我 相 信 这 么 好 看 的 你 已 经 置 顶了 我 转自：爱可可-爱生活 The logic / control flow of AlphaGo itself is not very complicated and is replicated here. The secret sauce of AlphaGo is in its various neural networks. (As I understand it) AlphaGo uses three neural networks during play. The first NN is a slow but accurate policy network. This network is trained to predict human moves (~57% accuracy), and it outputs a list of plausible moves, with probabilities attached to each move. This first NN is used to seed the Monte Carlo tree search with plausible moves. One of the reasons this first NN is slow is because of its size, and because the inputs to the neural network are various expensive-to-compute properties of the Go board (liberty counts; ataris; ladder status; etc.). The second NN is a smaller, faster but less accurate (~24% accuracy) policy network, and doesn't use computed properties as input. Once a leaf node of the current MCTS tree is reached, the second faster network is used to play the position out to the end with vaguely plausible moves, and score the end position. The third NN is a value network: it outputs an expected win margin for that board, without attempting to play anything out. The results of the monte carlo playout using NN #2 and the value calculation using NN #3 are averaged, and this value is recorded as the approximate result for that MCTS node. Using the priors from NN #1 and the accumulating results of MCTS, a new path is chosen for further Monte Carlo exploration. 链接： https://github.com/brilee/MuGo 原文链接： http://weibo.com/1402400261/EpP8YmzoQ?from=page_1005051402400261_profile |
|
声明：文章版权归原作者所有 部分文章转自互联网 如有侵权请联系
[邮箱地址] 删除
|
注册
登录
