# Interpretable Control Policies: What They Mean and How to Evolve Them _by Giorgia Nadizar (University of Trieste, Italy)- 2025.10.16_ ###### tags: `VAADER` `Seminar` ![image](https://hackmd.io/_uploads/BJqkKUd5gg.png) ## Abstract Interpretability is often treated as an objective property of models, yet what counts as "interpretable" is inherently subjective and depends on the user and task. In this talk, I focus on interpretability in control, where understanding a policy’s behavior is essential for trust, debugging, and discovery. I use graph-based genetic programming to evolve compact, structured policies with intuitive semantics, and then apply quality-diversity optimization to discover a diverse set of such policies, offering multiple perspectives for solving the problem and improving human understanding. Finally, I extend these ideas to visual control, where interpretable policies outperform conventional XAI techniques in revealing what the controller "really does". Together, these methods enable controllers that are both effective and insightful.