No abstract

No abstract

This paper proposes a novel technique to approximate functions for time series and signal processing using the special type of neural network, called Critical Support Vector Machine (CSVM). CSVM is a combination of the Support Vector Machine, the Nearest Neighbor Algorithm and the Perceptron. The CSVM has been shown to be an effective method for classification problems. In this work, we generalize CSVM so that it can be used for tbe application of time series prediction. The experiment on the chaotic MackeyGlass time series significantly verifies the performance of our algorithm.

The drawback of SVM technique leads to a positive semidefinite quadratic programming problem with a dense, structured, positive semi-definite matrix, and also requires a set of kernel functions. We propose the learning algorithms that do not need any kemel functions. The separability is based on the critical Support vectors (CSV) essential to determine the locations of all separating hyperplanes. The algorithms give better performance compared with the other proposed SVM-based algorithms when they are tested with 2 spirals problem, Sonar, Ionosphere, Mushroom, Liver disorder, Cleveland heart, Pima diabetes, Tic Tac Toe, and Votes.

No abstract