Problems of Further Development of GMDH Algorithms: Part 2
Alexey G. Ivakhnenko, G. A. Ivakhnenko, Evgeniya A. Savchenko, Donald C. Wunsch · 2002
To increase accuracy in interpolation problems of artificial intelligence (pattern recognition, depen- dence detection, object identification, stepwise forecasting of random processes, etc.), the inductive algorithms are integrated in more extensive algorithms which consist of several gradually complicated stages of searching for the output value optimization. At the first stage, a simple threshold analysis of the efficiency of the input variables is performed. At the final stage, the twofold and threefold multirow neural networks with active neu- rons are self-organized. Geometrically, the steps of complication can be represented as a gradient descent along the axis of accuracy. Each step of descent should increase the accuracy of problem solving, which is controlled by the depth of the minimum of the external accuracy criterion. At each step of descent, the main problems that need to be further developed and investigated are considered. Beginning with the second step, each step of descent can include both the search GMDH algorithms and the search for the analogs from the history accord- ing to some external criterion.