{"id":43032,"date":"2023-06-09T16:56:18","date_gmt":"2023-06-09T08:56:18","guid":{"rendered":"https:\/\/swarma.org\/?p=43032"},"modified":"2023-06-09T16:56:18","modified_gmt":"2023-06-09T08:56:18","slug":"%e6%9f%b3%e6%98%80%e5%93%b2%ef%bc%9a%e5%ad%a6%e4%b9%a0%e6%99%ba%e8%83%bd-%e5%a6%82%e4%bd%95%e9%ab%98%e6%95%88%e5%9c%b0%e5%ad%a6%e4%b9%a0%ef%bc%9f-neuroai-%e8%af%bb%e4%b9%a6%e4%bc%9a","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=43032","title":{"rendered":"\u67f3\u6600\u54f2\uff1a\u5b66\u4e60\u667a\u80fd\u2014\u2014\u5982\u4f55\u9ad8\u6548\u5730\u5b66\u4e60\uff1f| NeuroAI \u8bfb\u4e66\u4f1a\u5206\u4eab\u6574\u7406"},"content":{"rendered":"
\n
\n
\n
\n


<\/p>\n


<\/p>\n

<\/p>\n


<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/p>\n

\u5bfc\u8bed\uff1a\u65e0\u8bba\u662f\u5bf9\u4e8e\u751f\u7269\u667a\u80fd\u8fd8\u662f\u4eba\u5de5\u667a\u80fd\uff0c\u5b66\u4e60\u90fd\u662f\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\u3002\u90a3\u4ec0\u4e48\u662f\u5b66\u4e60\u5462\uff1f\u5728\u96c6\u667a\u4ff1\u4e50\u90e8\u300cNeuroAI \u8bfb\u4e66\u4f1a<\/a>\u300d\u4e2d\uff0c\u67f3\u6600\u54f2\u8001\u5e08\u8ba4\u4e3a\uff0c\u4ece\u5f3a\u5316\u5b66\u4e60\u7684\u89d2\u5ea6\uff0c\u5b66\u4e60\u5373\u662f\u6784\u5efa\u72b6\u6001\u7a7a\u95f4\uff0c\u5e76\u57fa\u4e8e\u72b6\u6001\u7a7a\u95f4\u505a\u51fa\u4e00\u7cfb\u5217\u9009\u62e9\uff0c\u76ee\u7684\u662f\u8d8b\u5229\u907f\u5bb3\u3002\u67f3\u8001\u5e08\u8fdb\u4e00\u6b65\u4ecb\u7ecd\u4e86\u4e24\u79cd\u5f3a\u5316\u5b66\u4e60\u65b9\u5f0f\uff1a\u65e0\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\uff08model-free RL\uff09\u4e0e\u57fa\u4e8e\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\uff08model-based RL\uff09\uff0c\u5e76\u8ba4\u4e3a\u5728\u4eba\u5de5\u667a\u80fd\u4e0e\u751f\u7269\u667a\u80fd\u4e2d\uff0c\u4e24\u8005\u4e92\u76f8\u7ed3\u5408\u53ef\u4ee5\u63d0\u9ad8\u5b66\u4e60\u8868\u73b0\u3002<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span>\u672c\u6587\u603b\u7ed3\u81ea2022\u5e7412\u670818\u65e5\u67f3\u6600\u54f2\u8001\u5e08\u5728\u96c6\u667a\u4ff1\u4e50\u90e8 NeuroAI \u8bfb\u4e66\u4f1a\u7684\u5206\u4eab\u62a5\u544a\u3002<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/p>\n


<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/p>\n

\u5173\u952e\u8bcd\uff1a\u4eba\u5de5\u667a\u80fd\uff0c\u751f\u7269\u667a\u80fd\uff0c\u5f3a\u5316\u5b66\u4e60\uff0c\u72b6\u6001\u7a7a\u95f4<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/strong><\/span><\/p>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

\u5f20\u94c4\u3001\u9b4f\u6668<\/span><\/span><\/span> <\/span>| \u4f5c\u8005<\/strong><\/span><\/span><\/p>\n

\u5218\u6cc9\u5f71\u3001\u674e\u4e39\u3001\u5f20\u5f1b<\/span><\/span><\/span> <\/span>| \u6821\u5bf9<\/span><\/strong><\/span><\/p>\n

\u795e\u7ecf\u8ba1\u7b97\u4e0e\u63a7\u5236\u5b9e\u9a8c\u5ba4<\/span><\/span> <\/span>| \u6765\u6e90<\/span><\/strong><\/span><\/span><\/strong><\/span><\/p>\n


<\/p>\n<\/section>\n

\n

<\/span><\/section>\n
\u8bb2\u8005\u4ecb\u7ecd\uff1a<\/span><\/strong><\/span>\u67f3\u6600\u54f2\u8001\u5e08\u662f\u5317\u4eac\u5e08\u8303\u5927\u5b66\u8ba4\u77e5\u795e\u7ecf\u79d1\u5b66\u4e0e\u5b66\u4e60\u56fd\u5bb6\u91cd\u70b9\u5b9e\u9a8c\u5ba4 & \u5317\u4eac\u8111\u79d1\u5b66\u4e0e\u7c7b\u8111\u7814\u7a76\u4e2d\u5fc3\u7814\u7a76\u5458\u3002\u67f3\u6600\u54f2\u5b9e\u9a8c\u5ba4\u81f4\u529b\u4e8e\u89e3\u6790\u4eba\u7c7b\u9ad8\u7ea7\u667a\u80fd\u884c\u4e3a\u7684\u8ba1\u7b97\u548c\u795e\u7ecf\u673a\u5236<\/span>\uff08\u76ee\u524d\u4e3b\u8981\u5173\u6ce8\u8ba4\u77e5\u5730\u56fe\u7684\u5f62\u6210\u4e0e\u53d1\u5c55\uff09<\/span>\u548c\u5f00\u53d1\u65b0\u578b\u7684\u795e\u7ecf\u7f16\u89e3\u7801\u6a21\u578b\u548c\u8111\u673a\u63a5\u53e3\uff0c\u4e3a\u8111\u75be\u75c5\u548c\u7cbe\u795e\u75be\u75c5\u7684\u8bca\u7597\u4e0e\u8c03\u63a7\u63d0\u4f9b\u65b0\u7684\u624b\u6bb5\u3002<\/span><\/section>\n
<\/span><\/section>\n

<\/section>\n<\/section>\n
\n
\n
\n
\n
<\/span><\/section>\n
\n

 <\/span>
<\/p>\n<\/section>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

\u7b80\u4ecb<\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

<\/section>\n
<\/span><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\u4ee5\u4e0b\u662f\u67f3\u6600\u54f2\u8001\u5e08\u5bf9\u672c\u6b21\u5206\u4eab\u7684\u7b80\u4ecb\uff1a<\/span><\/section>\n

