Artificial neural nets and genetic algorithms proceedings of the international conference in norwich, u k , 1997

I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I


George D. Smith
Nigel C. Steele
Rudolf F. Albrecht
Artificial Neural Nets
and Genetic Algorithms
Proceedings of the International Conference
in Norwich, U.K., 1997

Springer-Verlag Wien GmbH

Dr. George D. Smith
School of Information Systems
University of East Anglia, Norwieh, U.K.

Dr. Nigel C. Steele
Division of Mathematics
School of Mathematical and Information Sciences
Coventry University, Coventry, u.K.

Dr. Rudolf F. Albrecht
Institut für Informatik
Universität Innsbruck, Innsbruck, Austria

This work is subject to copyright.
All rights are reserved, whether the whole or part of the material is concerned, specifieally those of
translation, reprinting, re-use of illustrations, broadcasting, reproduction by photocopying machines
or similar means, and storage in data banks.
© 1998 Springer-Verlag Wien
Originally published by Springer-Verlag Wien 1998

Camera-ready copies provided by authors and editors
Graphie design: Ecke Bonk
Printed on acid-free and chlorine-free bleached paper
SPIN 10635776

With 384 Figures

ISBN 978-3-211-83087-1
ISBN 978-3-7091-6492-1 (eBook)
DOI 10.1007/978-3-7091-6492-1

Preface

This is the third in a series of conferences devoted primarily to the theory and applications of artificial neural
networks and genetic algorithms. The first such event was held in Innsbruck, Austria, in April 1993, the
second in Ales, France, in April 1995. We are pleased to host the 1997 event in the mediaeval city of Norwich,
England, and to carryon the fine tradition set by its predecessors of providing a relaxed and stimulating
environment for both established and emerging researchers working in these and other, related fields.
This series of conferences is unique in recognising the relation between the two main themes of artificial
neural networks and genetic algorithms, each having its origin in a natural process fundamental to life on
earth, and each now well established as a paradigm fundamental to continuing technological development
through the solution of complex, industrial, commercial and financial problems. This is well illustrated in
this volume by the numerous applications of both paradigms to new and challenging problems.
The third key theme of the series, therefore, is the integration of both technologies, either through the use
of the genetic algorithm to construct the most effective network architecture for the problem in hand, or,
more recently, the use of neural networks as approximate fitness functions for a genetic algorithm searching
for good solutions in an 'incomplete' solution space, i.e. one for which the fitness is not easily established for
every possible solution instance.
Turning to the contributions, of particular interest is the number of contributions devoted to the development
of 'modular' neural networks, where a divide and conquer approach is adopted and each module is trained to
solve a part of the problem. Contributions also abound in the field of robotics and, in particular, evolutionary
robotics, in which the controllers are adapted through the use of some evolutionary process. This latter field
also provided a forum for contributions using other related technologies, such as fuzzy logic and reinforcement
learning.
Furthermore, we note the relatively large number of contributions in telecommunications related research,
confirming the rapid growth in this industry and the associated emergence of difficult optimisation problems.
The increasing complexity of problems in this and other areas has prompted researchers to harness the power
of other heuristic techniques, such as simulated annealing and tabu search, either in their 'pure' form or
as hybrids. The contributions in this volume reflect this trend. Finally, we are also pleased to continue to

provide a forum for contributions in the burgeoning and exciting field of evolutionary hardware.
We would like to take this opportunity to express our gratitude to everyone who contributed in any way
to the completion of this volume. In particular, we thank the members of the Programme Committee for
reviewing the submissions and making the final decisions on the acceptance of papers, Romek Szczesniak
(University of East Anglia) for his unenvious task of preparing the LaTeX source file, Silvia Shilgerius
(Springer-Verlag) for the final stages of the publication process and, not least, to all researchers for their
submissions to ICANNGA97.
We hope that you enjoy and are inspired by the papers contained in this volume.
George D. Smith
Norwich

Nigel C. Steele
Coventry

Rudolf F. Albrecht
Innsbruck


Contents

Advisory and Programme Committees

xvi

Robotics and Sensors
Obstacle Identification by an Ultrasound Sensor Using Neural Networks
D. Diep, A. Johannet, P. Bonnefoy and F. Harroy
A Modular Reinforcement Learning Architecture for Mobile Robot Control
R. M. Rylatt, C. A. Czarnecki and T. W. Routen
Timing without Time - An Experiment in Evolutionary Robotics

H. H. Lund
Incremental Acquisition of Complex Behaviour by Structured Evolution
S. Perkins and G. Hayes
Evolving Neural Controllers for Robot Manipulators
R. Salama and R. Owens
Using Genetic Algorithms with Variable-length Individuals for Planning
Two-Manipulators Motion
J. Riquelme, M. Ridao, E. F. Camacho and M. Toro

1
6

11

16

21
26

ANN Architectures
Ensembles of Neural Networks for Digital Problems
D. Philpot and T. Hendtlass
A Modular Neural Network Architecture with Additional Generalization Abilities for
Large Input Vectors
A. Schmidt and Z. Bandar
Principal Components Identify MLP Hidden Layer Size for Optimal Generalisation
Performance
M. Girolami
Bernoulli Mixture Model of Experts for Supervised Pattern Classification
N. Elhor, R. Bertrand and D. Hamad

31
35
40

44

Power Systems
Electric Load Forecasting with Genetic Neural Networks
F. J. Marin and F. Sandoval
Multiobjective Pressurised Water Reactor Reload Core Design Using a Genetic Algorithm
G. T. Parks
Using Artificial Neural Networks to Model Non-Linearity in a Complex System
P. Weller, A. Thompson and R. Summers

49
53
58


viii

'Iransit Time Estimation by Artificial Neural Networks

T. Tambouratzis, M. Antonopoulos-Domis, M. Marseguerra and E. Padovani

62

Evolware
Evolving Asynchronous and Scalable Non-uniform Cellular Automata

M. Sipper, M. Tomassini and M. S. Capcarrere
One-Chip Evolvable Hardware: 1C-EHW

H. de Garis

66
71

Vision
Evolving Low-Level Vision Capabilities with the GENCODER Genetic Programming
Environment

78

P. Ziemeck and H. Ritter
NLRFLA: A Supervised Learning Algorithm for the Development of Non-Linear
Receptive Fields
S. L. Funk, 1. Kumazawa and J. M. Kennedy
Fuzzy-tuned Stochastic Scanpaths for AGV Vision
1. J. Griffiths, Q. H. Mehdi and N. E. Gough
On VLSI Implementation of Multiple Output Sequential Learning Networks

