Learning and Adaptation Among the Machines: The Biologically-Inspired Path to Intelligent Robots
Collegium de Lyon, 5 parvis René-Descartes, 69007 Lyon, salle R 143
Directly programming a robot to perform some specific task, such as avoiding obstacles while navigating in the environment, can be a difficult challenge for a human designer. And the difficulty only increases as the tasks become more sophisticated and open-ended. An alternative approach is for the robot to learn how to perform the task on its own, either guided by training examples provided by a teacher, or autonomously through its own experiences. Inspired by a variety of biological structures and processes capable of learning, adaptation, and self-organization, the field of machine learning has made exciting progress in recent years toward the long-term goal of creating fully autonomous, intelligent robots (although there is still far to go). In this talk, I will give an overview of several topics in machine learning, including artificial neural networks, which are computer programs loosely modeled on the structure of the human brain, and genetic algorithms, in which a population of programs is randomly generated and allowed to "evolve" over time, in a manner similar to biological evolution.
I will also describe the newly emerging field of developmental robotics (also called epigenetic robotics), which lies at the intersection of machine learning, robotics, and developmental psychology. A key challenge of developmental robotics is to create systems capable of self-driven, open-ended learning, in which the motivation for learning a particular task or behavior emerges from within the system itself, rather than being imposed on the system from the outside by the human designer. Recent work by myself and several colleagues has explored an approach to creating intrinsically-motivated systems by combining neural networks with a form of artificial curiosity. Currently we are working on developing a new algorithm that enables a robot to learn higher-level patterns of behavior in a self-motivated way, by recording in the robot's memory hierarchical sequences of basic experiences (or "interactions") between the robot and its environment. This algorithm has been successfully tested in a simple simulated environment, but we are currently attempting to implement it on a physical robot operating in a real environment.