Unsupervised learning of control surfaces based on B-spline models

Based on our earlier work on construction of fuzzy controllers with B-spline models, we propose an automatical learning approach for generating control vertices of such a type of fuzzy controller. For supervised learning, we point out that rapid convergence of this learning procedure can be guaranteed, which is confirmed by diverse examples of approximating non-linear functions and interpolating training data. For unsupervised learning, we employ a type of state evaluation functions which can be found for a large amount of control problems. Using such an evaluation function, a learning algorithm is devised which modifies the local control action efficiently to guide the system to the desired state. Implementations with the cart-pole balancing and a sensor-based mobile robot validate this learning approach.