A. Bermak and H. Poulard

83
88

93

Speech/Hearing
Automated Parameter Selection for a Computer Simulation of Auditory Nerve Fibre
Activity using Genetic Algorithms
C. P. Wong and M. J. Pont
Automatic Extraction of Phase and Frequency Information from Raw Voice Data

S. McGlinchey and C. Fyfe

98
103

A Speech Recognition System using an Auditory Model and TOM Neural Network

107

Fahlman-Type Activation Functions Applied to Nonlinear PCA Networks Provide
a Generalised Independent Component Analysis

112

E. Hartwich and F. Alexandre
M. Girolami and C. Fyfe

Blind Source Separation via Unsupervised Learning

B. Freisleben, C. Hagen and M. Borschbach

116

Signal/Image Processing and Recognition

Neural Networks for Higher-Order Spectral Estimation

F.-L. Luo and R. Unbehauen
Estimation of Fractal Signals by Wavelets and GAs

H. Cai and Y. Li
Classification of 3-D Dendritic Spines using Self-Organizing Maps

G. Sommerkorn, U. Seiffert, D. Surmeli, A. Herzog, B. Michaelis and K. Braun
Neural Network Analysis of Hue Spectra from Natural Images

C. Robertson and G. M. Megson

121
126
129
133


ix

Detecting Small Features in SAR Images by an ANN
1. Finch, D. F. Yates and L. M. Delves
Optimising Handwritten-Character Recognition with Logic Neural Networks
G. Tambouratzis

138
143

Medical Applications

Combined Neural Network Models for Epidemiological Data: Modelling Heterogeneity
and Reduction of Input Correlations
M. H. Lamers, J. N. Kok and E. Lebret
A Hybrid Expert System Architecture for Medical Diagnosis
L. M. Brasil, F. M. de Azevedo and J. M. Barreto
Enhancing Connectionist Expert Systems by lAC Models through Real Cases
N. A. Sigaki, F. M. de Azevedo and J. M. Barreto

147
152
157

G A Theory and Operators
A Schema Theorem-Type Result for Multidimensional Crossover
M.-E. Balazs
Mobius Crossover and Excursion Set Mediated Genetic Algorithms
S. Baskaran and D. Noever
The Single Chromosome's Guide to Dating
M. Ratford, A. Tuson and H. Thompson
A Fuzzy Taguchi Controller to Improve Genetic Algorithm Parameter Selection
C.-F. Tsai, C. G. D. Bowerman, J. l. Tait and C. Bradford
Walsh Functions and Predicting Problem Complexity
R. B. Heckendorn
Migration through Mutation Space: A Means of Accelerating Convergence
in Evolutionary Algorithms
H. Copland and T. Hendtlass

161
166

171
175
179
183

GA Models/Representation

Dual Genetic Algorithms and Pareto Optimization
M. Clergue and P. Collard
Multi-layered Niche Formation
C. Fyfe
Using Hierarchical Genetic Populations to Improve Solution Quality
J. R. Podlena and T. Hendtlass
A Redundant Representation for Use by Genetic Algorithms on Parameter
Optimisation Problems
A. J. Soper and P. F. Robbins

188
193
198
202

GA Applications

A Genetic Algorithm for Learning Weights in a Similarity Function
Y. Wang and N. Ishii

206

x

Learning SCFGs from Corpora by a Genetic Algorithm
B. Keller and R. Lutz
Adaptive Product Optimization and Simultaneous Customer Segmentation:
A Hospitality Product Design Study with Genetic Algorithms
E. Schifferl
Genetic Algorithm Utilising Neural Network Fitness Evaluation for Musical Composition
A. R. Burton and T. Vladimirova

210
215
219

Parallel GAs
Analyses of Simple Genetic Algorithms and Island Model Parallel Genetic Algorithms
T. Niwa and M. Tanaka
Supervised Parallel Genetic Algorithms in Aerodynamic Optimisation
D. J. Doorly and J. Peiro

224
229

Combinatorial Optimisation
A Genetic Clustering Method for the Multi-Depot Vehicle Routing Problem
S. Salhi, S. R. Thangiah and F. Rahman
A Hybrid Genetic / Branch and Bound Algorithm for Integer Programming
A. P. French, A. C. Robinson and J. M. Wilson
Breeding Perturbed City Coordinates and Fooling Travelling Salesman Heuristic

Algorithms
R. Bradwell, L. P. Williams and C. L. Valenzuela
Improvements on the Ant-System: Introducing the MAX-MIN Ant System
T. Stiitzle and H. Hoos
A Hybrid Genetic Algorithm for the 0-1 Multiple Knapsack Problem
C. Cotta and J. M. Troya
Genetic Algorithms in the Elevator Allocation Problem
J. T. Alander, J. Herajiirvi, G. Moghadampour, T. Tyni and J. Ylinen

234
238
241

245
250
255

Scheduling/Timetabling
Generational and Steady-State Genetic Algorithms for Generator Maintenance
Scheduling Problems
K. P. Dahal and J. R. McDonald
Four Methods for Maintenance Scheduling
E. K. Burke, J. A. Clarke and A. J. Smith
A Genetic Algorithm for the Generic Crew Scheduling Problem
N. Ono and T. Tsugawa
Genetic Algorithms and the Timetabling Problem
B. C. H. Turton
Evolutionary Approaches to the Partition/Timetabling Problem
D. Corne

259

264
270

275
281


xi

Telecommunications -

General

Discovering Simple Fault-Tolerant Routing Rules by Genetic Programming
I. M. A. Kirkwood, S. H. Shami and M. C. Sinclair
The Ring-Loading and Ring-Sizing Problem
J. W. Mann and G. D. Smith
Evolutionary Computation Techniques for Telephone Networks Traffic Supervision
Based on a Qualitative Stream Propagation Model
I. Servet, L. Trave-Massuyes and D. Stern
NOMaD: Applying a Genetic Algorithm/Heuristic Hybrid Approach to Optical
Network Topology Design
M. C. Sinclair
Application of a Genetic Algorithm to the Availability-Cost Optimization of a
Transmission Network Topology
B. Mikac and R. Inkret
Telecommunications -

289
294
299
304

FAP

Breeding Permutations for Minimum Span Frequency Assignment
C. L. Valenzuela, A. Jones and S. Hurley
A Practical Frequency Planning Technique for Cellular Radio
T. Clark and G. D. Smith
Chaotic Neurodynamics in the Frequency Assignment Problem
K. Dorkofikis and N. M. Stephens
A Divide-and-Conquer Technique to Solve the Frequency Assignment Problem
A. T. Potter and N. M. Stephens
Applications -

