{"id":13430,"date":"2016-12-20T09:00:36","date_gmt":"2016-12-20T00:00:36","guid":{"rendered":"http:\/\/www.techscore.com\/blog\/?p=13430"},"modified":"2018-11-14T16:33:43","modified_gmt":"2018-11-14T07:33:43","slug":"%e6%a9%9f%e6%a2%b0%e5%ad%a6%e7%bf%92%e3%81%ae%e3%83%8f%e3%82%a4%e3%83%91%e3%83%bc%e3%83%91%e3%83%a9%e3%83%a1%e3%83%bc%e3%82%bf%e6%8e%a2%e7%b4%a2-%e3%83%99%e3%82%a4%e3%82%ba%e6%9c%80%e9%81%a9","status":"publish","type":"post","link":"https:\/\/www.techscore.com\/blog\/2016\/12\/20\/%e6%a9%9f%e6%a2%b0%e5%ad%a6%e7%bf%92%e3%81%ae%e3%83%8f%e3%82%a4%e3%83%91%e3%83%bc%e3%83%91%e3%83%a9%e3%83%a1%e3%83%bc%e3%82%bf%e6%8e%a2%e7%b4%a2-%e3%83%99%e3%82%a4%e3%82%ba%e6%9c%80%e9%81%a9\/","title":{"rendered":"\u6a5f\u68b0\u5b66\u7fd2\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u63a2\u7d22 : \u30d9\u30a4\u30ba\u6700\u9069\u5316\u306e\u6d3b\u7528"},"content":{"rendered":"<p>\u3053\u3093\u306b\u3061\u306f\u3001\u5ca1\u5d0e\u3067\u3059\u3002<br \/>\n\u3053\u306e\u8a18\u4e8b\u306f TECHSCORE Advent Calendar 2016 \u306e 20 \u65e5\u76ee\u306e\u8a18\u4e8b\u3067\u3059\u3002<\/p>\n<p>\u5148\u65e5\u3001<a href=\"http:\/\/ibisml.org\/ibis2016\/\" target=\"_blank\">IBISML2016<\/a>(\u7b2c19\u56de\u60c5\u5831\u8ad6\u7684\u5b66\u7fd2\u7406\u8ad6\u30ef\u30fc\u30af\u30b7\u30e7\u30c3\u30d7) \u306e<a href=\"http:\/\/ibisml.org\/ibis2016\/tutorial-detail\/\" target=\"_blank\">\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb<\/a>\u306b\u53c2\u52a0\u3057\u307e\u3057\u305f\u3002<br \/>\n\u305d\u306e\u4e2d\u3067\u300c\u30d9\u30a4\u30ba\u6700\u9069\u5316\u306b\u3088\u308b\u6a5f\u68b0\u5b66\u7fd2\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u300d\u3068\u3044\u3046\u5185\u5bb9\u304c\u9762\u767d\u304b\u3063\u305f\u306e\u3067\u7d39\u4ecb\u3057\u307e\u3059<a href=\"#ref1\"><em>*1<\/em><\/a>\u3002<\/p>\n<h2>\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u30fb\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0<\/h2>\n<p>\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u306b\u304a\u3044\u3066\u306f\u3001\u826f\u3044\u30e2\u30c7\u30eb\u3092\u4f5c\u308b\u305f\u3081\u306b\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf(\u5916\u90e8\u304b\u3089\u4e0e\u3048\u308b\u30e2\u30c7\u30eb\u306e\u30d1\u30e9\u30e1\u30fc\u30bf)\u306e\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u304c\u5fc5\u9808\u3067\u3059\u3002<\/p>\n<p>\u76ee\u7684\u306e\u7cbe\u5ea6\u3084\u6c4e\u5316\u6027\u80fd\u3092\u3082\u3068\u3081\u3066\u3001\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u3092\u7e70\u308a\u8fd4\u3057\u6700\u9069\u306a\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u63a2\u7d22\u3059\u308b\u306e\u3067\u3059\u304c\u3001\u9078\u629e\u3059\u308b\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3084\u30c7\u30fc\u30bf\u91cf\u3001\u8a08\u7b97\u74b0\u5883\u306b\u3088\u3063\u3066\u306f\u4e00\u56de\u306e\u8a66\u884c\u304c\u6570\u6642\u9593\u306b\u53ca\u3076\u3053\u3068\u3082\u3042\u308a\u3001\u52b9\u7387\u3088\u304f\u63a2\u7d22\u3059\u308b\u3053\u3068\u304c\u5927\u4e8b\u306b\u306a\u308a\u307e\u3059<a href=\"#ref2\"><em>*2<\/em><\/a>\u3002<\/p>\n<p>\u5f93\u6765\u3088\u308a\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3067\u306f\u300c\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u300d\u3068\u3044\u3046\u624b\u6cd5\u304c\u3088\u304f\u7528\u3044\u3089\u308c\u3066\u3044\u307e\u3059<a