Evolution and Ecology Program - Evolutionary Algorithms and Artificial Evolution

Evolution and Ecology Program

Evolutionary Algorithms and Artificial Evolution

Overview • Illustrations • Publications

Overview

Evolutionary Algorithms utilize the principles of biological evolution for the solution of computational problems. For tasks in which the space of potential solutions is large (usually due to combinatorical explosion), using an evolutionary process as a heuristic often yields excellent results. Although not necessarily 'optimal,' these evolutionary solutions tend to be robust and to match prescribed performance criteria.

As a case study in evolutionary computation, the ADN program is exploring evolutionary algorithms for the artificial synthesis of controllers that are represented by neural networks. A major feature of such a scheme is that the evolutionary 'learning' in populations of these networks does not only modify weights and thresholds within a given network architecture, but at the same time alters and improves architectures themselves. Especially for problems that require recurrent (instead of standard feed-forward) architectures, this is of critical importance for obtaining viable solutions.

In view of the intricate relation between the structure of a neural network and its function, the evolution of neural controllers also serves as an interesting example of complex genotype-phenotype mappings. Processes of artifical evolution can therefore serve as testbeds for understanding more realistic types of evolutionary dynamics. For this purpose, new paradigms need to be incorporated into evolutionary theory: among others, concepts like 'neutral clusters' or 'holey adaptive landscapes' are promising topics of current discussions.

Illustrations

Artificial evolution of a neural network that solves the odd-parity-4 problem.


(a) Generation 0	(b) Generation 2000	(c) Generation 4000	(d) Generation 6000

Neurons are depicted by circles, connections by curves.
Excitatory connections are hollow, inhibitory connections are filled,
the thickness of curves measures connection strength.
Horizontal bars at neurons indicate their firing thresholds.
Notice that the evolutionary algorithm adds and deletes neurons
as well as connections and modifies connection strengths as well as threshold values.
The resulting network (d), automatically constructed by the evolutionary algorithm,
is the minimal solution to the odd-parity-4 problem.

Publications

1.	Brandt H, Dieckmann U: Correlation Analysis of Fitness Landscapes. IIASA Interim Report IR-99-052 (1999).

2.	Brandt H: Correlation Analysis of Fitness Landscapes. IIASA Interim Report IR-01-058 (2001).

3.	Pasemann F, Dieckmann U: Evolved Neurocontrollers for Pole Balancing. Mira J, Moreno-Diaz R, Cabestany J (eds): Biological and Artificial Computation: From Neuroscience to Technology, Lecture Notes in Computer Science 1240, Springer-Verlag, Heidelberg, pp. 1279-1287 (1997).

4.	Pasemann F, Steinmetz U, Dieckmann U: Evolving Structure and Function of Neurocontrollers. Proceedings of the 1999 Congress on Evolutionary Computation, July 6-9 1999, Madison, Washington DC, USA. IEEE Press, pp. 1973-1978 (1999).

Responsible for this page: Melanie Wenighofer
Last updated: 24 Feb 2006

International Institute for Applied Systems Analysis (IIASA)

Schlossplatz 1

A-2361 Laxenburg, Austria
Phone: (+43 2236) 807 0

Fax: (+43 2236) 71 313

Web: www.iiasa.ac.at

Disclaimer