README.md

ZeroGo

ZeroGo is an attempt to use machine learning techniques to develop agents for the game of Go using limited hardware resources. It uses the AlphaGo style Monte-Carlo Tree Search to combine the results of a policy agent and a value agent. By reducing the dimension of board encoder and the number of layers/units of the network, the ZeroGo agent is able to operate on macbooks and PCs. Depending on the search depth and rollout times, ZeroGo agent reaches up to 1-dan player level.

Click to see all deliverables of ZeroGo

To run a human vs bot game:

• $python3 dl_app.py

ZeroGo is developed using tensorflow version 1.13, you might need to use tensorflow 1.x to run the program

• Go to http://127.0.0.1:5000/static/play_predict_19.html

Current models:

• models/AC: actor critic models based on 5x5 board
• models/AlphaGo: policy and value agents on 19x19 board. Policy v0-0-0 is based on the previous NN model with 27% accuracy

midterm_agents

Contains all agents before midterm, including:

• random (5x5 and 19x19)
• greedy (5x5 and 19x19)
• depth_pruned (5x5)
• alpha_beta (5x5)
• mcts (5x5)
• actor_critic (5x5)
• NN (19x19)

2 game hosts

• __5x5_host.py
• __19x19_host.py

rl.py

• process to simulate games and train the RL agents

alphago.py

• process to debug and test the alphago MCTS agent

dlgo bug fixes

• AlphaGo MCTS bugs

maxpumperla/deep_learning_and_the_game_of_go#55

• dlgo-rl-simulate.py line 50: white_player, black_player = agent1, agent2

NN_end_to_end.py

• end to end script from downloading data to open web app
• steps:

1. set encoder
2. set data processor/generator
3. construct NN layers
4. compile model and train
5. save model
6. initiate deep learning agent
7. start web app