Genetic Algorithm Optimized Training for Neural Network Spectrum Prediction

Spectrum prediction forecasts future channel status based on history data, which partly solves the problem of robustness and reliability in spectrum sensing. A genetic algorithm optimized back propagation (GA-BP) training has been proposed to solve the problem that the neural network based spectrum prediction model always trapped in local optimal solution. Selection, crossover and mutation are performed to increase the randomness, which ensures the population converge to the set that contains the global optimal solution. Then the model continuously performs local searching with back propagation (BP) training. Simulation results show that the performance of GA-BP training outperforms BP training, and SU should choose training method according to his own requirements. The improvement of prediction accuracy will promote the application of spectrum prediction in cognitive radio networks, and maybe helpful to solve the problem in robustness and reliability of spectrum sensing.