This is a AI based Chess Game
- Make a sklearn based model
- Make a UI in pygame or js or something (using llm)
- Make a basic model with pytorch, maybe ANN
- Improve the chess engine so that it can consistently beat me
- Reinforcement Learning based ?
The model overfits.
Model with Smooth L1 loss, l2 and dropout regularisation, AdamW, Trapezoidal scheduler, 32 embedding, train-val = 1M, 10K, bs=10K, warmup=30
Model with MSE loss, l2 and dropout regularisation, AdamW, Trapezoidal scheduler, 32 embedding, train-val = 1M, 10K, bs=10K, warmup=300
I've tried
- diff loss(mse, smooth l1)
- diff scheduler(trapezoidal, cosing, exponential)
- adjusted model size
- dataset is very big and has large variety
- changing embedding size
- added dropout, tried with many values
- AdamW has L2 regularisation inbuild, played around the parameter
No luck in preventing overfitting
Clone the repo, Go to GUI, install all requirements and run the python file.