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

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