Tino and Sajda (1995)

(http://www.dlsi.ua.es/~mlf/nnafmc/papers/tino95learning.pdf) use a first-order DTRNN which is basically an augmented version of the recurrent error propagation network of Robinson and Fallside (1991), first-order DTRNN (see section 3.2.1), with an extra layer to compute the output, to learn the transduction tasks performed by Mealy machines (see section 2.3.1). The network is trained using an online algorithm similar to RTRL (see section 3.4.1); weights are updated after each symbol presentation (online learning). In addition to being one of the few papers dealing with transduction instead of recognition tasks, it introduces a new FSM extraction method based on Kohonen's self-organizing maps (see section 5.4.3; see also Haykin (1998, 408)).