Adversarial Random Forest Classifier for Automated Game Design

This research investigates the use of an adversarial random forest classifier to automate game design, specifically addressing the limitation of relying on hand-authored fitness functions in search-based procedural content generation. The study explores whether a machine learning model can learn to approximate human-like game quality by iteratively training a classifier to distinguish between human-authored games and algorithmically generated ones. The scope of the work is limited to games developed within the Video Game Description Language (VGDL) framework, commonly used in the General Video Game Artificial Intelligence competition.

The methodology involves a greedy hill-climbing search guided by a random forest classifier, which is retrained in an adversarial loop using eight human-authored games as a baseline. The researchers tested two versions of the system—one including termination conditions and one without—to evaluate the impact of game representation on the classifier's ability to learn structural patterns. Performance was assessed by comparing the success rates and behaviors of a Monte Carlo tree search agent against a random agent across five iterations of the system.

The findings indicate that while the system failed to produce games of consistent human quality, the random forest successfully identified specific structural features of human-authored games, such as the prevalence of certain sprite types. However, the classifier’s utility as a fitness function diminished over time as it became less effective at discriminating between human and generated content in later iterations. The authors conclude that the performance of the agent was heavily influenced by the evaluation methodology and the limitations of the input representation, suggesting that future research should focus on more advanced search algorithms and more robust feature extraction to better capture designer intent.