Recent work has shown that genetic algorithms are a good choice for use in game design, particularly for finding improved versions of a game’s parameters to better fit a designer’s requirements. A significant issue with this approach to game optimisation is the very long time it can take to evaluate fitness, since this requires running the target game many times. In this work we test the use of several different fitness approximators, all used in a similar manner, to greatly reduce the number of times a game has to be played for the purpose of fitness evaluation. The approximators use data generated online by the genetic algorithm to train an underlying model. When the model is ready, it is invoked to provide an estimate of the fitness of each newly created individual. If this is worse than a given threshold, it is taken to be the fitness of the individual. Otherwise, the original fitness function is invoked. We assess this approach on two video games Ms PacMan and TORCS. Results are positive and move us one step closer to the goal of a games balancing tool usable in industry.
Available at Springer.com. Also available on direct request.
I had the opportunity to present this work in front of a wonderful audience at EvoStar 2018, during the EvoGames track of EvoApps.
Here are the presentation slides themselves.