Neural Learning Schemes Applied in Behavior-Based Robotics
Description: The degree of autonomy of an agent is related to the ability to decide by itself how to relate the data of the sensors with the commands to the actuators in their efforts to reach the objectives for which it was designed. In this way, the agent’s ability to learn and adapt is closely related to his or her degree of autonomy. Within the paradigm of biological inspiration adopted in the PiramidNet control architecture, Artificial Neural Networks are the tools used to implement the intelligence and control of robotic systems. However, mechanisms capable of making the learning permanent and in time of operation in robotic systems controlled by Neural Networks are scarce or are still in the initial phase of development. In this work, we intend to study, propose and implement methods that enable the real-time learning of Intelligent Mobile Robots controlled by Artificial Neural Networks. For this, a neural control architecture was proposed, capable of presenting adequate plasticity and stability characteristics, using ART-Adaptive Ressonance Theory networks and MLP-Multi-Layer Perceptron networks, associated with a reinforcement learning scheme as a methodology for real-time learning. The main goal of this work was to implement an intelligent, autonomous and adaptive agent capable of feeling and reacting to changes in the environment and learning from its experience through real-time reinforcement learning methods. In order to achieve this, there were still other objectives to be achieved: In-depth study of the learning methodology by reinforcement, types of algorithms and problem verification; In-depth study on the implementation of reinforcement learning in artificial neural networks; Study and development of artificial neural network architectures capable of implementing behaviors and being adapted by the methods studied; Implementation of architecture.
Status: Completed.
Nature: Research.
Students involved: Academic Master’s: (1).
Members: Mauro Roisenberg – Coordinator / Glaucio Adriano Fontana – Member / Thiago Henrique da Silva – Member / Marcelo de Souza – Member.
Years: 2002-2003.