href=\"#ref3\"><em>*3<\/em><\/a>\u3002<br \/>\n\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u306f\u305d\u306e\u540d\u306e\u901a\u308a\u3001\u53d6\u308a\u3046\u308b\u7bc4\u56f2\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u5168\u3066\u63a2\u7d22\u3057\u3001\u305d\u306e\u4e2d\u304b\u3089\u6700\u9069\u306a\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u9078\u3073\u51fa\u3059\u624b\u6cd5\u3067\u3059\u3002<br \/>\n\u30b7\u30f3\u30d7\u30eb\u3067\u5206\u304b\u308a\u3084\u3059\u3044\u306e\u3067\u3059\u304c\u3001\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u7a2e\u985e\u304c\u5897\u3048\u308b\u3068\u7d44\u307f\u5408\u308f\u305b\u6570\u304c\u7206\u767a\u3057\u3066\u3057\u307e\u3044\u3001\u6642\u9593\u3084\u30ea\u30bd\u30fc\u30b9\u5360\u6709\u306a\u3069\u304c\u554f\u984c\u3068\u306a\u3063\u3066\u304d\u307e\u3059\u3002<\/p>\n<p>\u6a5f\u68b0\u5b66\u7fd2\u306f\u30ed\u30b8\u30c3\u30af\u3092\u9078\u629e\u30fb\u69cb\u7bc9\u3059\u308b\u3060\u3051\u304c\u30b9\u30b3\u30fc\u30d7\u3067\u306f\u3042\u308a\u307e\u305b\u3093\u3002<br \/>\n\u5be7\u308d\u3001\u30b3\u30b9\u30c8\u3068\u82e6\u52b4\u3092\u304b\u3051\u3066\u69cb\u7bc9\u3057\u305f\u30e2\u30c7\u30eb\u3092\u6700\u5927\u9650\u6d3b\u7528\u3057\u3066\u3044\u304f\u4e0a\u3067\u306f\u3001\u30e2\u30c7\u30eb\u306e\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u306a\u3069\u306e\u904b\u7528\u9762\u304c\u30e1\u30a4\u30f3\u306e\u30bf\u30b9\u30af\u306b\u306a\u308a\u3001\u3053\u3046\u3044\u3063\u305f\u554f\u984c\u306f\u3088\u308a\u5927\u5207\u306b\u3001\u3088\u308a\u5207\u5b9f\u306b\u306a\u3063\u3066\u304d\u307e\u3059\u306d\u3002<\/p>\n<p>\u4eca\u56de\u7d39\u4ecb\u3059\u308b\u306e\u306f\u3001\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u306b\u4ee3\u308f\u3063\u3066\u30d9\u30a4\u30ba\u6700\u9069\u5316\u3092\u7528\u3044\u308b\u3068\u3088\u308a\u52b9\u7387\u7684\u306b\u6700\u9069\u306a\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u63a2\u7d22\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u308b\u3088\u3001\u3068\u3044\u3046\u8a71\u3067\u6a5f\u68b0\u5b66\u7fd2\u30b5\u30fc\u30d3\u30b9\u306e\u904b\u7528\u306b\u304a\u3044\u3066\u3082\u529b\u3092\u767a\u63ee\u3059\u308b\u3082\u306e\u3067\u3059\u3002<\/p>\n<h2>\u30d9\u30a4\u30ba\u6700\u9069\u5316\u306b\u3088\u308b\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u6700\u9069\u5316<\/h2>\n<p>\u3053\u306e\u8a18\u4e8b\u3067\u306f\u7406\u8ad6\u7684\u306a\u80cc\u666f\u306b\u3064\u3044\u3066\u306f\u6df1\u304f\u89e6\u308c\u305a\u3001\u30b7\u30f3\u30d7\u30eb\u306a\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u5b9f\u88c5\u3092\u901a\u3057\u3066\u30d9\u30a4\u30ba\u6700\u9069\u5316\u306e\u6d3b\u7528\u30a4\u30e1\u30fc\u30b8\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002<\/p>\n<p>\u7406\u8ad6\u7684\u80cc\u666f\u306b\u3064\u3044\u3066\u8a73\u7d30\u3092\u77e5\u308a\u305f\u3044\u65b9\u306f<a href=\"#ref1\"><em>*1<\/em><\/a>\u3092\u53c2\u8003\u306b\u3057\u3066\u304f\u3060\u3055\u3044\u306d\u3002<\/p>\n<p>\u8b1b\u7fa9\u3067\u7d39\u4ecb\u3055\u308c\u3066\u3044\u305f\u30d9\u30a4\u30ba\u6700\u9069\u5316\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306f\u3001\u4fe1\u983c\u533a\u9593\u6226\u7565(Gaussian Processs Upper Confidence Bound (GP-UCB) \u30a2\u30eb\u30b4\u30ea\u30ba\u30e0)<a