Fuzzy cognitive networks with functional weights for time series and pattern recognition applications

Abstract

Over the last decades, a number of remarkable pattern recognition algorithms have been proposed, as a result of the continuous raising of pattern classification in one of the major application areas of artificial intelligence. Many real world problems require the use of efficient classification models pointing out the need for continuous research and study of new techniques. Such a proposal is to use Fuzzy Cognitive Maps (FCMs) and their extensions to solve distinct classification tasks. The extension named Fuzzy Cognitive Network (FCN) has the clear advantage of guaranteed convergence to equilibrium points, which in turn makes it more suitable for pattern recognition applications. However, in order to store the broad range of associations using FCN, large fuzzy rule databases have to be built. In this work, the FCNs with functional weights are introduced and their use in pattern recognition and time series prediction is proposed. The new scheme keeps the nice convergence properties of FCN but is alleviated from the memory and computational requirements of using large fuzzy rule databases, as well as from the inevitable human intervention. The training of the classifier is performed by using a combination of a gradient descent like procedure which uses either a linear or a bilinear parametric model of the network and a least squares method for the estimation of the functional weights. The efficiency and reliability of the proposed classifier is supported by its high overall performance on a set of publicly available time series and pattern recognition datasets outperforming other well-known machine learning models, as well as the most efficient FCM based classifiers.