acbraith / go_player

AlphaZero inspired Go bot

Go Experiments

Overview

• go.py contains a Go class, which can play games of Go
• players.py, mcts.py and model.py are the files for creating Go playing agents
• download_*.py downloads SGF files
• train.py trains and saves a CNN (from model.py) on SGF files
• play.py logs on to an OGS account using details in config.yml, and plays games using the model saved by train.py

TODO

• Go Game
  • Play SGF files
• Create an Agent
  • Random Agent
  • Webscrape SGF files
    • Use proxy and multithreading
  • Build AlphaZero style CNN
    • https://github.com/AppliedDataSciencePartners/DeepReinforcementLearning
  • Train policy and value heads
    • Convert SGF files into network inputs / targets
  • Greedy policy and value head agents
  • MCTS
    • https://github.com/pbsinclair42/MCTS
  • MCTS agent
  • Self-play to generate training data
  • Full AlphaZero : Self-Play / Train / Evaluate
    • https://medium.com/applied-data-science/alphago-zero-explained-in-one-diagram-365f5abf67e0
    • https://applied-data.science/static/main/res/alpha_go_zero_cheat_sheet.png
• Play Online
  • https://online-go.com/api/
  • https://ogs.docs.apiary.io/
  • https://forums.online-go.com/t/ogs-api-notes/17136
  • Create account for AI
  • Log on to API
  • Place a move in a game
  • Load a game into Go object
  • Plug in simple random agent
  • Handle multiple games
    • Play multiple games at once
    • Pick up existing games on startup
    • Pick up new games whilst running (bug?)
  • Search for game when no games going on
  • Multiprocessing for multiple games
  • Stone removal phase
  • Resign based on win chance
  • Chat
    • Pre-programmed messages
    • More complex chat-bot model
• Optimisations
  • Benchmark and optimise Go game
  • Train on AWS
  • Download more data and train from only dan level games
  • Optimise train data selection
    • https://stackoverflow.com/questions/15993447/python-data-structure-for-efficient-add-remove-and-random-choice