Increasing Motorway Operational Capacity by Dynamic Adaptation of Variable Speed Limit Zones Based on Cooperative Learning Agents

Continuously increasing demand for mobility, especially in road transport, and new transport policies that prioritize optimizing the capacity of existing roads over building new ones, pose major challenges in the field of traffic control towards a competitive and resource-efficient transport system. This thesis is primarily concerned with urban motorways and their characteristic problems, the presence of bottlenecks. Variable speed limit (VSL) as an effective traffic control strategy is used to optimize traffic on motorways. Due to the stochastic nature of traffic congestion, static VSL zones, i.e., the VSL application area remains unchanged, may not always respond accurately to spatiotemporal spreading congestion. Accordingly, a model for dynamic spatiotemporal adaptation of VSL zones that complements the prevailing VSL with static zones is proposed. To address the control problem for the proposed dynamic VSL zones, a novel distributed multi-agent reinforcement learning VSL (DWL-ST-VSL) controller, capable of dynamically adapting the length and position of VSL zones during VSL operation, is suggested. To ensure rigorous evaluation of the controller, a novel paradigm in motorway traffic modeling is proposed and demonstrated by using the continuously synchronized digital twin (DT) model of the Geneva motorway (DT-GM). Thus, DWL-ST-VSL applicability is verified through a virtual field test, which uses DT-GM microscopic simulation of actual traffic received in real-time from the Geneva motorway. Such functional integration presents a novelty in evaluating learningbased traffic controllers using traffic scenarios that more closely reflect the uncertainty in real traffic. Accordingly, DT-GM advances technological capabilities of traffic simulations to a more sophisticated and accurate level by enabling simulation-based control optimization during system run-time. Furthermore, DWL-ST-VSL outperforms all baseline control approaches with static VSL zones. Different VSL zones configurations per traffic scenarios are, thus, learned without requiring a manual setup. This is of practical relevance as it can guide traffic engineers in optimizing VSL zones design in complex motorway environments. Thus, demonstrated functional integration of the DWL-ST-VSL with the DT technology provides the basis for further development of predictive real-time analytics of complex autonomous control processes as a step toward safety-critical decision-making in a nonstationary traffic setting and provides an experimental environment for developing control system prototypes in the context of emerging connected vehicles technologies.