OpenAIR @ RGU >
Design and Technology >
Reports (Engineering) >
Please use this identifier to cite or link to this item:
|Title: ||Evolution and devolved action: towards the evolution of systems.|
|Authors: ||MacLeod, Christopher|
Capanni, Niccolo Francesco
Maxwell, Grant M.
|Keywords: ||Artificial neural networks|
|Issue Date: ||2001|
|Publisher: ||Robert Gordon University|
|Citation: ||MACLEOD, C., MCMINN, D., REDDIPOGU, A. B., CAPANNI, N. F. and MAXWELL, G. M., 2001. Evolution and devolved action: towards the evolution of systems. In: Appendix B of MCMINN, D. Using evolutionary artificial neural networks to design hierarchical animat nervous systems, Ph. D. thesis. Aberdeen : Robert Gordon University.|
|Abstract: ||The Artificial Neural Networks group at the Robert Gordon University has, over the
last six years, built up considerable knowledge and practical experience in
Evolutionary Artificial Neural Networks. This experience started with the PhD project
by C MacLeod  and continued with the work of D McMinn  which is due for
submission in June 2001. These are being followed up by the work of research
students A B Reddipogu and N F Capanni.
Initial work concentrated on Neural Networks that could grow to fulfil their function.
C MacLeod, in his thesis on this topic, proposed a model of a robotic control system
to be used as a vehicle for further research. This model formed the basis of D
McMinn’s project. The robotic control system [2, 3, 4] has a defined modular
structure that enables the researcher to create networks for particular tasks and so
allow the robot to function. McMinn has used this structure successfully as a basis to
develop Evolutionary ANNs implementing Central Pattern Generators and Reflexes
for robot locomotion.
It has become apparent, over the course of these projects, that a network that can
evolve into a modular structure without the need for designed partitioning would be
the next step forward for the group’s research. This should allow the network to
develop naturally and in an open-ended way without the need to artificially constrain
it. Such an approach needs an evolutionary algorithm that can automatically and
naturally evolve a “system”: that is, a modular network rather than a fully
interconnected homogenous structure. No acceptable Genetic or Evolutionary
Techniques are currently available to do this. The group therefore needed to look to
nature and discover the reasons why natural systems allowed such modularity to
evolve and how it might be exploited in the course of future work.
It was felt that this work would be a culmination of the group’s research to date, both
furthering the work on robotic control systems and also including ideas from previous
practical work, by MacLeod and McMinn, in terms of Evolutionary Networks and
Incremental Evolution (Embryological Algorithms) . This is because such a
network will need to evolve and to “grow” from a simple to complex structure. The
research and ideas that lay the foundations for this work are contained in this report.
The report also considers two related aspects (neural learning and neural function),
but as explained above, it is the layout topology or wiring which is felt to be the most
important topic. After all, complex systems may be built up from simple elements,
such as transistors or gates (or indeed, at the most basic level, atoms), providing that
they are interconnected in the correct manner.
Therefore, to summarise: We are concerned with discovering an evolutionary method
capable of evolving systems. To establish a method capable of this, we must look again at the action of biological evolution and compare it with current artificial
|Appears in Collections:||Reports (Engineering)|
All items in OpenAIR are protected by copyright, with all rights reserved.