benzyx / generalscnn

Convolutional neural network for playing generals.io

This is a bot designed for playing generals.io using a convolutional neural network (http://bot.generals.io/profiles/%5BBot%5D%20%5Buw%5D%20zxqfl).

The program is undocumented and probably difficult to understand. When writing it I didn't plan for it to be read by other people. However, here is some information that will help you understand it if you're interested:

• The most important files are featurize.py, which converts a game state to a feature vector, and tf_model.py, which defines and trains a model to map feature vectors to move probabilities.
• The paths in 1v1_sorted_paths.txt refer to files from http://dev.generals.io/replays.

The workflow looks like this:

1. Get some replays and feed them to simulator.js.
2. Pipe the output of the simulator into featurize.py, which will write features to files. (This takes a while.)
3. Train a model using tf_model.py with the features produced in the previous step.
4. Once the model is trained, use it on the server with run_bot.py.

You will want a GPU for step 3 and possibly step 4 too.