Ergodic capacity estimation with deep learning methods in noma based cognitive radio communication system

In this thesis, the total ergodic capacity of the near user in cognitive radio (CR) based non-orthogonal multiple access technique (NOMA) system models are estimated by the proposed feed-forward back propagation artificial neural network and nonlinear autoregressive exogenous model (NARX). The data set used in the neural network is obtained from the CR-NOMA system model, which is modeled with the exponential fading channel characteristic. The dataset entries collected through the CR-NOMA system model discussed consist of the path loss coefficient, power allocation coefficient, SNR, the distance between source-relay-destination, and the ratio of the power value of the second user to the power value of the first user. By entering and training output data into the artificial neural network designed using the supervised learning method, the ergodic capacity of the near user is estimated via test data that the network has never been introduced to before. While evaluating the performance of ANN and NARX, the training time, the number of iterations and the saturation of the network were taken into account. The actual ergodic capacity value of the near user is compared with the predicted values of feed-forward backpropagation ANN and NARX networks. The epoch value of 1000 was taken for the training process of the neural networks. The performance analysis of the proposed neural networks under Levenberg-Marquardt, Bayesian, and Scaled-Conjugate training algorithms, epoch value graph, error histogram analysis, and training state analysis where the error reaches the minimum are examined. For the feed-forward, backpropagation ANN model, the estimation of the ergodic capacity value of the near user was obtained with the Bayesian algorithm with the highest accuracy. In addition, considering the training times, it was concluded that the fastest algorithm is Levenberg-Marquardt, and the slowest algorithm is the Bayesian algorithm. From the simulation results, it is concluded that the ergodic capacity estimation is obtained by using the Levenberg-Marquardt training algorithm with feed-forward backpropagation ANN, considering that the network does not under-fitting overfitting and the algorithm works fast. When the trained system model is tested, total ergodic capacity is estimated at 95,934% accuracy for training data, 95,830% for validation data, and 95,511% for test data. With feedforward, backpropagation ANN, training dataset accuracy was 95.934%, validation dataset accuracy was 95.830, and test dataset accuracy was 95.511%. When the analysis and all the results obtained are examined, it is concluded that the proposed neural network models achieve the desired high accuracy and benefit from increasing the algorithm calculation speed. In addition, it is predicted that it will reduce the computation time for real-time systems by allowing the algorithm complexity to be reduced and provide guidance to researchers for solving many nonlinear communication problems.