Quantised kernel least mean square with desired signal smoothing. Issue 18 (1st September 2015)

• Record Type: Journal Article
• Authors: Xu, Xiguang; Qu, Hua; Zhao, Jihong; Yang, Xiaohan; Chen, Badong
• Is Part Of: Electronics letters. Volume 51:Issue 18(2015)
• Page Start: 1457
• Page End: 1459
• Publication Date: 2015-09-01
• Subjects: smoothing methods -- least mean squares methods -- learning (artificial intelligence) -- Monte Carlo methods -- signal processing; quantised kernel least mean square -- desired signal smoothing -- QKLMS -- online learning algorithm -- input space compression -- Monte Carlo simulations; Electronics -- Periodicals; 621.381
• DOI: 10.1049/el.2015.1757
• Languages: English
• ISSNs: 0013-5194

Abstract

Abstract : The quantised kernel least mean square (QKLMS) is a simple yet efficient online learning algorithm, which reduces the computational cost significantly by quantising the input space to constrain the growth of network size. The QKLMS considers only the input space compression and assumes that the desired outputs of the quantised data are equal to those of the closest centres. In many cases, however, the outputs in a neighbourhood may have big differences, especially when the underlying system is disturbed by impulsive noises. Such fluctuation in desired outputs may seriously deteriorate the learning performance. To address this issue, a simple online method is proposed to smooth the desired signal within a neighbourhood corresponding to a quantisation region. The resulting algorithm is referred to as the QKLMS with desired signal smoothing. The desirable performance of the new algorithm is confirmed by Monte Carlo simulations.