Generalization of RL-Assisted Evolutionary Optimization Methods

Establish generalization of reinforcement-learning-assisted evolutionary optimization methods across diverse target problems rather than restricting them to specific problem classes and evolutionary operators suited to a single problem type.

Background

The paper surveys reinforcement-learning-assisted evolutionary optimization methods and identifies generalization as a major unresolved limitation. Many existing approaches are designed for particular target problems and rely on evolutionary operators whose behavior is suitable only for a narrow problem type. Consequently, it remains unresolved whether such methods can transfer effectively across diverse optimization landscapes and applications, including structural model updating tasks.

The proposed DRL-DCO algorithm is motivated in part by this limitation and is evaluated on both mathematical benchmark functions and the IASC-ASCE finite-element model-updating problem. However, the broader generalization challenge for RL-assisted evolutionary optimization methods is not established by the paper as a general result, so it remains an explicit open problem.

References

First, generalization remains largely unresolved, since many methods are developed for specific target problems, constrained by EA operators suited only to a single problem type.

— Learning to Orchestrate Evolutionary Search: Progression-Aware Deep Reinforcement Learning for Dynamic DE-CMA-ES Coordination in Optimization and Structural Model Updating  (2610.11546 - Li et al., 8 Oct 2026) in Section 1.1, paragraph discussing challenges in RL-assisted optimization of evolutionary algorithms