Issue No.02 - April (1996 vol.11)
DOI Bookmark: http://doi.ieeecomputersociety.org/10.1109/64.491280
Learning is a research topic often pursued by researchers in Robotics, as results in this area may significantly reduce the time required to teach complex tasks to a robot. This paper introduces an approach to the reinforcement learning of robotic tasks. Besides leading to learning algorithms with no special complexity, reinforcement learning schemes require underlying performance measures applicable to a wide category of algorithms usually implemented in robotic systems. One such measure, balancing the reliability and computational cost of an algorithm, is introduced here. The measure is applied to learn the best among a set of alternative tasks capable of executing a command communicated to an intelligent machine. Furthermore, the measure definition reduces the significance of the learning scheme slow convergence. We assume that the alternative tasks are pre-defined by an expert, thus enhancing the fact that a priori knowledge of the steps composing a robotic task is often relevant, even though other learning approaches minimize its use.
Pedro U. Lima, George N. Saridis, "Learning Optimal Robotic Tasks", IEEE Intelligent Systems, vol.11, no. 2, pp. 38-45, April 1996, doi:10.1109/64.491280