REINFORCEMENT LEARNING OR TRACKING OF INPUT-OUTPUT MAPS

Three approaches are discussed for learning a slowly time-varying and smooth input-output map in a real-world environment: query smoothing, update smoothing, and radial basis functions. The main restrictions in a real-world environment are (1) some points of the map are more frequently updated than other points, and (2) the updating information may be noisy. A new algorithm is presented called update smoothing with linear interpolation. This learning algorithm allows immediate response (flash map) to a query of the map and is also very fast in updating the map. The advantages and limitations of the algorithm are analyzed. The presented numerical example and discussion of industrial applications show the relevance of these self-learning flash maps for the industry.