Improved blind deconvolution scheme for system identification

This paper proposes an improved blind signal deconvolution (BSD) scheme that fits into system identification for control. An existing BSD scheme based on FIR (finite impulse response) approximation is effective but requires a large number of taps since the mixer (system to be identified) has finite invariant zeros. In contrast, the paper represents the demixer (inverse system) via polynomial matrix fraction and computes its gradient with respect to mutual information, thereby deriving a learning law with a given modest number of parameters. Numerical simulation is carried out to verify the effecitveness of the proposed scheme.