Chess RL

Roadmap

• Chess Environment & Encoding [done]

  • Board stepping [done]
  • Observataion encoding [done]
  • Action encoding [done]

• Neural Nework Model [done]

  • Shared resnet with policy and value head [done]

• Monte Carlo Tree search [done]

• Self play loop [done]

  • Play [done]
  • Store data [done]
  • Export data as PyTorch dataset-compatible format [done]

• Training pipeline [done]

• Evaluation and Rating TODO

• Optimization and Scaling TODO

Interresting Papers:

• Monte-Carlo tree search as regularized policy optimization https://arxiv.org/abs/2007.12509

Improvement of MCTS for low Nsim values