285

308
312
317
321

General Heuristics

Genetic Algorithm Based Software Testing
J. T. Alander, T. Mantere and P. Turunen
An Evolutionary /Meta-Heuristic Approach to Emergency Resource Redistribution
in the Developing World

A. Tuson, R. Wheeler and P. Ross
Automated Design of Combinational Logic Circuits by Genetic Algorithms
C. A. Coello Coello, A. D. Christiansen and A. Hernandez Aguirre
Forecasting of the Nile River Inflows by Genetic Algorithms
M. E. EI-Telbany, A. H. Abdel- Wahab and S. I. Shaheen
Evolutionary ANN s I -

325
329
333
337

RBFs

A Comparative Study of Neural Network Optimization Techniques
T. Ragg, H. Braun and H. Landsberg
GA-RBF: A Self-Optimising RBF Network
B. Burdsall and C. Giraud-Carrier
Canonical Genetic Learning of RBF Networks Is Faster
R. Neruda

341
346
350


xii
Evolutionary ANNs II

The Baldwin Effect on the Evolution of Associative Memory

A. Imada and K. Araki
Using Embryology as an Alternative to Genetic Algorithms for Designing Artificial
Neural Network Topologies
C. MacLeod and G. Maxwell

354
359

Evolutionary ANNs III

Empirical Study of the Influences of Genetic Parameters in the Training of a
Neural Network
P. Gomes, F. Pereira and A. Silva
Evolutionary Optimization of the Structure of Neural Networks by a Recursive Mapping
as Encoding
B. SendhoJJ and M. Kreutz
Using Genetic Engineering To Find Modular Structures for Architectures of
Artificial Neural Networks
C. M. Friedrich
Evolutionary Learning of Recurrent Networks by Successive Orthogonal Inverse
Approximations
C. Gegout

364
368
373
378

Reinforcement Learning

Evolutionary Optimization of Neural Networks for Reinforcement Learning Algorithms
H. Braun and T. Ragg
Generalising Experience in Reinforcement Learning: Performance in Partially
Observable Processes
C. H. C. Ribeiro

384
389

Genetic Programming

Optimal Control of an Inverted Pendulum by Genetic Programming: Practical Aspects
F. Gordillo and A. Bernal
Evolutionary Artificial Neural Networks and Genetic Programming: A Comparative
Study Based on Financial Data
S.-H. Chen and C.-C. Ni
A Canonical Genetic Algorithm Based Approach to Genetic Programming
F. Oppacher and M. Wineberg
Is Genetic Programming Dependent on High-level Primitives?
D. Heiss-Czedik
DGP: How To Improve Genetic Programming with Duals
J.-L. Segapeli, C. Escazut and P. Collard
Fitness Landscapes and Inductive Genetic Programming
V. Slavov and N. I. Nikolaev

393
397
401
405
409

414


xiii

Discovery of Symbolic, Neuro--Symbolic and Neural Networks with Parallel
Distributed Genetic Programming
R. Poli

419

ANN Applications

A Neural Network Technique for Detecting and Modelling Residential Property
Sub-Markets
O. M. Lewis, J. A. Ware and D. Jenkins
Versatile Graph Planarisation via an Artificial Neural Network
T. Tamboumtzis
Artificial Neural Networks for Generic Predictive Maintenance
C. Kirkham and T. Harris
The Effect of Recurrent Networks on Policy Improvement in Polling Systems
H. Sato, Y. Matsumoto and N. Okino
EXPRESS - A Strategic Software System for Equity Valuation
M. P. Foscolos and S. Nilchan
Virtual Table Tennis and the Design of Neural Network Players
D. d'Aulignac, A. Moschovinos and S. Lucas
Investigating Arbitration Strategies in an Animat Navigation System
N. R. Ball

424

428
432
436
440
445
449

Sequences/Time Series

Sequence Clustering by Time Delay Networks
N. Allott, P. Halstead and P. Fazackerley
Modeling Complex Symbolic Sequences with Neural Based Systems
P. Tino and V. Vojtek
An Unsupervised Neural Method for Time Series Analysis, Characterisation and Prediction
C. Fyfe
Time-Series Prediction with Neural Networks: Combinatorial versus Sequential Approach
A. Dobnikar, M. Trebar and B. Petelin
A New Method for Defining Parameters to SETAR(2;k1 ,k2)-models
J. Kyngiis
Predicting Conditional Probability Densities with the Gaussian Mixture-RVFL Network
D. Husmeier and J. G. Taylor

454
459
464
468
473
477

ANN Theory, Thaining and Models

An Artificial Neuron with Quantum Mechanical Properties
D. Ventura and T. Martinez
Computation of Weighted Sum by Physical.Wave Properties-Coding Problems by
Unit Positions
1. K umazawa and Y. K ure
Some Analytical Results for a Recurrent Neural Network Producing Oscillations
T. P. Fredman and H. Saxen

482
486
491


xiv
Upper Bounds on the Approximation Rates of Real-valued Boolean Functions by
Neural Networks
K. Hlavackova, V. Kurkova and P. Savicky
A Method for Task Allocation in Modular Neural Network with an Information Criterion
H.-H. Kim and Y. Anzai
A Meta Neural Network Polling System for the RPROP Learning Rule
C. McCormack
Designing Development Rules for Artificial Evolution
A. G. Rust, R. Adams, S. George and H. Bolouri
Improved Center Point Selection for Probabilistic Neural Networks
D. R. Wilson and T. R. Martinez
The Evolution of a Feedforward Neural Network trained under Backpropagation
D. McLean, Z. Bandar and J. D. O'Shea

495

500
505
509
514
518

Classification
Fuzzy Vector Bundles for Classification via Neural Networks
D. W. Pearson, G. Dray and N. Peton
A Constructive Algorithm for Real Valued Multi-category Classification Problems
H. Poulard and N. Hernandez
Classification of Thermal Profiles in Blast Furnace Walls by Neural Networks
H. Saxen, L. Lassus and A. Bulsari
Geometrical Selection of Important Inputs with Feedforward Neural Networks
F. Rossi
Classifier Systems Based on Possibility Distributions: A Comparative Study
S. Singh, E. L. Hines and J. W. Gardner

523
527
532
535
539

Intelligent Data Analysis/Evolution Strategies
Learning by Co-operation: Combining Multiple Computationally Intelligent Programs
into a Computational Network
H. L. Viktor and 1. Cloete
Comparing a Variety of Evolutionary Algorithm Techniques on a Collection of
Rule Induction Tasks

