Use this URL to cite or link to this record in EThOS:
Title: Evolving hierarchical visually guided neural network agents to investigate complex interactions
Author: Schlessinger, Ehud
ISNI:       0000 0004 2745 5857
Awarding Body: University of London
Current Institution: University College London (University of London)
Date of Award: 2008
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
A complex system is a system with a large number of interacting components without any mechanism for central control that displays self organisation. Understanding how these interactions affect the overall behaviour of a system is of great interest to science. Indeed, researchers use a wide variety of models to investigate complex systems. The problem with most models is that they disregard the hierarchical nature of complex systems: they ignore the fact that components of real world systems tend to be complex systems as well. This prevents researchers from investigating the interactions taking place between the lower and the higher levels of the model which may be crucial in order to gain a full understanding of the examined phenomena and of complex systems in general. Therefore, this thesis introduces Mosaic World, a multi-agent model for the purpose of investigating interactions (focusing on 'complex' multilevel interactions) within a hierarchical complex system, in addition to other computational and biological hypotheses. Mosaic World comprises a population of evolving neural network agents that inhabit a changing visual environment. By analysing the interactions that occur within Mosaic World, this thesis demonstrates the importance of incorporating hierarchical complexity into a model, and contributes to our understanding of hierarchical complex systems by showing how selective pressures cause differentiation across levels. Additionally, the study of multilevel interactions is used to probe several hypotheses and provides the following contributions among others: Analysis of agent evolvability as affected by the usage of different types of structural mutations in the evolutionary process. Demonstration that agents controlled by modular neural networks are fitter than agents that are controlled by non-modular neural networks the improvement in fitness occurs through specialisation of modules. Empirical support for a biological theory suggesting that colour vision evolved as a method of dealing with ambiguous stimuli.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available