Machine learning can increase the appeal of video games by allowing agents to adapt in response to the player. Therefore, methods need to be developed specifically for video games that adapt agent behaviors in real-time. For example, the real-time NeuroEvolution of Augmenting Topologies (rtNEAT) method evolves artificial neural networks (ANNs) fast enough so that improvements can be perceived by the player. However, video game developers are accustomed to relying on prescripted behaviors, frequently encoded in finite state machines (FSMs). It is difficult to incorporate agents that develop behaviors on their own into the current practice. Such learned behaviors might be undesirable, violating the designer's intentions. This problem could be avoided if game designers could specify an initial behavior using an FSM and allow adaptation. This paper describes such a method, Knowledge-Based NEAT (KB-NEAT), which converts a FSM into an ANN using a KBANN-based technique. In this paper, KB-NEAT is tested in the game of blackjack, demonstrating that the FSM successfully converts into an ANN with identical behavior and further improves its performance during the game using NEAT. KB-NEAT can help the game industry utilize machine learning methods with minimal change to current practices.

PDF

Technical Report HR-05-01, Department of Computer Sciences, The University of Texas at Austin, 2005.

@techreport{cornelius:ugthesis05, title={Improving Prescripted Agent Behavior with Neuroevolution}, author={Ryan Cornelius}, number={HR-05-01}, school={Department of Computer Sciences, The University of Texas at Austin}, institution={Department of Computer Sciences, The University of Texas at Austin}, type={Undergraduate Honors Thesis}, url="http://nn.cs.utexas.edu/?cornelius:ugthesis05", year={2005} }