In most high-risk applications, interpretability is crucial for ensuring system safety and trust. However, existing research often relies on hard-to-understand, highly parameterized models, such as neural networks. In this paper, we focus on the problem of policy search in continuous observations and actions spaces. We leverage two graph-based Genetic Programming (GP) techniques—Cartesian Genetic Programming (CGP) and Linear Genetic Programming (LGP)—to develop effective yet interpretable control policies. Our experimental evaluation on eight continuous robotic control benchmarks shows competitive results compared to state-of-the-art Reinforcement Learning (RL) algorithms. Moreover, we find that graph-based GP tends towards small, interpretable graphs even when competitive with RL. By examining these graphs, we are able to explain the discovered policies, paving the way for trustworthy AI in the domain of continuous control.