<\/span><\/section>\n
\u4ec0\u4e48\u662f\u5b66\u4e60\uff1f\u4ece\u5f3a\u5316\u5b66\u4e60\u7684\u89d2\u5ea6\uff0c\u5373\u6784\u5efa\u72b6\u6001\u7a7a\u95f4<\/span><\/span><\/strong>\uff08state space\uff09<\/span><\/span>\uff0c\u5e76\u57fa\u4e8e\u72b6\u6001\u7a7a\u95f4\u505a\u51fa\u7684\u4e00\u7cfb\u5217\u9009\u62e9\uff0c\u76ee\u7684\u662f\u8d8b\u5229\u907f\u5bb3<\/span><\/strong>\u3002\u8fd9\u4e5f\u662f\u751f\u7269\u667a\u80fd\u6700\u6838\u5fc3\u7684\u529f\u80fd<\/strong>\u3002\u672c\u6b21\u5206\u4eab\u6211\u5c06\u4ece\u5f3a\u5316\u5b66\u4e60\u5c55\u5f00\uff0c\u89e3\u8bfb\u5b66\u4e60\u7684\u6a21\u5f0f\uff0c\u4ecemodel-free\u5230model-based\uff1b\u6211\u5c06\u8ba8\u8bba\u5982\u4f55\u6784\u5efa\u9ad8\u6548\u7684\u72b6\u6001\u7a7a\u95f4\uff0c\u4ee5\u4fc3\u8fdb\u5b66\u4e60\u3001\u51b3\u7b56\u548c\u6cdb\u5316\u3002\u5168\u7a0b\u5c06\u6ce8\u91cd\u795e\u7ecf\u673a\u5236<\/span>\uff08\u7279\u522b\u662f\u795e\u7ecf\u8868\u5f81\uff09<\/span>\u548c\u5b66\u4e60\u667a\u80fd\u7684\u5bf9\u5e94\u3002<\/p>\n

<\/span><\/section>\n

\u5728\u62a5\u544a\u6700\u5f00\u5934\uff0c\u67f3\u6600\u54f2\u8001\u5e08\u4e3e\u4e86\u56db\u4e2a\u5fc3\u7406\u5b66\u7684\u7ecf\u5178\u7814\u7a76\u6765\u5f15\u51fa\u672c\u6b21\u62a5\u544a\u201cHow to learn efficiently\u201d\u7684\u4e3b\u8981\u5185\u5bb9\u3002\u8fd9\u56db\u4e2a\u7ecf\u5178\u7814\u7a76\u5206\u522b\u5bf9\u5e94\u4e86\u672c\u6b21\u62a5\u544a\u7684\u56db\u4e2a\u65b9\u9762\uff1a<\/span><\/section>\n

<\/span><\/section>\n
1. Learning actions directly from reward<\/span>\uff08\u76f4\u63a5\u4ece\u5956\u52b1\u5b66\u4e60\uff09<\/span><\/section>\n
2. Learning from reward in complex state spaces<\/span> (\u5728\u590d\u6742\u7684\u72b6\u6001\u7a7a\u95f4\u4e2d\u5b66\u4e60)<\/span><\/section>\n
3. What do those state spaces look like\uff1f<\/span>(\u72b6\u6001\u7a7a\u95f4\u7684\u7ed3\u6784\u662f\u600e\u6837\u7684)<\/span><\/section>\n
4. Can we make better state spaces to make learning easy\uff1f<\/span>(\u5982\u4f55\u6784\u5efa\u66f4\u597d\u7684\u72b6\u6001\u7a7a\u95f4\u4f7f\u5b66\u4e60\u66f4\u52a0\u9ad8\u6548)<\/span><\/section>\n

<\/span><\/section>\n
5. Making state spaces that are good for other things too<\/span> (like planning) (\u72b6\u6001\u7a7a\u95f4\u7528\u4e8e\u5176\u4ed6\u4efb\u52a1\uff0c\u5982\u89c4\u5212\u4efb\u52a1)<\/span><\/section>\n
6. Making state spaces from bits of other state spaces<\/span> (\u4ece\u5176\u4ed6\u72b6\u6001\u7a7a\u95f4\u6784\u5efa\u65b0\u7684\u72b6\u6001\u7a7a\u95f4)<\/span><\/section>\n
 <\/span><\/section>\n
\u4ee5\u4e0b\u6211\u4eec\u4e5f\u5c06\u4ece\u8fd9\u516d\u4e2a\u65b9\u9762\u6765\u5bf9\u8fd9\u6b21\u67f3\u6600\u54f2\u8001\u5e08\u5206\u4eab\u62a5\u544a\u8fdb\u884c\u603b\u7ed3\u3002<\/span><\/section>\n<\/section>\n
\n
\n
\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n
\n
<\/span><\/section>\n
\n

 <\/span>
<\/p>\n<\/section>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

\u62a5\u544a\u5206\u4eab<\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
1. Learning actions directly from reward (\u76f4\u63a5\u4ece\u5956\u52b1\u5b66\u4e60)<\/span><\/strong><\/span><\/section>\n

<\/span><\/section>\n
\u5728\u62a5\u544a\u4e2d\u67f3\u8001\u5e08\u63d0\u5230\u4e86\u5f3a\u5316\u5b66\u4e60[2]<\/span>\uff08Reinforcement Learning\uff0c RL\uff09<\/span>\u4e0e\u9a6c\u5c14\u79d1\u592b\u51b3\u7b56\u8fc7\u7a0b<\/span>\uff08Markov Decision Process\uff0cMDP\uff09<\/span>\uff0c\u5e76\u4e3e\u4e86\u4e00\u4e2a\u201cPI\u664b\u5347\u201d\u7684\u4f8b\u5b50\u6765\u8bf4\u660e\u73b0\u5b9e\u4e2d\u5982\u4f55\u5229\u7528\u5956\u52b1<\/span>\uff08reward\uff09<\/span>\u6765\u5b66\u4e60\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u8fd9\u91cc\u6211\u4eec\u5bf9\u6bd4\u4e00\u4e0b\u76d1\u7763\u5b66\u4e60\u4e0e\u5f3a\u5316\u5b66\u4e60\uff0c\u6765\u7406\u89e3\u2018\u76f4\u63a5\u5229\u7528\u5956\u52b1\u6765\u5b66\u4e60\u2019\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u76d1\u7763\u5b66\u4e60<\/span><\/strong>\u7684\u5b66\u4e60\u4fe1\u53f7\u662f\u2018\u4f8b\u5b50\u2019<\/span>\uff08examplar\uff09<\/span>\uff0c\u901a\u8fc7\u6700\u5c0f\u5316\u635f\u5931\u51fd\u6570\u6765\u4f18\u5316\u6a21\u578b\u3002\u6700\u597d\u7684\u76d1\u7763\u5b66\u4e60\u6a21\u578b\u5e94\u8be5\u80fd\u591f\u5b8c\u7f8e\u6a21\u4eff\u8bad\u7ec3\u6570\u636e<\/span>\uff08\u5373\uff0c\u5f97\u5230\u4e0e\u8bad\u7ec3\u6570\u636elabel\u4e00\u81f4\u7684\u8f93\u51fa\uff09<\/span>\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5f3a\u5316\u5b66\u4e60<\/span><\/strong>\u7684\u5b66\u4e60\u4fe1\u53f7\u662f\u2018\u5956\u52b1\u2019<\/span>\uff08reward\uff09<\/span>\uff0c\u901a\u8fc7\u8d1d\u5c14\u66fc\u65b9\u7a0b\u6765\u8fdb\u884cvalue\u8fed\u4ee3\u3001\u6700\u5c0f\u5316\u8bef\u5dee<\/span>\uff08value\u51fd\u6570\u4e0e\u771f\u5b9e\u7d2f\u79ef\u5956\u52b1\u7684\u8bef\u5dee\uff09<\/span>\u4ece\u800c\u4f18\u5316\u6a21\u578b\u3002\u6700\u597d\u7684\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\u4e0d\u662f\u80fd\u591f\u5b8c\u7f8e\u6a21\u4eff\u8bad\u7ec3\u6570\u636e\uff0c\u800c\u662f\u80fd\u591f\u901a\u8fc7\u8bad\u7ec3\u6570\u636e\u7684\u5956\u52b1\u6765\u627e\u5230\u4e00\u4e2a\u53ef\u4ee5\u6700\u5927\u5316\u7d2f\u79ef\u5956\u52b1\u7684\u884c\u52a8\u7b56\u7565<\/span>\uff08\u7c7b\u4f3c\u4e8e\u5f3a\u5316\u5b66\u4e60\u6a21\u578b\u4f1a\u66f4\u52a0\u504f\u5411\u4e8e\u9009\u62e9\u80fd\u83b7\u5f97\u66f4\u591a\u5956\u52b1\u7684action\uff09<\/span>\u3002<\/span><\/section>\n
 <\/span><\/section>\n
\u3010NCC Lab \u6ce8\u3011<\/span><\/strong><\/span><\/section>\n
\u8d1d\u5c14\u66fc\u65b9\u7a0b<\/span>\uff08Bellman Equation\uff09<\/span>\uff0c\u4e5f\u88ab\u79f0\u4f5c\u52a8\u6001\u89c4\u5212\u65b9\u7a0b<\/span>\uff08Dynamic Programming Equation\uff09<\/span>\uff0c\u7531\u7406\u67e5\u00b7\u8d1d\u5c14\u66fc<\/span>\uff08Richard Bellman\uff09<\/span>\u63d0\u51fa\u3002\u8d1d\u5c14\u66fc\u65b9\u7a0b\u662f\u52a8\u6001\u89c4\u5212<\/span>\uff08Dynamic Programming\uff09<\/span>\u6570\u5b66\u4f18\u5316\u65b9\u6cd5\u80fd\u591f\u8fbe\u5230\u6700\u4f18\u5316\u7684\u5fc5\u8981\u6761\u4ef6\u3002\u6b64\u65b9\u7a0b\u628a\u201c\u51b3\u7b56\u95ee\u9898\u5728\u7279\u5b9a\u65f6\u95f4\u600e\u4e48\u7684\u503c\u201d\u4ee5\u201c\u6765\u81ea\u521d\u59cb\u9009\u62e9\u7684\u62a5\u916c\u6bd4\u4ece\u521d\u59cb\u9009\u62e9\u884d\u751f\u7684\u51b3\u7b56\u95ee\u9898\u7684\u503c\u201d\u7684\u5f62\u5f0f\u8868\u793a\u3002\u56e0\u6b64\uff0c\u501f\u6b64\u8fd9\u4e2a\u65b9\u5f0f\uff0c\u8d1d\u5c14\u66fc\u65b9\u7a0b\u628a\u52a8\u6001\u6700\u4f18\u5316\u95ee\u9898\u53d8\u6210\u7b80\u5355\u7684\u5b50\u95ee\u9898\uff0c\u7528\u6765\u7b80\u5316\u5f3a\u5316\u5b66\u4e60\u6216\u8005\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u95ee\u9898\u3002\u8d1d\u5c14\u66fc\u65b9\u7a0b\u5982\u4e0b\uff1a<\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