D. Corne
An Investigation into the Performance and Representations of a Stochastic, Evolutionary
Neural Tree
K. Butchart, N. Davey and R. G. Adams
Experimental Results of a Michigan-like Evolution Strategy for Non-stationary Clustering
A. 1. Gonzalez, M. Grana, J. A. Lozano and P. Larranaga
Excursion Set Mediated Evolutionary Strategy
S. Baskaran and D. Noever
Use of Mutual Information to Extract Rules from Artificial Neural Networks
T. Nedjari
Connectionism and Symbolism in Symbiosis
N. Allott, P. Fazackerley and P. Halstead

543
547
551
555
560
565
570


xv

Coevolution and Control
Genetic Design of Robust PID Controllers
A. H. Jones and P. B. de Moura Oliveira
Coevolutionary Process Control
J. Paredis
Cooperative Coevolution in Inventory Control Optimisation

R. Eriksson and B. Olsson

575
579
583

Process Control/Modelling
Dynamic Neural Nets in the State Space Utilized in Non-Linear Process Identification
R. C. L. de Oliveira, F. M. de Azevedo and J. M. Barreto
Distal Learning for Inverse Modeling of Dynamical Systems
A. Toudeft and P. Gallinari
Genetic Algorithms in Structure Identification for NARX Models
C. K. S. Ho, I. G. French, C. S. Cox and 1. Fletcher
A Model-based Neural Network Controller for a Process Trainer Laboratory Equipment
B. Ribeiro and A. Cardoso
MIMO Fuzzy Logic Control of a Liquid Level Process
I. Wilson, 1. G. French, 1. Fletcher and C. S. Cox

588

592
597
601
606

LCS/Prisoner's Dilemma
A Practical Application of a Learning Classifier System in a Steel Hot Strip Mill
W. Browne, K. Holford, C. Moore and J. Bullock
Multi-Agent Classifier Systems and the Iterated Prisoner's Dilemma
K. Chalk and G. D. Smith

Complexity Cost and Two Types of Noise in the Repeated Prisoner's Dilemma
R. Hoffman and N. C. Waring

611

615
619

Workshop Summary

624

Plenary Lectures

626

Subject Index

628


ICANNGA 97
International Conference on Artificial Neural Networks and Genetic Algorithms
Norwich, UK, April 2 - 4, 1997

International Advisory Committee
Professor R. Albrecht, University of Innsbruck, Austria
Dr. D. Pearson, Ecole des Mines d'Ales, France
Professor N. Steele, Coventry University, England (Chairman)
Dr. G. D. Smith, University of East Anglia, England

Programme Committee
Thomas Baeck, Informatik Centrum, Dortmund, Germany
Wilfried Brauer, TU Munchen, Germany
Gavin Cawley, University of East Anglia, Norwich, UK
Marco Dorigo, Universite Libre de Bruxelles, Belgium
Simon Field, Nortel, Harlow, UK
Terry Fogarty, Napier University, Edinburgh, UK
Jelena Godjevac, EPFL Laboratories, Switzerland
Dorothea Heiss, TU Wien, Austria
Michael Heiss, Neural Net Group, Siemens AG, Austria
Tom Harris, BruneI University, London, UK
Anne Johannet, EMA-EERlE, Nimes, France
Helen Karatza, Aristotle University of Thessaloniki, Greece
Sami Khuri, San Jose State University, USA
Pedro Larranaga, University Basque Country, Spain
Francesco Masulli, University of Genoa, Italy
Josef Mazanec, WU Wien, Austria
Janine Magnier, EMA-EERIE, Nimes, France
Christian Omlin, NEC Research Institute, Princeton, USA
Franz Oppacher, Carleton University, Ottawa, Canada
Ian Parmee, University of Plymouth, UK
David Pearson, EMA-EERIE, Nimes, France
Vic Rayward-Smith, University of East Anglia, Norwich,UK
Colin Reeves, Coventry University, Coventry, UK
Bernardete Ribeiro, Universidade de Coimbra, Portugal
Valentina Salapura, TU Wien, Austria
V. David Sanchez A., University of Miami, Florida, USA
Henrik Saxen, Abo Akademi, Finland
George D. Smith, University of East Anglia, Norwich, UK (Chairman)

Nigel Steele, Coventry University, Coventry, UK
Kevin Warwick, Reading University, Reading, UK
Darrell Whitley, Colorado State University, USA


Obstacle Identification by an Ultrasound Sensor Using Neural Networks

1

D. Diepl, A. Johannet 1 , P. Bonnefoy2 and F. Harroy2
LGI2P - EMA/EERlE, Parc Scientifique G. Besse, 30000 Nimes, FRANCE.
2 IMRA Europe, 220 rue Albert Caquot, 06904 Sophia Antipolis, FRANCE
Email:

Abstract

2

This paper presents a method for obstacle recognition to
be used by a mobile robot. Data are made of range measurements issued from a phased array ultrasonic sensor,
characterized by a narrow beam width and an electronically controlled scan. Different methods are proposed:
a simulation study using a neural network, and a signal analysis using an image representation. Finally, a
solution combining both approaches has been validated.

Ultrasound sensors are usually used as proximity
sensors, but they lack bearing directivity which generally prevents us from obtaining any accurate information. In order to reduce this drawback we
have proposed an original sensor including several
individual ultrasound emitter-receivers [3,4]. The
ultrasonic sensor concerned consists of an array
of 7 transmitters simultaneously emitting acoustic

waves at the frequency of 40 kHz (Figure 1).
The phase of each emitter can be adjusted individually, so that the beam width of the resultant
wave will have a restricted size, and its bearing direction may be fixed (Figure 2).
Echoes coming from reflectors are detected by two
receivers, and the reflectors' range and orientation
can be determined by measuring the time of flight,
i.e. the "time duration between the transmission and
the reception of a signal. The sensor is thus analogous to a sonar system, upon whose main principles
the ultrasound system was developed.

1

Introduction

The development of an autonomous mobile robot
is still a difficult task. Generally three types of
problems are studied: the first deals with locomotion (stability, efficiency) the second deals with reflex actions (obstacle avoidance) and the third with
navigation in order to reach a goal. The major difficulties encountered in such a task is the extreme
variability of the environment with which the robot
interacts, and the noise inherent in the real world.
Obviously nobody tries to develop a robot able to
evolve in all types of environment but. the variability intrinsic to even a specific type of environment
is sufficient to lead to a relative failure of the traditional methods of modelling [1]. In this context,
the neural networks approach appears to be an alternative solution in which the robot learns to adapt
to the environment rather than learns all the reactions to each possible event. Within the wide field
of research dealing with the development of mobile robots, starting from works centred on obstacle
avoidance [9], this study focuses on the neural identification of obstacles using an original ultrasound
sensor.

