Emerging Technologies '97

Evolving Biological Neural Networks Using Genetic Algorithms

Adam V. Adamopoulos1, Spiridon D. Likothanassis1,2 and Efstratios F. Georgopoulos1

1 University of Patras, School of Engineering, Dept. of Computer Engineering and Informatics, 265 00, Patras, Greece
2 Computer Technology Institute (CTI), 3 Kolokotroni str., 261 00, Patras, Greece

In the present work, a new computer simulation methodology in introduced considering the structure/function problem of the Central Nervous System (CNS). The proposed methodology is based on and inspired from the areas of Evolutionary Computation (EC) and Genetic Algorithms (GA) and incorporates the benefits of these effective tools. To be more explicit, the methodology presented here deals with the design and implementation of specific EC-GA programs that account for the structural development of Biological Neural Networks (BNN) with desired structural and/or functional characteristics. A description of the basic elements that construct the developed GA is given. Those elements refer to the individuals of the GA population (which are BNNs) and the specialized genetic operators (of crossover and mutation type) that were implemented to handle the BNN population. Furthermore, results of a first application of the described methodology with respect to the effect  of neural network structure and the appearance of cycling neural activity are reported.

The Revolution of Evolution for Real-World Applications

Peter Bentley

Intelligent Systems Group, Department of Computer Science University College London, Gower St., London WC1E 6BT, UK. email: P.Bentley@cs.ucl.ac.uk

This paper describes an evolutionary search method known as the genetic algorithm (GA) and examines its application to real-world problems. A description of the algorithm itself and its history is provided. A general review of GA theory and analysis is given, and many of the newer, more advanced types of GA are introduced. Some of the hundreds of different applications tackled by GAs are then described. Two more detailed case-studies are included, to outline how GAs can be applied to real-world applications, and to illustrate some of the difficulties such problems present to the designers of GA-based systems. The paper concludes that the GA is a highly significant and useful tool, suitable for a variety of different applications.

Rethinking the Appeal of Evolution: Empirical Evidences from the Financial Applications of Genetic Algorithms

Shu-Heng Chen1 and Wei-Yuan Lin2

1 AI-ECON Research Group, Department of Economics, National Chengchi University, Taipei, Taiwan 11623
2 AI-ECON Research Group, Department of Economics, Soochow University, Taipei, Taiwan 100

In this paper, some reinterpretations of the appeal of evolution are provided. These reinterpretations are based on the empirical
evidences from a financial application of the genetic algorithm. The simulations conducted here is a typical portfolio selection problem. When applying GAs to this problem, the landscape is not fixed; therefore, it is interesting to see how the schema theorem can be applied in this application with dynamic landscape. Using box-and-whisker analysis, we show that the traditional schema theorem can fail in many cases. However, the analysis of skewness and outliers in the box-and-whisker plot provides us with some new interpretations of the schema theorem. These interpretations can be very useful for the design of GA-based portfolios.

The Impact of External Dependency in Genetic Programming Primitives

Una-May O'Reilly

The Artificial Intelligence Lab, Massachusetts Institute of Technology, 545 Technology Sq, Cambridge MA, 02139. email: unamay@ai.mit.edu

Both control and data dependencies among primitives impact the behavioural consistency of subprograms in genetic programming solutions. Behavioural consistency in turn impacts the ability of genetic programming to identify and promote appropriate subprograms. We present the results of modelling dependency through a parameterized problem in which a subprogram exhibits internal and external dependency levels that change as the subprogram is successively incorporated into larger subsolutions. We find that the key difference between non-existent and "full'' external dependency is a longer time to solution identification and a lower likelihood of success as shown by increased difficulty in identifying and promoting correct subprograms.

A Review of Theoretical and Experimental Results on Schemata in Genetic Programming

Riccardo Poli and W. B. Langdon

School of Computer Science, The University of Birmingham, Birmingham B15 2TT, UK. email: {R.Poli,W.B.Langdon}@cs.bham.ac.uk

Schemata and the schema theorem, although criticised, are often used to explain why genetic algorithms (GAs) work. A considerable research effort has been produced recently to extend the GA schema theory to Genetic Programming (GP). In this paper we review the main results available to date in the theory of schemata for GP and some recent experimental work on schemata.

Phenotype-Object Programming, Phenotype-Array Datatype, and an Evolutionary Signal-Ensemble FX Trading Model

Poomjai Nacaskul

Department of Computing, Imperial College, University of London, 180 Queen's Gate, London SW7 2BZ, England. email: pn2@doc.ic.ac.uk