<\/section>\n
2. Learning from reward in complex state spaces (\u5728\u590d\u6742\u7684\u72b6\u6001\u7a7a\u95f4\u4e2d\u5b66\u4e60)<\/span><\/strong><\/span><\/section>\n

<\/span><\/section>\n
\u5728\u590d\u6742\u7684\u72b6\u6001\u7a7a\u95f4\u4e2d\u5b66\u4e60\uff0c\u9700\u8981\u9488\u5bf9\u4e0d\u540c\u7684\u73af\u5883\u548c\u9700\u6c42\uff0c\u91c7\u7528\u4e0d\u540c\u7684\u5b66\u4e60\u65b9\u5f0f\u3002<\/span><\/section>\n
\u5728\u62a5\u544a\u4e2d\u67f3\u8001\u5e08\u8bf4\u660e\u4e86\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u65e0\u6a21\u578b\u5f3a\u5316\u5b66\u4e60<\/span> (model-free RL) <\/span>\u4e0e\u57fa\u4e8e\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60<\/span> (model-based RL)<\/span>\uff0c\u5e76\u5f15\u7528\u591a\u7bc7\u6587\u7ae0\u6765\u8bf4\u660e\u5927\u8111\u7684\u5b66\u4e60\u4e2d\u540c\u65f6\u5e94\u7528\u4e86\u8fd9\u4e24\u79cd\u5f3a\u5316\u5b66\u4e60\u65b9\u5f0f\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u4e0b\u9762\u6211\u4eec\u7b80\u5355\u4ecb\u7ecd\u4e0bmodel-free RL\u4e0emodel-based RL\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u65e0\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60<\/span><\/strong>\uff1a\u5b83\u53ea\u7f16\u7801\u72b6\u6001-\u52a8\u4f5c\u5bf9\u4e0e\u4ef7\u503c\u51fd\u6570\u7684\u6620\u5c04\u5173\u7cfb (s,a)\u2192Q(s,a)\uff0c\u5373\u53ea\u4f1a\u57fa\u4e8e\u5956\u52b1R\uff0c\u5b66\u4e60\u5728\u67d0\u4e2a\u72b6\u6001s\u4e0b\u505a\u51fa\u67d0\u4e2a\u52a8\u4f5ca\u6240\u5bf9\u5e94\u7684\u52a8\u4f5c\u72b6\u6001\u4ef7\u503c\u51fd\u6570Q(s,a)\u3002\u5728model-free RL\u5b66\u4e60\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u53ea\u662f\u4f1a\u901a\u8fc7\u65f6\u95f4\u5dee\u5206\u5b66\u4e60(Temporal Differences learning\uff0cTD learning)\u6765\u66f4\u65b0\u4ef7\u503c\u51fd\u6570\uff0c\u6bcf\u6b21\u5b66\u4e60\u53ea\u66f4\u65b0\u76f8\u5e94\u72b6\u6001\u7684\u4ef7\u503c\u51fd\u6570\u3002<\/span><\/section>\n

<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n
 <\/span><\/section>\n
\u57fa\u4e8e\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60<\/span><\/strong>\uff1a\u9664\u4e86\u7f16\u7801\u4ef7\u503c\u51fd\u6570\u4e4b\u5916\uff0c\u8fd8\u7f16\u7801\u4e86\u73af\u5883\u7684\u72b6\u6001\u8f6c\u79fb\u6982\u7387\u3002\u5728model-based RL\u5bf9\u5e94\u7684\u5b66\u4e60\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u72b6\u6001\u8f6c\u79fb\u6982\u7387\u6765\u5bf9\u6240\u6709\u72b6\u6001\u7684\u4ef7\u503c\u51fd\u6570\u8fdb\u884c\u66f4\u65b0\u3002<\/span><\/section>\n<\/section>\n
\n
 <\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

<\/section>\n
\u8fd9\u4e24\u79cd\u5b66\u4e60\u65b9\u5f0f\u6709\u5176\u5404\u81ea\u7684\u4f18\u52a3\u3002<\/span><\/section>\n

<\/span><\/section>\n
Model-free RL\u5341\u5206\u7b80\u5355<\/span>\uff08\u8ba4\u77e5\u8d1f\u8377\u4f4e\uff09<\/span>\uff0c\u4f46\u4ef7\u503c\u51fd\u6570\u7684\u66f4\u65b0\u8f83\u6162\uff0c<\/span>\uff08\u6570\u636e\uff09<\/span>\u6548\u7387\u4f4e\uff0c\u5728\u7a33\u5b9a\u7684\u73af\u5883\u66f4\u52a0\u51c6\u786e\uff0c\u4f46\u662f\u4e0d\u9002\u7528\u4e8e\u4e0d\u65ad\u53d8\u5316\u7684\u73af\u5883\u3002<\/span><\/section>\n

