Abstract
This paper describes two automated methods to identify binary and continuous systems from their operating data. Both methods employ data clustering and localized modeling. For binary functions, the Karnaugh map is employed to transform an arbitrarily high dimensional distribution into a two-dimensional one where an elliptical clustering algorithm can be performed to identify a set of local neural net models to approximate the data. For real continuous functions, the input space is partitioned into regions that are small enough that a simple neural network can approximate the data well in each region. The partitioning method is inspired by an automatic mesh generation algorithm for Finite Element Method (FEM). A 7-bit Exclusive-OR (XOR), a Van der Pol oscillator, and a pendulum, respectively, are used to test the two proposed algorithms, and they are found to be satisfactory in generating models that can simulate the systems well.
