A Self-Organizing Autonomous Prediction System for Controlling Mobile Robots

摘要

No single innovation is likely to precipitate the sudden emergence of a truly general and robust artificial intelligence. Instead, a careful study of the history of efforts to build AI systems and consideration of factors responsible for evolved general intelligence suggest point toward several promising lines of research. In the research reported here we explore the hypothesis that the ability to learn autonomously from patterns of success and failure at making predictions about what will happen next in a perception-action loop may be a necessary condition for robust AI, even if it is not sufficient. We present a novel learning architecture, the Self-Organizing Autonomous Prediction System, which learns to predict the consequences of actions within a perception-action loop. Two simulation studies demonstrate the success of the approach and the value of prediction-based learning as an important feature of highly adaptive intelligence. Future elaborations of this approach are also discussed.