Decoupling Numerical and Structural Parameters: An Empirical Study on Adaptive Genetic Algorithms via Deep Reinforcement Learning for the Large-Scale TSP

A research team led by Hongyu Wang has published a study demonstrating a novel method for supercharging traditional Genetic Algorithms (GAs) using Deep Reinforcement Learning (DRL). Their framework tackles a core challenge in evolutionary computing: parameter tuning. Instead of treating all algorithm settings equally, they decouple them into two categories. Numerical parameters, like crossover and mutation rates, are tuned for local refinement, while structural parameters, like population size and operator selection, are dynamically reconfigured to prevent stagnation. The system uses a Recurrent Proximal Policy Optimization (PPO) agent as an intelligent controller to manage these settings in real-time.

Experimental results on large-scale Traveling Salesman Problem (TSP) instances are compelling. The AI-optimized GA significantly outperformed static baselines, reducing the optimality gap—the difference from a known best solution—by approximately 45% on the largest tested instance (rl5915 with 5915 cities). Crucially, the study's ablation analysis revealed a fundamental insight: while fine-tuning numerical probabilities helps, the ability to dynamically change the algorithm's structure is the decisive factor for escaping local optima and achieving scalability. This suggests a paradigm shift for automated algorithm design, prioritizing high-level structural plasticity over low-level probability tweaks. The source code is publicly available, facilitating further research and application in complex optimization domains like logistics, chip design, and network routing.