On the Performance of Autoencoder-Based Space Optical Communications

In this paper, we propose a deep learning autoencoder (AE) to model and design space optical communications (SOC) systems from end-to-end performance. The proposed AE is based on multiple-decoders and a new layered structure for constructing both encoders and decoders. The use of multiple-decoders can increase the receiver diversity, which allows the gradient descent to minimize the cost function compared to state-of-the-art models and therefore improve the bit error rate (BER) performance. In addition, we utilize system tool kit (STK) simulator for a realistic SOC channel modeling in Log-normal fading and additive white Gaussian channel. Numerical results reveal that the traditional convolution codes-based SOC systems are outperformed by the proposed learning-based AE for various code rates. Moreover, the proposed AE outperforms the benchmark AE learning-based models in terms of BER performance metric.