mawise / alphahex

Toy implementation of AlphaGo for the game Hex

AlphaHex

I enjoyed learning about AlphaGo, so I decided to try building a simple move predictor for my favorite game, Hex.

I decided to use Apache Spark because I have some experience with it and the Spark machine learning libraries are pretty dead-simple to use.

Status

You can train the model against a database of games. The games don't need to be complete. You can play against the model, the computer must go first and you cannot swap. (yet). The model is predicting what a human player would do based on the data it has trained. This means that a board state as input generates a move as output. The model doesn't have any idea of the rules of Hex, so it doesn't know it cannot play in a cell that is occupied. Unlike AlphaGo, the library I'm using only generates a move, it doesn't generate a probability distribution over moves, so there is no quick-and-dirty way to get the computer to choose a different cell, since it can't play in an occupied cell. My game database existed in SQLite, so I have some code for converting that to a list of comma-separated moves. That's the format you'll want to use.

Update: I implemented the logic again with an MLPC, and the results are much better! Downside is that as of Spark 1.6.1 there is no functionality to save off the model used for that classifier, so the UseMultiLayerPredictor class has to spend some time training the model before you can play hex against the model. Still no logic to esure that the model plays on an empty cell.

Update: I've added functionality to persist a model to disk with a sloppy wrapper around the MultiLayerPredictor. I've also checked in a serialized version of one trained model. The UseMultiLayerPredictor class is configured to load that model and let you play Hex against the AI. The wrapper also exposes all the output values, so it will always make a legal move.

alphahex's Issues

Training set doesn't spin the board

When loading in training data, we should also rotate the board state 180 degrees so the model will be trained by symmetric data.