MSc theses 2001
[2005] [2004] [2003] [2002] [2001] [2000] [1999] [1998] [1997] [< 1997]
- A. J. Balaam, Full Text
Analysing homeostatic adaptation to sensorimotor disruptions
Simulated agents are evolved in circumstances which require both simple phototaxis and homeostasis in the neurons of their plastic neural network controllers. Some of these evolved agents are found to be capable of adaptation to radical sensorimotor disruptions (during their lifetimes) such as inversion of vision and reversal of effector direction, regaining phototactic behaviour after a period of maladaptation. This adaptation occurs despite no specific selection for adaptability being applied. The processes involved in agents which succeed and fail to adapt are analysed, and some evidence is found to support the hypothesis that the dynamics of internal homeostasis and external phototactic behaviour are coupled in successful agents by the evolutionary process. Some preliminary analysis of the space of possible behaviours for agents with plastic neural network controllers is described, which suggests that the adaptation observed may not be as difficult as it appears. - Alex Bridge
Credit scoring using genetic programming
Determining the credit-worthiness of a client is a task which is increasingly being automated in the financial industry. A genetic programming system is presented, designed to learn how to do this from decisions made by humans. Several different configurations of the system are discussed, varying the rates of control variables in order to find near-optimal settings. The system is then modified, with new mutation functions to improve average performance, and to force all solutions to produce valid classifications. The ability of solutions produced to generalise to unseen data is tested, and it is found that the overall performance of GP compares favourably with other methods. - Shenghong Chenm, Full Text
Foot-and-mouth disease spread in this small world
Foot-and-Mouth disease spread as desvastating storms that swept UK and Taiwan in recent epidemics. Many observations attributed these disaters to the extensive countrywide movements of animals, and the intensive nature of the livestock industry. However, there is no mathematical model incorporating these two factors in its framework. In this project, we combined formal analytical models, Artificial Life, and small world networks to build a simulation model trying to define these causative factors explicitly, and found that high density farming and long-distance transportation could not only greatly facilitate the virus transmission, but also assisted the specific adaptation of Foot-and-Mouth Disease virus in Taiwan evolving an extremely short incubation period. - Graham Clarke
Genetic algorithm determination of spacecraft trajectories
Spacecraft trajectories can be improved by subtle harnessing of the gravitational forces present in the solar system. However, no general analytic solution to the n-body gravitational system exists and so traditional trajectory design schemes split the path into separate manageable pieces and join these sections together. This work examines the possibility of creating trajectories by using a genetic algorithm. A schedule of rocket firings describes each candidate trajectory, which is calculated in a solar system simulator. The genetic algorithm then scores each path according to certain mission objectives and produces a new generation of potential trajectories. - Michael Ducker
Homeostatic mechanisms in spiking neurons and spike-time dependent plasticity
Homeostatic adaptability is studied in spiking neural networks undergoing spike-time dependent plasticity. Variable intensity lights are presented as input to a spiking network evolved purely to retain individual neuronal firing rates. Testing of evolved individuals under normal, depleted and high intensity conditions reveals interesting dynamics within individual neurons and between connected groups of neurons in the networks, often resulting in retained firing rates. Ultimately, limitations in the mechanisms are discussed and alternative potential candidates for biologically plausible homeostatic adaptation are suggested. - Thomas Hope
Over-reaching outselves: problems and responses in artificial coevolutionary design
Complex, synthetic intelligence implies both structural and behavioural complexity. There are limits to the complexity that human designers can cope with - limits that may render 'real' artificial intelligence impossible if it must be designed explicitly. An alternative is inspired by the coevolutionary arms race in nature - offering the potential for 'open-ended' optimisation, without explicit direction. Suggestive though the coevolutionary arms race may be, its potential is severely constrained in current simulations. Optimisation relative to a coevolutionary opponent may not be optimising in an objective sense. This is the problem of Relativism, explored through a set of experiments with a simple, coevolutionary model. The results provide the context for a discussion of the role of biological inspiration in the search for intelligent synthetic agents. - Ian Macinnes
An evolutionary framework for the exploration of perceptual embodiment and other aspects of embodied cognition
Simulated-physical world architecture for the study of embodied cognition is developed, concentrating on the problems ofthe hardware aspects of evolutionary robotics. The concept of perceptual embodiment is introduced, discussed and demonstrated. - Vinicius Pereira
