You are here : University of Gloucestershire > Research > Research Repository

A New Method for Artificial Neural Network Reasoning Using Case Based Reasoning System

Tools

Appuhamy Liyanage, Don SM (2022) A New Method for Artificial Neural Network Reasoning Using Case Based Reasoning System. PhD thesis, University of Gloucestershire. doi:10.46289/YZKU4765

Text (Final thesis)
15102 Appuhamy, Liyanage (2022) A new method for artificial neural network reasoning using case based reasoning system.pdf - Accepted Version
Restricted to Repository staff only until 18 August 2027. (Public release would prejudice commercial interests).
Available under License All Rights Reserved.
Download (119MB)

Official URL: https://eprints.glos.ac.uk/15102

Abstract

Problem Statement: Automated decision-making systems that decision-making use Artificial Neural Networks are challenged to show why a decision is made. The formation of this reasoning process is a governing requirement in most regulated industries. Moreover, when an automated decision system uses a deep neural network within a complex problem space, it creates a myriad of problems when trying to generate a reason or a rationalisation explanation. With the exploration of the technical literature and the empirical studies: current attempts to resolve the above problem, particularly in artificial neural networks, are not appropriate for deep neural networks, systems with complex problem spaces, nor when the decision requires a reasoning solution, not just an input weighting based rationalisation solution. The thesis focuses on creating a mathematical metamodel for generating a reasoning resolution specific to artificial neural network applications. Research Design and Methodology: The research encompasses positive epistemological mathematical artefacts that can be simulated using mathematical model simulation techniques such as mathematical proofing, formal specificationbased simulations and algorithmic observations. Moreover, the thesis uses an inferential study method for artefact creation, model simulation and data collection. The research strategy tails the following phases: 1 – Theory, 2 – Objectives and theory, 3 – Empirical Studies related to the research, 4 – a creation of mathematical artefacts, and 5 – validation. Results: The outcomes of the research, including mathematical modelling using cross simulation techniques, establish that the proposed methods and models are effective for determining reasons for generating solutions using artificial neural networks within a complex problem space. A tri-phase simulation technique was used to cross validate the models, including 1 – mathematical proofing, 2 – formal specification based testing and 3 – algorithmic simulations. The mathematical proofs and separately other two simulation methods provide the necessary evidence that the new reasoning architecture and related artefacts offer great potential for creating reasoning systems. Furthermore, all the parameters, parameter types and parameter limitations related to each artefact are detailed as part of the research. Contributions: Foremost contributions of the thesis are from the theoretical perceptions, and the development of the theories on knowledge representation as case based reasoning cases, case versioning, case indexing, case retrieval, and case adaptation can be used not only on reasoning solutions but also in general information storage, distributed information processing and large data indexing solutions. The main contribution of the thesis from an applied standpoint is a new method for causal reasoning for a particular artificial neural network system decision/output with the application of models formed in the thesis. Moreover, the new Synapses Logger based neural network architecture will contribute to the knowledge by creating more biologically inspired artificial neural networks.

Item Type:

Thesis (PhD)

Thesis Advisors: