andylolu2 / Alpha-Reversi

This model implements the Alpha Zero paper on the game of Reversi/Orthello.

Aphla Reversi

This model implements the Alpha Zero paper on the game of Reversi/Orthello.

The overall algorithm is divided into three parts:

1. Self-play
   • The current best model plays itself to generate training data.
   • Each game takes about 10 seconds to complete.
2. 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.
3. 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.

---

Model elo over time

Training loss over time