<\/span><\/section>\n
Model-based RL\u8f83\u4e3a\u590d\u6742<\/span>\uff08\u8ba4\u77e5\u8d1f\u8377\u9ad8\uff09<\/span>\uff0c\u6a21\u578b\u7684\u590d\u6742\u5ea6\u4f1a\u968f\u7740\u72b6\u6001\u7684\u589e\u52a0\u4ee5\u6307\u6570\u7ea7\u7684\u589e\u52a0\uff0c\u4f46<\/span>\uff08\u6570\u636e\uff09<\/span>\u6548\u7387\u9ad8\uff0c\u53ef\u4ee5\u6210\u529f\u5e94\u7528\u4e8e\u4e0d\u65ad\u53d8\u5316\u7684\u73af\u5883\u4e2d\u3002<\/span><\/section>\n
 <\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

\u56fe1\uff1aMastering the game of Go with deep neural networks and tree search, Silver et al (2016)[4]<\/span><\/p>\n


<\/span><\/section>\n
\u5728\u4eba\u5de5\u667a\u80fd\u4e0e\u751f\u7269\u667a\u80fd\u4e2d\uff0c\u65e0\u6a21\u578b\u7684\u5b66\u4e60\u4e0e\u57fa\u4e8e\u6a21\u578b\u7684\u5b66\u4e60\u90fd\u53ef\u4ee5\u4e92\u76f8\u7ed3\u5408\u6765\u63d0\u9ad8\u8868\u73b0\u3002<\/span><\/section>\n
 <\/span><\/section>\n
\u5728AlphaGo [Silver et al 2016, Nature][4]\u4e2d\uff0c\u57fa\u4e8e\u6a21\u578b\u7684\u6811\u641c\u7d22\u4e0e\u65e0\u6a21\u578b\u7684Q-learning\u76f8\u7ed3\u5408\uff0c\u901a\u8fc7\u6811\u641c\u7d22N\u6b65\u4e4b\u540e\u505c\u6b62\uff0c\u4e4b\u540e\u7684\u7528\u4ef7\u503c\u51fd\u6570\u66ff\u4ee3\uff0c\u5982\u56fe1\uff0c\u6210\u529f\u5c06\u57fa\u4e8e\u6a21\u578b<\/span>\uff08\u6570\u636e\u6548\u7387\u9ad8\uff0c\u4ef7\u503c\u51fd\u6570\u66f4\u65b0\u901f\u5ea6\u5feb\uff09<\/span>\u548c\u65e0\u6a21\u578b<\/span>\uff08\u590d\u6742\u5ea6\u4f4e\uff0c\u8868\u8fbe\u6027\u5f3a\uff09<\/span>\u7684\u4f18\u52bf\u7ed3\u5408\u8d77\u6765\u3002<\/span><\/section>\n
 <\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

\u56fe2\uff1a\u524d\u989d\u53f6\u548c\u80cc\u5916\u4fa7\u7eb9\u72b6\u4f53\u7cfb\u7edf\u4e4b\u95f4\u57fa\u4e8e\u4e0d\u786e\u5b9a\u6027\u7684\u884c\u4e3a\u63a7\u5236\u7ade\u4e89\uff0cDaw et al (2005)[3]<\/span><\/p>\n


<\/span><\/section>\n
\u5728\u751f\u7269\u667a\u80fd\u4e2d\uff0c\u6709\u8bb8\u591a\u7814\u7a76\u53d1\u73b0\u5927\u8111\u540c\u65f6\u5229\u7528\u4e86\u57fa\u4e8e\u6a21\u578b\u7684\u548c\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\uff0c\u5e76\u4e14\u4e24\u79cd\u7b97\u6cd5\u5728\u5927\u8111\u5f53\u4e2d\u540c\u65f6\u5b58\u5728\u7ade\u4e89\u4e0e\u5408\u4f5c\u7684\u5173\u7cfb\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5728\u56fe2\u7684\u8fd9\u7bc7\u6587\u7ae0\u4e2d<\/span>[Daw et al 2005, Nature Neuroscience]<\/span>[3]\uff0c\u4f5c\u8005\u7528\u5c0f\u9f20\u7684\u884c\u4e3a\u6570\u636e\u8bba\u8bc1\u4e86\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5\u4e0e\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\u7684\u7ade\u4e89\u5173\u7cfb\u662f\u4e0e\u4e0d\u786e\u5b9a\u6027\u76f8\u5173\u7684\uff0c\u5373\u5927\u8111\u4f1a\u5229\u7528\u4e0d\u786e\u5b9a\u6027\u4f4e\u7684\u7b97\u6cd5\uff1a\u5728\u4efb\u52a1\u521d\u671f\uff0c\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5\u80fd\u591f\u66f4\u5feb\u7684\u5b66\u4e60\u5230\u4ef7\u503c\u51fd\u6570\u4e5f\u5c31\u6709\u7740\u66f4\u5c0f\u7684\u4e0d\u786e\u5b9a\u6027\uff1b\u5728\u4efb\u52a1\u540e\u671f\uff0c\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\u80fd\u591f\u5728\u8fd9\u79cd\u7a33\u5b9a\u7684\u73af\u5883\u4e2d\u5b66\u4e60\u5230\u66f4\u52a0\u7a33\u5b9a\u3001\u51c6\u786e\u7684\u4ef7\u503c\u51fd\u6570\uff0c\u6b64\u65f6\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\u7684\u4e0d\u786e\u5b9a\u6027\u66f4\u5c0f\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u5728\u56fe3\u7684\u8fd9\u7bc7\u6587\u7ae0\u4e2d[5]\uff0c\u4f5c\u8005\u7528\u4eba\u7c7b\u88ab\u8bd5\u7684\u884c\u4e3a\u6570\u636e\u8bba\u8bc1\u4e86\u5927\u8111\u5f53\u4e2d\u540c\u65f6\u4f7f\u7528\u4e86\u65e0\u6a21\u578b\u4e0e\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5\uff0c\u56e0\u4e3a\u88ab\u8bd5\u7684\u884c\u4e3a\u6570\u636e\u7684\u6a21\u5f0f\u4e0d\u540c\u4e8e\u7eaf\u7684\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\u6216\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5<\/span>\uff08\u6a21\u62df\u6570\u636e\uff09<\/span>\uff0c\u800c\u66f4\u63a5\u8fd1\u4e8e\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\u4e0e\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5\u7684\u6df7\u5408\u3002<\/span><\/section>\n
 <\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

\u56fe3\uff1a\u57fa\u4e8e\u6a21\u578b\u7684\u5bf9\u4eba\u7c7b\u9009\u62e9\u7684\u5f71\u54cd\u548c\u7eb9\u72b6\u4f53\u9884\u6d4b\u7684\u8bef\u5dee, Daw et al (2011)[5]<\/span><\/p>\n


<\/span><\/section>\n
\u5728\u5927\u8111\u4e2d\uff0c\u65e0\u6a21\u578b\u7684\u7b97\u6cd5\u7684\u5b9e\u73b0\u4e0e\u524d\u989d\u76ae\u8d28\u7cfb\u7edf\u6709\u5173\uff0c\u800c\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5\u7684\u5b9e\u73b0\u4e0e\u80cc\u5916\u4fa7\u7eb9\u72b6\u4f53\u7cfb\u7edf\u6709\u5173\uff0c\u9664\u6b64\u4e4b\u5916\u8fd8\u6709\u8bb8\u591a\u7814\u7a76\u8868\u660e[Miller et al 2017] [6]\u6d77\u9a6c\u4f53\u4e0e\u57fa\u4e8e\u6a21\u578b\u7684\u7b97\u6cd5\u7684\u5b9e\u73b0\u4e5f\u6709\u7d27\u5bc6\u7684\u8054\u7cfb\u3002<\/span><\/section>\n
 <\/span><\/section>\n
3. What do those state spaces look like\uff1f(\u72b6\u6001\u7a7a\u95f4\u7684\u7ed3\u6784\u662f\u600e\u6837\u7684\uff1f)<\/span><\/strong><\/span><\/section>\n

