A Grid-Based Competitive Mixture of Experts Approach for Partition Optimization of the Multidimensional Magnitude-Selective Affine-Function Behavioral Model

In this article, a grid-based competitive mixture of experts (MoE) approach is proposed for partition optimization of the multidimensional magnitude-selective affine-function (MD-MSA) behavioral model in multifrequency transmitters. The core operation of the MD-MSA model is to partition the space spanned by the amplitude of input signals and then fit the output signals of power amplifiers (PAs) in each partition. However, little attention is paid to how to perform optimal partition of the amplitude space, which is usually an empirical uniform partitioning. Fortunately, the MoE approach has been attempted to perform soft segmentation on the piecewise models and has achieved excellent success. Inspired by this work, we first propose a MoE-MD-MSA model, in which a probability model is adopted to model the contribution of each partition, and then the partitions are adjusted through posterior probability. Furthermore, facing high complexity problem caused by each expert working simultaneously in the MoE model, a competitive MoE-MD-MSA (CMoE-MD-MSA) model is proposed to reduce complexity by introducing a competition mechanism, which ensures that only the most valuable expert works at each moment. Finally, to further reduce the complexity, a grid-based CMoE-MD-MSA (GCMoE-MD-MSA) model is proposed, and it avoids real-time probability calculation by dividing the amplitude space into grids and prestoring partition information. Numerous experimental results have revealed that the optimized partition can effectively improve the performance of the MD-MSA model when a Doherty (DHT) PA is excited by dual-band and tri-band signals, respectively.