Abstract
The aim is to perform to realize a learning system for active perception using a neural network. It obtains inputs only from a movable visual sensor and learns both appropriate recognition and sensor motions for effective perception. The proposed learning method is based on reinforcement learning using reinforcement signals calculated from the recognition results. We conclude from simulations that it enables the system to move a visual sensor to the appropriate location and finally classify presented patterns correctly. The recognition rate was better than that in a simulation for comparison where the sensor was fixed at the initial location.
