Local evolvability of statistically neutral GasNet robot controllers

Smith, Tom, Husbands, Phil and O'Shea, Michael (2003) Local evolvability of statistically neutral GasNet robot controllers. BioSystems, 69 (2-3). pp. 223-243. ISSN 0303-2647

Full text not available from this repository.


In this paper we introduce and apply the concept of local evolvability to investigate the behaviour of populations during evolutionary search. We focus on the evolution of GasNet neural network controllers for a robotic visual discrimination problem, showing that the evolutionary process undergoes long neutral fitness epochs. We show that the local evolvability properties of the search space surrounding a group of statistically neutral solutions do vary across the course of an evolutionary run, especially during periods of population takeover. However, once takeover is complete there is no evidence for further increase in local evolvability across fitness epochs. We also see no evidence for the neutral evolution of solution robustness, but show that this may be due to the ability of evolutionary algorithms to focus search on volumes of the fitness landscape with above average robustness.

Item Type: Article
Additional Information: Originality: Applied new measures for local evolvability (introduced in previous work by same authors) to a highly complex evolutionary robotics search space in order to probe the difference in evolvability between GasNets and other forms of neural network. Introduced notion of statistical neutrality in relation to noisy fitness. Rigour: Statistically rigorous methods used to build evolvability portraits. Very large samples from a high dimensional search space used to ensure robust results. Significance: At time of publication the most detailed investigation and analysis of the dynamics of evolutionary search on a complex noisy real problem (probably still the case). Demonstrated a range of significant aspects of search dynamics including that the population tends to occupy volumes of the search space that contain far more robust solutions than the average. Impact:10 Google scholar citations, 3 Web of Science. Total Google scholar citations for this journal paper and its preceding conference paper are 33.
Schools and Departments: School of Engineering and Informatics > Informatics
Depositing User: Phil Husbands
Date Deposited: 06 Feb 2012 18:37
Last Modified: 11 Jun 2012 10:57
URI: http://sro.sussex.ac.uk/id/eprint/17435
📧 Request an update