u53d6\u6837\u5faa\u73af<\/span><\/span>\n    <\/span>while<\/span> <\/span>True<\/span>:<\/span><\/span>\n<\/span>\n        <\/span># rendering<\/span><\/span>\n        <\/span>if<\/span> (<\/span>not<\/span> finished_rendering_this_epoch) <\/span>and<\/span> render:<\/span><\/span>\n            env.render()<\/span><\/span>\n<\/span>\n        <\/span># save obs<\/span><\/span>\n        batch_obs.append(obs.copy())<\/span><\/span>\n<\/span>\n        <\/span># act in the environment<\/span><\/span>\n        act <\/span>=<\/span> get_action(torch.as_tensor(obs, <\/span>dtype<\/span>=<\/span>torch.float32))<\/span><\/span>\n        <\/span># old version<\/span><\/span>\n        <\/span># obs, rew, done, _ = env.step(act)<\/span><\/span>\n  <\/span># version 0.26.2<\/span><\/span>\n  obs, reward, done, truncated, info <\/span>=<\/span> env.step(act)<\/span><\/span>\n<\/span>\n        <\/span># save action, reward<\/span><\/span>\n        batch_acts.append(act)<\/span><\/span>\n        ep_rews.append(rew)<\/span><\/span>\n<\/span>\n        <\/span>if<\/span> done:<\/span><\/span>\n            <\/span># if episode is over, record info about episode<\/span><\/span>\n            <\/span># \u672c\u5c40\u6e38\u620f\u5df2\u7ecf\u6253\u5b8c\u4e86\uff0c \u5f00\u59cb\u7edf\u8ba1\u6548\u679c<\/span><\/span>\n            ep_ret, ep_len <\/span>=<\/span> <\/span>sum<\/span>(ep_rews), <\/span>len<\/span>(ep_rews)<\/span><\/span>\n            batch_rets.append(ep_ret)<\/span><\/span>\n            batch_lens.append(ep_len)<\/span><\/span>\n<\/span>\n            <\/span># the weight for each logprob(a|s) is R(tau)<\/span><\/span>\n            batch_weights <\/span>+=<\/span> [ep_ret] <\/span>*<\/span> ep_len<\/span><\/span>\n<\/span>\n            <\/span># reset episode-specific variables<\/span><\/span>\n            obs, done, ep_rews <\/span>=<\/span> env.reset(), <\/span>False<\/span>, []<\/span><\/span>\n<\/span>\n            <\/span># won't render again this epoch<\/span><\/span>\n            finished_rendering_this_epoch <\/span>=<\/span> <\/span>True<\/span><\/span>\n<\/span>\n            <\/span># end experience loop if we have enough of it<\/span><\/span>\n            <\/span># \u5982\u679c\u5df2\u7ecf\u5230\u53d6\u6837\u7684step\u6570\u91cf\uff0c\u9000\u51fa\u53d6\u6837\u5faa\u73af\uff0c\u66f4\u65b0policy<\/span><\/span>\n            <\/span>if<\/span> <\/span>len<\/span>(batch_obs) <\/span>><\/span> batch_size:<\/span><\/span>\n                <\/span>break<\/span><\/span>\n<\/span>\n    <\/span># take a single policy gradient update step<\/span><\/span>\n    <\/span># \u66f4\u65b0policy<\/span><\/span>\n    optimizer.zero_grad()<\/span><\/span>\n    batch_loss <\/span>=<\/span> compute_loss(<\/span>obs<\/span>=<\/span>torch.as_tensor(batch_obs, <\/span>dtype<\/span>=<\/span>torch.float32),<\/span><\/span>\n                              <\/span>act<\/span>=<\/span>torch.as_tensor(batch_acts, <\/span>dtype<\/span>=<\/span>torch.int32),<\/span><\/span>\n                              <\/span>weights<\/span>=<\/span>torch.as_tensor(batch_weights, <\/span>dtype<\/span>=<\/span>torch.float32)<\/span><\/span>\n                              )<\/span><\/span>\n    batch_loss.backward()<\/span><\/span>\n    optimizer.step()<\/span><\/span>\n    <\/span>return<\/span> batch_loss, batch_rets, batch_lens<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
    \u4ee5\u4e0a\u5c31\u662f\u6700\u6700\u5355\u7eaf\u7684policy gradient\u7684\u5b9e\u73b0\uff0c\u4e0a\u9762\u8d34\u51fa\u6765\u7684\u7f51\u9875\u4e2d\u8fd8\u544a\u8bc9\u6211\u4eec\u4e00\u4e9btricks\uff0c\u6709\u5174\u8da3\u53ef\u4ee5\u81ea\u5df1\u770b\u4e0b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
    \u5148\u6765\u56de\u5fc6\u4e00\u4e0b\u51e0\u4e2a\u53d8\u91cf\u7684\u5b9a\u4e49\uff0cPolicy Gradient\u7684\u5173\u952e\u662f\u901a\u8fc7Gradient\u6765\u66f4\u65b0Policy $$\\theta_{k+1} = \\theta_{k} + a \\nabla _{\\theta}J(\\pi_{\\theta})|_{\\theta_k}$$ \u5176\u4e2d$\\pi_{\\theta}$\u662f\u53c2\u6570\u8bdd\u7684policy\uff0c$\\theta$\u662f\u5b83\u7684\u7cfb\u6570\uff0c$J(\\pi_{\\theta})$\u7528\u6765\u8861\u91cf\u5f53\u524dpolicy $\\pi_{\\theta}$\u7684\u6027\u80fd\uff0c\u54b1\u4eec\u8fd9\u91cc\u7528$\\pi_{\\theta}$\u7684\u671f\u671b\u6536\u76ca$E_{\\tau \\sim \\pi_{\\theta}}[R(\\tau)]$\u4f5c\u4e3apolicy\u7684\u6027\u80fd\uff0c$R(\\tau)$\u8868\u793a\u4e00\u5c40\u6e38\u620f\u7684\u6536\u76ca,$\\tau \\sim \\pi_{\\theta}$\u8868\u793a\u662f\u5728\u5f53\u524dpolicy $\\pi_{\\theta}$\u4e0b\u3002 $\\nabla _{\\theta}J(\\pi_{\\theta})$\u7b49\u4e8e\u4e0b\u9762\u8fd9\u4e00\u4e32 $$\\nabla _{\\theta}J(\\pi_{\\theta}) = E_{\\tau \\sim \\pi_{\\theta}} \\left [ \\sum_{t=0}^{T} \\nabla_{\\theta} […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[18],"tags":[],"class_list":["post-567","post","type-post","status-publish","format-standard","hentry","category-reinforcement-learning"],"_links":{"self":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/567","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=567"}],"version-history":[{"count":13,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/567\/revisions"}],"predecessor-version":[{"id":582,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=\/wp\/v2\/posts\/567\/revisions\/582"}],"wp:attachment":[{"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=567"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=567"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/tensorzen.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=567"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}