{"id":38129,"date":"2020-06-24T00:00:15","date_gmt":"2020-06-23T16:00:15","guid":{"rendered":"https:\/\/case.ntu.edu.tw\/blog\/?p=38129"},"modified":"2021-12-28T01:17:23","modified_gmt":"2021-12-27T17:17:23","slug":"%e7%9b%ae%e6%a8%99%e5%8d%9a%e5%bc%88%e2%94%80%e2%94%80%e4%ba%ba%e5%b7%a5%e6%99%ba%e6%85%a7%e5%89%b5%e9%80%a0%e5%8a%9b%e7%9a%84%e7%bc%ba%e9%99%b7","status":"publish","type":"post","link":"https:\/\/case.ntu.edu.tw\/blog\/?p=38129","title":{"rendered":"\u76ee\u6a19\u535a\u5f08\u2500\u2500\u4eba\u5de5\u667a\u6167\u5275\u9020\u529b\u7684\u7f3a\u9677"},"content":{"rendered":"
\n
\"\u76ee\u6a19\u535a\u5f08\u2500\u2500\u4eba\u5de5\u667a\u6167\u5275\u9020\u529b\u7684\u7f3a\u9677\"<\/div>\n<\/div>\n
\n
\n

\u7de8\u8b6f\uff0f\u8a31\u5b88\u5091<\/strong><\/p>\n

\u76ee\u6a19\u535a\u5f08<\/strong><\/p>\n

\u5e0c\u81d8\u795e\u8a71\u4e2d\u6709\u9019\u9ebc\u4e00\u500b\u6545\u4e8b\uff1a\u570b\u738bMidas\u5354\u52a9\u9152\u795e\u627e\u56de\u4e86\u9152\u9189\u5931\u8e64\u7684\u8001\u5e2b\uff0c\u9152\u795e\u70ba\u4e86\u5831\u7b54\u4ed6\uff0c\u61c9Midas\u7684\u5fc3\u9858\u8d08\u4e88\u4ed6\u9ede\u77f3\u6210\u91d1\u7684\u8d85\u80fd\u529b\u2500\u2500\u4f46\u4ed6\u5f88\u5feb\u5730\u767c\u73fe\u9019\u500b\u734e\u52f5\u5176\u5be6\u662f\u5834\u707d\u96e3\uff1aMidas\u6240\u89f8\u78b0\u7684\u4efb\u4f55\u6771\u897f\uff0c\u7121\u8ad6\u98df\u7269\u3001\u98f2\u6599\u548c\u5973\u5152\uff0c\u90fd\u8b8a\u6210\u4e86\u91d1\u5b50\u3002\u5728\u73fe\u5be6\u4e16\u754c\u4e2d\uff0c\u985e\u4f3c\u7684\u4e8b\u60c5\u4e5f\u5e38\u5e38\u767c\u751f\uff0c\u4f8b\u5982\u5b78\u751f\u53ef\u80fd\u6703\u70ba\u4e86\u907f\u514d\u53d7\u5230\u8655\u7f70\u6216\u76e1\u5feb\u53ef\u4ee5\u73a9\u5230\u96fb\u73a9\uff0c\u76f4\u63a5\u7167\u6284\u53e6\u4e00\u500b\u540c\u5b78\u7684\u4f5c\u696d\uff0c\u884d\u751f\u51fa\u50c5\u628a\u4f5c\u696d\u7c3f\u4e0a\u7684\u7a7a\u767d\u8655\u586b\u6eff\uff0c\u5be6\u969b\u4e0a\u6c92\u6709\u5b78\u7fd2\u5230\u4efb\u4f55\u6771\u897f\u7684\u60c5\u6cc1\uff0c\u6211\u5011\u5c31\u7a31\u4e4b\u70ba\u300c\u76ee\u6a19\u535a\u5f08\u300d\uff08Specification gaming\uff09\u3002<\/p>\n