<\/section>\n
\u67f3\u8001\u5e08\u5411\u6211\u4eec\u9610\u8ff0\u4e86\u751f\u7269\u5982\u4f55\u5bf9\u7269\u7406\u7a7a\u95f4\u8fdb\u884c\u5efa\u6a21\uff0c\u4ee5\u89e3\u91ca\u72b6\u6001\u7a7a\u95f4\u7684\u6982\u5ff5\u3002\u5bfb\u8def\u8fc7\u7a0b\u4e2d\uff0c\u751f\u7269\u9700\u8981\u5c06\u81ea\u5df1\u6240\u5904\u7684\u7269\u7406\u7a7a\u95f4\u8fdb\u884c\u5efa\u6a21\uff0c\u9700\u8981\u5bf9\u7279\u5b9a\u5730\u70b9\u8fdb\u884c\u8bb0\u5fc6\uff0c\u5e76\u5224\u65ad\u8ddd\u79bb\uff0c\u8fdb\u884c\u8def\u5f84\u89c4\u5212<\/strong>\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e9b\u76ee\u7684\uff0c\u751f\u7269\u7684\u5927\u8111\u9700\u8981\u62e5\u6709\u4e00\u79cd\u9ad8\u6548\u7684\u795e\u7ecf\u673a\u5236\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u7814\u7a76\u53d1\u73b0\uff0c\u6d77\u9a6c<\/span>\uff08Hippocampus\uff09<\/span>\u4e2d\u7684\u4f4d\u7f6e\u7ec6\u80de<\/strong><\/span>\uff08Place cell\uff09<\/span>\u4e0e\u8be5\u529f\u80fd\u5bc6\u5207\u76f8\u5173\u3002\u968f\u540e\uff0c\u7814\u7a76\u8005\u53c8\u53d1\u73b0\u4e86\u591a\u79cd\u7ec6\u80de\u7c7b\u578b\uff0c\u5305\u62ec\u7f51\u683c\u7ec6\u80de<\/strong><\/span>\uff08Grid Cell\uff09<\/span>\u548c\u65b9\u5411\u7ec6\u80de<\/strong><\/span>\uff08Object Vector Cell\uff09<\/span>\u7b49<\/span>\uff08\u56fe4\uff09<\/span>\u3002\u5728\u8fd9\u4e9b\u7ec6\u80de\u4e2d\uff0c\u4f4d\u7f6e\u7ec6\u80de\u548c\u7f51\u683c\u7ec6\u80de\u662f\u6700\u91cd\u8981\u7684\u4e24\u79cd\uff0c\u5206\u522b\u8d1f\u8d23\u7279\u5b9a\u4f4d\u7f6e\u7684\u8868\u5f81\u548c\u76f8\u5bf9\u4f4d\u7f6e\u7684\u8868\u5f81\u3002\u540e\u7eed\u7684\u7814\u7a76\u53d1\u73b0\uff0c\u8fd9\u4e9b\u7ec6\u80de\u4e0d\u4ec5\u53ef\u4ee5\u8868\u5f81\u7269\u7406\u7a7a\u95f4\uff0c\u8fd8\u53ef\u4ee5\u8868\u5f81\u62bd\u8c61\u7a7a\u95f4\uff0c\u4f8b\u5982\u4eba\u9645\u5173\u7cfb\u548c\u7269\u4f53\u76f8\u4f3c\u6027\u7b49\u3002<\/span><\/section>\n
 <\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n
\u56fe4 \u7269\u7406\u7a7a\u95f4\u8868\u5f81\u76f8\u5173\u7684\u795e\u7ecf\u7ec6\u80de[7][8][9][10][11][12][13][14]
<\/span><\/section>\n

<\/section>\n
4. Can we make better state spaces to make learning easy?\uff08\u5982\u4f55\u6784\u5efa\u66f4\u597d\u7684\u72b6\u6001\u7a7a\u95f4\u4f7f\u5b66\u4e60\u66f4\u52a0\u9ad8\u6548\uff09<\/span><\/strong><\/span><\/section>\n

<\/span><\/section>\n
\u5728\u8fd9\u4e00\u5c0f\u8282\uff0c\u67f3\u8001\u5e08\u4e3b\u8981\u8bb2\u4e86\u6211\u4eec\u5982\u4f55\u5efa\u7acb\u8d77\u66f4\u597d\u7684\u72b6\u6001\u7a7a\u95f4\u3002\u5728\u771f\u5b9e\u4e16\u754c\u4e2d\uff0c\u72b6\u6001\u3001\u52a8\u4f5c\u7a7a\u95f4\u90fd\u662f\u8fde\u7eed\u7684\uff0c\u6211\u4eec\u4e0d\u53ef\u80fd\u4f7f\u7528\u8868\u683c\u6765\u5bf9\u6bcf\u4e00\u4e2a\u72b6\u6001\u8fdb\u884c\u8868\u5f81\u3002\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u5bf9\u4e8e\u8fde\u7eed\u72b6\u6001\u3001\u52a8\u4f5c\u7a7a\u95f4\u7684\u5904\u7406\u5c31\u662f\uff1a\u65b9\u7a0b\u7ea6\u5316<\/span> (Function approximation)<\/span>\u3002\u76ee\u524d\u6700\u706b\u7684\u65b9\u7a0b\u7ea6\u5316\u7684\u65b9\u5f0f\u5c31\u662f\u795e\u7ecf\u7f51\u7edc\uff0c\u795e\u7ecf\u7f51\u7edc+\u5f3a\u5316\u5b66\u4e60=\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60<\/span>\uff08\u56fe5\uff09<\/span>\u3002<\/span><\/section>\n

<\/span><\/section>\n
\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u5b9e\u9645\u4e0a\u5c31\u662f\u6df1\u5ea6\u5b66\u4e60\u5f3a\u5927\u7684\u8868\u5f81\u80fd\u529b\u4e0e\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u7ed3\u5408\uff0c\u4f7f\u5f97\u5f3a\u5316\u5b66\u4e60\u80fd\u591f\u5e94\u7528\u5728\u66f4\u591a\u9886\u57df<\/span><\/strong>\u3002\u4f46\u662f\u6df1\u5ea6\u5b66\u4e60\u4e0e\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u5b66\u4e60\u51fa\u6765\u7684\u72b6\u6001\u8868\u5f81\u5b9e\u9645\u4e0a\u662f\u4e0d\u540c\u7684\uff0c\u539f\u56e0\u5728\u4e8e\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u4ecd\u7136\u4e5f\u662f\u4ece\u5956\u52b1\u4fe1\u53f7\u6765\u5b66\u4e60\u7684\u3002\u76f8\u5173\u7814\u7a76<\/span>[Cross, et al, 2021, Neuron]<\/span>[15]\u4e2d\u53d1\u73b0\uff1a\u4e0e\u57fa\u4e8e\u6df1\u5ea6\u5b66\u4e60\u5f97\u5230\u7684\u72b6\u6001\u8868\u5f81\u76f8\u6bd4\uff0c\u57fa\u4e8e\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60\u5f97\u5230\u7684\u72b6\u6001\u8868\u5f81\u53ef\u80fd\u66f4\u52a0\u63a5\u8fd1\u771f\u5b9e\u5927\u8111\u4e2d\u7684\u72b6\u6001\u8868\u5f81\u3002<\/span><\/section>\n

