The thesis proposes a new methodology for identification of a DES (Discrete Event System) using a HRNN (Hybrid Recurrent Neural Network). The identification of an unknown DES is recognition of characteristic functions of a DEVS (Discrete EVent systems Specification) model which validly represents the system.
A DES is a useful abstraction for behavior modeling of various man-made systems, such as communication networks, manufacturing systems, traffic systems and others, at a high level. A DES, besides the only concern of the UDES (Untimed Discrete Event System) abstraction is the logical sequence of events, also has considerations of its temporal behavior. The DEVS formalism specifies such a DES system in a hierarchical and modular manner. There are many researches and results for UDES identification problems but little research has been reported concerning the identification of the timed DES. We extend the sight of the identification problem limited to the field of UDES to the DES abstraction using neural networks.
Such identification consists of two major steps: behavior learning using a specially designed neural network called HRNN and extraction of a DEVS model from HRNN which is trained using observed input/output events of an unknown DES.
We designed a HRNN architecture so that a general structure preservation is con-served between a DEVS model being identified and the HRNN architecture. The HRNN has an interconnected structure of two classes of neural networks and a timer working as a spontaneous event scheduler. The BPTW (BackPropagation through Time with Working set) algorithm was used to train the HRNN. The algorithm employs a concept of dynamic training set and is appropriate for logical training. Algorithms for DEVS model exploration from a trained HRNN and for minimization of the explored DEVS model were derived.
Identification experiments were performed with three types of unknown DESs, the results of which verified the validity of the proposed mod...