To meet the actual traffic demand, this work applies machine learning-based flexible payload power resource-allocation for non-orthogonal SATCOM. Specifically, a tailored deep neural network (DNN) architecture with a customized loss function is trained to intelligently allocate payload power resources among both the beams and users, by learning the undercover structure of its input (i.e., unsupervised learning). Since the DNN-based scheme doesn't need signaling and real-time information exchange between the gateways and the users, it can significantly decrease the implementation complexity by employing the channel statistics of users in multibeam SATCOM. Moreover, the DNN-based scheme can be trained as a universal approximator of the payload power resource-allocation agent for any unseen satellite channel and has the potential for a real-time operation with reduced implementation complexity, compared to the mathematical optimization-based scheme. Numerical results show the DNN-based scheme achieves comparable performance.    «

To meet the actual traffic demand, this work applies machine learning-based flexible payload power resource-allocation for non-orthogonal SATCOM. Specifically, a tailored deep neural network (DNN) architecture with a customized loss function is trained to intelligently allocate payload power resources among both the beams and users, by learning the undercover structure of its input (i.e., unsupervised learning). Since the DNN-based scheme doesn't need signaling and real-time information exchange...    »