Title: Moving Past Fixed Priors: Implementing Dynamic Neural Guidance in Large-Scale Ant Colony Optimization

Abstract:

Neural-guided Ant Colony Optimization (ACO) encounters a core challenge known as the training-inference misalignment. Typically, these policies are trained to produce static priors, such as heatmaps, but are subsequently deployed to steer iterative, long-horizon search processes. To address this, we introduce DyNACO, a new framework that enables dynamic neural guidance by regularly monitoring both the incumbent solution and the pheromone distribution.

To ensure DyNACO remains computationally feasible at a large scale, we integrate the policy with a perturbation-based ACO backend and a scope-restricted refinement mechanism. This combination guarantees both effectiveness and stable credit assignment. In tests on the Traveling Salesman Problem (TSP), DyNACO successfully scaled to instances with 100,000 nodes. It surpassed neural baselines and frequently decreased total runtime when compared to the unguided solver.

We also adapted DyNACO for the Capacitated Vehicle Routing Problem (CVRP) by employing a capacity-aware backend. This approach consistently enhanced the unguided baseline while incurring less than 1% additional overhead from the neural component. Furthermore, we offer a detailed analysis that confirms the model’s generalization potential and explains the reasons behind the superior performance of dynamic guidance over static priors. Our findings highlight the critical importance of aligning neural training with the dynamics of iterative search in learning-guided optimization. The source code can be accessed at https://github.com/shoraaa/DyNACO.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC