Deep RL discovers interpretable control policies for evolutionary algorithms

A team led by Tai Nguyen from Sorbonne University and partners has demonstrated how deep reinforcement learning can overcome a long-standing barrier in evolutionary computation: deriving interpretable multi-parameter control policies. Their work, submitted to arXiv in June 2026, uses the (1+(λ,λ))-genetic algorithm optimizing the OneMax problem as a case study—a rare setting where dynamic parameter control has proven super-constant speedup. The researchers found that standard deep RL approaches (like PPO) often collapse in this multi-parameter space, but Double Deep Q-Networks (DDQN) with carefully designed action-space decomposition, reward shaping, and discounting produce stable trajectories.

The key breakthrough is moving beyond the black-box nature of neural networks. By distilling learned behaviors into a transparent, symbolic control policy, the team achieved both interpretability for theoretical analysis and exceptional performance—consistently beating all existing baselines across a wide range of problem sizes. This opens the door to formally studying how multiple parameters should be dynamically controlled in evolutionary algorithms, a problem previously limited to single-parameter settings. The paper is significant for researchers in evolutionary computation, neural architecture search, and automated algorithm configuration, offering a principled method to create policies that are both high-performing and amenable to rigorous mathematical analysis.