Learning to Speed Up Evolutionary Content Generation in Physics-Based Puzzle Games

Abstract

Procedural content generation (PCG) systems are designed to automatically generate content for video games. PCG for physics-based puzzles requires one to simulate the game to ensure feasibility and stability of the objects composing the puzzle. The major drawback of this simulation-based approach is the overall running time of the PCG process, as the simulations can be computationally expensive. This paper introduces a method that uses machine learning to reduce the number of simulations performed by an evolutionary approach while generating levels of Angry Birds, a physics-based puzzle game. Our method uses classifiers to verify the stability and feasibility of the levels considered during search. The fitness function is computed only for levels that are classified as stable and feasible. An approximation of the fitness that does not require simulations is used for levels that are deemed as unstable or unfeasible by the classifiers. Our experiments show that naively approximating the fitness values can lead to poor solutions. We then introduce an approach in which the fitness values are approximated with the average fitness value of the levels' parents added to a penalty value. This approximation scheme allows the search procedure to find good-quality solutions much more quickly than a competing approach—we reduce from 43 to 25 minutes the running time required to generate one level of Angry Birds.