National Technical University of Athens (NTUA)
Decentralized approaches to perimeter control in urban road networks using artificial intelligence methods
Abstract
dc:descriptionIn this doctoral dissertation, a complete methodological framework is presented for developing a decentralized traffic management system incorporating both learning-based perimeter control and internal traffic signal control. The overall system is able to regulate exogenous inflows and/or interregional transfers to prevent over-saturation, maximizing regional travel efficiency, while internal traffic signal control supports this by increasing the network’s operational capacity and achieving traffic homogenization. First, a general decentralization framework is developed that uses multi-agent imitation learning to transfer the behavior of a model-free and centralized reinforcement learning-based perimeter controller to decentralized agents which operate at different spatial levels, i.e., regions, boundaries, and individual intersections. To achieve high imitation efficiency at the most granular form of decentralization, a learnable communication layer is introduced where intersection-level agents communicate through compact messages in order to recover the global context that is lost under local observation. Also, communication sparsification is examined to relieve communication overhead. An integrated framework is introduced that couples perimeter control with Max Pressure (MP) control applied to the internal signalized intersections to assess the combined effect of the two control layers. A partial MP deployment is considered during integration following an aggregated-based logic accounting for optimality while post-integration, to account for the implementation costs, a simulation-based heuristic further reduces the number of intersections that are equipped with MP within a given tolerance. All proposed methods were tested in large-scale microsimulation environments that reflect many characteristics of real urban networks, including complex topologies and demand patterns, and realistic intersection layouts. The results are promising and highlight the efficiency of combining reinforcement and imitation learning as the retrieved decentralized perimeter controllers were able to closely match the performance of a fully centralized expert while operating with limited information and no communication, and communication-enhanced variants further improved imitation behavior and coordination at the intersection-level decentralization case. Moreover, our integration logic produced the best overall performance, and we show that different integration logics can lead to distinct behavioral patterns of the perimeter controller. Finally, targeted deployment can outperform full-scale implementation while using only a fraction of the available intersections.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kampitakis, Emmanouil
- Καμπιτάκης, Εμμανουήλ
Subjects
dc:subject × 16- Διαχείριση κυκλοφορίας
- Αποκέντρωση
- Ενισχυτική μάθηση
- Μιμητική μάθηση
- Περιμετρικός έλεγχος
- Τraffic management
- Decentralization
- Reinforcement learning
- Imitation learning
- Perimeter control
- Επιστήμη Πολιτικού Μηχανικού
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη και Τεχνολογία των μεταφορών
- Civil Engineering
- Engineering and Technology
- Transportation science and Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/60729