Expressive Real-time Intersection Scheduling: New Methods for Adaptive Traffic Signal Control

Abstract

Traffic congestion is a widespread problem throughout global metropolitan areas. In this thesis, we consider methods to optimize the performance of traffic signals to reduce congestion. We begin by presenting Expressive Real-time Intersection Scheduling (ERIS), a schedule-driven intersection control strategy that runs independently on each intersection in a traffic network. For each intersection, ERIS maintains separate estimates of the vehicles on each lane, allowing it to more accurately estimate the effects of scheduling decisions than previous schedule-driven control approaches. ERIS outperforms a less expressive schedule-driven control approach and a fully actuated control method in a variety of simulated traffic environments.

We examine several limitations to ERIS and discuss methods to improve schedule quality. First, we consider that when each intersection minimizes a local objective based on delay, global delay is not necessarily minimized. We propose Centralized Agent Reducing Intersection Congestion (CARIC), a centralized method that augments ERIS. CARIC generates potential green wave coordination plans, asks individual intersections to score these plans, and selects a plan that minimizes global delay. As a result of coordination, vehicles are able to travel across the network while encountering successive green lights, ultimately reducing delay and fuel usage when compared against a decentralized control strategy. Additionally, we examine the impact of green wave coordination plans across a variety of demand profiles and train a model estimating when green wave coordination plans are beneficial. Integrating this model with CARIC improves performance when demand is highly variable.

Second, we present two additional limitations of ERIS that lead to inefficient decision making: its finite sensing horizon and its objective function. We examine three extensions to ERIS to address these limitations: generating expected vehicle clusters based on average arrival rates, adapting the objective function to include a term for schedule makespan, and adapting the objective function to weight vehicles travelling along each movement (direction) differently. We notice considerable improvement when adapting the objective function to weight vehicles and examine this extension in further detail. We run additional experiments on a larger network demonstrating improvement, present a library of rules for when adapting the objective function to weight vehicles is beneficial, and develop a queuing model that supports the results of these experiments.