I started with trying to do k-Means, ended with k-NearestNeighbours? and simplified it to simple NN, which surprisingly works best. The only special thing is, that I develop the weights for the different features (distance, velocity, etc) via a simple genetic algorithm.
My other bot Sparse? uses something totally different. I tried to adapt a classification method called Sparse Grids to Robocode. It is especially suited for high dimensional problems and large training datasets. In comparison to other regression/classification methods it's fast, but still to slow for Robocode. It works ok, but not as good as I expected. Unfortunately I can't tell too much about it because we are still waiting for an opportunity to publish the method. -- Ebo