ML-Optimized Stochastic Vehicle Routing for Urban Delivery Services

Abstract

Last-mile delivery routing in cities is a complex problem mainly due to two main sources of uncertainty: random demand arrivals and traffic conditions that change over time and influence travel times while the delivery is happening. This paper presents a practical replication and extension study of the SVRPBench stochastic vehicle routing framework, adapted and evaluated in a new urban context–Chandigarh, India. We create a complete city-wide pipeline that: (1) gets spatio-temporal order arrivals from open data sources, (2) learns demand and traffic prediction models (like ARIMA, LSTM, and decision-focused SPO+ approaches), and (3) combines these forecasts with optimisation and metaheuristic methods (like OR-Tools, ALNS, ACO, ABC, and nearest_neighbor heuristics) to make strong delivery routes. A scenario-based simulator tests how well a route holds up under real-world random changes, such as traffic jams, noise in travel times, and accidents. Swarm-based metaheuristics are consistently outperformed by improvement-based heuristics and classical solvers, according to experimental evaluation across representative instances. While ALNS shows better service quality with 12.34% less mean lateness than ABC, OR-Tools achieves the lowest mean cost per delivery (4.09% improvement over the lowest-performing solver). Lower-cost solutions are typically associated with better service performance, according to a moderately positive correlation (r = 0.667) between cost and lateness. Nearest_neighbor achieves about 65% less runtime than ALNS while maintaining competitive solution quality, according to computational efficiency analysis. For urban delivery networks, the study shows that deterministic and improvement-based routing strategies are still very competitive, especially when balancing cost effectiveness and service dependability in stochastic environments. An accessible platform for operational decision support is provided by a lightweight web prototype that displays routes, interruptions, and KPIs.