Evolving provably explainable fuzzy pattern tree classifiers using grammatical evolution

The central hypothesis of this thesis is that Fuzzy Pattern Trees (FPTs) can be considered a powerful explainable artificial intelligence (XAI) technique and that highly accurate FPTs can be evolved using Grammatical Evolution (GE), an evolutionary computation technique. 

While no single definition of what constitutes an XAI system is agreed upon, we investigate the suitability of a system to be deemed as XAI based on four core criteria: transparency, accuracy, trustworthiness and the ability to incorporate domain knowledge.

We start with a bottom-up system to identify useful subtrees, i.e., building blocks, in a GE run, which were subsequently made available for later generations. While this improved the performance, the complicated and unintuitive nature of the full solutions found using GE were not useful for XAI.

We then pivot strategy and aim to create an intrinsically interpretable model. Specifically, we evolved FPTs using GE, which we call FGE. We systematically explore their ability to satisfy the core criteria, described above. We first show FGE meets the accuracy requirement and investigate the effect ensemble methods have on performance, improving it in half of benchmarks. Parsimony pressure was shown to reduce the size of the trees with no compromise in performance.

Next, the transparency and trustworthiness of FPTs are directly investigated by a domain expert. This is done using a selection of real work benchmark problem sets. Models with sensible logic, as judged by the expert, outperform models with poor logic, validating that FGE creates interpretable models.