Chapter 8. Understanding and Explaining Drn

In this chapter we discuss how symbolic knowledge in the form of DFAs can be extracted from trained DRNs. We give an overview of various methods that have been proposed for DFA extraction and give a in-depth discussion of two methods that use fundamentally diierent approaches: Dynamic space exploration uses a cluster analysis for identifying DFA states and transitions whereas a proposed machine learning approach only uses the obserations of a trained network's input/output behavior for inferring DFAs.