DyNACO: Dynamic neural guidance boosts ant colony optimization to 100k nodes
New AI framework solves training-inference misalignment in neural-guided ACO with dynamic pheromone observation.
DyNACO addresses a key flaw in neural-guided ACO: policies trained to generate static priors (e.g., heatmaps) are deployed in iterative, long-horizon search, leading to a misalignment between training and inference. The framework introduces periodic observation of pheromone distribution and the incumbent solution, enabling truly dynamic guidance. By pairing this policy with a perturbation-based ACO backend and a scope-restricted refinement mechanism, DyNACO ensures stable credit assignment and scalability.
Empirically, DyNACO scales to 100,000-node Traveling Salesman Problem (TSP) instances, outperforming static neural baselines while reducing total runtime compared to the unguided solver. It extends naturally to the Capacitated Vehicle Routing Problem (CVRP) via a capacity-aware backend, improving the unguided baseline with less than 1% neural overhead. Accepted at KDD 2026, the paper provides in-depth analysis of generalization and why dynamic guidance outperforms static priors, underscoring the need to align neural training with iterative search dynamics.
- DyNACO dynamically observes pheromone distribution and incumbent solution to guide iterative search, overcoming static prior misalignment.
- Scales up to 100,000-node TSP instances, outperforming neural baselines while reducing runtime versus unguided solvers.
- Extended to CVRP with less than 1% neural overhead; accepted at KDD 2026 with publicly available code.
Why It Matters
Dynamic neural guidance for combinatorial optimization could slash routing and logistics costs at scale.