Real-time optimization of traffic flow addresses important practical problems:
reducing a driver's wasted time, improving city-wide efficiency, reducing gas
emissions, and improving air quality. Many current implementations and research
studies that address traffic signal control construct a light controller's program
(whether adaptive or static) by segmenting the day into divisions in which distinct
traffic patterns are expected: rush hours, weekends, nights, etc. We consider the
problem of automatically adapting a set of traffic lights to changing conditions
based upon the distribution of observed traffic-density in surrounding areas.
Unlike previous techniques which specify an a priori set number of unique flow
patterns, we assume an over-complete set of traffic patterns. A combination of
machine learning approaches are used to create a diverse set of traffic-light
programs that can be instantiated when new traffic flow patterns are recognized. We
have observed significant reduction in expected emissions and delays, while being
agnostic to the number of underlying distinct patterns in the traffic.