Q-learning baseado em pedágios com pagamento circunstancial

Abstract

Congestion is a recurring problem in large cities, leading to productivity loss, pollution, and decreased quality of life. Existing techniques for traffic congestion resolution are not always effective or economically viable. However, implementing toll systems to control traffic flow in busy areas has shown observable improvements. Mathematical analysis and virtual simulation emerge as useful tools to evaluate the cost-effectiveness of each approach. Mitigating congestion involves balancing the optimal system performance and user equilibrium, requiring incentives to align individual driver behavior with system improvements. Toll-based approaches have a theoretical foundation in effectively addressing this issue. However, the assumption that all drivers pay tolls may limit the real-world efficiency of the models due to non-compliance or economic limitations. Addressing these challenges, this work explores the impacts of different levels and modes of participation in toll systems. We adapt an existing toll-based approach to handle diverse scenarios of selective toll payment and investigate the viability of the gradual adoption of toll systems. Having implemented a variation of the TQ-learning algorithm with conditional payment, we can control parameters such as the proportion of toll-paying drivers or the proportion of roads where tolling is obligatory. Through experiments at multiple proportions, we present results that expand the knowledge base for practical decision-making in congestion resolution. Our findings demonstrate that when the toll system is gradually implemented through an increasing proportion of regularly-paying users the gains are steady without introducing chaotic behavior. However, when introducing tolling on a per-route or per-link basis the results were at best inconclusive and, at worst, they caused a deterioration of system performance compared to no implementation.