Abstract
A multiinput-multioutput type GMDH algorithm using regression-principal component analysis is described. In conventional GMDH algorithms, estimated values of output variables are used as intermediate variables and partial polynomials are constructed by using these intermediate variables. So, multiinput-multioutput type nonlinear system can not be indentified by using conventional GMDH algorithms because a large number of intermediate variables are generated in each selection layer and GMDH calculations can not be continued. The GMDH algorithm in this paper uses total characteristic variables, which can explain variation of all output variables, and optimal partial polynomials are constructed by combining these total characteristic variables.
