This model implements the Alpha Zero paper on the game of Reversi/Orthello.
The overall algorithm is divided into three parts:
- Self-play
- The current best model plays itself to generate training data.
- Each game takes about 10 seconds to complete.
- Training
- A model is fed the training data generated from self-play.
- The input to the model is the current board state with history.
- The labels are the policy evaluated by MCTS, and the value given by the final winner of the game.
- This model is trained for ~30hours.
- Evaluation
- The newly trained model competes with the current best model. It replaces the current best if it wins.
- Each games takes about 20 seconds to complete.
