4 October 2021 - 14h00
Learning Automata over Large Alphabets as an Alternative to Recurrent Neural Networks
by Hadi DAYEKH from VERIMAG, UGA
Abstract: We present an attempt toward learning automata and Moore machines to model Recurrent Neural Networks with real input and output. Our algorithm is based on a combination of the Angluin's L*-variant algorithm of learning Mealy machines, as well as an improved version of Irini Mens' and Oded Maler's algorithm for learning symbolic automata defined over a large input alphabet such as subsets of R. The algorithm was tested on an RNN used in control systems. This work is part of an M1 internship carried at VERIMAG in spring-summer 2020, under the supervision of Nicolas Basset and Thao Dang.