Environment Setup

[rickytan-AA]

We need to:

• [x] construct an environment that lets me play on a chessboard
  • [x] reset() - resets the board with standard pieces
  • [x] step(a) - returns the new state and done boolean
    • [x] step_seq() - takes a sequence of actions
  • [x] stalemates and insufficient material cases
• [ ] (optional) ELO reward function

Helpful links:

• CharlesChessAI
• gym-chess (can't render pieces)
• python chess module
• Juliani RL medium blog

[rickytan-AA]

Doesn't seem like there are many good RL-based environments for chess.

That's fine as it gives me a chance to experiment around with value functions and defining the reward states.

[rickytan-AA]

[rickytan-AA]

Should reward be a function from the environment or from the agent?

The environment should at least return game outcomes (+1 for a win, -1 for a loss, 0 for draws).

Probably need to read the AlphaZero paper to understand how they did their reward functions.