We set out to optimise a financial trading model which utilises an ensemble of time-series trend indicating signal-models. The optimisation search is combinatorial over the combination of signal-model classes as well as parametric over the parameterisation of individual signal-model objects. Because the optimisation objective (net gain from simulated transactions against a historical price series) is not differentiable w.r.t. our trading model space, we look to evolutionary optimisation [EO] methodologies [Fogel, 1994], e.g. Genetic Algorithm [GA] [Holland, 1975/92], which rely on direct solution performance evaluation. However, because the multi-stage nature of our solution space prohibits stochastic-evolutionary convergence, we need to engineer a new EO paradigm to implement combinatorial and parametric search processes concurrently. This, we accomplish by exploiting the inherently object-oriented [OO] [Booch, 1994; Stroustrup, 1991] nature of an EO algorithm and of the combinatorial-parametric solutions. We propose Phenotype-Object Programming [POP] as a generalised OO model and implementation of an EO algorithm and Phenotype-Array Datatype [PAD] as a generalised OO model and implementation of a combinatorial-parametric solution [Nacaskul, 1997]. We apply this to our signal-ensemble trading model and discuss experimental results on DEM/JPY and USD/DEM data.

Soft Computing And Re-Engineering

Conor Ryan, J.J. Collins, Jim Buckley, Tony Cahill, Niall Griffith and Paddy Healy

Department of Computer Science and Information Systems, University of Limerick, Ireland. email: Conor.Ryan@ul.ie

The Soft Computing and Re-engineering groups at the University of Limerick have recently been granted significant funding for collaborative work. The work is based on developing and integrating the concepts for serial to parallel code transformation. The traditional, hard computing, approaches of Re-Engineering will be combined with Soft Computing approaches with two goals in mind. Firstly to produce a general and practical serial to parallel code translator, and secondly, to compare and contrast the different approaches, specifically to investigate the advantages of combining the two approaches.

Genetic Algorithms for Industrial Planning

Thomas Stidsen

Dept. of Computer Science, University of Aarhus, Ny Munkegade, Bldg. 540, DK-8000 Aarhus C, Denmark

Genetic Algorithms have been an active research area for more than three decades, but the industrial applications of this search technique have been scarce. There may be several reasons for this. The EVALIA 1 project (EVolutionary ALgorithms for Industrial Applications) attempts to test the value of Genetic Algorithms on realistic industrial problems. Further a general framework is developed to ease the implementation of optimisation programs for industrial problems. This article reports on the first results of this project, when testing the framework on a real-world planning problem at Odense Steel Shipyard (OSS). Further this article will report on the possibilities of specialised Genetic Algorithm techniques: Adaptive Operators which are used in what we call the shotgun approach, the Pareto technique for multi-objective optimisation and Co-evolutionary Constraint Satisfaction.

An Investigation of Sexual Selection as a Mechanism for Obtaining Multiple Distinct Solutions

Michael Ratford, Andrew Tuson, and Henry Thompson

Department of Artificial Intelligence, University of Edinburgh, 80 South Bridge, Edinburgh EH1 1HN, U.K. email: {miker,andrewt,htg}@dai.ed.ac.uk

The ability to obtain multiple distinct solutions in a single run is an important, though often forgotten, practical advantage of genetic algorithms, and therefore methods that enhance this ability are desirable. The approach taken here is inspired by nature: sexually reproducing organisms do not mate indiscriminately --- the choice of mate has a large impact upon the fitness of the organism's offspring, and by balancing exploration and exploitation, mate (or sexual) selection can lead to speciation effects which may enhance the genetic algorithm's ability to locate multiple distinct solutions. The investigation described here shows that sexual selection was able to enhance the genetic algorithm's ability to locate and maintain multiple distinct solutions, although it can interact detrimentally with other techniques designed for the same purpose.

Fitness Causes Bloat: Mutation

W. B. Langdon and R. Poli

School of Computer Science, University of Birmingham, Birmingham B15 2TT, UK. email: {W.B.Langdon,R.Polig}@cs.bham.ac.uk http://www.cs.bham.ac.uk/~wbl, ~rmp. Tel: +44 (0) 121 414 4791, Fax: +44 (0) 121 414 4281

The problem of evolving, using mutation, an artificial ant to follow the Santa Fe trail is used to study the well known genetic programming feature of growth in solution length. Known variously as "bloat'', "fluff '' and increasing "structural complexity'', this is often described in terms of increasing ``redundancy'' in the code caused by "introns''. Comparison between runs with and without fitness selection pressure, backed by Price's Theorem, shows the tendency for solutions to grow in size is caused by fitness based selection. We argue that such growth is inherent in using a fixed evaluation function with a discrete but variable length representation. With simple static evaluation search converges to mainly finding trial solutions with the same fitness as existing trial solutions. In general variable length allows many more long representations of a given solution than short ones. Thus in search (without a length bias) we expect longer representations to occur more often and so representation length to tend to increase. I.e. fitness based selection leads to bloat.

GPsys: A Scalable Genetic Programming System 

Adil Qureshi

Department of Computer Science University College London, Gower St., London WC1E 6BT, UK. email: A.Qureshi@cs.ucl.ac.uk

GPsys is a portable, scalable Genetic Programming system written in the programming language Java.  It is based on a steady state engine, and supports automatically defined functions (ADFs), strong typing and distributed populations (demes).  This paper describes its architecture and demonstrates its use through examples.