Computational discovery for nonlinear classifiers

The author describes an approach to computer-aided discovery of nonlinear classifiers and cluster finders for complex distributions of multidimensional data. Special attention is given to the discovery and selection of features for piecewise linear classifiers and large-scale tree classifiers. The feature discovery mechanisms are in the form of networks of three types of elements: feature selectors (FSs), feature expanders (FEs), and trainable linear arrays (TLAs). The FSs are built on selection techniques such as relaxed branch-and-bound and genetic selectors. Each FE produces an advanced set of features from a primitive set, using mating and mutation processes of genetic algorithms in combination with algebraic operations. Each TLA transforms a set of features into another set via a set of trainable linear weights. The output of the network of FSs, FEs and TLAs is a set of discovered features of a parallel nonlinear classifier. The network discovers these features by combining algebraic transformations, hill climbing, and genetic search