Highlighting the Human Side: Covering the Last Mile of Route Optimization

Highlighting the Human Side: Covering the Last Mile of Route Optimization Route optimization is a core component of decision-support systems in logistics. Although decades of research have produced increasingly efficient algorithms and richer models, routes recommended by optimization systems are often substantially modified or disregarded by human planners before implementation. Case studies report deviations between planned and driven routes ranging from 58.6% to 93.1% [1]. These deviations are commonly attributed to contextual knowledge, operational constraints, or planner preferences that are not represented in the optimization model [2]. However, attempts to capture such information through richer models and preference-learning methods have only partially reduced the gap between optimized and implemented routes. We investigate an additional explanation: planners may systematically prefer routes with cognitively salient topological properties. Rather than interpreting every deviation from a model-optimal route as evidence of a missing constraint, we examine whether observed routes reflect recurring perceptual and behavioral patterns. This perspective distinguishes among the numerically optimal solution produced by the model, the solution preferred by the planner, the solution ultimately implemented, and the theoretically best solution under the complete set of relevant real-world objectives and constraints. The aim of a decision-support system is not merely to produce a mathematical optimum, but to favor decisions that are as close as possible to this real-world optimum. Our analysis draws on Gestalt theories of perception, which explain how people organize visual elements into coherent structures whose meaning goes beyond their individual components. Prior studies, which focus on the manual resolution of Travelling Salesman Problems from scratch, have identified consistent preferences for features in the Euclidean space, such as proximity, familiarity, and the avoidance of visually salient crossings [3,4]. However, these findings have largely been obtained in controlled experimental settings rather than from routes produced in real logistics operations. We expand these investigations in two ways. First, we identify cognitive-topological heuristics which are defined on the road-network graph, instead of the Euclidean space alone. Second, we consider a real-world setting, namely 9,184 routes from the Amazon Last Mile Routing Research Challenge (ALMRRC) [5], we compare observed stop sequences with distance-minimizing counterparts generated from the same instances. Our preliminary results show a significant effect of proximity (adherence to the Nearest Neighbors on street graph: ALMRRC 70.15%, optimized 59.01% p < .001, in euclidean space ALMRRC 62.169%, optimized 50.995% p < .001), avoidance of visually salient crossings in euclidean space (mean number of crossings per route: ALMRRC 23.205, optimized 49.826 p < .001), and familiarity on street graph (mean road-network overlap in the single route: ALMRRC 32.03%, optimized 22.99% p < .001). That is, observed routes exhibit systematic structural properties that are not captured by distance optimization alone. Euclidean features are confirmed to have an impact in real-world settings. Graph-based representations further reflect the operational environment in which planners make decisions. Overall, planner deviations cannot be explained solely by omitted operational information: they may also reflect stable cognitive preferences for particular route structures. Following our findings, we propose a methodology to incorporate these patterns into the optimization process, to bring the model optimum closer to the planner-perceived one, thereby improving solution acceptance without significantly increasing business costs. [1] Li Y. et al. Learning from route plan deviation in last-mile delivery. 2018 [2] Nourmohammadi Z. et al. A data-driven preference learning approach for multi-objective vehicle routing problems in last-mile delivery. Transportation Research Part C: Emerging Technologies 2025; 174:105101 [3] Kong X. et al. Global vs. local information processing in visual/spatial problem solving: The case of traveling salesman problem. Cognitive Systems Research 2007; 8:192–207 [4] MacGregor JN. Effects of cluster location and cluster distribution on performance on the traveling salesman problem. Attention, Perception, & Psychophysics 2015; 77:2491–501 [5] Merchán D. et al. 2021 Amazon last mile routing research challenge: Data set. Transportation Science 2024; 58:8–11