AI-Driven Dynamic Routing Under Uncertainty: Integrating Machine Learning and Optimization for Cost-Efficient and Sustainable Logistics Networks

Abstract

Vehicle routing sits at the center of logistics cost and logistics emissions, yet the routes that carriers execute are usually planned on deterministic estimates of travel time and demand that are wrong by the time a vehicle leaves the depot. Machine learning now predicts travel times, demand and request arrivals with useful accuracy, and optimization methods now solve large routing problems close to optimality, but the two are often joined only loosely: forecasts are reduced to point values before they reach the optimizer, and the optimizer is rerun from scratch when conditions change. This paper reviews the literature on dynamic and stochastic vehicle routing, learning-based routing, and green routing, and organizes it around the question of how prediction, uncertainty and optimization should be connected. It classifies uncertainty sources and solution paradigms and compares classical stochastic and robust models, approximate dynamic programming and rollout, end-to-end deep reinforcement learning, and hybrid learn-and-optimize methods. On this basis it proposes the Uncertainty-Aware Learn-and-Optimize Routing framework, which passes calibrated predictive distributions rather than point forecasts to the optimizer, represents uncertainty through scenarios, budgets and chance constraints, embeds cost, emissions and tail risk in one objective, and re-plans through an event-triggered rollout or learned policy. The paper closes with a research agenda on calibration, benchmark realism, electrified fleets and governance.