Online Traffic Signal Control through Sample-Based Constrained Optimization

Abstract

Traffic congestion reduces productivity of individuals by increasing time spent in traffic and also increases pollution. To reduce traffic congestion by better handling dynamic traffic patterns, recent work has focused on online traffic signal control. Typically, the objective in traffic signal control is to minimize expected delay over all vehicles given the uncertainty associated with the vehicle turn movements at intersections. In order to ensure responsiveness in decision making, a typical approach is to compute a schedule that minimizes the delay for the expected scenario of vehicle movements instead of minimizing expected delay over the feasible vehicle movement scenarios. Such an approximation degrades schedule quality with respect to expected delay as vehicle turn uncertainty at intersections increases.

We introduce TUSERACT (TUrn-SamplE-based Real-time trAffic signal ConTrol), an approach that minimizes expected delay over samples of turn movement uncertainty of vehicles. Specifically, our key contributions are: (a) By exploiting the insight that vehicle turn movements do not change with traffic signal control schedule, we provide a scalable constraint program formulation to compute a schedule that minimizes expected delay across multiple vehicle movement samples for a traffic signal; (b) a novel mechanism to coordinate multiple traffic signals through vehicle turn movement samples; and (c) a comprehensive experimental evaluation to demonstrate the utility of TUSERACT over SURTRAC, a leading approach for online traffic signal control which makes the aforementioned approximation. Our approach provides substantially lower (up to 60%) mean expected delay relative to SURTRAC with very few turn movement samples while providing real-time decision making on both real and synthetic networks.