G. D. Smith et al., Artificial Neural Nets and Genetic Algorithms

© Springer-Verlag Wien 1998

3

The Ultrasonic Sensor

Simulation Study

The first part of the work consists of modelling the
sensor and the echoes in order to find out by simula-

,,1 I:t:l{*...

gem

...

transmitter receiver

Figure 1: Configuration of the transducers.


2

.. .

...

9 input

6 utputs

dl-+O
d2 ....

wall

0

door

pillar
len part
of wall
right p rt

10.

of wall

none

I t layer :
5 neurons

output layer :
3 neuron

Figure 4: Architecture of the network.
Figure 2: Directivity diagram for a transmission at

_10° and 0°: (a) theoretical, (b) experimental.

Figure 3: Simulated situations for a mobile robot.

tion the best way to identify simple obstacles such as
walls, doors and pillars. Assuming that the distance
between the obstacle and the sensor can be computed from the time of Hight, a multilayer network
was used in order to classify the obstacles used. The
inputs which seem to be relevant are the distance
between the obstacle and the sensor for 9 emission
directions in front of the robot, stepping from -320
to +32°.
Data collected were issued from a software program simulating the dynamical behavior of a mobile robot equipped with the ultrasonic sensor [7].
Figure 3 shows different situations encountered by
the robot when moving along in a room.

The learning was performed with a hundred examples by standard backpropagation in order to
classify 6 types of obstacles including the particular
scene where there is no obstacle. Inputs called dl to
d9 on Figure 4 were the distances measured along
each direction of transmission.
The results obtained were quite good with 92%
well classified and 3% of error evaluated on a test
set [2]. Nevertheless, this, simulation allowed us to
demonstrate one principal limitation: the problem
of the apparent size of the obstacle, which increases
when the obstacle is nearer to the sensor. This problem cannot be solved by the neural net and has to
be treated beforehand. Secondly when we tried to
compare the results obtained with the true signals,
it appeared that it was not possible to compute the

distance between the obstacle and the sensor in the
case of a large angle of bearing, without additional
information on the amplitude of signals. In conclusion, in spite of the good results, the modelling
approach of this first treatment was not sufficiently
realistic to be applied to a real concrete case.

4

Signal Analysis

The second study we carried out took into account
the problems listed above and had two goals: firstly
to estimate the distance to the obstacle and secondly to find a way to characterise a type of obstacle independently of its distance. For this, a method
based on a signal modeling method used in [6] was


3

-

--- - - ~

Figure 6: Simulated image of a corner, original image,
simulated image of a wall/edge.
Figure 5: Images from walls, corners and edges.

employed: first the distance is estimated including
all the angular reflections, afterwards the signal is
compared to a simulated reference signal computed
from the previously estimated range [8]. In practice, the array of transmitters was programmed to

make an acquisition at each degree between -300
and +30 0 for 512 samples (the acquisition for each
direction was done at 50 kHz, so 512 samples gave a
visibility window of 1.8 m). All the values collected
were gathered together to form an image of 61x512
pixels. (Figure 5).
According to the nature, the orientation, and the
distance of the obstacle, the images are very different, be it for the number of echoes or for their position. Furthermore, each type of obstacle studied
does not always give the same response, depending
on its orientation and its distance. Then, these 'images' were analysed in order to extract some kind of
constant pattern for each obstacle. Then, for a few
simple obstacles (wall, corner, edge as classified in
[6]) the reflection pattern could be easily explained
depending on the height of the sensor and the distance between the sensor and the obstacle. Based
on this analysis, a simulation generates an artificial
reflection image for each type of obstacle, which is
then compared to the real image (Figure 6).
Operating on the real image, the mean amplitude
of each of the 512 vectors is computed (mean amplitude versus distance). Hence, the darkest echo
on an image corresponds to the minimum of this
mean amplitude, which gives the distance between
the obstacle and the sensor. A similar operation is
performed for the angle to obtain the direction of

the obstacle. Once the distance and the angle have
been found, the recognition is performed :.'.1 making
a comparison between the real image and the simulated image for the three types of obstacles considered. A series of 26 measurements was performed
in a room, . the sensor being located at various distances and orientation angles from the obstacles. In
all cases, the distance to the obstacle was accurately
estimated by the sensor with a margin of error less

than 1 cm. Among the different kinds of obstacles,
21 shapes (Le. 81% of the total number) were correctly recognised. The estimation of the angle was
correct for 18 obstacles (69%). In some cases, the
values found by this method were incorrect, so two
ways were used to empirically improve the performances: the first was based on the comparison of
the values found for the two channels (one for a left
sensor, the other for the right sensor), and the second calculates the disparity in the distance for the
two channels to find the angle.

5

Recognition with Neural Network

The logical follow-up to the previous study was to
integrate neural networks in order to: flrst implement the computation of various thresholds intervening during the recognition process, and second
to enable adaptations to various wall coverings. The
problem was the following: starting from the previously described images (61x512), we want to classify the scene viewed by a robot in three categories:
wall, edge or corner. Using the estimation of the
distance D between the sensor and the obstacle described above, and assuming in the case of a corner that the sensor is located roughly at the same
distance from both walls, several features were ex-


4

tracted from the image in order to represent the
information independently of the distance:
• energy (Le. the integral value) of the first peek
(Le. the first echo received) located at the distance D, which is in any case issued from a
wall,
• energy at the distance ..tiD (location of a possible comer).

