HELSINKI UNIVERSITY OF TECHNOLOGY
ABSTRACT OF MASTER'S THESIS 

Author:			Ella Bingham  
Title of thesis:	Neurofuzzy Traffic Signal Control 
Finnish title:		Neurosumea liikennevalo-ohjaus 
Date:			1st September, 1998 
Pages:			107
Department:		Department of Engineering Physics and Mathematics 
Chair:			Mat-2 Applied Mathematics  
Supervisor:		Professor Harri Ehtamo 
Instructor:		Lic.Tech. Jarkko Niittymäki
Keywords:		fuzzy logic, neural networks, neurofuzzy systems,
			reinforcement learning, traffic signal control

The aim of this work was to create an adjustable fuzzy traffic signal 
controller. An existing fuzzy traffic signal controller was enhanced
with a learning algorithm. This adjustable controller can modify its
parameters in different traffic situations, and thus reach a better 
control result. The performance of the traffic signal controller is 
measured by the delay of vehicles.

A fuzzy traffic signal controller uses linguistic rules such as ``if the 
approaching traffic volume is large and the queuing traffic volume is small, 
then the green signal is long''. The fuzzy concepts large, small and long
are presented using membership functions. 

Neural networks consists of simple processing elements interconnected as a 
structured network. In a neurofuzzy controller, the parameters of the fuzzy 
membership functions are adjusted using a neural network. The neural 
learning algorithm in this work is reinforcement learning. The neurofuzzy 
system under consideration is such that the most usual neural learning 
algorithms cannot be used.

The adjustable traffic signal controller is studied in a traffic simulation
system which includes a fuzzy signal controller. The neural learning
algorithm is realized in a Matlab program which interacts with the traffic
simulation system.

The learning algorithm is found successful at constant traffic volumes. 
Starting from the initial membership functions, the learning algorithm 
modifies the parameters of the membership functions in different ways at
different but constant traffic volumes. The membership functions after the
learning produce smaller delays than the initial membership functions.
The learning algorithm is not found successful in situations where the
traffic volume changes rapidly.
 
An additional contribution of this thesis is a small manual modification in 
the rule base of the fuzzy traffic signal controller. This modification 
reduces the delays significantly at low traffic volumes.