Monotonic Normalized Radial Basis Function Networks

Abstract

In the paper we address the problem of deriving monotonicity conditions for normalized radial basis function networks. For general shape of the kernels the necessary conditions are expressed as trivial inequalities imposed on the kernel weights together with set of linear inequalities on elements of matrices describing the kernels. If the shape is considered to be the same for all kernels the conditions become simple and intuitive. Two examples are given to demonstrate benefit of incorporation of information about monotonicity.