Design of Codebooks for Nonorthogonal Multiple Access with Noncoherent Reception: Comparison of Autoencoder-Based Approaches

Three architectures of autoencoder neural networks are presented, intended for optimization of codebooks in an uplink nonorthogonal multiple access (NOMA) system with non-coherent reception. All three networks simulate NOMA system with receiver based on a generalized likelihood ratio test, the differences are in the optimization goal: minimization of the rate of combined errors, of average error rate, or of the individual error rate. The codebooks obtained by different methods are compared in terms of power efficiency and performance spread between users. It is shown that minimization of combined and individual error rates give close results, the former performing better in the case of large total number of bits transmitted by all users, and latter providing better results at high “overload” factor with low bitload of users.