Halite Challenge
Based on https://github.com/brianvanleeuwen/Halite-ML-starter-bot
python train_bot.py ./replays
Model
For every point we look at 15*15 patch around it and use a small convolutional network for classification.
The model uses 2D Convolutions with border_mode as 'same' and kernel size of 3.
The exact model looks like
Conv2D(32)
ParametricReLU
MaxPool(2, 2)
Conv2D(64)
ParametricReLU
MaxPool(2, 2)
Flatten()
Linear(256)
ParametricReLU
Linear(128)
ParametricReLU
Linear(5)
No post processing was applied, I believe a small post processing could have improved results by a good amount, but due to limited time I could put in for this, my major concern was learning.
I used adam with cross entropy loss function.
The replays were downloaded using scrapeData.py script.
Issues
- The bot timed out on large board sizes at times because of unavailability of GPUs and convolutions being calculated on CPU.
Results
The model finished 94th on the final list.