A simple chess AI.
- The dataset used was downloaded from FICS
- A total of 97,000 games were used, with a sample size of 71,00,000.
- The chess board was encoded to a 8x8x12 matrix (one hot encoded for each piece type - color variant).
- For every white move, both the board and the move were flipped horizontally and vertical, to "pretend" as if it was black's move. (This improved the validation accuracy from 37% -> 49%)
- The move was split into a from-square vector and a to-square vector (both of length 64, one-hot encoded), which acted as labels for the problem.
- The model was trained on colab for 13 epochs.
- Validation accuracy achieved :
- "From" output - 49%
- "To" output - 37%
- To render the chess board, chessboard.js was used.
- This UI was connected to the model through an API of Flask backend.
- Finally the whole application was deployed on Heroku.
- Improve the engine by introducing minimax (partially).
- [IMP] Fix the bugs on Heroku.