• energy at the distance J D2 + H2, where H is
the height of the sensor above the floor level
(echo reflecting from the ground at the foot of
a wall).
• energy at the distance ..ti..;D2 + H2 (echo reflecting from the ground at the foot of a comer).
These characteristics, called E 1 , ~, Ea, E4, plus
the estimated distance D for each ultrasonic receiver
(right and left) led to a total amount of 10 inputs
for the network (Figure 8).
A first study showed that, with the chosen coding, the classes (walls, comers, edges) were not linearly separable, so a multilayer neural network was
necessary. Nevertheless, because of the well known
problems of convergence inherent in the use of the
backpropagation learning rule, we begin with a simpler network where the learning operates only on the
first layer, whereas the second layer computes logical combinations. This type of network had been
used for the recognition of zip code [5] and gave in
this case very surprising and satisfactory results.
The principle of the method is the following: we
consider that the classes to separate are non linearly
separable one from all the others, but their representation is good, and the classes are linearly separable
one class from another one. Then it is possible to
compute the separation with several straight lines
rather than one more complicated curve. This type
of configuration can be illustrated in a smaller dimension with only two inputs in Figure 7.
The learning is performed on the first layer of
the network: each neuron defines a straight line
which separates one class from another using a simple learning rule (such as perceptron learning rule).
For example the line 8 1 in Figure 7 separates the
class of 'Comers' from the class of 'Edges'. The final interpretation is computed by a logical function:

Figure 7: Example of classification with combination of

straight lines. The classes are separated one from another because the separation of one class from all classes
is not possible using straight lines.
30utpul

~~ _ _ wml
:~

comer

O..-. ..... edge
~o

I t I yer:
3 neur n

Ulput layer:
3 neur n

Figure 8: Architecture of the network.

for example in the Figure 7, the class of 'comers' is
identified in the upper part of the line 8 1 , AND in
the right part of the line 8 2 • This logical combination operating on the responses of the neurons of
the first layer can be implemented using a neural
formalism and leads to a multilayer neural network
(Figure 8).
Real tests were performed on the same measurements as previously and the neural network behaves
very satisfactorily, because 100% of the learning examples, which were the same 26 measurements as in
section 5, were well classified. During the test phase
the network worked well on straightforward obsta-

5

cles. Nevertheless the main problem encountered
was, for several measurements, the interpretation of
what the obstacle was: for instance the extremity
of a wall was perhaps considered as an edge, and,
depending on the angle, a part of a corner might
be considered as a wall. During the generalisation
phase such ambivalence has to be tolerated.

6

Conclusion

In conclusion, for the identification of obstacles by
ultrasound sensors no direct method can work well
because of the complexity of the problem and the
presence of noise. Therefore we proposed a method
which takes into account the behaviour of reflected
ultrasound waves in order to extract some features
from the signals, and then to take a decision using
a neural network. This method had proved efficient
for a small set of data. Further work will have to
be done in order to generalize this result to more
complex environments.

7

Acknowledgements

The authors would like to thank M. Denis Roux
and M. Gerard Cauvy, students from the University
of Montpellier for their enthusiasm and their work
on this difficult problem, including hardware and
software difficulties.

References
[1] R.A. Brooks. Intelligence without representation.
Artificial Intelligence, 47:139, 1991.
[2] G. Cauvy. Etude par reseau de neurones d'un sonar
pour robot mobile. Technical report, DEA-USTL,
Montpellier, 1995.
[3] D. Diep and K. EI Kherdali. Un radar ultra-sons
pour la localisation d'un robot mobile. In Joumees
SEE Capteurs en Robotique, 1993.
[4] K. EI Kherdali. Etude, conception et realisation d'un
mdar ultm-sonore. PhD thesis, USTL, Montpellier,
1992.
[5] S. Knerr, L. Personnaz, and G. Dreyfus. Handwritten digit recognition by neural networks with singlelayer training. IEEE 7rans. Neuml Networks, 1992.
[6] R. Kuc and M.W. Siegel. Physically based simulation model for acoustic sensor robot navigation.
IEEE 7rans., PAMI 9(6), November 1987.

[7] C. Moschetti. Neural network - a connectionist way
for artificial intelligence & application to acoustic
recognition of shapes. Technical report, IMRA-ESSI
DESS, Sophia-Antipolis, 1994.
[8] D. Roux, D. Diep, P. Bonnefoy, and F. Harroy.
Reconnaissance d'obstacles avec un capteur ultrasonore. In ·feme Congres Fhm~ais d'Acoustique,

Marseille, 1997.
[9] 1. Sarda and A. Johannet. Behaviour learning by
ARP: From Gait learning to obstacle avoidance by
neural networks. In D. W. Pearson, N. C. Steele,
R. F. Albrecht (editors), Artificial Neural Networks
and Genetic Algorithms, pages 464-467. SpringerVerlag, Wien New York, 1995.


A Modular Reinforcement Learning Architecture for Mobile Robot Control
R. M. Rylatt, C. A. Czarnecki and T. W. Routen
Department of Computer Science, De Montfort University,
Leicester, LEI 9BH, UK
Email: {rylatt.cc.twr}@dmu.ac.uk

Abstract
The paper presents a way of extending complementary reinforcement backpropagation learning (CRBP) to
modular architectures using a new version of the gating network approach in the context of reactive navigation tasks for a simulated mobile robot. The gating
network has partially recurrent connections to enable
the co-ordination of reinforcement learning across both
modules' 'successive time steps. The experiments reported explore the possibility that architectures based
on this approach can support concurrent acquisition of
different reactive navigation related competences while
the robot pursues light-seeking goals.

1

Introduction

Schemes for the control of mobile robots based on
a stimulus-response view of behaviour offer an alternative to traditional AI approaches that relied

on much more computationally demanding representational structures. The aim is to achieve effective autonomous real-time performance in unstructured and uncertain domains. As a representative
example, Brooks' subsumption architecture [2] relies on the idea of multiple behavioural layers concurrently active and competing for control of the
robot or agent, mediated by some kind of arbitration scheme that is often based on simple prioritisation. However, the problems of co-ordinating behaviours, or action selection, is a central concern for
this branch of adaptive autonomous agent research.
It can be argued that schemes like subsumption offer ad hoc engineering solutions conceived too prescriptively in observer space. For example, Rylatt
et al. [9], and MoHand et al. [7] have discussed
respectively the role of learning and of short-term

G. D. Smith et al., Artificial Neural Nets and Genetic Algorithms
© Springer-Verlag Wien 1998

memory in achieving run-time adaptivity. Rylatt
et al. [8] also survey approaches based on neural
networks to explore the argument that this kind of
substrate is an inherently more promising basis for
achieving the necessary flexibility of behaviour. Another issue is whether this alternative substrate also
implies architectural modularity. Ziemke [13] argues that a monolithic neural network can acquire
modular features (learn its own control structure)
during the process of adapting to an environment
at run-time. However, a contra-indication is provided by our knowledge of brain structure, where
there is good evidence for predetermined functional
modularity. Obviously this kind of modularity is
the result of phylogenetic adaptation, or evolution,
rather than the kind of ontogenetic changes that
could be compared to the run- time adaptation of
an artificial autonomous agent. Taking broad inspiration from the biological existence proof, our initial
approach was to define modules in relation to distinct sensory modalities of the agent. More details
of the architecture are given in Section 2. Section 3
discusses some experimental results. Section 4 concludes with a summary of the achievements to date,
some reflections on their implications and an outline

of further work.

