ExTree - Explainable Genetic Feature Coupling Tree using Fuzzy Mapping for Dimensionality Reduction with Application to NACA 0012 Airfoils Self-Noise Data Set

EasyChair Preprint 5049

Abstract

This research presents an AI-based tool (ExTree) that provides high explainability in prediction problems that involve multiple continuous inputs. The algorithm uses an input cou-pling tree that gradually reduces the dimension of the system. The desired dimension reduction is achieved developing a network of fuzzy inference systems (FIS) wherein in each layer of the network, two inputs get combined to yield a single outcome. These outcomes are then submit-ted to the same procedure at the following layer until we arrive at a single output, thereby re-ducing the dimensionality of the problem in every step. Hence, large scale problems with more inputs require more layers. The final outcome is that we obtain a set of FIS nodes across the network, where each FIS may be characterized using an explainable control surface. The struc-ture of the tree is optimized using a genetic algorithm that gets the best hierarchy of fuzzy fea-tures to minimize the dispersion of the final outcome. This tool has been benchmarked using NASA’s wind tunnel testing database of NACA 0012 Airfoils. The results, demonstrating accurate validation, are of value not only from the perspective of a high performing AI-based algorithm, but also because of the substantial amount of interpretability and traceability that the algorithm offers.

Keyphrases: Airfoils, ExTree, Explainability, Fuzzy Feature Mapping, Genetic Algorithm, Input Coupling Tree, dimensionality reduction, dispersion

@booklet{EasyChair:5049,
  author    = {Javier Viaña and Kelly Cohen},
  title     = {ExTree - Explainable Genetic Feature Coupling Tree using Fuzzy Mapping for Dimensionality Reduction with Application to NACA 0012 Airfoils Self-Noise Data Set},
  howpublished = {EasyChair Preprint 5049},
  year      = {EasyChair, 2021}}