{"id":3756,"date":"2017-12-02T15:52:38","date_gmt":"2017-12-02T07:52:38","guid":{"rendered":"https:\/\/www.zlaire.net\/events\/previous\/?p=3756"},"modified":"2018-10-14T22:46:19","modified_gmt":"2018-10-14T14:46:19","slug":"thomas-bolander-data-2017-12-7","status":"publish","type":"post","link":"https:\/\/www.zlaire.net\/events\/previous\/2017\/12\/thomas-bolander-data-2017-12-7\/","title":{"rendered":"Thomas Bolander: Learning to Plan from Raw Data in Grid-based Games [2017-12-7]"},"content":{"rendered":"

\u897f\u6eaa\u903b\u8f91\u8bba\u575b\u7b2c76\u671f<\/strong><\/p>\n

\"\"<\/a>Speaker:<\/strong>\u00a0Thomas Bolander\u00a0(Technical University of Denmark)<\/a>
\nDate & Time:<\/strong>\u00a07 December 2017 (Thursday), 18:30 \u2013 20:30
\nPlace:<\/strong> Room 259, Main Teaching Building, Zhejiang University
\nTitle:<\/strong> Learning to Plan from Raw Data in Grid-based Games<\/p>\n

Abstract:<\/strong>
\nAn agent that autonomously learns to act in its environment\u00a0must acquire a model of the domain dynamics. This can be\u00a0a challenging\u00a0task, especially in real-world domains, where\u00a0observations are high-dimensional (e.g. pixels) and noisy. Recent methods in deep reinforcement learning for\u00a0games learn from vector representations of pixel observations. However, they typically do not acquire an environment\u00a0model, but a policy for one-step action selection. Even when\u00a0a model is\u00a0learned, it cannot generalize to unseen instances\u00a0of the training domain. Here we propose a neural network-based method that learns\u00a0from high-dimensional visual observations an approximate, compact, implicit representation\u00a0of the domain dynamics, which can be\u00a0used for planning with\u00a0standard search algorithms, and generalizes to novel domain\u00a0instances. We evaluate our approach on visual versions of\u00a0the\u00a0standard domain Sokoban, and show that it learns a transition model that can be successfully used to solve new levels\u00a0of the game.<\/p>\n

\"\"<\/a>\"\"<\/a><\/strong>\"\"\"\"<\/a><\/strong><\/p>\n

\u4f1a\u540e\u62a5\u9053\uff1a<\/strong><\/p>\n

12\u67087\u65e5\uff0c\u4f5c\u4e3aBRaD\u76842017\u5e74\u7cfb\u5217\u6d3b\u52a8\u4e4b\u4e00\u7684\u7b2c76\u671f \u201c\u897f\u6eaa\u903b\u8f91\u8bba\u575b\u201d\uff0c\u6765\u81ea\u4e39\u9ea6\u6280\u672f\u5927\u5b66\u5e94\u7528\u6570\u5b66\u548c\u8ba1\u7b97\u673a\u79d1\u5b66\u7684\u526f\u6559\u6388Thomas Bolander\uff0c\u5206\u4eab\u4e86\u4ed6\u6765\u81ea\u795e\u7ecf\u7f51\u7edc\u548c\u673a\u5668\u5b66\u4e60\u7684\u7814\u7a76\u6210\u679c\uff0c\u5728\u6d59\u6c5f\u5927\u5b66\u8bed\u8a00\u4e0e\u8ba4\u77e5\u7814\u7a76\u4e2d\u5fc3\u4e3e\u529e\u4e86\u4e00\u573a\u9898\u4e3a\u201c\u57fa\u4e8e\u641c\u7d22\u89c4\u5212\u548c\u673a\u5668\u5b66\u4e60\u7684\u7f51\u683c\u6e38\u620f\u6c42\u89e3\u201d\u7684\u8bb2\u5ea7\u3002<\/p>\n

