Spectrum Prediction For Cognitive Radio Based On Echo State Network

The rapid development of wireless communication technology has brought increasing scarcity of spectrum resources. Cognitive Radio (CR) is considered to be the key technology to solve this problem. The core idea of CR is to make the wireless communication devices with capacity of sensing spectrum surroundings, finding available spectrum and using them rationally. The CR can realize dynamic spectrum allocation and spectrum sharing. In order to maximize the spectrum utilization, and minimize the rate of collisions between authorized users (Primary User, PU) and cognitive users (Second User, SU), the spectrum prediction mechanism is introduced into the cognitive radio by professor Acharya in2006. At present, the spectrum predicted methods mainly include Markov chain, time series analysis and neural networks.As a new type of recurrent neural network, echo state network (ESN) has caused wide public concern in academia because of its unique dynamic reservoir (DR) structure as well as simple learning algorithm, and has been successfully used in many fields. Compared with traditional recurrent neural networks, the main features of the ESN include following aspects:random hidden layer (DR) generation mode; simple weights training algorithm; and well short-term memory capacity.This paper firstly study the principle and characteristic of the ESN, and then use the ESN for cognitive radio spectrum prediction problem. The main work is as follows.First, the principle and characteristic of the ESN is studied in detail. Mathematical model and training algorithm of the ESN is introduced, and its short-term memory capacity is explained from a mathematical perspective. Experimental study on the activation functions of the DR neurons impact on network performance, and the prediction performance of the ESN, the classical feed-forward neural network and the traditional recurrent neural network is compared. Second, convenient for hardware implementation, a new type of ESN that different from the randomly generated pattern of dynamic reservoir in classical ESN is constructed. The proposed ESN is constructed by the reservoir with deterministic feedback connection, which could reduce the parameters required for reservoir optimization design, and its performance is verified by simulation experiments on benchmark problems.Third, the ESN is used for wireless spectrum prediction. Experiment is designed to predict the duration of the spectrum occupied by authorized users in the CR systems, the prediction performance of the classical ESN, the proposed ESN and the traditional recurrent neural network is compared. Experimental results show that, no matter from prediction accuracy but also from the training speed, the classical ESN, and the proposed ESN are better than the traditional recurrent neural networks.