Comparison of Evolution Strategy, Genetic Algorithm and Their Hybrids on Evolving Autonomous Game Controller Agents

Researchers have been applying artificial/computa-tional intelligence (AI/CI) methods to computer games. In this research field, further researches are required to compare AI/CI methods with respect to each game application. In this paper, we report our experimental results on the comparison of two evolutionary algorithms (evolution strategy and genetic algorithm) and their hybrids, applied to evolving autonomous game controller agents. The games are the CIG2007 simulated car racing and the MarioAI 2009. In the application to the simulated car racing, premature convergence of solutions was observed in the case of ES, and GA outperformed ES in the last half of generations. Besides, a hybrid which uses GA first and ES next evolved the best solution among the whole solutions being generated. This result shows the ability of GA in globally searching promising areas in the early stage and the ability of ES in locally searching the focused area (fine-tuning solutions). On the contrary, in the application to the MarioAI, GA revealed its advantage in our experiment, whereas the expected ability of ES in exploiting (fine-tuning) solutions was not clearly observed. The blend crossover operator and the mutation operator of GA might contribute well to explore the vast search space.