Abstract. The paper introduces a reinforcement learning-based methodology for performance improvement of Intelligent Controllers. The translation interfaces of a 3-level Hierarchical Goal-Directed Intelligent Machine (HGDIM) are modeled by a 2-stage Hierarchical Learning Stochastic Automaton (HLSA). The decision probabilities at the two stages are recursively updated from the success and failure signals received by the bottom stage whenever a primitive action of the HGDIM is applied to the environment where the machine operates. The top translation stage and the use of regular stochastic grammars to accomplish the translation of commands into tasks are described here. Under this framework, subsets of con icting grammar productions represent di erent task strategies to accomplish a command. At this stage, an LSA is associated to each subset of con icting grammar productions. Results of simulations show the application of the methodology to an Intelligent Robotic System.
Pedro U. Lima, George N. Saridis