Complex-valued neural network using magnitude encoding technique for real-valued classification problems & time series prediction

Abstract

In this paper a new conversion technique is proposed for complex-valued neuron (CVN) to convert real value into complex value in order to solve real-valued classification problems & Time series analysis. Previously phase encoding system was used to solve these types of problems. In this proposed encoding system, each real-valued input is converted into complex value according to the input real value with a fixed phase. In this model the input magnitude ranges from the lowest and highest value of the given input. The converted value is then multiplied by complex weight and then they sums up to feed into an activation function. The activation function converts the complex value into real value within a certain range. We used this encoding system in solving different Boolean problems. Some real world benchmark problems are also tested by this process. Different time series analysis is performed to test the prediction ability of this encoding system. The result shows that this magnitude encoding provides better accuracy in different benchmark problems. Especially in case of predicting time series data, this encoding system provides better result than the phase encoding system.