Dynamic partitioning of heterogeneously loaded road networks: A two-level regionalization scheme with Monte Carlo tree search

Abstract

This paper proposes a novel dynamic road network partitioning framework tailored for hierarchical network control based on macroscopic fundamental diagrams. The framework establishes a subregion-region system that can be used for both dynamic road network partitioning and perimeter control strategies through a two-level regionalization model. The first level of regionalization is formulated as a mixed-integer quadratic programming (MIQP) problem, and a specialized max-p region algorithm is designed to solve it. An adaptive large neighborhood search (ALNS) algorithm is introduced to optimize the road network partitioning at the subregion level. Treating each subregion as a fundamental geographic unit, the second level of regionalization is modeled as a mixed-integer linear programming (MILP) model. Due to the significant reduction in the problem size, this model can be solved exactly using a solver. Subsequently, dynamic road network partitioning is achieved by performing multiple boundary subregion movements at discrete time points, based on past network partitioning solutions. This partitioning update process is described using a Markov decision process (MDP), and a Monte Carlo tree search (MCTS) algorithm is designed to iteratively determine the optimal movement actions. The performance of the two-level regionalization method in static road network partitioning is analyzed using the urban road network of Yuelu District in Changsha, China. The dynamic road network partitioning method is tested through simulations on a grid network and the urban road network of Bilbao, Spain. The results validate the effectiveness of the proposed framework, which provides valuable insights and practical support for embedding dynamic road network partitioning methods into network-level traffic control strategies.