Predictive Category Learning in Mobile Robots

Josh de Leeuw, Vassar College ’08, Jacqueline Kory, Vassar College ’11 and Prof. Ken Livingston

The Predictive Category Learner (PCL) allows an autonomous robot to learn how to predict the events it will encounter, thus enabling the development of the skill needed to navigate successfully. The model was tested in simulations built using Microsoft Robotics Studio (MSRS) using two distinct approaches. The first implemented a previous version of the model in a different robot to establish generalizability of the procedures across platforms. Performance was remarkably similar for the two different robots. The model has now been ported to a real version of the robot for further testing. The second approach used a genetic algorithm to evolve variations on the basic architecture as a way of exploring its optimal configuration for use in a much more complex environment than those used in previous simulations. Performance improved across the 15 generations of evolution, converging on a system capable of predicting the results of its actions as it navigates.