Abstract
Urban logistics and on-demand mobility increasingly rely on routing decisions that must be executed under uncertainty, limited communication, and strict service commitments. Requests arise unpredictably, travel conditions fluctuate, and vehicles operate with finite capacities while facing time-windowed obligations whose violations incur disproportionate operational risk. Centralized solvers can be effective when complete information and stable horizons are available, yet their sensitivity to delays, communication bottlenecks, and model misspecification constrains performance in volatile environments. Distributed control based on local autonomy offers resilience, but naïve decentralization can fragment global objectives and induce inefficient myopic behavior. This paper develops a robust agent-based metaheuristic framework tailored to dynamic vehicle routing with time windows in decentralized logistics networks. The design integrates local optimization, market-mediated coordination, and adaptive learning to balance short-horizon responsiveness with long-horizon regularization. Agents maintain compact state, submit value-consistent bids for tasks, and update shadow prices that reconcile capacity, time-window feasibility, and geographical coupling across vehicles. The metaheuristic layer orchestrates destroy–repair neighborhoods, route perturbations, and acceptance criteria with online operator selection that tracks both immediate and deferred impacts. Stability and performance are analyzed using primal–dual regularization and smoothness arguments that link local incentives to global regret. Experiments across synthetic and city-scale instances examine sensitivity to demand burstiness, travel-time volatility, and time-window tightness. Results indicate improved feasibility preservation under congestion and reduced reoptimization burden without reliance on heavy central computation, yielding a practical pathway for sustained performance under uncertainty and partial observability.