A Multiobjective Tuning of a Procedural Content Generator for Game Level Design via Evolutionary Algorithms

This work introduces a new multiobjective modeling approach for fine-tuning the parameters of a procedural game level generator in the platform game Infinite Mario Bros. The optimization problem aims to maximize three objectives related to game difficulty, including enemy placement, types of movements required, and time limits. The multiobjective problem is solved using two well-known evolutionary algorithms, NSGA-II and C-TAEA. In order to evaluate candidate parameter configurations, the averaged values of indicators returned by three artificial intelligent agents playing the levels are considered. A comprehensive computational experiment is conducted, and a statistical comparison using the Wilcoxon test is performed based on hypervolume values. The results include a nondomination analysis, an exploration of the distribution of final solutions, and the illustration of three levels from the final Pareto front. The key contribution of this work lies in the development of a multiobjective methodology that leverages evolutionary algorithms and incorporates agent-based evaluation, providing an effective approach for tuning procedural game level generators.