A learning method for recurrent networks based on minimization of states of finite state machines
Itsuki Noda · Systems and Computers in Japan · 1995
Abstract A new learning method for simple recurrent networks (SRN) is proposed. The correspondence between SRNs and finite stage machines (FSM) is examined and then a learning method for FSMs is constructed based on the state‐minimization technique of FSMs. This method comprises three stages: (1) generating states according to input histories; (2) grouping states according to outputs and next states; and (3) combining states within the same groups and reconstructing state transitions. Then, three networks, each behaving in the same manner as each stage through learning, are considered. A unified model of these networks and its learning method are proposed. The effectiveness of the proposed model and the learning method are verified by experimental results.