The Invisible Referee

21st Century Traffic Control: The Invisible Referee University of Southampton

Artificial Intelligence in Signal Control

E-mail Print PDF

Mobile Users

Watch video...      View images...

Computer Game

Researchers in the Transportation Research Group at the University of Southampton have developed a computer game that allows the player to control a simulation of a traffic light controlled junction. The computer game was built by adapting traffic simulation software (S-Paramics) to have a human interface. In fact an updated version of this game is going to be on display at our Royal Society Summer Science exhibit.

One interesting aspect about the Junction Control computer game is that while it is being played it can collect statistics and analyse how good the player is at controlling a junction. It can even be used to compare the player's performance to real world junction control systems that are in use today. The, perhaps surprising, result is that human beings can be very good junction controllers and in many cases players of the game significantly outperform today's automated traffic light controllers.

The Human Trainer

This motivated the Transportation Research Group to investigate Machine Learning technology to see if the computer game could be used to learn what it was that made the human players such good traffic light controllers. A Neural Network Classification computer program that was originally developed to teach a computer how to play backgammon was adapted for use in the Junction control game.

The program works by observing the human playing the computer game and picking out statistical patterns between the situation on the roads around the junction and the decisions that the human is making about who to give the light to. In this way the computer program becomes trained. Then the trained computer program can control the junction itself using what it has learned and in many cases the trained computer program can match the performance of its human trainer.

Reinforcement Learning

Current research in the Transportation Research Group is focused on seeing if the learning computer program can be improved so that it can beat the human's performance. The approach to doing this is to try and build a junction controller that can learn through experience. To do this a technique called Reinforcement Learning is used. Here when the learning junction controller makes a decision we have the computer game analyse if the decision was good or bad and supply feedback to the junction controller so it can decide whether to try this decision again in the future or try something else instead. This way, over time the junction controller can learn for itself the best way to control the traffic lights.

Mobile Users

Watch video...      View images...