This paper examines opponent modeling in the
real-time strategy game StarCraft. Actual game
replays are used to identify similar player strategies
via unsupervised QT clustering as an alternative
to relying on expert knowledge for identifying
strategies.
We then predict the strategy of a human
player using two well-known classifiers, artificial
neural networks and Bayesian networks, in addition
to our own novel approach called Action-
Trees. Finally we look at the classifiers’ ability
to accurately predict player strategies given
both complete and incomplete training data,
and also when the training set is reduced in size.