Illuminating meaningful diversity in complex feature spaces through adaptive grid-based genetic algorithms

Overbury, Peter Charles (2020) Illuminating meaningful diversity in complex feature spaces through adaptive grid-based genetic algorithms. Doctoral thesis (PhD), University of Sussex.

[img] PDF - Published Version
Download (20MB)

Abstract

In many fields there exist problems for which multiple solutions of suitably high performance may be found across distinct regions of the search space. Optimisation of the search towards including these distinct solutions is important not only to understanding these spaces but also to avoiding local optima. This is the goal of a type of genetic algorithms called illumination algorithms. In Chapter 2, we demonstrate the use of an illumination algorithm in the exploration of networks sharing only a given set of structural features (valid networks). This method produces a population of valid networks that are more diverse than those produced using state of the art methods, however, it was found to be too inefficient to be usable in real-world problems. Additionally, setting an appropriate resolution of the search requires some amount of prior knowledge of the space of solutions. Addressing this problem is the focus of Chapter 3, in which we develop three extensions to the method: a) an exact method of mutation whereby only valid networks are explored, b) an adaptive mechanism for setting the resolution of the search, c) a principle for tuning mutations parameters to the search’ s resolution. We show that with these additions our method is able to increase the diversity of solutions found in significantly fewer iterations. Finally, in Chapter 4 we expand our method for use in more general problem spaces. We benchmark it against the state of the art. In all tested landscapes, we show that our method is able to identify more meaningful niches in the spaces in the same number of iterations. We conclude by highlighting the limits of our framework and discuss further directions.

Item Type: Thesis (Doctoral)
Schools and Departments: School of Engineering and Informatics > Informatics
Subjects: Q Science > QA Mathematics > QA0075 Electronic computers. Computer science > QA0076.87 Neural computers. Neural networks
Q Science > QH Natural history > QH0301 Biology > QH0426 Genetics
Depositing User: Library Cataloguing
Date Deposited: 30 Sep 2020 13:14
Last Modified: 30 Sep 2020 13:14
URI: http://sro.sussex.ac.uk/id/eprint/93352

View download statistics for this item

📧 Request an update