Multimodal adaptive traffic signal control: A decentralized multiagent reinforcement learning approach

• A novel multimodal decentralized MARL ATSC is developed. • The proposed method simultaneously improves transit and traffic performance. • The results of vehicle-based and person-based approaches were compared. • The results of the decentralized approach and independent agents were compared. Public transit is considered a compelling alternative to the car, renowned for its affordability and sustainability, given that a single transit vehicle can accommodate a substantially higher number of passengers compared to regular passenger vehicles. In urban areas, a significant portion of the travel time spent by street-running transit vehicles is consumed waiting at traffic signals. Thus, transit signal priority (TSP) strategies have evolved over the years to give preference to transit vehicles at signalized intersections. Traffic signals are usually optimized for the general vehicular traffic flow, with TSP logic subsequently inserted as an add-on to modify the underlying signal timing plans, thereby granting priority to transit vehicles. However, one major issue associated with the implementation of TSP is its negative impact on the surrounding traffic, creating a conflict between prioritizing passenger vehicles versus transit vehicles. This paper proposes a novel decentralized multimodal multiagent reinforcement learning signal controller that simultaneously optimizes the total person delays for both traffic and transit. The controller, named embedding communicated Multi-Agent Reinforcement Learning for Integrated Network-Multi Modal (eMARLIN-MM), consists of two components: the encoder that is responsible for transforming the observations into latent space and the executor that serves as the Q-network making timing decisions. eMARLIN-MM establishes communication between the control agents by sharing information between neighboring intersections. eMARLIN-MM was tested in a simulation model of five intersections in North York, Ontario, Canada. The results show that eMARLIN-MM can substantially reduce the total person delays by 54% to 66% compared to pre-timed signals at different levels of bus occupancy, outperforming the independent Deep Q-Networks (DQN) agents. eMARLIN-MM also outperforms eMARLIN which does not incorporate buses and bus passengers in the signal timing optimization process.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.