Smart City: Application of Multi-agent Reinforcement Learning Systems in Adaptive Traffic Management

Traffic congestion on urban road networks has increased substantially during the last decade, characterized by slower speeds, longer travel times, increased vehicular queuing, and increased pollution. The main pain point in traffic management is the static nature of our city structures that cannot adapt to the traffic dynamics changing throughout the day. This work focuses on a futuristic smart city design where making structural changes to the city graph is possible. These changes include modifying number of lanes, opening or closing the ramps, and changing signal timings on road segments. We also assume the local observability of the system where sensors can provide all the data needed for decision making. Under these assumptions, we propose a multi agent reinforcement learning framework for improving traffic flow in city networks. Our learning agents observe their assigned environment and find the best structural changes based on a set of features that represent the recent traffic condition, dependent road segment characteristics, recent structural modifications, etc. Our results show that the proposed framework improves the total travel time (TTT) by 31.5% during rush hours in Salt Lake City area.