黃博士今日演講內容 - 圍棋
By Frederica
at 2017-11-10T13:23
at 2017-11-10T13:23
Table of Contents
簡單提一下今天黃博士演講重點
演講標題是深度學習與強化學習的勝利
認為zero是最佳的deepmind電腦圍棋這部分最佳的收尾
黃博士對於一開始擊敗樊輝就發nature有些不解,我們要挑戰李世石結果把所有技術都透漏給所有人,但deepmind的想法是我們需要分享技術讓世界一起進步
google對alphago團隊最大幫助是TPU
認為Master已經完美解決李世石第四盤的bug,解決方式與神經網路架構(dual res)和訓練都有關,並且以他多年的電腦圍棋經驗與測試過後,認為不會再出現此類bug
Master是20block res-net,並改進了training pipeline和MCTS,也解決了模仿棋和循環劫(沒說怎麼做),能讓lee版本3子並超過50%勝率
master年初60連勝每一步4-8秒,在台灣,吃泡麵配黑松沙士下的,是黃博士積極鼓吹要出來測試,Hassabis說要低調並使用韓國國籍,一開始不得透漏身分
Hassabis說要挑強的下,但是第一天職業沒人願意跟0勝0負的下,都被拒絕,等到第一天10連勝之後第二天開始拒絕別人邀請
master下的時候可以看勝率隨步數的圖,基本上50手之前斜率很高並且確立極大優勢,唯一例外是柯潔烏鎮第二盤
4月的時候已經有zero,但由於要發nature所以不能拿來下
當初開發zero沒預料到會超過master
master年初開發完畢之後,zero由其他人負責開發,黃博士繼續想方法增強master
zero不是放在那邊增強學習就會變強,中間需要做很多優化,否則有bug不會進步,其中一個重大bug發生在第三天(紀錄人表示:所以看來絕藝有得忙了
AntiAlphaGo,不是像大家想的那樣有新的技術,就是左右互搏,也不是gan(生成式對抗網路)
master是否被人類棋譜拖累?答案是不確定,因為master訓練時間較短,deepmimd也沒有針對同等條件去比較。
以上,有其他疏漏請其他人補充,並歡迎轉載,但請說明作者是Hetercompute
-----
Sent from JPTT on my Samsung SM-A710Y.
--
Tags:
圍棋
All Comments
By Hazel
at 2017-11-13T01:29
at 2017-11-13T01:29
By Catherine
at 2017-11-16T09:20
at 2017-11-16T09:20
By Jacob
at 2017-11-20T16:59
at 2017-11-20T16:59
By George
at 2017-11-25T07:23
at 2017-11-25T07:23
By Dora
at 2017-11-27T12:31
at 2017-11-27T12:31
By Irma
at 2017-11-28T14:25
at 2017-11-28T14:25
By David
at 2017-12-02T23:58
at 2017-12-02T23:58
By Tom
at 2017-12-04T01:21
at 2017-12-04T01:21
By David
at 2017-12-06T18:40
at 2017-12-06T18:40
By Edith
at 2017-12-07T06:56
at 2017-12-07T06:56
By Caitlin
at 2017-12-08T08:40
at 2017-12-08T08:40
By Joe
at 2017-12-13T06:36
at 2017-12-13T06:36
By Tracy
at 2017-12-16T18:56
at 2017-12-16T18:56
By Frederic
at 2017-12-17T02:22
at 2017-12-17T02:22
By Jessica
at 2017-12-19T07:38
at 2017-12-19T07:38
By Dorothy
at 2017-12-22T19:24
at 2017-12-22T19:24
By Iris
at 2017-12-24T00:03
at 2017-12-24T00:03
By Ida
at 2017-12-26T08:42
at 2017-12-26T08:42
By Heather
at 2017-12-27T21:38
at 2017-12-27T21:38
By Lucy
at 2017-12-28T04:46
at 2017-12-28T04:46
By Connor
at 2017-12-30T11:50
at 2017-12-30T11:50
By Yedda
at 2018-01-01T11:38
at 2018-01-01T11:38
By Callum
at 2018-01-02T01:31
at 2018-01-02T01:31
By Gary
at 2018-01-07T00:42
at 2018-01-07T00:42
By Hamiltion
at 2018-01-07T09:16
at 2018-01-07T09:16
By Carolina Franco
at 2018-01-08T11:58
at 2018-01-08T11:58
By Bethany
at 2018-01-12T18:45
at 2018-01-12T18:45
By Madame
at 2018-01-14T13:21
at 2018-01-14T13:21
By Jacky
at 2018-01-15T10:53
at 2018-01-15T10:53
By Sarah
at 2018-01-17T06:50
at 2018-01-17T06:50
By Wallis
at 2018-01-21T18:22
at 2018-01-21T18:22
By Christine
at 2018-01-25T15:10
at 2018-01-25T15:10
By Kelly
at 2018-01-25T19:20
at 2018-01-25T19:20
By Harry
at 2018-01-29T11:26
at 2018-01-29T11:26
By Vanessa
at 2018-01-31T18:51
at 2018-01-31T18:51
Related Posts
如果有機會問黃士傑博士問題
By Frederica
at 2017-11-10T09:19
at 2017-11-10T09:19
黃士傑返台分享AlphaGo Zero開發過程
By Edith
at 2017-11-10T01:32
at 2017-11-10T01:32
黃士傑返台分享AlphaGo Zero開發過程
By Hamiltion
at 2017-11-09T23:06
at 2017-11-09T23:06
黃士傑返台分享AlphaGo Zero開發過程
By Belly
at 2017-11-09T14:31
at 2017-11-09T14:31
CGI死活強度
By Joe
at 2017-11-08T21:33
at 2017-11-08T21:33