Speeding up top-down attention control learning by using full observation knowledge

We present a general mathematical description of the top-down attention control problem. Three important components are identified in the model: context extraction, attention focus and decision making. The context gives a coarse blurry representation of the whole input; the attention module models the focus of attention on a limited part of input, and the decision making component accounts the final decision of the agent for its motory actions. In order to achieve a faster convergence of attention learning in the online phase, an offline optimization step is performed in advance. To do so, we incorporate the knowledge of a full observer agent that has approximately learned the optimal decision making of the task. The simulation results show that by employing our algorithm, the learning speed is improved.