Improving Prescripted Agent Behavior with Neuroevolution
Ryan Cornelius · 2005
Abstract : Machine learning can increase the appeal of video gamesby allowing agents to adapt in response to the player.Therefore, methods need to be developed specifically forvideo games that adapt agent behaviors in real-time. Forexample, the real-time NeuroEvolution of AugmentingTopologies (rtNEAT) method evolves artificial neuralnetworks (ANNs) fast enough so that improvements can beperceived by the player. However, video game developersare accustomed to relying on prescripted behaviors,frequently encoded in finite state machines (FSMs). It isdifficult to incorporate agents that develop behaviors ontheir own into the current practice. Such learned behaviorsmight be undesirable, violating the designer's intentions.This problem could be avoided if game designers couldspecify an initial behavior using an FSM and allowadaptation. This paper describes such a method,Knowledge-Based NEAT (KB-NEAT), which converts aFSM into an ANN using a KBANN-based technique. Inthis paper, KB-NEAT is tested in the game of blackjack,demonstrating that the FSM successfully converts into anANN with identical behavior and further improves itsperfonnance during the game using NEAT. KB-NEATcan help the game industry utilize machine learningmethods with minimal change to current practices.