Abstract
This seminar will look at how pattern recognition techniques can be used to completely define signalized junction control strategies by a classification of state space (training). Two methods for training junction controllers will be discussed. A supervised learning technique where the junction controller is trained by a human expert using a computer game interface. And a reinforcement learning technique where the junction controller learns strategies through experience using temporal difference (TD) learning. We present simulation results comparing the performance of “Machine Learning” junction controllers with the SCOOT system, which they outperform significantly.
Biography
Simon Box is a Senior Research Fellow in the Transportation Research Group at the University of Southampton. He has a research background in dynamic stochastic computer simulation of engineering systems. He has worked in both industry and academia where he has developed simulation and modelling software for emissions measurement devices, passively guided rockets, capital asset management and most recently Urban Traffic Control.