\u79d1\u5b78\u5bb6\u5728\u8a2d\u8a08\u4eba\u5de5\u667a\u6167\u6642\u4e5f\u6703\u9047\u5230\u9019\u500b\u554f\u984c\u3002\u8b93\u6211\u5011\u4f86\u770b\u4e00\u500b\u4f8b\u5b50\uff1a\u5728\u758a\u6a02\u9ad8\u7a4d\u6728\u7684\u4efb\u52d9\u4e2d\uff0c\u9810\u671f\u7684\u76ee\u6a19\u662f\u8b93\u7d05\u8272\u7a4d\u6728\u6700\u7d42\u758a\u5728\u85cd\u8272\u7a4d\u6728\u7684\u9802\u90e8\uff1b\u734e\u52f5\u662f\u7576\u667a\u6167\u9ad4\u672a\u89f8\u6478\u7a4d\u6728\u6642\uff0c\u5982\u679c\u7d05\u8272\u7a4d\u6728\u7684\u5e95\u9762\u96e2\u5730\u9ad8\u5ea6\u8d8a\u9ad8\uff0c\u4fbf\u7d66\u4e88\u66f4\u5927\u7684\u734e\u52f5\u3002\u7136\u800c\u53ef\u4ee5\u7531\u4e0b\u5716\u770b\u51fa\uff1a\u667a\u6167\u9ad4\u4e26\u672a\u5982\u9810\u671f\u4f7f\u7528\u6a5f\u68b0\u624b\u81c2\u64bf\u8d77\u7d05\u8272\u7a4d\u6728\u4e26\u5c07\u5176\u653e\u7f6e\u5728\u85cd\u8272\u7a4d\u6728\u7684\u9802\u90e8\uff0c\u800c\u50c5\u662f\u7c21\u55ae\u5730\u5c07\u7d05\u8272\u7a4d\u6728\u7ffb\u9762\u4ee5\u6536\u96c6\u734e\u52f5\u3002\u9019\u6a23\u7684\u884c\u70ba\u78ba\u5be6\u9054\u6210\u4e86\u4efb\u52d9\u5b57\u9762\u4e0a\u7684\u6558\u8ff0\uff08\u4f7f\u7d05\u8272\u7a4d\u6728\u7684\u5e95\u9762\u8b8a\u9ad8\uff09\uff0c\u4f46\u537b\u72a7\u7272\u4e86\u8a2d\u8a08\u8005\u5be6\u969b\u95dc\u5fc3\u7684\u76ee\u6a19\uff08\u5c07\u5176\u5806\u758a\u5728\u85cd\u8272\u7a4d\u6728\u7684\u9802\u90e8\uff09\u3002<\/p>\n

\"\"<\/p>\n

\u5716\u4e00\u3001\u5f37\u5316\u5b78\u7fd2\u667a\u6167\u9ad4\u900f\u904e\u5c07\u7a4d\u6728\u7ffb\u9762\uff0c\u9a19\u53d6\u734e\u52f5\u3002\uff08\u5716\u7247\u4f86\u6e90\uff1aPopov et al., 2017.\uff09<\/p>\n<\/div>\n

\u9019\u6a23\u7684\u884c\u70ba\u5176\u5be6\u76f8\u7576\u5e38\u898b\uff0c\u5c24\u5176\u5728\u5f37\u5316\u5b78\u7fd2\uff08reinforcement learning\uff09\u7684\u667a\u6167\u9ad4\u4e0a\u3002DeepMind\u4fbf\u641c\u96c6\u4e86\u5927\u7d0460\u500b\u5be6\u969b\u767c\u751f\u7684\u6848\u4f8b<\/a>\uff0c\u4e26\u4e14\u8a66\u5716\u5f9e\u9019\u4e9b\u6848\u4f8b\u4e2d\u627e\u51fa\u76ee\u6a19\u535a\u5f08\u7684\u53ef\u80fd\u539f\u56e0\u3001\u5be6\u969b\u767c\u751f\u7684\u60c5\u6cc1\uff0c\u4e26\u63d0\u51fa\u53ef\u80fd\u7684\u89e3\u6c7a\u65b9\u6cd5\u3002<\/p>\n

