Analysis of the Effectiveness of Wang-Mendel Fuzzy Network Learning Algorithms

When training fuzzy networks, they often face the problem of choosing a learning algorithm. To solve the problem of choosing the most effective learning algorithm, various gradient and stochastic learning algorithms of the Wang-Mendel network were studied. The gradient descent algorithm and the conjugate gradient algorithm were chosen as gradient algorithms. Stochastic algorithms are represented by a genetic algorithm, an annealing simulation algorithm, a particle swarm simulation algorithm, and a differential evolution algorithm. The effectiveness of learning algorithms is investigated by the example of solving the problem of classifying two sets of model data: Fischer's irises and Italian wines. For the study, several modifications of the Wang-Mendel network with various fuzzy logic algebras are studied: Gödel algebra, Gauguin algebra and Lukashevich algebra. For wine classification, the best results were shown by a network model with Gauguin algebra trained by a genetic algorithm. For the classification of irises, the best results were also shown by a model with a Gauguin algebra trained by a particle swarm algorithm and a model with a Lukashevich algebra trained by a genetic algorithm.