Interpretable Control with Graph-Based Genetic Programming

Interpretable decision-making remains a central challenge in both discrete and continuous control, where black-boxblack-boxmodels like Artificial Neural Networks (ANNs) dominate despite their lack of transparencytransparency. In this work, we present a systematic empirical study benchmarking Graph-based Genetic Programming (GGP) as a practical alternative for learning control policies that are both effective and interpretable. We apply two variants—Cartesian Genetic Programming (CGP) and Linear Genetic Programming (LGP)—to a diverse set of benchmark tasks spanning discrete and continuous control. Compared to standard state-of-the-art ANN-based Reinforcement Learning (RL) methods, our approach produces policies with generally competitive performance and significantly smaller, human-readable representations. To assess interpretabilityinterpretability, we conduct a systematic analysis of evolved policypolicygraphs, identifying the rationales driving decision-making. While the evolved policies are indeed compact, we find that meaningful interpretation often requires symbolic post-processing, simplificationsimplification, and visualization, to be truly accessible. Additionally, we highlight the computational inefficiency of current GGP methods relative to RL, suggesting a need for future improvements in optimization and scalability. In summary, our findings on a comprehensive empirical assessment underscore the promise and limitations of GGP for evolving transparent control policies.