<\/span><\/section>\n<\/section>\n
\n

<\/p>\n<\/section>\n

\n
\u56fe5\uff1a\u6df1\u5ea6\u5f3a\u5316\u5b66\u4e60<\/span><\/section>\n

<\/span><\/section>\n
5. Making state spaces that are good for other things too (like planning)\uff08\u72b6\u6001\u7a7a\u95f4\u7528\u4e8e\u5176\u4ed6\u4efb\u52a1\uff0c\u5982\u89c4\u5212\u4efb\u52a1\uff09<\/span><\/strong><\/span><\/section>\n

<\/span><\/section>\n

\u5728\u8fd9\u4e00\u5c0f\u8282\uff0c\u67f3\u8001\u5e08\u63d0\u51fa\u4e86\u4e00\u4e2a\u597d\u7684\u72b6\u6001\u8868\u5f81\u7684\u51e0\u4e2a\u7279\u70b9\uff1a<\/span><\/p>\n

1. \u9884\u6d4b\u6027\uff1a\u9884\u6d4b\u7c7b\u4f3c\u7ed3\u679c\u7684\u72b6\u6001<\/span>\uff08\u4f8b\u5982\uff0c\u7ee7\u627f\u8868\u5f81\uff09<\/span><\/p>\n

2. \u4f4e\u7ef4\uff1a\u7ef4\u5ea6\u8d8a\u5c11\uff0c\u5b66\u4e60\u8d8a\u597d<\/span><\/p>\n

3. \u5206\u5c42\uff1a\u5206\u522b\u8868\u5f81\u7c97\u7c92\u5ea6\u6216\u7ec6\u7c92\u5ea6<\/span><\/p>\n

4. \u53ef\u8fc1\u79fb\uff1a\u540c\u4e00\u8868\u5f81\u652f\u6301\u591a\u79cd\u4e0d\u540c\u4efb\u52a1<\/span><\/p>\n

5. \u53ef\u89e3\u8026\/\u53ef\u7ec4\u5408\uff1a\u53ef\u72ec\u7acb\u63a8\u7406\u7684\u7ef4\u5ea6\u662f\u6b63\u4ea4\u7684<\/span><\/p>\n

6. \u5173\u7cfb\uff1a\u8868\u5f81\u4efb\u52a1\u201c\u8fde\u901a\u6027\u201d<\/span><\/section>\n

<\/section>\n
\u5176\u4e2d\u9884\u6d4b\u6027\u8fd9\u4e00\u6807\u51c6\uff1a\u9884\u6d4b\u6027\u7684\u8868\u5f81<\/span> (Successor representations) <\/span>\u80fd\u591f\u66f4\u5feb\u7684\u9002\u5e94\u5956\u52b1\u6761\u4ef6\u53d8\u5316\u7684\u73af\u5883\uff1b\u540c\u65f6\u6709\u7740\u65e0\u6a21\u578b\u5b66\u4e60\u4e0e\u57fa\u4e8e\u57fa\u4e8e\u6a21\u578b\u7684\u4f18\u52bf<\/span>\uff08\u53ef\u4ee5\u901a\u8fc7TD-learning\u66f4\u65b0\u7684\u62d3\u5c55\u6027\uff0c\u5bf9\u72b6\u6001\u8fdb\u884c\u5efa\u6a21\u7684\u7075\u6d3b\u6027\uff09<\/span>\u3002
<\/span><\/section>\n

<\/span><\/section>\n<\/section>\n
\n

<\/p>\n<\/section>\n

\n
\u56fe6\uff1a\u6d77\u9a6c\u4f53\u662f\u4e00\u79cd\u9884\u6d4b\u6027\u5730\u56fe, Stachenfeld et al (2017)[16]<\/span><\/section>\n

<\/section>\n
\u5728\u76f8\u5173\u7814\u7a76\u4e2d\u8868\u660e\uff0c\u89c1<\/span>[Stachenfeld et al, 2017, Nature Neuroscience]<\/span>[16]\uff0c\u4e0e\u5b9e\u65f6\u7684\u8868\u5f81\u76f8\u6bd4\uff0c\u5927\u8111\u5bf9\u4e8e\u4f4d\u7f6e\u7684\u8868\u5f81\u66f4\u63a5\u8fd1\u4e8e\u9884\u6d4b\u6027\u7684\u8868\u5f81<\/span>\uff08\u7f16\u7801\u5f53\u524d\u72b6\u6001\u7684\u7d2f\u8ba1\u9891\u7387\uff09\uff08\u56fe6\uff09<\/span>\u3002<\/span><\/section>\n
 <\/span><\/section>\n
\u5728\u62a5\u544a\u4e2d\u67f3\u8001\u5e08\u8fd8\u63d0\u5230\u4e86\u4e00\u4e2a\u95ee\u9898\uff1aHow do we get these state spaces that our input is sensory?<\/span><\/section>\n

<\/span><\/section>\n
\u4ece\u8d1d\u53f6\u65af\u7edf\u8ba1\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u611f\u5b98\u6570\u636e\u6765\u63a8\u6d4b\u51fa\u9690\u53d8\u91cf<\/span> (Latent variables) <\/span>:<\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

<\/section>\n
\u53e6\u4e00\u65b9\u9762\u6211\u4eec\u8fd8\u53ef\u4ee5\u4f7f\u7528\u751f\u6210\u6a21\u578b\u6765\u5f97\u5230\u72b6\u6001\u8868\u5f81\uff0c\u4f8b\u5982VAE<\/span> [Kingma & Welling 2013]<\/span>[17]<\/span>\uff08\u56fe7\uff09<\/span>\uff0c\u6216GAN <\/span>[Goodfellow et al 2014]<\/span>[18]\u3002
<\/span><\/section>\n

<\/section>\n
\u3010NCC Lab \u6ce8\u3011<\/span><\/strong><\/section>\n
\u751f\u6210\u6a21\u578b<\/span>\uff08Generative model\uff09<\/span>\u662f\u4e00\u79cd\u53ef\u4ee5\u751f\u6210\u6570\u636e\u7684\u6a21\u578b<\/span>\uff08\u751f\u6210\u7684\u6570\u636e\u5e94\u7c7b\u4f3c\u4e8e\u771f\u5b9e\u6570\u636e\uff09<\/span>\u3002\u4f8b\u5982\uff0c\u6211\u4eec\u63a5\u6536\u4e86\u4ece\u5206\u5e03P_data\u53d6\u6837\u7684\u82e5\u5e72\u6837\u672c\u6784\u6210\u6211\u4eec\u7684\u8bad\u7ec3\u96c6\uff0c\u6211\u4eec\u7684\u6a21\u578b\u53ef\u4ee5\u5b66\u4e60\u5230\u4e00\u4e2a\u80fd\u591f\u6a21\u62df\u8fd9\u4e00\u5206\u5e03\u7684\u6982\u7387\u5206\u5e03P_model\u3002\u4ece\u800c\u901a\u8fc7\u5bf9P_model\u91c7\u6837\uff0c\u53ef\u4ee5\u751f\u6210\u4ee5\u5047\u4e71\u771f\u7684\u6570\u636e\u3002\u5e38\u89c1\u7684\u751f\u6210\u6a21\u578b\u6709autoregressive\u6a21\u578b\u3001VAE\u3001GAN\u3001flow\u6a21\u578b\u53ca\u5176\u53d8\u79cd\u7b49\u7b49\u3002<\/span><\/section>\n

<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n
\u56fe7\uff1aVAE (Variational auto-encoder)\u793a\u610f\u56fe, Kingma & Welling (2013)[17]
<\/span><\/section>\n

