OD-DEAL: Dynamic Expert-Guided Adversarial Learning with Online Decomposition for Scalable Capacitated Vehicle Routing

Solving large-scale capacitated vehicle routing problems (CVRP) is hindered by the high complexity of heuristics and the limited generalization of neural solvers on massive graphs. We propose OD-DEAL, an adversarial learning framework that tightly integrates hybrid genetic search (HGS) and online barycenter clustering (BCC) decomposition, and leverages high-fidelity knowledge distillation to transfer expert heuristic behavior. OD-DEAL trains a graph attention network (GAT)-based generative policy through a minimax game, in which divide-and-conquer strategies from a hybrid expert are distilled into dense surrogate rewards. This enables high-quality, clustering-free inference on large-scale instances. Empirical results demonstrate that OD-DEAL achieves state-of-the-art (SOTA) real-time CVRP performance, solving 10000-node instances with near-constant neural scaling. This uniquely enables the sub-second, heuristic-quality inference required for dynamic large-scale deployment.

💡 Research Summary

OD‑DEAL (Dynamic Expert‑Guided Adversarial Learning with Online Decomposition) tackles the scalability‑optimality dilemma of large‑scale Capacitated Vehicle Routing Problems (CVRP). The authors first construct a high‑fidelity expert oracle by augmenting the state‑of‑the‑art Hybrid Genetic Search (HGS) with an online Barycenter Clustering (BCC) decomposition. Starting from an initial feasible solution, the barycenters of each route are computed, k‑means clustering partitions the routes into geographically coherent sub‑problems, and HGS is run in parallel on each sub‑problem. The locally optimized sub‑solutions are then merged, yielding a “divide‑and‑conquer” expert solution that captures both global routing logic and fine‑grained local improvements.

The neural generator is a Graph Attention Network (GAT) encoder followed by an autoregressive decoder. The encoder projects node features (coordinates, demand) and edge features (travel cost) into a common hidden space, then applies multi‑head self‑attention to model long‑range spatial, demand, and residual capacity interactions. The decoder builds a route step‑by‑step, scoring each candidate transition with a multilayer perceptron that consumes concatenated source‑target embeddings. A dynamic mask enforces the CVRP constraints (single visit, capacity) and a softmax over masked logits yields a stochastic policy.

Training departs from conventional reinforcement learning. OD‑DEAL adopts the GFlowNet paradigm, using the Trajectory Balance (TB) objective to align the generative flow with the expert distribution. An adversarial discriminator evaluates the divergence between trajectories sampled from the generator and those produced by the expert oracle. The generator receives dense surrogate rewards derived from the expert’s decomposition logic (e.g., reduction in sub‑problem cost, balance of cluster loads), which mitigates reward sparsity and encourages the policy to internalize the divide‑and‑conquer strategy without explicit clustering at inference time.

Empirical evaluation spans CVRP instances from 1 000 to 10 000 customers. OD‑DEAL consistently outperforms recent neural baselines (Attention Model, POMO, GFN variants) in solution quality, achieving 2–5 % lower average cost. Compared with pure heuristic solvers (HGS, LKH‑3), it reaches near‑optimal costs while delivering sub‑second inference, even on the 10 k‑node benchmark. Notably, the computational cost grows only marginally with problem size, demonstrating near‑constant neural scaling—a direct benefit of training with the decomposition‑augmented expert rather than scaling the neural architecture itself. Ablation studies confirm that (i) removing BCC from the expert degrades performance, (ii) replacing the GAT encoder with a standard GCN reduces solution quality, and (iii) training without the adversarial discriminator leads to higher variance and poorer convergence.

The paper acknowledges limitations: the current framework is tailored to CVRP and would require redesign of the expert oracle and reward shaping for problems with time windows, heterogeneous fleets, or multi‑objective criteria. Moreover, the approach inherits the dependence on the expert’s quality; a weak or computationally expensive oracle could diminish the benefits. Nonetheless, OD‑DEAL establishes a compelling blueprint for integrating high‑performance heuristic knowledge into adversarial flow‑based learning, opening pathways for real‑time, large‑scale combinatorial optimization in logistics, transportation, and beyond.