Traffic Flow Optimization
In 2000, road traffic congestion in USA alone caused 3.6 billon vehicle-hours of delay, 21.6 billion liters of wasted fuel, and US$67.5 billion in lost productivity. Yearly estimates on economic, health, and environmental cost of traffic congestion in New Zealand is in excess of NZ$1 billion [Hazelton, 2010]. Traditionally, traffic modelling has concentrated on simulating traffic behaviour. The science of traffic analysis, modelling, and optimization aims to estimate traffic load, to detect and prevent traffic congestion, and to optimize the flow of traffic. Optimization of traffic flow not only reduces drivers stress levels, but also reduces air pollution [Angleno, 1999] and controls fuel consumption with respect to the environment and the economy. This proposal directly addresses the later to optimise vehicular gasoline consumption in urban centers by regulating the flow of traffic using smart traffic lights.
Classical traffic models are mostly based on the treatment of vehicles on the road, their statistical distribution, or their density and average velocity as a function of space and time. Most models employ techniques ranging from cellular automata, particle-hopping, car-following, gas-kinetics, through to fluid dynamics present a passive approach to traffic optimization. That is, traffic data is collated apriori and the models are validated posthoc. In a compelling argument for the need to change the manual adjustments to traffic signals, Thorpe [1997] showed, using limited simulation models, that the best traffic signal performance could be achieved using Reinforcement Learning.
Thorpe [1997] reports the re-timing of a major artery in Denver, CO, USA, from 90 seconds to 100 seconds, in the heavy-flow direction, to yield 87% reduction in times vehicles stopped at light. Many urban centers now employ traffic lights that respond to real-time data obtained from devices such as road loops, video cameras, and other traffic detectors [Olsson, 1996]. In contemporary models, traffic situations are represented by statistical or mathematical abstractions and traffic control is exerted by methods that utilize information gleaned from these abstractions. This proposal focuses primarily on loosely modelling the causality of traffic. The causal model then drives the state changes in traffic control. Such a causal model approaches a fully-informed solution; that is, the more we know at real-time about vehicles on the road the smoother the flow of traffic and the better the gasoline usage. The proposed method will be accurate enough to capture the exact nature of undesirable traffic outcomes (e.g., traffic jams, longer wait period, higher number of stops) as well as to model the causality of these undesirable traffic outcomes. It is also possible to direct the traffic to enact a desirable traffic outcome, such as one that clears a pathway for an ambulance or a VIPs convoy. Further, the causal model is updated at real-time and hence the state changes are real-time responses to the dynamics of the causal model.
