Researchers Use Deep Reinforcement Learning to Discover Interpretable Multi-Parameter Control Policies for Evolutionary Algorithms

A new preprint on arXiv presents a method for using deep reinforcement learning (deep-RL) to derive interpretable, symbolic control policies for evolutionary algorithms operating under multiple simultaneous parameters. The authors identify that while deep-RL has been widely applied to parameter control in evolutionary algorithms, theoretical analysis has remained largely confined to single-parameter settings due to the difficulty of producing interpretable multi-parameter policies. To address this, they introduce several algorithm-agnostic enhancements—including action-space decomposition, reward shifting, and long-horizon discounting—that enable convergence in the multi-parameter regime where standard approaches fail. Comparing common deep-RL methods, they find that Double Deep Q-Networks (DDQN) avoid the policy collapse seen in Proximal Policy Optimization (PPO), producing trajectories amenable to downstream analysis. The learned neural network policies are then distilled into transparent, symbolic rules, providing both interpretability for future theoretical work and exceptional empirical performance that consistently outperforms existing baselines across a wide range of problem sizes. The study uses the (1+(λ,λ))-genetic algorithm optimizing OneMax as its case study, one of the few settings where a super-constant speedup from dynamic control has been formally proven.