Use this URL to cite or link to this record in EThOS:
Title: A quantitative approach to the analysis of memory requirements for autonomous agent behaviours using evolutionary computation
Author: Kim, DaeEun
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2002
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
When building controllers for autonomous agents, such as real or simulated robots, there is very little theoretical guidance on how complex the controller must be in terms of internal state to be able to achieve good performance in a particular task. The underlying problem is that it is hard to quantify the amount of internal state — memory — the controller should have based on knowledge of the task to be tackled. This problem manifests itself in two related questions: how complex is a given task in terms of the amount of memory a controller needs to enable a particular agent to complete it? and how does this complexity vary depending on the sensorimotor abilities of the agent? In this thesis I address these questions empirically using evolutionary computation techniques, by attempting to evolve controllers with good performance for a task while simultaneously minimizing the amount of internal memory used. Boolean logic networks with memory elements (registers) allow a natural progression from purely reactive (memoryless) controllers while being straightforward to use with evolutionary computations. Controllers are evolved using three kinds of representational structure — finite state machines, rule-based state machines and tree state machines — equivalent to Boolean logic networks with memory, for which the internal state can easily be quantified. Various evolutionary computational techniques are adapted to support analysis of memory, including evolutionary multiobjective optimisation, a sample selection method, the design of suitable genetic operators and an elitism strategy. Performance evaluation and evolutionary algorithm issues for noisy environments are also considered. The evolutionary computational approaches are applied to a variety of standard grid world and simple robotic tasks to quantify the memory needed for adequate performance. The results demonstrate, as expected, that internal state plays a significant role in improving performance of agent behaviours when sensors or motor actions are restricted, and is related to the amount of perceptual aliasing present in the interaction between agent and environment. New results concerning the "difficulty" of the standard problems are obtained. The investigation of these problems also demonstrates the utility of the tools developed for quantifying and analysing memory use.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available