\u6aa2\u8996\u554f\u984c\u8207\u6210\u56e0<\/strong><\/p>\n

\u6211\u5011\u53ef\u4ee5\u5f9e\u5169\u500b\u4e0d\u540c\u7684\u89d2\u5ea6\u4f86\u6aa2\u8996\u76ee\u6a19\u535a\u5f08\uff1a\u5728\u5efa\u7acb\u5f37\u5316\u5b78\u7fd2\u6f14\u7b97\u6cd5\u6642\uff0c\u4eba\u985e\u7684\u76ee\u6a19\u662f\u8a13\u7df4\u667a\u6167\u9ad4\u5b8c\u6210\u6307\u5b9a\u4efb\u52d9\uff0c\u5982\u4f55\u9054\u6210\u5247\u7121\u95dc\u7dca\u8981\u3002\u56e0\u6b64\u5f9e\u9019\u500b\u89d2\u5ea6\u4f86\u770b\uff0c\u76ee\u6a19\u535a\u5f08\u9019\u500b\u73fe\u8c61\u5176\u5be6\u662f\u4e00\u500b\u597d\u5146\u982d\u2500\u2500\u4ee3\u8868\u7cfb\u7d71\u5df2\u7d93\u627e\u5230\u4e86\u5be6\u73fe\u76ee\u6a19\u7684\u65b9\u6cd5\u3002\u4f46\u662f\u56de\u5230\u524d\u8ff0\u758a\u6a02\u9ad8\u7a4d\u6728\u7684\u4efb\u52d9\uff0c\u540c\u6a23\u7684\u5275\u9020\u529b\u537b\u5e36\u4f86\u9ebb\u7169\uff0c\u56e0\u70ba\u667a\u6167\u9ad4\u662f\u5229\u7528\u6f0f\u6d1e\u4f86\u300c\u9a19\u53d6\u300d\u734e\u52f5\u3002\u7136\u800c\u9019\u6a23\u7684\u884c\u70ba\u4e26\u4e0d\u662f\u5f37\u5316\u5b78\u7fd2\u6f14\u7b97\u6cd5\u672c\u8eab\u7684\u7f3a\u9677\u5c0e\u81f4\uff0c\u800c\u662f\u4efb\u52d9\u8a2d\u8a08\u7684\u554f\u984c\u3002\u4e5f\u5c31\u662f\u8aaa\uff0c\u958b\u767c\u8005\u5982\u4f55\u6b63\u78ba\u5730\u63cf\u8ff0\u4efb\u52d9\uff0c\u5c07\u6210\u70ba\u8a13\u7df4AI\u7cfb\u7d71\u7684\u95dc\u9375\u3002<\/p>\n

