Alternating between Surrogate Model Construction and Search for Configurations of an Autonomous Delivery System (Hot off the Press at GECCO 2025)
Chin-Hsuan Sun, Thomas Laurent, Paolo Arcaini, Fuyuki Ishikawa · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2025
Demand for goods delivery services keeps growing, creating problems with last mile delivery. While autonomous delivery robots provide a solution to many of these problems, configuring their deployment to balance cost, performance, and safety is complex. Our industry partner Panasonic uses a multi-objective search approach to find configurations for their system, leveraging a simulator to assess the quality of the different configurations. However, due to the complexity of the simulator, this approach proved very computationally expensive. In this work, we propose an approach that uses a surrogate model to speed up fitness computation. Since building the surrogate model has a cost itself, the approach trains the model incrementally during the search: first, an imprecise model is used, while, as the search progresses, more precise models are used. The search stops when no improved solutions can be found. Experimental results show that this new approach, while it produces configurations of slightly lower quality, requires significantly less computing time. This is an extended abstract of "Alternating Between Surrogate Model Construction and Search for Configurations of an Autonomous Delivery System" by C.-H. Sun, T. Laurent, P. Arcaini, and F. Ishikawa, IEEE International Conference on Software Analysis, Evolution and Reengineering (SANER 2024).