Canonical representation in genetic programming
Canonical representation in genetic programming
This paper explores the effect of different forms of representation and different forms of genetic operations on the performance of a Genetic Programming (GP) system. The GP system is based on a Genetic Algorithm (GA) that evolves software agents, represented by tree structures, which are then applied to some problems in robotics. In their evolved form, the trees are not well-formed, and many of them include a considerable amount of intronic material that plays no part in the performance of the underlying agent. We introduce an alternative way of representing the agent, which eliminates the intronic material and reflects more clearly the decisions that the agent needs to make in different situations as it attempts to solve problems in spatial awareness. We use the term canonical to refer to this alternative tree structure. We then extend the standard GP crossover operator to perform canonical crossover, in which the parents exchange canonical sub-trees.
This paper shows that using canonical evolution alongside conventional evolution can result in substantial improvements in the performance of the GP system and thus the resulting agents, the level of improvement being dependent on the mix of canonical and non-canonical members of the population, and on the rates and types of crossover.
Keane, A.J.
26d7fa33-5415-4910-89d8-fb3620413def
Bearpark, K.
cf87d373-6c01-48e2-8070-0a6199ce2cd2
29 April 2008
Bearpark, K.
cf87d373-6c01-48e2-8070-0a6199ce2cd2
Keane, A.J.
26d7fa33-5415-4910-89d8-fb3620413def
Keane, A.J.
(2008)
Canonical representation in genetic programming.
Bearpark, K.
(ed.)
Adaptive Computing in Design and Manufacture 2008, Bristol, United Kingdom.
29 Apr - 01 May 2008.
Record type:
Conference or Workshop Item
(Paper)
Abstract
This paper explores the effect of different forms of representation and different forms of genetic operations on the performance of a Genetic Programming (GP) system. The GP system is based on a Genetic Algorithm (GA) that evolves software agents, represented by tree structures, which are then applied to some problems in robotics. In their evolved form, the trees are not well-formed, and many of them include a considerable amount of intronic material that plays no part in the performance of the underlying agent. We introduce an alternative way of representing the agent, which eliminates the intronic material and reflects more clearly the decisions that the agent needs to make in different situations as it attempts to solve problems in spatial awareness. We use the term canonical to refer to this alternative tree structure. We then extend the standard GP crossover operator to perform canonical crossover, in which the parents exchange canonical sub-trees.
This paper shows that using canonical evolution alongside conventional evolution can result in substantial improvements in the performance of the GP system and thus the resulting agents, the level of improvement being dependent on the mix of canonical and non-canonical members of the population, and on the rates and types of crossover.
More information
Published date: 29 April 2008
Venue - Dates:
Adaptive Computing in Design and Manufacture 2008, Bristol, United Kingdom, 2008-04-29 - 2008-05-01
Identifiers
Local EPrints ID: 148241
URI: http://eprints.soton.ac.uk/id/eprint/148241
PURE UUID: f48328bf-9069-42ab-9965-4519c791dcb2
Catalogue record
Date deposited: 27 Apr 2010 13:53
Last modified: 14 Mar 2024 02:39
Export record
Contributors
Editor:
K. Bearpark
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics
