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

2026-03-24 · Source: cs.NE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Combinatorial Optimization · Depth: Advanced, long

Summary

A new study introduces DRLGA, a dual-level Deep Reinforcement Learning Genetic Algorithm framework, to address the challenge of parameter configuration in Evolutionary Algorithms (EAs) for large-scale Traveling Salesman Problems (TSPs). This framework decouples control variables into numerical parameters (e.g., crossover and mutation rates) and structural parameters (e.g., population size and operator switching), hypothesizing distinct roles. Utilizing a Recurrent PPO agent with LSTM, DRLGA dynamically regulates these parameters, acting as a probe into evolutionary dynamics. Experiments show DRLGA policies outperform static baselines, reducing the optimality gap by approximately 45% on the largest tested instance (rl5915). Ablation analysis reveals that structural plasticity is the decisive factor in preventing stagnation and escaping local optima, while numerical tuning offers local refinement. The source code is available on GitHub for reproducibility.

Key takeaway

For research scientists developing optimization algorithms for NP-hard problems, this work suggests a paradigm shift. You should prioritize designing systems that dynamically adapt structural parameters, such as population size and operator modes, over merely fine-tuning numerical rates. Focusing on structural plasticity will yield greater scalability and prevent stagnation in high-dimensional search spaces, leading to more robust and effective solutions for complex problems like the TSP.

Key insights

Dynamic structural adaptation is more critical than numerical fine-tuning for scalable evolutionary algorithms.

Principles

• Structural plasticity prevents stagnation.
• Numerical tuning offers local refinement.
• Zero-shot generalization is achievable with scale-invariant state representation.

Method

A dual-level DRL framework with a Recurrent PPO agent and LSTM dynamically controls numerical and structural parameters in Genetic Algorithms, treating optimization as a POMDP.

In practice

• Prioritize dynamic structural reconfiguration in algorithm design.
• Use LSTM-based DRL for non-Markovian evolutionary histories.
• Employ curriculum learning for robust generalization.

Topics

• Deep Reinforcement Learning
• Genetic Algorithms
• Traveling Salesman Problem
• Adaptive Parameter Control
• Combinatorial Optimization

Code references

• StarDream1314/DRLGA-TSP

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Large Language Model-Driven Full-Component Evolution of Adaptive Large Neighborhood Search — LLM-driven evolution can automate and enhance complex metaheuristic algorithm design by decoupling and optimizing core components.
• CADO: From Imitation to Cost Minimization for Heatmap-based Solvers in Combinatorial Optimization — Objective mismatch in heatmap-based CO solvers is resolved by RL fine-tuning that directly optimizes post-decoded solution cost.
• Model-Driven Policy Optimization in Differentiable Simulators via Stochastic Exploration — MDPO enhances differentiable planning by adaptively injecting stochastic noise for improved exploration in complex optimization landscapes.
• Discovering Lattice Reduction Strategies via Self-Play — Deep reinforcement learning can discover superior, generalizable lattice reduction strategies beyond traditional algorithms like LLL.
• A Model-Driven Approach for Developing Families of Reinforcement Learning Environments — A model-driven approach using a hybrid genetic algorithm automates the creation of diverse reinforcement learning environment families.
• Beyond Single-Model Optimization: Preserving Plasticity in Continual Reinforcement Learning — Preserving diverse policy neighborhoods in a stable latent space enhances continual reinforcement learning plasticity.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.NE updates on arXiv.org.