Snijders, P.I.J. (2005) Incorporating frequency dependent selection and sexual selection in genetic algorithms. Doctoral, Artificial Intelligence.
|
Text
AI-DKI-2005-P.I.J.SNIJDERS.pdf - Published Version Download (1MB) | Preview |
Abstract
In this project, I investigated whether the inclusion of frequency dependent selection in genetic algorithms may overcome the problem of “standard” genetic algorithms that often get stuck at local fitness peaks and do not find the global optimum. I have implemented and tested four methods to maintain diversity in the population of a genetic algorithm by frequency dependent selection. With frequency dependent selection individuals with a rare trait are in the advantage of individuals without the rare trait. When the rare trait becomes common the advantage disappears. Frequency dependent selection is incorporated in the genetic algorithm by selecting individuals that are different from the average population more often than individuals that are less different and have the same fitness value. I tested 4 different methods, a bonus method based on frequency dependent selection in nature, the sharing method as proposed by Goldberg [12], double-objective diversity maintenance as used in genetic programming by de Jong [16], and the fitness uniform selector as proposed by Hutter [15]. These methods were tested on a NK landscape and a series of two-dimensional problems. Clear winner is the double-objective diversity maintenance method [16] for promoting diversity. This method outperformed the other methods. The bonus method and the sharing method [11] did also perform better than the baseline genetic algorithm, but not as much as the double-objective method. The FUSS method [15] performed worse than the baseline genetic algorithm. Besides frequency dependent selection I had a look at two other principles from biology, sexual selection and a selective arena. Sexual selection in a genetic algorithm, using different selection pressure and mutation rates for the genders did perform a little better than the baseline genetic algorithm. By the selective arena in a genetic algorithm the idea of making a lot of offspring and using the ones that seemed best was used. This method was not efficient. The principle of ‘low investment’ and ‘quick and dirty testing’ did not seem to be useful in a genetic algorithm and with the NK landscape problem.
| Item Type: | Thesis (Doctoral) |
|---|---|
| Degree programme: | Artificial Intelligence |
| Thesis type: | Doctoral |
| Language: | English |
| Date Deposited: | 15 Feb 2018 07:28 |
| Last Modified: | 15 Feb 2018 07:28 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/8464 |
Actions (login required)
 |
View Item |