AN ALGORITHM TO DETERMINE THE FEASIBILITIES AND WEIGHTS OF 2-LAYER PERCEPTRONS FOR PARTITIONING AND CLASSIFICATION

Necessary and sufficient conditions for implementing classification pr oblems by two-layer perceptrons have been presented in the viewpoints of mathematics and geometry. [Zwietering, Arts and Wessels, Int. J. Ne ural Systems 3(2), 143-156 (1992)]. This paper, in an engineering view point, provides an algorithm, called the Weight Deletion/Selection Alg orithm, to examine the feasibility of implementation of a decision reg ion by two-layer perceptrons, and to select the weights of the second layer in a two-layer perceptron without any training process if a deci sion region is implementable for two-layer perceptrons. For the purpos e of visualization, we explain the algorithm by the examples of two-in put-two-class cases, and then discuss the generalization to multi-inpu t-multi-class cases. Finally we present a three-class classification e xample with four features selected as the inputs. (C) 1998 Published b y Elsevier Science Ltd on behalf of the Pattern Recognition Society. A ll rights reserved.