Dynamic Variable Speed Limit Zones Allocation Using Distributed Multi-Agent Reinforcement Learning

Variable Speed Limit (VSL) has been proven to be an effective motorway traffic control strategy. However, VSL strategies with static VSL zones may operate suboptimally under traffic conditions with spatially and temporally varying congestion intensities. To enable efficient operation of the VSL system under varying congestion intensities, we propose a novel Distributed Spatio-Temporal multi-agent VSL (DWL-ST-VSL) strategy with dynamic adjustment of the VSL zone configuration. According to the current traffic conditions, DWL-ST-VSL continuously adjusts not only the speed limits but also the length and position of the VSL zones. Each agent uses Reinforcement-Learning (RL) to optimize two goals: maximizing travel speed and resolving congestion. Cooperation between VSL agents is performed using the Distributed W-Learning (DWL) algorithm. We evaluate the proposed strategy using two collaborative agents controlling two segments upstream of the congestion area in SUMO microscopic simulation on two traffic scenarios with medium and high traffic load. The results show a significant improvement in traffic conditions compared to the baselines (W-learning based VSL and simple proportional speed controller) with static VSL zones.