Comparing Learning Attention Control in Perceptual and Decision Space

The first question answered in this paper is whether or not learning attention control in the decision space is feasible and how to develop an online as well as interactive learning approach for such control in this space, in case of feasibility. Here, decision space is formed by the decision vector of the agents each has allowed to dynamically observe just a subset of all available sensors. Attention control in this new space means active and dynamic selection of these decision agents to contribute in making final decision. The second debate is verifying the advantages of attention control in decision space over that in perceptual space. According to the tight coupling of attention control and motor action selection, in order to answer above mentioned questions, attention control and motor action selection are formulated in a unified optimization problem and reinforcement learning is utilized to solve it. In addition to the theoretic comparison of learning attention control in perceptual and decision space in terms of computational complexity, two proposed approaches are tested on a simple traffic sign recognition task.