Evolutionary Synthesis of Interpretable Control Policies for Complex Systems

Automated control policies that determine actions in response to changing conditions play a central role in many real-world systems. Reinforcement learning methods based on artificial neural networks have achieved strong performance in such tasks, but their decision-making processes are often opaque, making them difficult to interpret, validate, and trust, particularly in safety-critical domains such as healthcare, transportation, or aerospace.
In my research, I investigate interpretable learning-based control methods using evolutionary computation, with a particular focus on genetic programming. This approach evolves control policies represented as symbolic expressions, graphs, or programs composed of human-readable building blocks. Such representations provide decomposability and transparency, enabling complex policies to be analyzed and understood in ways that are typically not possible with neural networks.
In this talk, I will present results obtained on several benchmark problems, including continuous robot control and visual decision-making tasks, where interpretable policies can achieve competitive performance. I will then discuss key limitations that remain, notably scalability to more complex environments and the computational cost of evolutionary search. Finally, I will outline my ongoing and future research directions aimed at improving the efficiency and scalability of interpretable policy synthesis, with the long-term goal of making interpretable controllers a practical alternative to black-box neural policies.