'In this project we empirically test various learning AI techniques for their appliance in computer games. This is done through a self-developed modular framework for solving the board game Risk. The framework covers tasks commonly needed in computer games such as analysis of the AIs environment, movement planning, etc.
The AI techniques tested in this project are scripting, decision trees, neural networks, Bayesian networks, and naive Bayes classifiers. We have implemented a scripted AI for Risk which is used to construct training data for the learning techniques.
We discuss different aspects of training the various techniques and the problems we encountered while doing so.
We present a way to test the importance of each module in the framework and discuss different ways of planning tests with a high number of AI participants. We come up with a testing scheme that will eventually present the best AI for playing Risk without having to play all possible AI compositions against each other.
The result of the tests is an AI composed of different AI techniques that each is the best at its task.
The tests show that learning AI techniques can learn how to solve different tasks involved in computer games. In some cases they even outperform their teachers.
Finally we reflect upon the test results, what most the likely appliance of each AI technique should be in computer games.'