Concurrent learning of task and attention control in the decision space

摘要

Learning attention control is a real need specifically when a robot tries to learn a sequential decision-making-type task. This is even more critical when learning directly in the perceptual space is not feasible mainly due to the high dimensionality thus non-homogeneity. Therefore, two learning problems are raised to be solved at the same time. In this paper, a novel approach with three learning phases is proposed to facilitate learning of these two coupled problems: 1) learning how to divide attention among multiple dimensions of robots perceptual space and also how to shift it efficiently inside one modality from one spatial part to another and 2) learning the main task. The main task is considered “driving in a simulated road using a miniature mobile robot” in order to demonstrate the necessity of attention control. An important new feature of the proposed learning method is that the attention is learned in the decision space rather than the original perceptual space and this brings some discussed advantages. Obtained results justify practicability and usefulness of learning attention control in the proposed alternate space.