Genome-wide, computer based palaeontology of retrotransposons in Arabidopsis thaliana
LTR-retrotransposons are mobile genetic elements related to (retro)viruses. They have colonised the genomes of all eukaryotic organisms on Earth. Evolution of the host genomes has in many cases been profoundly affected by their presence. This project developed a computer-based, automated methodology for i) large-scale mining of DNA sequence data for genomic fossils of LTR-retrotransposons, and ii) analysis of their diversity, physical distribution and temporal patterns of activity. This methodology was applied to the complete genomic sequence of the plant Arabidopsis thaliana, where LTR-retrotransposons inserted mostly in the last 6 million years, and were found to be highly concentrated near centromeres. - Adam Russell
The body as an active motor project in philosophical and practical models of cognition
The phenomenology of embodied action is compared to recent developments in cognitive science and situated robotics, leading to a focus on Merleau-Ponty's notion of the body as a motor intentionality. It is argued that this notion cannot be accommodated by the body models currently employed in synthetic studies of situated action. The construction of a truly active body through coupling of an inverse controller to a vehicle subject to nonholonomic constraints is discussed in the light of a complex 3D car driving model. - Nicholas Simons
Assessing the characteristics of a self-organizing network architecture for robot control
This report describes an experiment in evolutionary robotics that has been carried out to examine the characteristics of a self-organizing style of artificial neural network, one which adjusts its synaptic weights in real time... Evolutionary fitness was measured by the robot's success in remaining close to a light source... In this experiment, adaptive synapse networks have not evolved solutions in fewer generations nor have they reached higher values of fitness than their fixed-weight counterparts; this is contrary to the findings of Floreano and Urzelai. This discrepancy in findings may result from the different approaches taken to selecting a mutation rate - Stu Smith
Evolution of communicative behaviour
This project examines some of the issues involved in evolving communicative behaviour, from a minimal cognition viewpoint. The paper examines clonal and aclonal approaches, methods to analyse evolved behaviour (from both behavioural and genetic viewpoints), and deception in communication. - Benjamin Taylor
Environmental influences on the development of altruism
The degree of homo sapiens' social intelligence and its consequent costs, in the bearing and rearing of children, set us apart from all other species including our closest relatives. While it is difficult to argue that our peculiar traits bring about evolutionary superiority what does need explaining is the nature of the benefits necessary to at least offset the obviously elevated costs. This is particularly true of our altruistic tendencies... This dissertation seeks to investigate through an evolutionary agent model what factors impact the development of a benign cooperative environment, from initially belligerent agents, that would prove sufficient to allow the development of altruism and what I will term 'proto-trade'. The effects of resource and distribution together with variable rearing costs of offspring are modelled, to seek illumination on which factors and to what degree might be important. - Robert Vickerstaff, Full text
Dynamic Forms: Evolving Spatially Distributed Asynchronous Boolean Networks
When modelling complex systems composed of multiple interacting components, simplifications must often be made in order to make analysis easier. One such simplification is the so called Boolean idealisation, which approximates states in a system to two discrete values. The Boolean Network class of model, consisting of a set of interacting components with Boolean states, can be used to model interactions between genes in a genetic regulatory network. An algorithmically simple form of Asynchronous Boolean Network (ABN) is hereargued to be a more satisfactory approximation of real genetic networks than the alternative synchronous form of the model. A spatially extended form of the ABN is implemented in simulation to test the ability of the model to display orderly behaviour and co-ordination as a spatially distributed system. The simulation is based on a metaphor with biological morphogenesis, and uses an evolutionary search method to generate ABNs with the ability to produce cell colonies of a desired shape. The simulation results are felt to be a further step in the validation of the ABN as an abstract model of genetic interactions. - David Wrathall
The concentration and dissipation of structure in hierarchically bounded self-organising systems
Two theories of general systems development are examined. The first, known as infodynamics, originating in evolutionary biology and natural philosophy. The second, known as highly optimised tolerance (HOT), from recent studies of chaos and complexity. A computer model is described that demonstrates the formation and subsequent dissolution of dissipative structure. The processes behind this creation of global order from local rules are examined... These investigations show that HOT and infodynamic theories have strong logical connections. However, aspects of terminology require further elaboration in infodynamics particularly with respect to senescence.