A Bayesian-guided aggregation-disaggregation algorithm for transit planning

Abstract

Public transit is essential for daily mobility, especially for individuals without access to private vehicles. At the same time, designing efficient transit systems remains challenging when demand is spatially dispersed, density is low, and operational resources are constrained. This paper introduces an integer programming model for transit system planning based on a many-to-many (M2M) matching framework. The framework captures detailed routing decisions of the transit fleet while accounting for vehicle capacity, transfer constraints, and user-perceived service quality. To address the scalability challenges of solving this model, we propose a hybrid solution approach that integrates Bayesian optimization (BO) with a structured aggregation-disaggregation (A-D) procedure. In this solution framework, aggregation abstracts the original network into a reduced representation, enabling efficient solution of a smaller and more tractable problem. Disaggregation then maps the resulting aggregated solution back to the original network. Bayesian optimization guides the search over the space of possible network aggregations, enabling efficient evaluation of candidate aggregations that yield high-quality disaggregated transit plans. We evaluate this approach in a real-world case study using data from Benton Harbor, Michigan, U.S.A. Experimental results show that our framework reliably identifies high-quality solutions in settings where the direct M2M solver is unable to produce feasible solutions or delivers consistently weak performance. In addition, the Bayesian-guided exploration markedly improves the efficiency of searching over candidate aggregations compared to random search, enabling efficient discovery of promising network aggregations.