href=\"http:\/\/www-stat.wharton.upenn.edu\/~skakade\/papers\/ml\/bandit_GP_icml.pdf\">[Srinivas+, 2010]<\/a>\u3067\u3001<\/p>\n<ol>\n<li>\u5165\u529b(\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf)\u3068\u51fa\u529b(\u6700\u5927\u5316\u3057\u305f\u3044\u3082\u306e\u3001\u7cbe\u5ea6\u306a\u3069)\u306e\u95a2\u4fc2\u3092\u30d6\u30e9\u30c3\u30af\u30dc\u30c3\u30af\u30b9\u95a2\u6570\u3068\u307f\u306a\u3057<\/li>\n<li>\u305d\u308c\u3092\u30ac\u30a6\u30b9\u904e\u7a0b(Gaussian Process)\u306b\u5f93\u3046\u3068\u4eee\u5b9a\u3057\u3066<\/li>\n<li>\u300c\u63a2\u7d22\u3068\u6d3b\u7528\u306e\u6226\u7565\u300d(Upper Confidence Bound)\u3067\u6700\u9069\u5316\u3057 y \u3092\u6700\u5927\u5316\u3059\u308b x \u3092\u63a2\u7d22\u3059\u308b<\/li>\n<\/ol>\n<p>\u3068\u3044\u3046\u3082\u306e\u3067\u3059\u3002<\/p>\n<p><a href=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bayesian_optimization.png\" rel=\"facebox\" rel=\"attachment wp-att-13432\"><img loading=\"lazy\" class=\"aligncenter wp-image-13432 size-large\" src=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bayesian_optimization-1024x502.png\" width=\"530\" height=\"260\" srcset=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bayesian_optimization-1024x502.png 1024w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bayesian_optimization-300x147.png 300w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bayesian_optimization-768x377.png 768w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bayesian_optimization.png 1600w\" sizes=\"(max-width: 530px) 100vw, 530px\" \/><\/a><\/p>\n<h2>\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0<\/h2>\n<p>\u4eca\u56de\u306f\u3001\u3053\u306e GP-UCB \u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u5b9f\u88c5\u3057\u307e\u3057\u305f\u3002<\/p>\n<ul>\n<li>x \u306f\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u3067\u3001y \u304c\u30e2\u30c7\u30eb\u5b66\u7fd2\u7d50\u679c\uff08\u6c4e\u5316\u7cbe\u5ea6\u3084\u8aa4\u5dee\u7b49\uff09\u3068\u7f6e\u304d\u63db\u3048\u3066\u4e0b\u3055\u3044\u3002<\/li>\n<li>x \u306f 0 \u304b\u3089 10 \u306e\u9593\u3092 0.01 \u523b\u307f\u3067\u53d6\u308a\u3046\u308b\u3068\u3057\u3001\u3053\u306e\u4e2d\u3067\u300cy \u3092\u6700\u5927\u5316\u3059\u308b x \u3092\u3067\u304d\u308b\u3060\u3051\u5c11\u306a\u3044\u63a2\u7d22\u3067\u6c42\u3081\u308b\u300d\u3068\u3044\u3046\u306e\u304c\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306e\u76ee\u7684\u3067\u3059\u3002<\/li>\n<li>\u30d6\u30e9\u30c3\u30af\u30dc\u30c3\u30af\u30b9\u95a2\u6570\u306f y = x sin(x) \u3068\u3057\u3066\u3044\u307e\u3059\uff08\u3082\u3061\u308d\u3093\u672c\u6765\u306f\u5185\u90e8\u69cb\u9020\u304c\u308f\u304b\u3089\u306a\u3044\u3082\u306e\u3067\u3059\u304c\u3001\u4f8b\u984c\u3068\u3057\u3066\u8b1b\u7fa9\u3068\u540c\u3058\u3082\u306e\u306b\u3057\u307e\u3059\uff09\u3002<\/li>\n<\/ul>\n<p><a href=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/initial.png\" rel=\"facebox\" rel=\"attachment wp-att-13431\"><img loading=\"lazy\" class=\"aligncenter size-full wp-image-13431\" src=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/initial.png\" alt=\"\" width=\"800\" height=\"600\" srcset=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/initial.png 800w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/initial-300x225.png 300w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/initial-768x576.png 