Real time input subset selection for linear time-variant MIMO systems

In this paper we propose an approach for multi-input multi-output (MIMO) system identification when the statistical relationship between input and output varies in input space as well as in time; i.e. nonstationary in space and time. An on-line variable selection algorithm, which has been recently developed for selecting a subset of input variables in real time by modifying least angle regression (LAR) with recursive estimators, is extensively applied to the linear time-variant MIMO systems. In our approach, a subset of input channels relevant with output is selected at every time instance based on the correlation between the filtering outcome of individual input channels and desired output. The on-line variable selection algorithm performs channel selection with weights using this real-time correlation. The proposed model is compared with a typical linear model in which only the least mean squares (LMS) is used to update system parameters. Tracking performances of these two models are demonstrated in a computer simulation and in a real-world application for tracking a linear relationship between neural firing rates of a primate and synchronously recorded hand kinematics. In both cases, our model demonstrates superior tracking performance.