Evolution strategy for optimizing parameters in Ms Pac-Man controller ICE Pambush 3

This paper describes an application of Evolutionary Strategy to optimizing ten distance parameters and seven cost parameters in our Ms Pac-Man controller, ICE Pambush 3, which was the winner of the IEEE CIG 2009 competition. Targeting at the first game level, we report our results from 14 possible optimization schemes; arising from combinations of which initial values to chose, those originally used in ICE Pambush 3 or those randomly assigned, and which parameter types to optimize first, the distance parameters, the cost parameters, or both. We have found that the best optimization scheme is to first optimize the distance parameters, with their initial values set to random values, and then the cost parameters with their initial values set to random values. The optimized ICE Pambush 3 using the resulting parameters from this optimization scheme has an improvement of 17% in the performance for the first game level, compared to the original ICE Pambush 3.