Inversion of a nonlinear communication channel by kriging
J.-P. Costa, Luc Pronzato, E. Thierry · 2003
We consider a problem of nonlinear system-inversion, with unknown underlying model structure. Classical parametric behavioural models (Volterra and NARMAX models) involve a lot of parameters, which are difficult to estimate from short training sequences. A similar difficulty is encountered when methods based on neural networks are used. We suggest in this paper to use a semi-parametric approach called kriging. We show on an example that good performances are obtained for short training sequences.