A Life-long Learning System based on Multiple Probabilistic Fuzzy Models for Effective Human-Robot Interaction

Human behavior understanding becomes essential in human-friendly human-robot interaction to recognize human intention. However, it is usually difficult to model and handle such interaction due to variability of the user's behavior with inconsistent and time-varying characteristics. In this paper, we shall show the benefits of a PFR (probabilistic fuzzy rule)-based learning system to handle inconsistent/time-varying data pattern in view of combining fuzzy logic, fuzzy clustering, probabilistic reasoning, and adaptation technique in a single system as an effective engineering solution. IFCS (Iterative Fuzzy Clustering with Supervision) learning algorithm has been successfully applied to extract a PFRB (Probabilistic Fuzzy Rule Base) from a given set of inconsistent training examples in short-term memory. In addition, we introduce a PFR-based life-long learning structure with multiple probabilistic fuzzy models for continual adaptation throughout incessant learning and control. A model base construction scheme with adaptation capability, in interim transition memory and long-term memory, improves control performance for continuously drawn time-varying data patterns by constructing and utilizing multiple probabilistic fuzzy models. To show the effectiveness of the proposed system, we introduce successful learning examples with benchmark data patterns and TV viewing data pattern.