Improvement of a Sorting-Out GMDH Algorithm Using Recurrent Estimation of Model Parameters

The paper presents and discusses a GMDH algorithm MULTI-R being an improved and revised version of the known multistage algorithm MULTI based on the successive search of model of the globally optimal structure. This means that the MULTI algorithm is intended for discovering the result of exhaustive search by the combinatorial algorithm COMBI GMDH with radically less computations. But this algorithm has some substantial drawbacks, for example, it tends to choose underfitted models in the searching process and is not optimized with respect to the parameter estimation procedures. The new revised version MULTI-R differs from the original algorithm MULTI by using a recurrent procedure of parameters estimation and additional optimizing the model structure to enhance both the computation speed and accuracy of discovering the globally optimal model. The comparative numerical characteristics of the processing speed and structural accuracy of this modified algorithm and the original one are given for several test tasks.