Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.662610
Title: Prescriptive formalism for constructing domain-specific evolutionary algorithms
Author: Surry, Patrick David
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 1998
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
It has been widely recognised in the computational intelligence and machine learning communities that the key to understanding the behaviour of learning algorithms is to understand what representation is employed to capture and manipulate knowledge acquired during the learning process. However, traditional evolutionary algorithms have tended to employ a fixed representation space (binary strings), in order to allow the use of standardised genetic operators. This approach leads to complications for many problem domains, as it forces a somewhat artificial mapping between the problem variables and the canonical binary representation, especially when there are dependencies between problem variables (e.g. problems naturally defined over permutations). This often obscures the relationship between genetic structure and problem features, making it difficult to understand the actions of the standard genetic operators with reference to problem-specific structures. This thesis instead advocates making the representation of solutions of the explicit focus, in order to highlight the way in which the genetic operators (and resulting search algorithms) form and test hypotheses about the relationship between observed problem structure and films. It is clear that any search algorithm must limit the class of hypotheses which it is able to learn (its bias), if it is to select the most accurate of those hypotheses efficiently. We demonstrate this in the context of evolutionary search by exploring the so-called "no free lunch" results, and argue that it is the chosen representation which determines what kinds of hypotheses can be formed and tested by the algorithm. To do this, we exploit a general formalism for generating a representation for an arbitrary instance of a given problem domain, using a characterisation of that problem domain which captures beliefs about its structure. Such a characterisation is simply an explicit set of mathematical statements about the relationship between features of solutions and their fitness values, making it clear that the resulting representations encapsulate all of the domain knowledge which is available to any search algorithm.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.662610  DOI: Not available
Share: