Performance Evaluation of Reinforcement Learning-Based Adaptive Signal Control Deployment in Urban Network

Abstract

The adaptive traffic control system (ATCS) has shown great potential to address traffic congestion, fuel consumption, and air pollution issues because of its strong adaption to traffic fluctuation. With the popularity of artificial intelligence, Reinforcement Learning (RL) has been applied to ATCS and achieved excellent performance. In RL-based ATCS, the controller collects traffic state information and learns a better signal control policy by minimizing travel delay. For ATCS of an area-wide network, most existing studies assumed that all intersections are installed with ATCS. However, fully ATCS networks are difficult to achieve in the real world due to the high upgrade cost. Therefore, a practical ATCS deployment strategy in a network is needed for agencies with limited budgets. This study proposes a concept of penetration rate, or the percentage of intersections with ATCS in an area, for intelligent intersections. The weighted PageRank algorithm determines a reasonable deployment plan based on road connectivity and link volume. The relationship between penetration rate and traffic performance is investigated in the SUMO simulation. It reveals that the average stopped and control delays gradually decreased with the penetration rate increase. When the penetration rate reaches 75%, the ATSC performance is about 90% of the optimum. However, inequity and inefficiency exist in non-intelligent intersections regarding average queueing time and length at the penetration rate below 40%. Consequently, the suitable range of penetration rate should be between 40% and 75% to balance the costs and benefits of ATCS deployment with budget constraints.