UCT for PCG

Abstract

This paper describes initial experiments in the use of UCT-based algorithms for procedural content generation in creative game-like domains. UCT search offers potential benefits for this task, as its systematic method of node expansion constitutes an inherent form of exhaustive local search. A new variant called upper confidence bounds for graphs (UCG) is described, suitable for bitstring domains with reversible operations, such as those to which genetic algorithms are typically applied. We compare the performance of UCT-based methods with known search methods for two test domains, with encouraging results.