Evolutionary, Neural, or LLM-Driven Heuristic Generation? A Unified Ant Colony Optimization Benchmark for...

Area of Science:

• Computational intelligence and optimization
• Biomimetic algorithms
• Operations research

Background:

• Biomimetic optimization leverages biological mechanisms for computational systems.
• Ant colony optimization (ACO) is a key example, mimicking ant foraging behavior.
• Evaluating different heuristic generation methods within a common ACO framework is needed.

Purpose of the Study:

• To conduct a cross-paradigm evaluation of routing-heuristic generation methods within an ACO solver.
• To compare human-designed rules, genetic programming (GHPP), a neural network proxy, and an offline evolutionary proxy.
• To assess performance based on solution quality, cost, interpretability, and scalability.

Main Methods:

• A standardized interface integrated diverse heuristic generation methods into a single ACO solver.
• Methods included human-designed rules, GHPP, a resource-constrained DeepACO-MLP, and an offline ReEvo-style proxy.
• Evaluations were performed on Traveling Salesperson Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP) instances.

Main Results:

• GHPP consistently produced the shortest routes across all tested scales.
• Offline evolutionary proxy and human-designed rules showed strong performance, forming a second tier.
• The resource-constrained neural proxy's performance degraded significantly with increasing problem size.

Conclusions:

• No single heuristic generation paradigm is universally superior; selection depends on operating constraints and evidence provenance.
• Longer offline search favors GHPP, while explicit rules characterize human and offline evolutionary proxies.
• This benchmark clarifies the interaction of evolutionary, neural, and LLM-style strategies with a fixed biomimetic substrate.