Multipath Component Clustering Based on Improved Self-organizing Feature Map for 5G Millimeter Wave Radio Channels

Abstract

Recently, clustering algorithm has become an important research hotspot, as cluster based wireless channel modeling can reduce the complexity of multiple-input multiple-output (MIMO) channel models. In this work, a clustering algorithm based on maximum-minimum distance algorithm (MMD) assisted Self-organizing Feature Map (SOM) is proposed, namely, MMD-SOM. Specifically, MMD algorithm is used to initialize the weight of the network, which solves the problem of inadequate network training in traditional SOM caused by randomly setting the initial weights. Furthermore, an output layer is added to the network, which gets over that SOM is easy to be over-trained in the unsupervised situation. The performance of the improved algorithm is evaluated based on the measured and simulated channel data, both numerical simulations and experimental clustering results are provided to demonstrate the effectiveness and robustness of the proposed algorithm.