Evolving Diverse Cellular Automata Based Level Maps

Abstract

This chapter generalizes a technique for creating terrain maps using a generative fashion based cellular automata representation. The original technique, using fashion based cellular automata, generated terrain maps that exhibit a consistent texture throughout. The generalization presented here co-evolves rules to permit a spatially varying type of map. Pairs of fashion based cellular automata rules are evaluated with objective functions that require connectivity within the terrain and encourage other qualities such as entropic diversity of terrain type, separation of the rule types, and a specified fraction of clear terrain pixels. These three encouraged properties are independently switchable yielding eight different possible fitness functions which are tested and compared. Use of the entropic diversity reward is found to strongly encourage good results while rewarding separation of the two rules without the entropic diversity reward was found to yield bad results with an excess of empty space. The matrix encoding of cellular automata rules yields a discrete granular space encoded with real parameters. Some properties of this space are provided.

The authors would like to thank the University of Guelph for its support of this work.