Comparison of deterministic and fuzzy finite automata extraction methods from Jordan networks

This paper compares two methods for the extraction of finite state automata from recurrent neural networks (RNNs). Neural networks store the knowledge implicit in the data in their weights, but do not provide an easy explanation of this knowledge to the user. This is a difficult task due to the spatial (distributed information in the network) and temporal (network states) relations built by the network among the data. One form to present the knowledge stored inside a RNN is using finite state automata, which shows explicitly the relations among the variables and their temporal causality. In this paper, we treat the nonlinear dynamical system inverted pendulum and controller and compare the performance of the extraction algorithm using two clustering methods: k-means and fuzzy clustering in terms of exactness and knowledge conciseness.