Piecewise Multivariate Polynomials Using a Four-Layer Perceptron

This paper proposes a new method for discovering piecewise polynomials to fit multivariate data containing numeric and nominal variables. Each polynomial is accompanied with the corresponding nominal condition stating a subspace where the polynomial is applied. Such a nominally conditioned polynomial is called a rule. A set of such rules can be represented as a single numeric function, which can be approximated by a four-layer perceptron. The method selects the best from those trained perceptrons, and finds the final rules from the best perceptron. In our experiments the proposed method finds the original rules for an artificial data set, and discovers succinct rules for a real data set.