<\/section>\n
6. Making state spaces from bits of other state spaces\uff08\u4ece\u5176\u4ed6\u72b6\u6001\u7a7a\u95f4\u6784\u5efa\u65b0\u7684\u72b6\u6001\u7a7a\u95f4\uff09<\/span><\/strong><\/span><\/section>\n

<\/span><\/section>\n
\u6700\u540e\u4e00\u90e8\u5206\uff0c\u67f3\u8001\u5e08\u8bb2\u8ff0\u4e86\u5982\u4f55\u901a\u8fc7\u5df2\u6709\u7684\u72b6\u6001\u7a7a\u95f4\u6765\u6784\u5efa\u65b0\u7684\u72b6\u6001\u7a7a\u95f4\u3002\u89e3\u8026\/\u7ec4\u5408<\/span>\uff08Disentangling\/Factorizing\uff09<\/span>\u8868\u5f81\u6307\u72b6\u6001\u7a7a\u95f4\u53ef\u4ee5\u88ab\u5206\u6210\u591a\u4e2a\u72ec\u7acb\u7684\u5b50\u72b6\u6001\u7a7a\u95f4\u3002<\/span><\/section>\n

<\/section>\n
\u5982\u56fe8\u6240\u793a[19]\uff0c\u53ef\u4ee5\u770b\u5230\uff0c\u7269\u4f53\u7684\u6574\u4f53\u7279\u5f81\u53ef\u4ee5\u7531\u8bb8\u591a\u4e2a\u5355\u72ec\u7684\u7279\u5f81\u5171\u540c\u6765\u63cf\u8ff0\u3002\u5728\u5177\u4f53\u5b9e\u73b0\u4e0a\u4e5f\u975e\u5e38\u7b80\u5355\uff0c\u5728\u4e0a\u8ff0\u7684VAE\u7684loss\u4e2d\u52a0\u5165\u5bf9\u9690\u53d8\u91cf\u89e3\u8026\u7a0b\u5ea6\u7684\u60e9\u7f5a\u9879\u5c31\u53ef\u4ee5\u4f7f\u5176\u4e2d\u7684\u5355\u4e2a\u9690\u53d8\u91cf\u80fd\u591f\u5bf9\u5e94\u5355\u4e2a\u7279\u5f81\u3002<\/span><\/section>\n
 <\/span><\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n
\u56fe8 \u89e3\u8026\u72b6\u6001\u7a7a\u95f4, Behrens et al (2018)[19]<\/span><\/section>\n

<\/section>\n<\/section>\n
\n

<\/section>\n
\n
\n
\n

\u53c2\u8003\u6587\u732e<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n
\n
\n

<\/span><\/p>\n

[1] \u3010\u8111\u5ba2\u4e2d\u56fd\u00b7\u79d1\u7814\u3011\u7b2c35\u4f4d\u8bb2\u8005\uff1a\u67f3\u6600\u54f2 \u4eba\u7c7b\u57fa\u4e8e\u8ba4\u77e5\u5730\u56fe\u7684\u7075\u6d3b\u5b66\u4e60\u548c\u51b3\u7b56\u673a\u5236\uff1ahttps:\/\/www.bilibili.com\/video\/BV1hQ4y1U77y\/<\/span><\/section>\n
[2] Sutton, R. S., & Barto, A. G. (1998). Introduction to reinforcement learning (Vol. 135, pp. 223-260). Cambridge: MIT press.<\/span><\/section>\n
[3] Daw, N. D., Niv, Y., & Dayan, P. (2005). Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nature neuroscience, 8(12), 1704-1711.<\/span><\/section>\n
[4] Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., … & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. nature, 529(7587), 484-489.<\/span><\/section>\n
[5] Daw, N. D., Gershman, S. J., Seymour, B., Dayan, P., & Dolan, R. J. (2011). Model-based influences on humans’ choices and striatal prediction errors. Neuron, 69(6), 1204-1215.<\/span><\/section>\n
[6] Miller, K. J., Botvinick, M. M., & Brody, C. D. (2017). Dorsal hippocampus contributes to model-based planning. Nature neuroscience, 20(9), 1269-1276.<\/span><\/section>\n
[7] O’keefe, J., & Nadel, L. (1979). Pr\u00e9cis of O’Keefe & Nadel’s The hippocampus as a cognitive map. Behavioral and Brain Sciences, 2(4), 487-494.<\/span><\/section>\n
[8] Hafting, T., Fyhn, M., Molden, S., Moser, M. B., & Moser, E. I. (2005). Microstructure of a spatial map in the entorhinal cortex. Nature, 436(7052), 801-806.<\/span><\/section>\n
[9] Omer, D. B., Maimon, S. R., Las, L., & Ulanovsky, N. (2018). Social place-cells in the bat hippocampus. Science, 359(6372), 218-224.<\/span><\/section>\n
[10] Danjo, T., Toyoizumi, T., & Fujisawa, S. (2018). Spatial representations of self and other in the hippocampus. Science, 359(6372), 213-218.<\/span><\/section>\n
[11] Taube, J. S., Muller, R. U., & Ranck, J. B. (1990). Head-direction cells recorded from the postsubiculum in freely moving rats. I. Description and quantitative analysis. Journal of Neuroscience, 10(2), 420-435.<\/span><\/section>\n
[12] Gauthier, J. L., & Tank, D. W. (2018). A dedicated population for reward coding in the hippocampus. Neuron, 99(1), 179-193.<\/span><\/section>\n
[13] Lever, C., Burton, S., Jeewajee, A., O’Keefe, J., & Burgess, N. (2009). Boundary vector cells in the subiculum of the hippocampal formation. Journal of Neuroscience, 29(31), 9771-9777.<\/span><\/section>\n
[14] Sarel, A., Finkelstein, A., Las, L., & Ulanovsky, N. (2017). Vectorial representation of spatial goals in the hippocampus of bats. Science, 355(6321), 176-180.<\/span><\/section>\n
[15] Cross, L., Cockburn, J., Yue, Y., & O\u2019Doherty, J. P. (2021). Using deep reinforcement learning to reveal how the brain encodes abstract state-space representations in high-dimensional environments. Neuron, 109(4), 724-738.<\/span><\/section>\n
[16] Stachenfeld, K. L., Botvinick, M. M., & Gershman, S. J. (2017). The hippocampus as a predictive map. Nature neuroscience, 20(11), 1643-1653.<\/span><\/section>\n
[17] Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114.<\/span><\/section>\n
[18] Goodfellow, I. J., Shlens, J., & Szegedy, C. (2014). Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572.<\/span><\/section>\n
[19] Behrens, T. E., Muller, T. H., Whittington, J. C., Mark, S., Baram, A. B., Stachenfeld, K. L., & Kurth-Nelson, Z. (2018). What is a cognitive map? Organizing knowledge for flexible behavior. Neuron, 100(2), 490-509.<\/span><\/section>\n


<\/span><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n

\uff08\u53c2\u8003\u6587\u732e\u53ef\u4e0a\u4e0b\u6ed1\u52a8<\/strong>\u67e5\uff09<\/p>\n<\/section>\n<\/section>\n


