D'Alterio, Pasquale
(2021)
Novel techniques for modelling uncertain human reasoning in explainable artificial intelligence.
PhD thesis, University of Nottingham.
Abstract
In recent years, there has been a growing need for intelligent systems that not only are able to provide reliable predictions but can also produce explanations for their outputs. The demand for increased explainability has led to the emergence of explainable artificial intelligence (XAI) as a specific research field. In this context, fuzzy logic systems represent a promising tool thanks to their inherently interpretable structure. The use of a rule-base and linguistic terms, in fact, have allowed researchers to design models with a transparent decision process, from which it is possible to extract human-understandable explanations. The use of interval type-2 fuzzy logic in the XAI field, however, is limited: the improved performances of interval type-2 fuzzy systems and their ability to handle a higher degree of uncertainty comes at the cost of increased complexity that makes the semantic mapping between the input and outputs harder to understand intuitively. The presence of type-reduction, in some contexts fail to preserve the semantic value of the fuzzy sets and rules involved in the decision process. By semantic value, we specifically refer to the capacity of interpreting the output of the fuzzy system in respect to the pre-defined and thus understood linguistic variables used for the antecedents and consequents of the system. An attempt at increasing the explainability of interval type-2 fuzzy logic was first established by Garibaldi and Guadarrama in 2011, with the introduction of constrained type-2 fuzzy sets. However, extensive work needs to be carried out to develop the algorithms necessary for their practical use in fuzzy systems. The aim of this thesis is to extend the initial work on constrained interval type-2 fuzzy sets to develop a framework that preserves the semantic value throughout the modelling and decision process. Achieving this goal would allow the creation of a new class of fuzzy systems that show additional interpretable properties, and could further encourage the use of interval type-2 fuzzy logic in XAI. After the formal definition of the required components and theorems, different approaches are explored to develop inference algorithms that preserve the semantic value of the sets during the input-output mapping, while keeping reasonable run-times on modern computer hardware. The novel frameworks are then tested in a series of practical applications from the real world, in order to assess both their prediction performances and show the quality of the explanations these models can generate. Finally, the original definitions of constrained intervals type-2 fuzzy sets are refined to produce a novel approach which combines uncertain data and represents them using intuitive constrained interval type-2 fuzzy sets.
Overall, as a result of the work presented here, it is now possible to design constrained interval type-2 fuzzy systems that preserve the enhanced semantic value provided by constrained interval-type-2 fuzzy sets throughout the inference, type-reduction and defuzzification stages. This characteristic is then used to improve the semantic interpretability of the system outputs, making constrained interval type-2 fuzzy systems a valuable alternative to interval type-2 fuzzy systems in XAI. The research presented here has resulted in three journal articles, two of which have already been published in IEEE Transactions on Fuzzy Systems, and four papers presented at the FUZZ-IEEE international conference between 2018 and 2020.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
