- author = "Jeffrey Kermes Bassett",
- title = "Methods for Improving the Design and Performance of Evolutionary Algorithms",
- school = "The Volgenau School of Engineering, George Mason University",
- year = "2012",
- address = "USA",
- month = "Fall",
- keywords = "genetic algorithms, genetic programming",
-
URL = "
http://hdl.handle.net/1920/8122",
-
URL = "
http://digilib.gmu.edu/dspace/bitstream/handle/1920/8122/Bassett_gmu_0883E_10215.pdf",
- size = "160 pages",
-
abstract = "Evolutionary Algorithms (EAs) can be applied to almost
any optimization or learning problem by making some
simple customizations to the underlying representation
and/or reproductive operators. This makes them an
appealing choice when facing a new or unusual problem.
Unfortunately, while making these changes is often
easy, getting a customized EA to operate effectively
(i.e. find a good solution quickly) can be much more
difficult.
Ideally one would hope that theory would provide some guidance here, but in these cases, evolutionary computation (EC) theories are not easily applied. They either require customization themselves, or they require information about the problem that essentially includes the solution. Consequently most practitioners rely on an ad-hoc approach, incrementally modifying and testing various customizations until they find something that works reasonably well.
The drawback that most EC theories face is that they are closely associated with the underlying representation of an individual (i.e. the genetic code). There has been some success at addressing this limitation by applying a biology theory called quantitative genetics to EAs. This approach allows one to monitor the behaviour of an EA by observing distributions of an outwardly observable phenotypic trait (usually fitness), and thus avoid modelling the algorithm's internal details. Unfortunately, observing a single trait does not provide enough information to diagnose most problems within an EA. It is my hypothesis that using multiple traits will allow one to observe how the population is traversing the search space, thus making more detailed diagnosis possible.
In this work, I adapt a newer multivariate form of quantitative genetics theory for use with evolutionary algorithms and derive a general equation of population variance dynamics. This provides a foundation for building a set of tools that can measure and visualize important characteristics of an algorithm, such as exploration, exploitation, and heritability, throughout an EA run. Finally I demonstrate that the tools can actually be used to identify and fix problems in two well known EA variants: Pittsburgh approach rule systems and genetic programming trees.",
- notes = "Supervisor Kenneth A. De Jong",