768w\" sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/a><\/p>\n<p>\u3082\u3057\u3001\u5f93\u6765\u624b\u6cd5\u306e\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u3067\u63a2\u7d22\u3057\u305f\u5834\u5408\u3001<strong>\u53d6\u308a\u3046\u308b\u7bc4\u56f2\u306e\u5168\u3066\u306e x \u3092\u63a2\u7d22\u3059\u308b\u305f\u3081\u30011000 \u56de\u306e\u8a66\u884c<\/strong>\u3092\u884c\u3044\u307e\u3059\u3002<\/p>\n<p>\u3053\u308c\u3092 GP-UCB \u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3067\u63a2\u7d22\u3057\u305f\u5834\u5408\u3001\u5927\u4f53<strong> 10 \u56de\u7a0b\u5ea6\u306e\u63a2\u7d22<\/strong>\u3067 y \u304c\u6700\u5927\u3068\u306a\u308b x \u3092\u6c42\u3081\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3057\u305f\u3002<\/p>\n<p><a href=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bo.gif\" rel=\"facebox\" rel=\"attachment wp-att-13433\"><img loading=\"lazy\" class=\"aligncenter wp-image-13433 size-full\" src=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/bo.gif\" width=\"800\" height=\"600\" \/><\/a><\/p>\n<p>\u63a2\u7d22\u306e\u4ed5\u65b9\u3092\u898b\u308b\u3068\u3001<\/p>\n<ul>\n<li>\u63a2\u7d22\u304c\u3042\u307e\u308a\u3055\u308c\u3066\u3044\u306a\u3044\u7bc4\u56f2(<strong>\u63a2\u7d22<\/strong>)\u3068<\/li>\n<li>\u65e2\u306b\u5206\u304b\u3063\u3066\u3044\u308b\u6700\u5927\u5024\u306b\u8fd1\u3044\u7bc4\u56f2(<strong>\u6d3b\u7528<\/strong>)\u3092<\/li>\n<\/ul>\n<p>\u30d0\u30e9\u30f3\u30b9\u3088\u304f\u9078\u629e\u3057\u306a\u304c\u3089\u63a2\u7d22\u3092\u9032\u3081\u308b\u3053\u3068\u304c\u898b\u3066\u53d6\u308c\u3001\u5f37\u5316\u5b66\u7fd2\u306e\u3088\u3046\u306a\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u306d\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u307e\u3068\u3081<\/h2>\n<p>\u4eca\u65e5\u7d39\u4ecb\u3057\u305f\u3088\u3046\u306a\u30b7\u30f3\u30d7\u30eb\u306a\u4f8b\u984c\u3067\u306f\u975e\u5e38\u306b\u52b9\u7387\u3088\u304f\u6700\u9069\u5024\u307e\u3067\u305f\u3069\u308a\u7740\u304f\u3053\u3068\u304c\u3067\u304d\u3066\u3044\u307e\u3059\u3002<\/p>\n<p>\u4f8b\u3067\u306f\u5206\u304b\u308a\u3084\u3059\u3055\u3092\u512a\u5148\u3057\u3001\uff11 \u7a2e\u985e\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u307f\u3092\u63a2\u7d22\u3057\u305f\u306e\u3067\u305d\u308c\u307b\u3069\u30a4\u30f3\u30d1\u30af\u30c8\u306f\u306a\u3044\u304b\u3082\u3057\u308c\u307e\u305b\u3093\u3002<br \/>\n\u3057\u304b\u3057\u3001\u3082\u3057 2 \u7a2e\u985e\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u304c\u305d\u308c\u305e\u308c 1000 \u306e\u7bc4\u56f2\u3092\u53d6\u308a\u3046\u308b\u5834\u5408\u3001\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u3067\u306f 1000x1000 \u56de\u306e\u8a66\u884c\u304c\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u3002<br \/>\n3 \u7a2e\u985e\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u304c\u305d\u308c\u305e\u308c 1000 \u306e\u7bc4\u56f2\u3092\u53d6\u308a\u3046\u308b\u5834\u5408\u30011000x1000x1000 \u56de\u306e\u8a66\u884c\u3068\u306a\u308a\u3001\u6700\u65e9\u73fe\u5b9f\u7684\u306a\u8a66\u884c\u56de\u6570\u306b\u306a\u3089\u306a\u3044\u3053\u3068\u304c\u308f\u304b\u308b\u3068\u601d\u3044\u307e\u3059\u3002<br \/>\n\u63a2\u7d22\u3059\u308b\u30d1\u30e9\u30e1\u30fc\u30bf\u6b21\u5143\u6570\u30fb\u53d6\u308a\u3046\u308b\u7bc4\u56f2\u304c\u5897\u3048\u308b\u307b\u3069\u3001\u30d9\u30a4\u30ba\u6700\u9069\u5316\u306f\u529b\u3092\u767a\u63ee\u3059\u308b\u3067\u3057\u3087\u3046\u3002<\/p>\n<p>\u3057\u304b\u3057\u3001\u73fe\u5b9f\u306e\u554f\u984c\u306b\u3053\u306e\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u9069\u7528\u3057\u3066\u5373\u5ea7\u306b\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u30fb\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u306e\u60a9\u307f\u304b\u3089\u89e3\u653e\u3055\u308c\u308b\u3001\u3068\u3044\u3046\u8a33\u306b\u306f\u3044\u304b\u306a\u3044\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n<p>\u306a\u305c\u306a\u3089 GP-UCB \u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u306b\u3082\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3059\u3079\u304d\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\uff08<a href=\"#ref4\"><em>*4<\/em><\/a>\uff09\u304c\u3042\u308a\u3001\u3053\u308c\u3089\u3092\u6700\u9069\u5316\u3057\u305f\u3044\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u5206\u5e03\u306b\u30d5\u30a3\u30c3\u30c8\u3055\u305b\u308b\u5de5\u592b\u306f\u8981\u308b\u3067\u3057\u3087\u3046\u3002<\/p>\n<p>\u6a5f\u68b0\u5b66\u7fd2\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3059\u308b\u305f\u3081\u306e\u30d9\u30a4\u30ba\u6700\u9069\u5316\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3001\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u30fb\u30fb\u30fb \u3068\u5c11\u3057\u304a\u5bd2\u3044\u611f\u3058\u304c\u3057\u3066\u304d\u307e\u3059\u304c(\u7b11) ...<br \/>\n\u305f\u3060\u3001\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u3067\u305f\u304f\u3055\u3093\u306e\u6b21\u5143\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u76f4\u63a5\u76f8\u624b\u306b\u3059\u308b\u5fc5\u8981\u304c\u306a\u304f\u306a\u308b\u3053\u3068\u306f\u671f\u5f85\u3067\u304d\u308b\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n<p>\u30d9\u30a4\u30ba\u6700\u9069\u5316\u306f\u6a5f\u68b0\u5b66\u7fd2\u306e\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3060\u3051\u3067\u306a\u304f\u3001\u69d8\u3005\u306a\u8ab2\u984c\u306b\u9069\u7528\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<br \/>\n\u8907\u6570\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u81ea\u52d5\u7684\u306b\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u3057\u306a\u304c\u3089\u3001\u5c11\u3057\u305a\u3064\u6700\u9069\u306a\u3068\u3053\u308d\u3092\u63a2\u3057\u3066\u3044\u304f\u30fb\u30fb\u30fb\u30fb\u69d8\u3005\u306a\u3068\u3053\u308d\u306b\u5fdc\u7528\u3067\u304d\u305d\u3046\u3067\u3059\u306d\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u30b3\u30fc\u30c9<\/h2>\n<p>\u6700\u5f8c\u306b\u5b9f\u88c5\u30b3\u30fc\u30c9\u3092\u8f09\u305b\u307e\u3059\u3002<\/p>\n<p>\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u304c\u7406\u89e3\u3067\u304d\u308c\u3070\u7279\u306b\u5b9f\u88c5\u4e0a\u6ce8\u610f\u3059\u308b\u3068\u3053\u308d\u306f\u306a\u3044\u3068\u601d\u3044\u307e\u3059\u304c\u3001\u79c1\u306f\u7372\u5f97\u95a2\u6570\u306b\u3088\u308b\u6b21\u306e x \u306e\u9078\u629e\u306e\u3068\u3053\u308d\u3067<\/p>\n<p><code><br \/>\na(x_next) = argmax \u03bc(x) + \u221a\u03b2 \u03c3(x)<br \/>\n--> \u7372\u5f97\u95a2\u6570\u304c\u6700\u5927\u3092\u4e0e\u3048\u308b\u3068\u304d\u306e x \u3092\u6b21\u306e x \u3068\u3059\u308b<br \/>\n<\/code><br \/>\n\u3092<br \/>\n<code><br \/>\nx_next = argmax \u03bc(x) + \u221a\u03b2 \u03c3(x)<br \/>\n--> \u7372\u5f97\u95a2\u6570\u306e\u6700\u5927\u5024\u3092\u6b21\u306e x \u3068\u3059\u308b<br \/>\n<\/code><br \/>\n\u3068\u8aad\u307f\u9593\u9055\u3048\u3066\u3057\u307e\u3044\u3001\u3057\u3070\u3089\u304f\u60a9\u307f\u307e\u3057\u305f\u3002<br \/>\n\u3053\u3046\u3044\u3063\u305f\u6570\u5f0f\u306e\u8aad\u307f\u9055\u3044\u30df\u30b9\u306f\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u5b9f\u88c5\u306b\u306f\u4ed8\u304d\u3082\u306e\u3067\u3059\u306d\u30fb\u30fb\u30fb\u3002<\/p>\n<pre class=\"lang:python decode:true\">#!\/usr\/bin\/env python\r\n# -*- coding: utf-8 -*-\r\n\"\"\"\r\n\u30d9\u30a4\u30ba\u6700\u9069\u5316(GP-UCB \u30a2\u30eb\u30b4\u30ea\u30ba\u30e0)\u306b\u3088\u308b\u63a2\u7d22.\r\n\r\n\u74b0\u5883\r\n  Python 2.7.12\r\n  scikit-learn==0.18.1\r\n  matplotlib==1.5.3\r\n  numpy==1.11.2\r\n\"\"\"\r\nimport sys\r\nimport os\r\nimport matplotlib.pyplot as plt\r\nimport numpy as np\r\nfrom sklearn.gaussian_process import GaussianProcessRegressor\r\n\r\n\r\ndef blackbox_func(x):\r\n    \"\"\"\r\n    \u30d6\u30e9\u30c3\u30af\u30dc\u30c3\u30af\u30b9\u95a2\u6570\uff08\u4f8b\u984c\u306a\u306e\u3067 x sin(x) \u3068\u3057\u3066\u3044\u307e\u3059\uff09\r\n    --&gt; \u672c\u6765\u306f\u30e2\u30c7\u30eb\u5b66\u7fd2\r\n        ex) y = svm(\u5b66\u7fd2\u30c7\u30fc\u30bf, x(\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf)) \u306e\u7d50\u679c\u306a\u3069\u6700\u5927\u5316\u3057\u305f\u3044\u5024\u3092\u8fd4\u3059.\r\n    \"\"\"\r\n    return x * np.sin(x)\r\n\r\n\r\ndef acq_ucb(mean, sig, beta=3):\r\n    \"\"\"\r\n    \u7372\u5f97\u95a2\u6570 (Upper Confidence Bound)\r\n    $ x_t = argmax\\ \\mu_{t-1} + \\sqrt{\\beta_t} \\sigma_{t-1}(x) $\r\n    \"\"\"\r\n    return np.argmax(mean + sig * np.sqrt(beta))\r\n\r\n\r\ndef plot(x, y, X, y_pred, sigma, title=\"\"):\r\n\r\n    fig = plt.figure()\r\n    plt.plot(x, blackbox_func(x), 'r:', label=u'$blackbox func(x) = x\\,\\sin(x)$')\r\n    plt.plot(X, y, 'r.', markersize=10, label=u'Observations')\r\n    plt.plot(x, y_pred, 'b-', label=u'Prediction')\r\n    plt.fill(np.concatenate([x, x[::-1]]),\r\n             np.concatenate([y_pred - 1.96 * sigma,(y_pred + 1.96 * sigma)[::-1]]),\r\n             alpha=.5, fc='b', ec='None', label='95% confidence interval')\r\n    plt.xlabel('$x$')\r\n    plt.ylabel('$y$')\r\n    plt.ylim(-10, 20)\r\n    plt.title(title)\r\n    plt.legend(loc='upper left')\r\n    plt.savefig('fig%02d.png' % (i))\r\n\r\n\r\n\r\n   \r\n# \u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u30a8\u30f3\u30c8\u30ea\u30dd\u30a4\u30f3\u30c8\r\nif __name__ == '__main__':\r\n    \r\n    # \u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u53d6\u308a\u3046\u308b\u7bc4\u56f2\r\n    x_grid = np.atleast_2d(np.linspace(0, 10, 1001)[:1000]).T\r\n    \r\n    # \u521d\u671f\u5024\u3068\u3057\u3066 x=1, 9 \u306e 2 \u70b9\u306e\u63a2\u7d22\u3092\u3057\u3066\u304a\u304f.