GECCO 2017 - Speeding up genetic algorithm-based game balancing using fitness predictors

Abstract

Genetic Algorithms (GAs) can find game parameters that fit a designer’s requirements. An issue with this is the long time taken to evaluate fitness, as this requires running the game many times. Here we use fitness predictors, currently neural networks, to speed up the process by reducing the number of fitness evaluations. The predictors are trained using data generated by the GA at runtime. After training, the model is invoked to estimate the fitness of newly created individuals. If the estimate is below a threshold, it is accepted. Otherwise, the original fitness function is invoked. We have used this approach on Ms PacMan with promising results.

Available Online

Available at ResearchGate.net. Also available on direct request.

GECCO 2017