REALT: Real-time Adaptive Learning-based Traffic Control

REALT is a 3-year project (May 2010 - April 2013) supported by Enterprise Ireland Technology Development Programme that is developing for commercialisation a self-organizing Urban Traffic Control (UTC) system that can use road-side sensor technology and vehicular communication capabilities to learn the appropriate traffic light sequences for the current (observed) traffic conditions.

Motivation

Currently deployed traffic control systems are relatively unsophisticated, using either fixed-time controllers with predefined phases (traffic light settings) and phase durations that are specific to the time of the day, or they provide a degree of adaptivity in terms of cycle time, phase length, or offset, but require a large amount of configuration and manual intervention in terms of phase specification, minimum and maximum cycle duration, and grouping of junctions.

Existing adaptive UTC systems (e.g., SCATS, SCOOT) rely on inductive loops for data input, which are expensive to install and maintain. For example, in a study performed in Heuston, TX, USA, the failure rate of inductive loops has been observed to be as high as up to 341 failures a year for an infrastructure of 600-1000 junctions. In addition to infrastructure cost, current UTC systems incur very high human resource costs, due to their complexity, steep learning curves and the time and expertise needed for configuration and management.

Deploying self-organizing UTC systems will remove the major part of such human resources costs, as traffic operators will move towards more high-level monitoring and supervising roles instead of low-level management of traffic systems. Instead of using inductive loops, using data from cheaper, more pervasive, sensors that are already or will be readily available and do not require installation costs, such as for example in-car GPS, can significantly reduce time and cost required for deployment or extension of a UTC system.

The REALT system is designed to exploit any available source of sensor data and, in particular, to take advantage of fine-grained sensor data from cars when available. REALT utilizes machine learning to learn appropriate behaviours for a variety of traffic conditions, in a fully decentralized self-organizing approach where data collection and analysis is performed by the intersections locally. It removes the need for extensive pre-configuration, the system can configure themselves based on the observed conditions and learnt behaviours, reducing the configuration, deployment, and operational time and costs. It also enables timely analysis of large amounts of sensor data and mapping of the current traffic conditions to deploy learnt optimal signal sequences for that given set of conditions.

Approach

REALT will be based on multi-policy optimization and pattern-change detection mechanisms resulting from our research within DWL and Soilse projects.

During the REALT project, we will be performing large-scale realistic simulations in commercial UTC simulators, as well as extending the capabilities of the system to include:

context-aware behaviour (i.e., switching from one learnt behaviour to another based on observed traffic patterns),
on-line tuning of the algorithm’s cooperation parameters to allow learning and exploitation of the relative importance of junctions,
automatic incident detection and intervention, and
improved fault tolerance capabilities to deal with failures and noisy sensors.

Field trials of the resulting UTC system, pending successful simulation studies, are planned for 2012-2013.