<\/section>\n

<\/section>\n
\n
\n
\n
\n

NeuroAI\u8bfb\u4e66\u4f1a<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/section>\n
\u795e\u7ecf\u79d1\u5b66\u548c\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u591a\u4f4d\u8457\u540d\u5b66\u8005\u8fd1\u65e5\u53d1\u8868 NeuroAI \u767d\u76ae\u4e66\u8ba4\u4e3a\uff0c\u795e\u7ecf\u79d1\u5b66\u957f\u671f\u4ee5\u6765\u4e00\u76f4\u662f\u63a8\u52a8\u4eba\u5de5\u667a\u80fd\uff08AI\uff09\u53d1\u5c55\u7684\u91cd\u8981\u9a71\u52a8\u529b\uff0cNeuroAI \u9886\u57df\u7684\u57fa\u7840\u7814\u7a76\u5c06\u63a8\u52a8\u4e0b\u4e00\u4ee3\u4eba\u5de5\u667a\u80fd\u7684\u8fdb\u7a0b\u3002\u6587\u7ae0\u53d1\u8868\u540e\u5f15\u53d1\u70ed\u8bae\uff1a\u795e\u7ecf\u79d1\u5b66\u662f\u5426\u63a8\u52a8\u4e86\u4eba\u5de5\u667a\u80fd\uff1f\u672a\u6765\u7684\u4eba\u5de5\u667a\u80fd\u662f\u5426\u9700\u8981\u795e\u7ecf\u79d1\u5b66\uff1f<\/span><\/section>\n

<\/section>\n
\u672c\u7740\u4fc3\u8fdb\u795e\u7ecf\u79d1\u5b66\u3001\u8ba1\u7b97\u673a\u79d1\u5b66\u3001\u8ba4\u77e5\u79d1\u5b66\u548c\u8111\u79d1\u5b66\u7b49\u4e0d\u540c\u9886\u57df\u7684\u5b66\u672f\u5de5\u4f5c\u8005\u7684\u4ea4\u6d41\u4e0e\u5408\u4f5c\uff0c\u96c6\u667a\u4ff1\u4e50\u90e8\u8054\u5408\u5317\u4eac\u5e08\u8303\u5927\u5b66\u67f3\u6600\u54f2\u3001\u5317\u4eac\u5927\u5b66\u9c8d\u5e73\u78ca\u548c\u660c\u5e73\u5b9e\u9a8c\u5ba4\u5415\u67c4\u6c5f\u4e09\u4f4d\u7814\u7a76\u5458\u5171\u540c\u53d1\u8d77\u4e86\u300cNeuroAI\u300d\u8bfb\u4e66\u4f1a<\/a>\uff0c\u805a\u7126\u5728\u89c6\u89c9\u3001\u8bed\u8a00\u548c\u5b66\u4e60\u9886\u57df\u4e2d\u795e\u7ecf\u79d1\u5b66\u4e0e\u4eba\u5de5\u667a\u80fd\u7684\u76f8\u5173\u7814\u7a76\uff0c\u671f\u5f85\u80fd\u591f\u67b6\u8d77\u795e\u7ecf\u79d1\u5b66\u4e0e\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u5408\u4f5c\u6865\u6881\uff0c\u6fc0\u53d1\u8de8\u5b66\u79d1\u7684\u5b66\u672f\u706b\u82b1\u3002<\/span><\/section>\n

<\/section>\n

<\/span><\/span><\/a><\/p>\n


<\/span><\/section>\n
\n

\u8be6\u60c5\u8bf7\u89c1\uff1a<\/span><\/p>\n

NeuroAI \u8bfb\u4e66\u4f1a\u542f\u52a8\uff1a\u63a2\u7d22\u795e\u7ecf\u79d1\u5b66\u4e0e\u4eba\u5de5\u667a\u80fd\u7684\u524d\u6cbf\u4ea4\u53c9\u9886\u57df<\/a><\/span><\/strong>
<\/span><\/strong><\/p>\n


<\/p>\n<\/section>\n<\/section>\n

\n

<\/span><\/strong><\/section>\n<\/section>\n

\u63a8\u8350\u9605\u8bfb<\/span><\/strong><\/strong><\/p>\n

1. <\/strong>\u673a\u5668\u610f\u8bc6\uff1a\u4eba\u5de5\u667a\u80fd\u7684\u7ec8\u6781\u6311\u6218<\/a><\/strong><\/span><\/strong><\/strong><\/section>\n
2.<\/strong> Nat. Commun. \u901f\u9012\uff1aNeuroAI \u50ac\u5316\u4e0b\u4e00\u4ee3\u4eba\u5de5\u667a\u80fd<\/a><\/strong><\/span><\/strong><\/section>\n
3. <\/strong><\/strong>\u5f53\u5927\u8111\u9047\u4e0aAI\uff1a\u4ece\u5927\u8111\u89e3\u7801\u6280\u672f\u5230\u901a\u7528\u4eba\u5de5\u667a\u80fd<\/a><\/strong><\/strong><\/span><\/strong><\/strong><\/section>\n
4. \u300a\u5f20\u6c5f\u00b7\u590d\u6742\u79d1\u5b66\u524d\u6cbf27\u8bb2\u300b\u5b8c\u6574\u4e0a\u7ebf\uff01<\/strong><\/a><\/strong><\/strong><\/strong><\/section>\n
<\/strong><\/a><\/strong><\/strong><\/strong><\/section>\n
<\/strong><\/a><\/strong><\/strong><\/strong><\/section>\n
5. \u6210\u4e3a\u96c6\u667aVIP\uff0c\u89e3\u9501\u5168\u7ad9\u8bfe\u7a0b\uff0f\u8bfb\u4e66\u4f1a<\/span><\/strong><\/a><\/strong><\/strong><\/strong><\/strong><\/section>\n
6.<\/strong><\/strong><\/strong> <\/strong><\/strong><\/strong>\u52a0\u5165\u96c6\u667a\uff0c\u4e00\u8d77\u590d\u6742\uff01<\/strong><\/a><\/strong><\/strong><\/strong><\/span><\/section>\n

<\/section>\n


<\/p>\n

\u70b9\u51fb\u201c\u9605\u8bfb\u539f\u6587\u201d\uff0c\u62a5\u540d\u8bfb\u4e66\u4f1a<\/strong><\/span><\/strong><\/strong><\/span><\/strong><\/section>\n

<\/mp-style-type><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

\u5bfc\u8bed\uff1a\u65e0\u8bba\u662f\u5bf9\u4e8e\u751f\u7269\u667a\u80fd\u8fd8\u662f\u4eba\u5de5\u667a\u80fd\uff0c\u5b66\u4e60\u90fd\u662f\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\u3002\u90a3\u4ec0\u4e48\u662f\u5b66\u4e60\u5462\uff1f\u5728\u96c6\u667a\u4ff1\u4e50\u90e8\u300cNeuroAI \u8bfb\u4e66\u4f1a\u300d\u4e2d\uff0c\u67f3\u6600\u54f2\u8001\u5e08\u8ba4\u4e3a\uff0c\u4ece\u5f3a\u5316\u5b66\u4e60\u7684\u89d2\u5ea6\uff0c\u5b66\u4e60\u5373\u662f\u6784\u5efa\u72b6\u6001\u7a7a\u95f4\uff0c\u5e76\u57fa\u4e8e\u72b6\u6001\u7a7a\u95f4\u505a\u51fa\u4e00\u7cfb\u5217\u9009\u62e9\uff0c\u76ee\u7684\u662f\u8d8b\u5229\u907f\u5bb3\u3002\u67f3\u8001\u5e08\u8fdb\u4e00\u6b65\u4ecb\u7ecd\u4e86\u4e24\u79cd\u5f3a\u5316\u5b66\u4e60\u65b9\u5f0f\uff1a\u65e0\u6a21\u578b\u5f3a\u5316\u5b66\u4e60\uff08model-free RL\uff09…<\/p>\n","protected":false},"author":0,"featured_media":43017,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[1],"tags":[],"special":[],"_links":{"self":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/43032"}],"collection":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=43032"}],"version-history":[{"count":0,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/43032\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/media\/43017"}],"wp:attachment":[{"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=43032"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=43032"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=43032"},{"taxonomy":"special","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fspecial&post=43032"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}