Discrete Memoryless Channel Capacity Estimation using Deep Learning and 256QAM

Abstract Channel capacity should be accurately estimated to design effective communication systems particularly in high-order modulation settings. In this paper, a deep learning-trained structure of the capacity estimation of discrete memoryless channels (DMCs) in the context of 256-Quadrature Amplitude Modulation (256QAM) is introduced. The presented approach is built upon Deep neural networks (DNNs) to train capacity directly based on input-output information, which is contrary to the classical methods like the Blahut-Arimoto algorithm which are both computationally intensive and depend on channel complexity. The simulation findings indicate that the CNN model has the capacity estimates that are close to the theoretical ones with low root mean square error (RMSE) and mean absolute error (MAE) at different signal-to-noise ratios (SNRs). The DNN-based capacity estimator takes 0.043 s (76% fast) to execute a test sample, which is much lower than the average execution time of the Blahut Arimoto algorithm of 0.184 s per SNR point, showing that the proposed approach is computationally efficient. The model is also more scalable and converges faster and hence can be used in real time applications. The article fits into the intersectional area of information theory and artificial intelligence, providing a viable and scalable capacity estimation methodology in next-generation wireless networks.