MA9 - Transport and Routing

Middle-mile delivery drives costs, congestion, and emissions, while underusing truck capacity. We study horizontal collaboration among consolidation carriers using time-dependent Service Network Design and a five-phase combinatorial auction framework. We show how decentralized cooperation can reduce costs while preserving carriers' planning autonomy.

Abstract
This paper presents a comprehensive parameter sensitivity analysis (PSA) of the advanced hybrid graph neural network-reinforcement learning (GNN-RL) optimization framework developed for locating high-speed rail (HSR) freight logistics hubs in the Ottawa–Quebec City corridor. Building upon the study’s innovative integration of fractal spatial analysis, fractional calculus, and a multi-criteria sustainability evaluation, this research systematically investigates the influence, interaction, and robustness of the model’s key input parameters. Utilizing a global sensitivity analysis (GSA) approach anchored by Sobol’s variance-based indices, we quantify the contribution of those parameters across geospatial, economic, operational, and sustainability domains to critical output variances, including total logistics cost, CO₂ emissions reduction, service coverage, and social equity scores. Our findings reveal that population density distribution and eco-network integrity indices are the primary first-order drivers of hub location efficiency. Crucially, the analysis uncovers significant higher-order interaction effects, particularly between fractal accessibility coefficients and carbon shadow pricing, which dictate trade-offs in the Pareto-optimal solution space. The model demonstrates robust performance within defined parameter bounds yet exhibits critical sensitivity thresholds in demand elasticity and infrastructure discount rates that warrant policy attention. This study not only validates the resilience of the interdisciplinary GNN-RL framework but also provides a replicable, data-driven methodology for prioritizing investment, refining data collection, and implementing adaptive governance in sustainable freight corridor planning.

We study how to set free-shipping thresholds dynamically in on-demand delivery. The threshold affects customers' basket sizes and operational efficiency. We model the problem as a Markov decision process and develop reinforcement-learning policies that adapt thresholds to real-time environment conditions. Numerical experiments show higher profits than static threshold policies.

We study methods to approximate road network distances on dense urban maps considering downstream optimization tasks such as route planning. In many applications, distances are obtained from a routing engine on demand, creating millions to billions of origin–destination queries and making distance retrieval a major computational bottleneck. We treat road distances as values of an arbitrary, possibly asymmetric oracle and build surrogate models that approximate these distances while preserving routing quality. We train neural networks by direct regression and decision-focused learning methods using grid-representation based network-level features. We consider the asymmetric traveling salesperson problem in a real last-mile service area in Montréal, using the Open Source Routing Machine as oracle. Our learned surrogates are more than 23 times faster than direct queries to the routing engine and also improve the solution quality by more than 92% compared to Euclidean baseline.