Self-Configuring Evolutionary Algorithms Based Design of Hybrid Interpretable Machine Learning Models

Abstract

The paper describes an approach in which the decision-making process of an artificial neural network is interpreted by a fuzzy logic system. A neural network and a fuzzy system are automatically designed with the use of the self-configuring evolutionary algorithms. Experiments are carried out on classification tasks. As a result, it is shown that the building of a fuzzy system on the inputs and outputs of a neural network allows one to build an interpreted rule base of a smaller size, as if this rule base were built on the data of the original problem. In addition, the accuracy of such a system is comparable to the accuracy of a fuzzy system trained on the original task. As a result, the researcher has a neural network with high accuracy of solving the problem, as well as a fuzzy system explaining the neural network's decision-making process. The article presents some constructed rule bases and neural networks for interpretation of which they were built.