\u9664\u4e86\u4efb\u52d9\u8a2d\u8a08\u4e4b\u5916\uff0c\u734e\u52f5\u8a2d\u8a08\u4e5f\u81f3\u95dc\u91cd\u8981\u3002\u76ee\u6a19\u535a\u5f08\u7684\u53ef\u80fd\u6210\u56e0\u4e5f\u5305\u542b\u4e0d\u826f\u7684\u734e\u52f5\u51fd\u6578\uff0c\u50cf\u662f\u4ee5\u667a\u6167\u9ad4\u9054\u6210\u4efb\u52d9\u7684\u904e\u7a0b\u4e2d\u67d0\u4e9b\u7279\u5b9a\u52d5\u4f5c\u4f5c\u70ba\u734e\u52f5\u4f9d\u64da\uff0c\u800c\u4e0d\u662f\u4e00\u9023\u4e32\u884c\u70ba\u7684\u6700\u7d42\u7d50\u679c\u3002\u5728\u758a\u6a02\u9ad8\u7a4d\u6728\u7684\u4efb\u52d9\u4e2d\uff0c\u53ea\u6709\u6307\u5b9a\u7d05\u8272\u7a4d\u6728\u7684\u5e95\u9762\u5fc5\u9808\u9ad8\u65bc\u5730\u9762\u662f\u4e0d\u5920\u7684\uff0c\u56e0\u70ba\u4ee3\u7406\u4eba\u6703\u7c21\u55ae\u5730\u7ffb\u8f49\u7d05\u8272\u7a4d\u6728\u4f86\u5b8c\u6210\u4efb\u52d9\uff0c\u66f4\u5168\u65b9\u4f4d\u7684\u734e\u52f5\u689d\u4ef6\u61c9\u8a72\u5305\u62ec\uff1a\u7d05\u8272\u7a4d\u6728\u7684\u9802\u90e8\u5fc5\u9808\u5728\u5e95\u9762\u4e4b\u4e0a\uff0c\u4e26\u4e14\u5e95\u9762\u8981\u8207\u85cd\u8272\u7a4d\u6728\u7684\u9802\u90e8\u5c0d\u9f4a\u3002\u7136\u800c\u8981\u6eff\u8db3\u5168\u90e8\u7684\u689d\u4ef6\u76f8\u7576\u4e0d\u5bb9\u6613\uff0c\u4e14\u904e\u591a\u7684\u8981\u6c42\u53ef\u80fd\u9069\u5f97\u5176\u53cd\uff0c\u4fc3\u4f7f\u7cfb\u7d71\u53bb\u9078\u64c7\u6b21\u4f73\u7684\u89e3\u6c7a\u65b9\u6848\u3002<\/p>\n

\u56e0\u6b64\uff0c\u8207\u5176\u5617\u8a66\u5efa\u7acb\u51fa\u6db5\u84cb\u6240\u6709\u53ef\u80fd\u7684\u689d\u4ef6\uff0c\u4e0d\u5982\u85c9\u7531\u4eba\u985e\u7684\u53cd\u994b\u4f86\u5b78\u7fd2\u5982\u4f55\u734e\u52f5\uff0c\u56e0\u70ba\u8a55\u4f30\u4efb\u52d9\u662f\u5426\u9054\u6210\u901a\u5e38\u6bd4\u660e\u78ba\u63cf\u8ff0\u4efb\u52d9\u8981\u5bb9\u6613\u5f97\u591a\uff0c\u53ea\u662f\u9019\u7a2e\u4f5c\u6cd5\u4e26\u7121\u6cd5\u5b8c\u5168\u907f\u514d\u76ee\u6a19\u535a\u5f08\u7684\u554f\u984c\uff0c\u7562\u7adf\u5982\u4f55\u64cd\u63a7\u4e26\u611a\u5f04\u4eba\u985e\u8a55\u5be9\u4e26\u975e\u96e3\u4e8b\u3002<\/p>\n

\u672a\u4f86\u7684\u6311\u6230<\/strong><\/p>\n

\u7d9c\u4e0a\u6240\u8ff0\uff0c\u70ba\u4e86\u89e3\u6c7a\u76ee\u6a19\u535a\u5f08\u7684\u554f\u984c\u6211\u5011\u81f3\u5c11\u8981\u514b\u670d\u4e09\u500b\u6311\u6230\uff1a<\/p>\n

    \n
  1. \u6211\u5011\u8981\u5982\u4f55\u8a2d\u8a08\u734e\u52f5\u4f86\u5fe0\u5be6\u5730\u53cd\u61c9\u4efb\u52d9\u7684\u7cbe\u9ad3\uff1f<\/li>\n
  2. \u6211\u5011\u8981\u5982\u4f55\u907f\u514d\u5c0d\u65bc\u4efb\u52d9\u505a\u51fa\u932f\u8aa4\u7684\u96b1\u542b\u5047\u8a2d\uff0c\u6216\u8005\u5982\u4f55\u8a2d\u8a08\u667a\u6167\u9ad4\u4f86\u7cfe\u6b63\u932f\u8aa4\u7684\u5047\u8a2d\u800c\u4e0d\u662f\u6295\u6a5f\u53d6\u5de7\uff1f<\/li>\n
  3. \u6211\u5011\u5982\u4f55\u907f\u514d\u667a\u6167\u9ad4\u947d\u734e\u52f5\u6a5f\u5236\u7684\u6f0f\u6d1e\uff1f<\/li>\n<\/ol>\n

    \u96d6\u7136\u76ee\u524d\u5df2\u7d93\u6709\u5404\u5f0f\u5404\u6a23\u7684\u65b9\u6cd5\u88ab\u63d0\u51fa\uff0c\u4f46\u6211\u5011\u5176\u5be6\u96e2\u5b8c\u5168\u89e3\u6c7a\u76ee\u6a19\u535a\u5f08\u5c1a\u6709\u5f88\u5927\u4e00\u6bb5\u8ddd\u96e2\u3002\u4e26\u4e14\u96a8\u8457AI\u7cfb\u7d71\u8d8a\u4f86\u8d8a\u6709\u80fd\u529b\u767c\u73fe\u4eba\u985e\u7121\u6cd5\u4e8b\u5148\u9810\u671f\u7684\u8a2d\u8a08\u6f0f\u6d1e\uff0c\u9019\u4e9b\u554f\u984c\u5728\u672a\u4f86\u6216\u8a31\u6703\u8b8a\u5f97\u66f4\u5177\u6709\u6311\u6230\u6027\u3002<\/p>\n

    \u00a0<\/strong><\/p>\n

    \u7de8\u8b6f\u4f86\u6e90<\/strong><\/p>\n

    V. Krakovna et al.,\u00a0\u201cSpecification gaming: the flip side of AI ingenuity\u201d<\/a>, DeepMind, 2020.<\/p>\n

    \u53c3\u8003\u8cc7\u6599<\/strong><\/p>\n

      \n
    1. J. Leike et al.,\u00a0\u201cLearning through human feedback\u201d<\/a>, DeepMind, 2017.<\/li>\n
    2. T. Everitt al.,\u00a0\u201cDesigning agent incentives to avoid reward tampering\u201d<\/a>, DeepMind, 2019.<\/li>\n<\/ol>\n

      (\u672c\u6587\u7531\u6559\u80b2\u90e8\u88dc\u52a9\u300cAI\u5831\u5831\u2500AI\u79d1\u666e\u63a8\u5ee3\u8a08\u756b\u300d\u57f7\u884c\u5718\u968a\u7de8\u8b6f)<\/p>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

      \u4eba\u5de5\u667a\u6167\u4e5f\u53ef\u80fd\u6703\u5077\u61f6\u9a19\u4eba\uff1f\u932f\u8aa4\u7684\u8a2d\u8a08\u53ef\u80fd\u6703\u4f7f\u5f97AI\u70ba\u4e86\u7372\u5f97\u734e\u52f5\u800c\u6284\u6377\u5f91\u812b\u8ecc\u6f14\u51fa\uff0c\u505a\u51fa\u4ee4\u4eba\u557c\u7b11\u7686\u975e\u7684\u884c\u70ba\u3002<\/p>\n","protected":false},"author":21,"featured_media":38130,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3772,3293],"tags":[5513,5514],"aioseo_notices":[],"jetpack_featured_media_url":"https:\/\/case.ntu.edu.tw\/blog\/wp-content\/uploads\/2021\/12\/speification-gaming-0.jpg","_links":{"self":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/38129"}],"collection":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/users\/21"}],"replies":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=38129"}],"version-history":[{"count":1,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/38129\/revisions"}],"predecessor-version":[{"id":38131,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/posts\/38129\/revisions\/38131"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=\/wp\/v2\/media\/38130"}],"wp:attachment":[{"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=38129"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=38129"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/case.ntu.edu.tw\/blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=38129"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}