Behavior implementation in autonomous agents using modular and hierarchical neural networks

This paper describes the development of a modular and hierarchical Artificial Neural Network (ANN) control architecture that is capable to implement behavior in Autonomous Agents (AAs). We make considerations about biological paradigms, as evolutionary mechanisms and animals' behaviors, trying to find solutions that, once applied to the development of artificial devices, provide more robust and useful autonomous agents to operate in the real world. This work investigates the relations between structure and function in both artificial and natural neural networks, and how increasingly complex behaviors can be achieved through the interaction of these neural structures, from the simple reflexive behavior to the most complex behaviors that need mapping and planning capabilities. The paper also proposes a special module for conversion of the inputs of the sensorial and control networks into propositional symbols to be processed at the highest level of the architecture, the symbolic level (in development).