Controlled observations of the genetic algorithm in a changing environment: case studies using the shaky ladder hyperplane-defined functions

• Authors: William Michael Rand;John H. Holland;Rick L. Riolo
• Venue: Controlled observations of the genetic algorithm in a changing environment: case studies using the shaky ladder hyperplane-defined functions
• Year: 2005

Though recently there has been interest in examining genetic algorithms (GA's) in dynamic environments, work still needs to be done. When studying the GA in static environments, a test suites of functions that are designed for the GA is often used so that the performance of the GA can be observed under systematic, controlled conditions. One example of these suites is the hyperplane-defined functions (hdf's) designed by Holland. This dissertation describes a new test suite of dynamic functions to the study of the GA, and explores three main variants of that test suite that exhibit different effects on the performance of the GA. This test suite, called the shaky ladder hyperplane-defined functions (sl-hdf's) is developed in accordance with a framework that calls for the GA and its behavior to be studied through systematic, controlled observations, thus providing recommendations to practitioners and possible hypotheses to theorists. I begin by describing the sl-hdf's: how they are created, what properties they have, and how they can be utilized for exploring the GA. I conduct examinations of the landscapes of the sl-hdf's and show that some standard metrics of ruggedness are inadequate for these landscapes. Next, I describe a suite of metrics that allow the behavior of the GA in dynamic environments to be observed and analyzed more readily. I then examine these measures on variations of the sl-hdf's. I show that the type and the frequency of changes in dynamic environments is important to the performance of the GA. For instance, in a very rugged environment, the GA is able to perform better when there are regular changes which prevent it from prematurely converging. In addition, I show that crossover greatly increases the performance of the GA on the sl-hdf's. I also investigate the use of self-adaptive mutation rates in dynamic environments. I show that in a variety of circumstances, self-adaptive mutation rates lead to worse performance on the sl-hdf's than a fixed mutation rate, even in a static environment. I conclude by detailing future work to further test the hypotheses presented in this thesis and investigate mechanisms that could be incorporated into the simple GA, like coevolution, to improve performance. I also show how the results of this dissertation could be applied to different fields, and end with a summary.