A New Kernel-Based Algorithm for Cluster Sparse System Identification

In this paper, to increase the performance of cluster sparse proportionate affine projection algorithm (CSPAPA), kernel extreme learning machine-based CSPAPA (KELM-CSPAPA) is proposed. An adaptive density-based clustering algorithm (ADBC) is used to decompose the training dataset into multiple subsets. Furthermore, the improved flower pollination algorithm (IFPA) is used to optimize the parameters of the KELM model. Initially, the performance of the ADBC algorithm is studied to estimate the optimal number of clusters using three different inputs. Then the KELM model is assessed with different machine learning models using various statistical indices. The time consumption of KELM for the colored noise and speech signal inputs are less compared to the white Gaussian noise input. Finally, a performance comparison of KELM-CSPAPA is presented with three benchmark algorithms: PAPA, block sparse PAPA (BS-PAPA), and CS-PAPA. The result shows that the proposed method outperformed existing algorithms. The simulation results demonstrate that KELM-CSPAPA fully exploits the sparsity and handled the clustered-sparse signal.