Redes de Petri coloridas para representação do conhecimento e aprendizado de redes neurais artificiais.

Abstract

Artificial neural networks (ANNs) comprise parallel and distributed computational tools that can learn from data and make inferences (i.e., predictions) for highly nonlinear systems. By its turn, Petri nets (PNs) consist of well established modeling tools for parallel and distributed discrete event systems with a number of successful contributions to automation and control of complex industrial tasks. Thus, motivated by a long lasting interest in using the formalism of PNs either to emulate or to design neural network architectures, we revisit this research topic by resorting to the powerful modeling framework of hierarchical timed colored PNs (HTCPNs) to introduce a novel approach that builds a fully adaptive one-hidden-layered multilayer perceptron (MLP) model trained by the famed backpropagation algorithm. The resulting proposed model is called HTCPNMLP and consists of a general structure capable of handling classification and regression tasks. In order to develop the HTCPN-MLP model, a perceptron-like colored PN (perceptron-CPN) model is first built upon a novel McCulloch-Pitts colored PN (McCulloch-Pitts-CPN) model of a neuron, this being another contribution of the current work. A pedagogical set of experiments is presented in order to highlight the learning capability of the proposed HTPCN-MLP model and its advantages with respect to alternative models available in the literature.