\u7f51\u683c\u6e38\u620f\u6a21\u62df\u7684\u662f\u5177\u6709\u56fa\u5b9a\u683c\u5c40\u3001\u52a8\u4f5c\u7ed3\u679c\u53d6\u51b3\u4e8e\u73a9\u5bb6\u7b56\u7565\u7684\u4e00\u7c7b\u573a\u666f\uff0c\u8fd9\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u7c7b\u4f3c\u4e8e\u81ea\u4e3b\u5b66\u4e60\u7684agent\u5728\u52a8\u6001\u73af\u5883\u4e2d\u7684\u5de5\u4f5c\u673a\u5236\u3002Bolander\u4ee5\u63a8\u7bb1\u5b50\u6e38\u620f\u4e3a\u4f8b\uff0c\u5c1d\u8bd5\u7528\u673a\u5668\u5b66\u4e60\u7684\u65b9\u6cd5\u5bf9\u8be5\u6e38\u620f\u5728\u4e0d\u540c\u683c\u5c40\u4e0b\u7684\u6c42\u89e3\u5efa\u7acb\u4e00\u4e2a\u901a\u7528\u6a21\u578b\u3002\u57fa\u4e8eagent\u5728\u4efb\u4f55\u72b6\u6001\u4e0b\u90fd\u4eb2\u77e5\u81ea\u5df1\u7f51\u683c\u5750\u6807\u7684\u5047\u8bbe\uff0c\u4ed6\u9996\u5148\u5efa\u7acb\u4e86\u6e38\u620f\u5e95\u5c42\u7684\u903b\u8f91\u6846\u67b6\uff0c\u8fd9\u662f\u4e00\u4e2a\u7531\u72b6\u6001\u96c6\u3001\u52a8\u4f5c\u96c6\u3001\u76ee\u6807\u96c6\u548c\u8f6c\u79fb\u51fd\u6570\u7ec4\u6210\u7684\u5143\u7ec4\u3002agent\u7684\u52a8\u4f5c\u89c4\u5212\u662f\u57fa\u4e8e\u51b3\u7b56\u641c\u7d22\u6280\u672f\u3002\u6839\u636e\u6e38\u620f\u89c4\u5219\uff0cagent\u7684\u6bcf\u4e00\u4e2a\u52a8\u4f5c\u53ea\u5f71\u54cd\u5230\u90bb\u8fd1\u5355\u4f4d\u7f51\u683c\u5176\u4e2d\u4e4b\u4e00\uff0c\u56e0\u6b64\uff0c\u53ea\u9700\u8981\u5bf9\u5168\u5c40\u4e2d\u5c40\u90e8\u7f51\u683c\u8fdb\u884c\u8ba1\u7b97\u5e76\u6bd4\u5bf9\u4e0e\u524d\u4e00\u4e2a\u52a8\u4f5c\u4e4b\u95f4\u7684\u5dee\u522b\u3002\u968f\u540e\uff0cBolander\u5f15\u5165\u4e86\u795e\u7ecf\u7f51\u7edc\u4f7fagent\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u3002\u4e3a\u4e86\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\uff0cBolander\u5bf9\u56fe\u50cf\u50cf\u7d20\u8fdb\u884c\u964d\u7ef4\u5e76\u5c1d\u8bd5\u4f7f\u7528\u5404\u79cd\u641c\u7d22\u7b56\u7565\u3002\u7ecf\u8fc7\u8fd0\u884c\u6d4b\u8bd5\uff0cBolander\u5efa\u7acb\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u81f3\u5c11\u53ef\u4ee5\u5b9e\u73b0\u5728\u516d\u4e07\u6b65\u4ee5\u5185\u5bf930*30\u7f51\u683c\u4e2d8\u4e2a\u7bb1\u5b50\u7684\u6e38\u620f\u767e\u5206\u767e\u6c42\u89e3\u3002Bolander\u8fd8\u53d1\u73b0\uff0c\u5c06\u4e00\u6b65\u5f0f\u641c\u7d22\u548c\u89c4\u5212\u641c\u7d22\u7ed3\u5408\u7684\u641c\u7d22\u7b56\u7565\u5c06\u5927\u5927\u63d0\u9ad8\u673a\u5668\u5b66\u4e60\u7684\u6548\u7387\u3002<\/p>\n

Bolander\u7684\u7814\u7a76\u548c\u53d1\u73b0\uff0c\u4e3a\u5927\u6570\u636e\u80cc\u666f\u4e0b\u7684\u63a8\u7406\u4e0e\u51b3\u7b56\u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u548c\u65b9\u6cd5\uff0c\u4e5f\u662f\u5f53\u524d\u70ed\u95e8\u7684\u673a\u5668\u5b66\u4e60\u5728\u6df1\u5316\u5b9e\u8df5\u548c\u5e94\u7528\u65b9\u5411\u4e0a\u7684\u4e00\u4e2a\u63a8\u8fdb\u3002
\n(\u674e\u5d07\u6167 \u62a5\u9053)<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"

\u897f\u6eaa\u903b\u8f91\u8bba\u575b\u7b2c76\u671f Speaker:\u00a0Thomas Bolander\u00a0(Technical University of Denmark) Date & Time:\u00a07 December 2017 (Thursday), 18:30 \u2013 20:30 Place: Room 259, Main Teaching Building, Zhejiang University Title: Learning to Plan from Raw Data in Grid-based Games Abstract: An agent that autonomously learns to …<\/span> Continue reading →<\/span><\/a><\/p>\n","protected":false},"author":11,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3756","post","type-post","status-publish","format-standard","hentry","category-invited-talks"],"_links":{"self":[{"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/posts\/3756","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/users\/11"}],"replies":[{"embeddable":true,"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/comments?post=3756"}],"version-history":[{"count":12,"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/posts\/3756\/revisions"}],"predecessor-version":[{"id":3844,"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/posts\/3756\/revisions\/3844"}],"wp:attachment":[{"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/media?parent=3756"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/categories?post=3756"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.zlaire.net\/events\/previous\/wp-json\/wp\/v2\/tags?post=3756"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}