Spectrum Prediction Based on Taguchi Method in Deep Learning With Long Short-Term Memory

Abstract

Spectrum prediction is a promising technology in cognitive radio networks, since it can reduce considerable time and energy consumed in spectrum sensing process. Many spectrum prediction algorithms have achieved good performance, but majority of them with shallow architecture cannot capture the inherent correlations of spectrum data very well. Long short-term memory (LSTM) neural network in deep learning has been validated to have strong capability of solving time series problems. In this paper, we develop a spectrum prediction framework with a deep learning approach on two real-world spectrum datasets. For the first dataset to predict channel occupancy states, we firstly employ the taguchi method to determine the best optimized configuration of neural network for certain spectrum point and then analyze the effect of each design hyper-parameter. Next, we build LSTM neural networks with two perspectives of regression and classification for spectrum prediction. For the second dataset to predict channel quality, we compare the prediction performance of the LSTM neural network and conventional multilayer perceptron (MLP) neural network. For both of our datasets, results show that the prediction performance varies with frequency bands. From the point of statistics, the LSTM neural network has better prediction performance than the MLP neural network and is more stable as well. Furthermore, we find that the performance of the LSTM neural network with classification perspective is slightly better than that with regression perspective in our first dataset.