2

Reinforcement Learning in
Modular Architectures

Different forms of reinforcement learning in neural networks have been described. The general approach in this paper is of the kind discussed by
Williams [12], known as associative reinforcement
learning: a neural network architecture reacts to
the environment by emitting a time varying vector
of effector outputs in response to a time-varying vec-


7

I1nCTOIlS

context
unit.

SINSOIlS
buap

Figure 1: Modular neural network architecture.

tor of sensor inputs and learns to maximise a timevarying scalar reinforcement signal that is some
task-dependent function of the input and output
patterns unknown to the controller. Meeden et
al. [6] applied complementary reinforcement ba.ckpropagation (CRBP), a form of associative reinforcement learning originally described by Ackley

and Littman [1]) to a simple monolithic neural network controller for a car-like mobile robot; we have
adapted it for use in modular neural network architectures, which presented a particular set of problems. In broad outline, the architectures are inspired by the Addam architecture [11] but as we
use trial and error rather than supervised learning
the principle of control is different. Our early work
used an explicitly algorithmic (if we regard neural
networks capable of simulation on Turing machines
as implicitly algorithmic) approach to the temporally extended credit and blame assignment problems [10]. In the work reported in this paper we have
been able to replace the arbitration algorithm with
a gating network [5], originally devised for static, or

time-implicit, problems, to which we have added a
partially recurrent connections [4] as a way of solving problems of credit assignment arising from both
the temporally extended nature of the domain and
the architectural structure, An example of the architecture is shown in Figure 1 - in this version,
although the modularity reflects the number of sensory modalities, each module has access to the whole
input space; another version assigns a different sensor grc;mp to each module. In each net, competence
in one of three modality-related tasks is expected to
develop through trial and error:
• light-seeking using light sensor data;
• wall avoidance using active-sonar range data;
• avoiding low obstacles ('invisible' to the sonars)
using bump detector data;

In each inchoate expert net a vector of sensor inputs i is propagated forward through the hidden
layer to reach the vector of sigmoid output units, 0,
each of which takes on a value in the range (0,1).
Each of the outputs for each net ~ multiplied by the
corresponding output from the gating network, nor-

8

malised as r:exP(Xt)
). Each of the resultant probai exp Xi
bilitstically weighted outputs is then summed with
the corresponding output from each of the other
expert nets to produce the continuous-valued output vector of the architecture in the range (0,1) termed the 'search vector', s. Independent Bernoulli
trials are then applied to the values in s so that
each is interpreted as a binary bit in a stochastic
output vector O. These two vectors are used to determine the error measure in the manner shortly described. In this way, initially random moves are suggested and, according to the reinforcement scheme,
either punished or rewarded. If a reward signal is received then, by analogy with the supervised learning
backpropagation algorithm, the error derivative can
be readily obtained, so we backpropagate (0 - s).
When a punishment signal is received however the
direction to force s is not so obvious. CRBP chooses
a somewhat stronger assumption than 'being like
not-o,' taking ((1 - 0) - s) as the desired direction,
but in our case this assumption can be considered
stronger still as we can use a little domain knowledge to ensure the encoding of our steering vectors
makes the binary complements equate to opposite
directions - reversing the direction of motion when
punished may often be a reasonable one to adopt.
Although this scheme may appear flawed (in the
sense that the agent is 'learning to run before it can
walk'), initially, the principle of a rich interaction
between control levels and sensory modalities needs
to be investigated in a search for flexible behaviour
patterns that are not excessively constrained in the
design time decision space. We also suggest that
there is biological evidence for this kind of learning in that imperfectly mastered neuro-motor skills

are gradually improved whilst the organism seeks
higher level goals - an animal does not wait until
it can walk perfectly before it moves to feed or flees
from danger.
The aim of our reinforcement learning scheme
can be rephrased as the intention that each module
should become an expert at mapping a particular
subset of the input domain onto the output range.
In static, or time-implicit domains, gating networks
of the kind described by Jacobs et al. [5] have
proved capable of selecting effective mixtures of 'experts'. Reinterpretation of the gating network error

measures in terms of CRBP is relatively straightforward. For example, competition between experts
should be induced by using the formulae (omitting
unnecessary superscripts):

(1)
and
E =

L Pi II (1 -

0) -

Si

112

(2)

where P represents the proportional contribution of
the ith expert to the proposed action on a given
time step. Equation (1) and (2) give the error measures when the agent is rewarded or punished, respectively. However, it is not obvious how such a
solution would handle the temporal aspect of the
modular credit assignment problem. The solution
adopted here is to provide the gating network with
Elman style [4] recurrent connections. It is well
known that such nets can solve context dependent
problems - in the temporal domain this can be interpreted as the ability to decide what happens next
on the basis of what has gone before.

3

Experiments

Referring to Figure 2, the mobile agent extinguishes
a light by coming into contact with it and this
remotely switches on another light some distance
away. The first light is positioned so that the agent
has to navigate around an obstacle to reach the
light source, thus overcoming the tendency· of the
first level module to be repelled. The next three
lights are located in situations that are relatively
straightforward or entail skirting obstacles and navigating through gaps between obstacles and walls.
The most difficult light seeking task entails navigation down a narrow corridor. The position of
the final light source goal requires the agent to return from the far end of the corridor back into open
space. Thus each level of competence is likely to
be exercised as the agent proceeds. To test the validity of using recurrent connections in the gating
network, a control experiment was run in which no
such connections were employed. Our observation

9

Figure 2: Experimental environment.

is that the presence of recurrent connections in the
gating network appears to be decisive in determining the gating network's ability to select inchoate
experts so as to assign credit and blame correctly
across time steps - without recurrent connections
the agent was unable to complete all the tasks and
usually failed at tasks requiring relatively complicated manoeuvring.

