State automata extraction from recurrent neural nets using k-means and fuzzy clustering

This paper presents the use of a recurrent neural network to learn the dynamical behavior of the inverted pendulum and from this network to extract a finite state automata. Two clustering methods are compared for the automata extraction: the K-means method, and the construction of fuzzy membership functions. It is shown that the number of states for the fuzzy clustering method induces much less states than the K-means method.

1. Introduction

The generic problem of the extraction of knowledge from neural nets is treated by the hybrid systems. These systems are based on neural nets, and allow a symbolic interpretation of its components or interact with symbolic components [22]. Some of these systems were investigated in,[1], [5], [6] and [9].