\r\n    X = np.atleast_2d([1., 9.]).T\r\n    y = blackbox_func(X).ravel()\r\n\r\n    \r\n    # Gaussian Processs Upper Confidence Bound (GP-UCB)\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\r\n    # --&gt; \u53ce\u675f\u3059\u308b\u307e\u3067\u7e70\u308a\u8fd4\u3059(\u53ce\u675f\u6761\u4ef6\u306a\u3069\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u30dd\u30a4\u30f3\u30c8)\r\n    n_iteration = 13\r\n    for i in range(n_iteration):\r\n    \r\n        # \u65e2\u306b\u5206\u304b\u3063\u3066\u3044\u308b\u5024\u3067\u30ac\u30a6\u30b9\u904e\u7a0b\u30d5\u30a3\u30c3\u30c6\u30a3\u30f3\u30b0\r\n        # --&gt; \u30ab\u30fc\u30cd\u30eb\u95a2\u6570\u3084\u30d1\u30e9\u30e1\u30fc\u30bf\u306f\u30c7\u30d5\u30a9\u30eb\u30c8\u306b\u3057\u3066\u3044\u307e\u3059(\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u30dd\u30a4\u30f3\u30c8)\r\n        gp = GaussianProcessRegressor()\r\n        gp.fit(X, y)\r\n        \r\n        # \u4e8b\u5f8c\u5206\u5e03\u304c\u6c42\u307e\u308b\r\n        posterior_mean, posterior_sig = gp.predict(x_grid, return_std=True)\r\n        \r\n        # \u76ee\u7684\u95a2\u6570\u3092\u6700\u5927\u5316\u3059\u308b x \u3092\u6b21\u306e\u30d1\u30e9\u30e1\u30fc\u30bf\u3068\u3057\u3066\u9078\u629e\u3059\u308b\r\n        # --&gt; \u03b2\u3092\u5927\u304d\u304f\u3059\u308b\u3068\u63a2\u7d22\u91cd\u8996(\u521d\u671f\u306f\u5927\u304d\u304f\u3057\u63a2\u7d22\u91cd\u8996\u3057\u30a4\u30c6\u30ec\u30fc\u30b7\u30e7\u30f3\u306b\u540c\u671f\u3057\u3066\u6e1b\u8870\u3055\u305b\u6d3b\u7528\u3092\u91cd\u8996\u3055\u305b\u308b\u306a\u3069\u3001\u30c1\u30e5\u30fc\u30cb\u30f3\u30b0\u30dd\u30a4\u30f3\u30c8)\r\n        idx = acq_ucb(posterior_mean, posterior_sig, beta=100.0)\r\n        x_next = x_grid[idx]\r\n    \r\n        plot(x_grid, y, X, posterior_mean, posterior_sig, title='Iteration=%2d,  x_next = %f'%(i+2, x_next))\r\n    \r\n        # \u66f4\u65b0\r\n        X = np.atleast_2d([np.r_[X[:, 0], x_next]]).T\r\n        y = np.r_[y, blackbox_func(x_next)]\r\n        \r\n    \r\n    print \"Max x=%f\" % (x_next)\r\n    \r\n<\/pre>\n<p>&nbsp;<\/p>\n<h2>\u53c2\u7167<\/h2>\n<ol>\n<li id=\"ref1\">\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u8b1b\u7fa9\u306f<a href=\"https:\/\/youtu.be\/pQHWew4YYao\">\u3053\u3061\u3089<\/a>\u3067\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u5185\u5bb9\u3068\u91cd\u306a\u308b\u3068\u3053\u308d\u3082\u591a\u3044\u3067\u3059\u306e\u3067\u662f\u975e\u53c2\u8003\u306b\u3057\u3066\u4e0b\u3055\u3044.<\/li>\n<li id=\"ref2\">\u7279\u306b DeepLearning \u306a\u3069\u306f\u4e00\u56de\u306e\u5b66\u7fd2\u304c\u6570\u6642\u9593\u3001\u5834\u5408\u306b\u3088\u3063\u3066\u306f\u6570\u65e5\u304b\u304b\u308b\u3053\u3068\u3082\u73cd\u3057\u304f\u3042\u308a\u307e\u305b\u3093<\/li>\n<li id=\"ref3\">\u30e9\u30f3\u30c0\u30e0\u306b\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u632f\u3063\u3066\u63a2\u7d22\u3059\u308b\u3068\u3044\u3046\u306e\u3082\u7acb\u6d3e\u306a\u63a2\u7d22\u65b9\u6cd5\u306e\u4e00\u3064\u3067\u3059\u3002<a href=\"http:\/\/www.jmlr.org\/papers\/volume13\/bergstra12a\/bergstra12a.pdf\">DeepLearning \u3067\u306f\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u3088\u308a\u3082\u30e9\u30f3\u30c0\u30e0\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u306e\u65b9\u304c\u826f\u3044\u3053\u3068\u304c\u5831\u544a\u3055\u308c\u3066\u3044\u307e\u3059\u3002<\/a><\/li>\n<li id=\"ref4\">\u4eca\u56de\u306f\u308f\u304b\u308a\u3084\u3059\u3055\u3092\u512a\u5148\u3057\u305f\u305f\u3081\u3001\u30e9\u30a4\u30d6\u30e9\u30ea\u306e\u30c7\u30d5\u30a9\u30eb\u30c8\u5024\u3092\u7528\u3044\u3001\u63a2\u7d22\u6226\u7565\u3092 UCB \u306e\u307f\u306b\u9650\u5b9a\u3057\u3066\u3044\u307e\u3059\u3002\u6050\u3089\u304f\u3001\u63a2\u7d22\u3068\u6d3b\u7528\u306e\u30d0\u30e9\u30f3\u30b9\u3092\u6c7a\u3081\u308b\u4fc2\u6570 \u03b2 \u306e\u9078\u3073\u65b9\u3084\u30ab\u30fc\u30cd\u30eb\u95a2\u6570\u306e\u7a2e\u985e\u3067\u5c40\u6240\u6700\u9069\u306b\u9665\u3063\u305f\u308a\u3082\u3059\u308b\u3067\u3057\u3087\u3046\u3002<\/li>\n<\/ol>\n<h2>\u6700\u5f8c\u306b<\/h2>\n<p>\u8ce2\u304f\u63a2\u7d22\u3001\u8ce2\u304f\u6700\u9069\u5316\u306f\u6020\u3051\u8005\u306e\u5473\u65b9\u3068\u3044\u3046\u3053\u3068\u3067\u3001\u3053\u308c\u3089\u306f\u732b\u306e\u597d\u7269\u3060\u3068\u601d\u3044\u307e\u3059\u3002<\/p>\n<div id=\"attachment_13528\" style=\"width: 473px\" class=\"wp-caption aligncenter\"><a href=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/2016-12-19-kawaii-becky.png\" rel=\"facebox\" rel=\"attachment wp-att-13528\"><img aria-describedby=\"caption-attachment-13528\" loading=\"lazy\" class=\" wp-image-13528\" src=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/2016-12-19-kawaii-becky.png\" alt=\"\" width=\"463\" height=\"304\" srcset=\"https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/2016-12-19-kawaii-becky.png 1302w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/2016-12-19-kawaii-becky-300x197.png 300w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/2016-12-19-kawaii-becky-768x505.png 768w, https:\/\/www.techscore.com\/blog\/wp\/wp-content\/uploads\/2016\/12\/2016-12-19-kawaii-becky-1024x673.png 1024w\" sizes=\"(max-width: 463px) 100vw, 463px\" \/><\/a><p id=\"caption-attachment-13528\" class=\"wp-caption-text\">\u304b\u308f\u3044\u3044\u30d9\u30c3\u30ad\u30fc<\/p><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u3053\u3093\u306b\u3061\u306f\u3001\u5ca1\u5d0e\u3067\u3059\u3002<br \/>\n\u3053\u306e\u8a18\u4e8b\u306f TECHSCORE Advent Calendar 2016 \u306e 20 \u65e5\u76ee\u306e\u8a18\u4e8b\u3067\u3059\u3002<\/p>\n<p>\u5148\u65e5\u3001IBISML2016(\u7b2c19\u56de\u60c5\u5831\u8ad6\u7684\u5b66\u7fd2\u7406\u8ad6\u30ef\u30fc\u30af\u30b7\u30e7\u30c3\u30d7) \u306e\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u306b\u53c2\u52a0\u3057\u307e\u3057\u305f\u3002<br \/><a href=\"https:\/\/www.techscore.com\/blog\/2016\/12\/20\/%e6%a9%9f%e6%a2%b0%e5%ad%a6%e7%bf%92%e3%81%ae%e3%83%8f%e3%82%a4%e3%83%91%e3%83%bc%e3%83%91%e3%83%a9%e3%83%a1%e3%83%bc%e3%82%bf%e6%8e%a2%e7%b4%a2-%e3%83%99%e3%82%a4%e3%82%ba%e6%9c%80%e9%81%a9\/\">\u7d9a\u304d\u3092\u8aad\u3080...<\/a><\/p>\n","protected":false},"author":48,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[239],"tags":[141,120,251],"_links":{"self":[{"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/posts\/13430"}],"collection":[{"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/users\/48"}],"replies":[{"embeddable":true,"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/comments?post=13430"}],"version-history":[{"count":22,"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/posts\/13430\/revisions"}],"predecessor-version":[{"id":17119,"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/posts\/13430\/revisions\/17119"}],"wp:attachment":[{"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/media?parent=13430"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/categories?post=13430"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.techscore.com\/blog\/wp-json\/wp\/v2\/tags?post=13430"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}