Saturation in Fuzzy Flip-Flop Neural Networks

The aim of this paper is to investigate the phenomenon of saturation in a recurrent neural network based on J-K Fuzzy Flip-Flop neurons, and to develop a modification of the learning process to reduce this phenomenon. The work on solving this problem consists of several stages. Firstly, the concept of neuron saturation in the FFF-type network was defined and an analysis of the influence of basic internal parameters of the neural network and the Particle Swarm Optimization algorithm on saturation reduction was performed. Next, modifications to the learning algorithm were investigated. In the paper, extensions to the basic learning algorithm were then presented. These allowed reduction of the saturation of neurons inside the network. In order to perform the above analysis, the issue of data classification was addressed and, on the basis of the obtained results, the effectiveness of the introduced changes was verified in terms of the achieved neural network learning errors and saturation indices. For the numerical analysis, five benchmark data sets were exploited: Iris, Seeds, Wine, Glass, Breast Cancer and Wisconsin - drawn from the well-known UCI Machine learning repository. The aim of the validation of the developed algorithms is to demonstrate the correctness of the proposed solution in reducing the saturation of neurons, a phenomenon that affects the efficiency of learning, and which translates into confidence in the results obtained.