4

Discussion

The specific contributions we have reported here are
the extension of CRBP learning to a modular architecture, and the introduction of partially recurrent
connections to a gating network in order to show
that this approach has potential for mediating the
actions of individual networks in a temporally extended domain. Our experiments show that architectures based on these principles are able to accomplish a series of tasks similar in type and arrangement to those reported in [11] and at a level of performance comparable to that achieved by our earlier
explicit algorithmic control scheme [10]. It remains
to be shown that the approach will scale well. The
divide and conquer approach to problem solving is
a universally accepted strategy in conventional software engineering but in the field of adaptive autonomous agents the questions of whether and how
it should be applied are still open to debate. Underlying these concerns is the need for our agents to
perform more complex and articulate tasks in uncertain and unstructured domains. Apart from its in-

herent lack of flexibility, the subsumption approach

to building individual agents leads to ad hoc engineering solutions to highly specific tasks - a useful
analogy might be that of a food processor with various task-oriented attachments - far from the emergent human-like intelligence promised at one time
by Brooks [3]. A lesson for neural net based approaches is therefore to avoid predetermined modularization at the task level. In our work, a flexible
approach to modularity that starts at the low level
of the agent's own sensory modalities has shown
some promise but, admittedly, the tasks we have
devised are each closely associated with a particular sensory modality. Further investigation of possible architectural and task variations and analysis
of the learning taking place in each module is now
being undertaken. The development of genuinely
autonomous agents entails extension of flexible control principles to higher cognitive levels; we hope
that our approach can support progress in this direction.

References
[1] D. H. Ackley and M. L. Littman. Generalisation
and scaling in reinforcement learning. In D. S.
Touretsky, editor, Advances in Neural Information
Processing Systems, pages 550-557. Morgan Kaufmann, San Mateo, CA, 1990.
[2] R. A. Brooks. A robust layered control system for
a mobile robot. IEEE Journal of Robotics and Automation, RA-2:14-23, 1986.

[3] R. A. Brooks. Intelligence without representation.
Artificial Intelligence, 47:131-159, 1991.

[4] J. Elman. Finding structure in time. Cognitive
Science, 14:179-192, 1990.

[5] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E.
Hinton. Adaptive mixtures of local experts. Neural
Computation, 3:337-345, 1991.

[6] L. Meeden, G. McGraw, and D. Blank. Emergent
control and planning in an autonomous vehicle. In

Proceedings of the Fifteenth Annual Conference of
the Cognitive Science Society, 1994.

[7] R. MoHand, T. Scutt, and P. Green. Extending
low-level reactive behaviours using primitive behavioural memory. In Proceedings of the International Conference on Recent Advances in Mechatronics, pages 510-516, 1995.


10

[8] R. M. Rylatt, C. A. Czarnecki, and T. W. Routen.

Connectionist learning in behaviour-based mobile
robots: A survey. In Artificial Intelligence Review.
Kluwer Academic Publishers. (to appear).

[9] R. M. Rylatt, C. A. Czarnecki, and T. W. Routen.

A perspective on the future of behaviour-based
robotics. In Mobile Robotics Workshop Notes Tenth Biennial Conference on Artificial Intelligence and Simulated Behaviour, 1995.

[10] R. M. Rylatt, C. A. Czarnecki, and T. W. Routen.
Learning behaviours in a modular neural net architecture for a mobile autonomous agent. In Proceedings of the First Euromicro Workshop on Advanced
Mobile Robots, pages 82-86, 1996.

[11] G. M. Saunders, J. F. Kolen, and J. B. Pollack.
The importance of leaky levels for behaviour based
A.1. In From Animals to Animats 3: Proceedings of

the Third International Conference on Simulation
of Adaptive Behaviour, pages 275-281. MIT Press,
1994.

[12] R. J. Williams. On the use of backpropagation in
associative reinforcement learning. In Proceedings
of the IEEE International Conference on Neural
Networks, pages 263-270, 1988.
[13] T. Ziemke. Towards adaptive perception in autonomous robots using second-order recurrent networks. In Proceedings of the First Euromicro W orkshop on Advanced Mobile Robots, pages 89-98,
1996.


Timing without Time -

An Experiment in Evolutionary Robotics

H. H. Lund
Department of Artificial Intelligence, University of Edinburgh,
5 Forrest Hill, Edinburgh EH1 2QL, Scotland, UK
Email:

Abstract
Hybrids of genetic algorithms and artificial neural
networks can be used successfully in many robotics applications. The approach to this is known as evolutionary robotics. Evolutionary robotics is advantageous because it gives a semi-automatic procedure to the development of a task-fulfilling control system for real robots.
It is disadvantageous to some extent because of its great
time consumption. Here, I will show how the time consumption can be reduced dramatically by using a simulator before transferring the evolved neural network
control systems to the real robot. Secondly, the time
consumption is reduced by realizing what are the sufficient neural network controllers for specific tasks. It is
shown in an evolutionary robotics experiment with the
Khepera robot, that a simple 2 layer feedforward neural

network is sufficient to solve a robotics task that seemingly would demand encoding of time, for example in
the form of recurrent connections or time input. The
evolved neural network controllers are sufficient for exploration and homing behaviour with a very exact timing, even though the robot (controller) has no knowledge
about time itself.

1

Introduction

When putting emphasis on developing adaptive
robots, one can either choose to develop single
robots with traditional learning techniques, or one
can develop a whole population of robots with a
simulated evolution process. The population based
approach named evolutionary robotics has the advantage of requiring only a specification of a taskdependent fitness formula as opposed to traditional
neural network learning techniques that demand a
learning set so that each single action of a robot

G. D. Smith et al., Artificial Neural Nets and Genetic Algorithms
© Springer-Verlag Wien 1998

can be evaluated. The disadvantage of the evolutionary robotics approach is the time that it uses to
reach a solution. This is because each single robot
has to be evaluated for a number of time steps (e.g.
1500 steps of 100 ms each). If the population is
large and the evolution has to run for many generation, then the time consumption when running online with real robots will be huge. Here, I describe
how to overcome this problem in specific robotics
tasks. This is done by designing an accurate simulator, in which the evolution of neural network control systems takes place before these evolved neural
network control systems are transferred to the real
robot in the real environment. The performances

of the simulated and real robots are almost equal.
This is due to the technique used to build the simulator. Sensory responses are simulated by using the
sensory inputs from the robot itself rather than using a mathematical or symbolic description of the
robot and its environment. Similarly, the possible
motor responses of the robot are recorded and used
in the simulator to determine the movement of the
simulated robot in the simulated environment.
Another way to decrease the time consumption in
evolutionary robotics is to determine the sufficient
complexity of a controller for a given task. Many researchers try to evolve complex structures in order
to have an open-ended evolutionary robotics, where
it is possible to evolve any kind of task-fulfilling behaviour. Yet this might mislead us to think that
the complex structures are necessary for the robot
to achieve the tasks. In many cases, a much simpler
structure can account for the behaviour, and the
time used to search for a solution can therefore be
reduced a lot by reducing the search space, when
allowing only evolution of simpler structures that