在线时间:8:00-16:00
迪恩网络APP
随时随地掌握行业动态
扫描二维码
关注迪恩网络微信公众号
开源软件名称:brilee/MuGo开源软件地址:https://github.com/brilee/MuGo开源编程语言:Python 100.0%开源软件介绍:UpdateThis repo is abandoned as of mid-2017. The code at https://github.com/tensorflow/minigo is a continuation of the work here. MuGo: A minimalist Go engine modeled after AlphaGoThis is a pure Python implementation of the essential parts of AlphaGo. 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. Getting StartedInstall TensorflowStart by installing Tensorflow along with GPU drivers (i.e. CUDA support for Nvidia cards). Get SGFs for supervised learningSecond, find a source of SGF files. You can find 15 years of KGS high-dan games at u-go.net. Alternately, you can download a database of professional games from a variety of sources. Preprocess SGFsThird, preprocess the SGF files. This takes all positions in the SGF files and extracts features for each position, as well as recording the correct next move. These positions are then split into chunks, with one test chunk and the remainder as training chunks. This step may take a while, and must be repeated if you change the feature extraction steps in
(This example takes advantage of bash wildcard expansion - say, if the KGS directories are named data/kgs-2006-01, data/kgs-2006-02, and so on.) Supervised learning (policy network)With the preprocessed SGF data (default output directory is
As the network is trained, the current model will be saved at
Additionally, you can follow along with the training progress with TensorBoard - if you give each run a different name (
Play against MuGoMuGo uses the GTP protocol, and you can use any gtp-compliant program with it. To invoke the raw policy network, use
To invoke the MCTS-integrated version of the policy network, use
One way to play via GTP is to use gogui-display (which implements a UI that speaks GTP.) You can download the gogui set of tools at http://gogui.sourceforge.net/. See also documentation on interesting ways to use GTP.
Another way to play via GTP is to play against GnuGo, while spectating the games
Another way to play via GTP is to connect to CGOS, the Computer Go Online Server. The CGOS server hosted by boardspace.net is actually abandoned; you'll want to connect to the CGOS server at yss-aya.com. After configuring your cgos.config file, you can connect to CGOS with Running unit testspython -m unittest discover tests |
2023-10-27
2022-08-15
2022-08-17
2022-09-23
2022-08-13
请发表评论