Procedural Game Content Generation using Machine Learning

Traditional game level design is a domain that utilizes one of two paths; using human creativity to construct meticulously designed game levels, or using predefined parameters and fixed algorithms to procedurally generate game levels. The consequence of using traditional methods is that most game levels eventually start to resemble each other, exhibiting similar rulesets and mechanical similarities. This issue is exacerbated in game genres like rougelikes, where all game levels are procedurally generated to keep the player’s experience fresh and challenging, despite the underlying monotony in the levels generated. This work uses machine learning techniques instead of fixed algorithms in procedural content generation, mitigating the issue of repetitiveness in levels. A generative adversarial network (GAN) variant is used, called conditional deep convolutional generative adversarial network (CDCGAN) to generate game levels from Super Mario Bros 2, with very little data available. The generation process is improved upon by applying latent space exploration (LSE) using covariance matrix adaptation (CMA) to retain key features like difficulty of the level or the flow of the level. The developed procedural generator achieved a 78% playability score and generated levels with 82% structural consistency compared to human-designed levels.