Acquisition of active perception and recognition through Actor-Q learning using a movable camera

In the previous work, it was shown in simulation that by employing Actor-Q learning, active perception and recognition problem can be learned. In this learning, a camera with non-uniform resolution was assumed. Appropriate camera motions and recognition timing as well as the correct recognition were acquired through the reinforcement learning using the final recognition results as rewards and punishments. In this paper, the authors tried to verify using a real camera whether appropriate continuous camera motions and identification of two patterns can be acquired simultaneously through Actor-Q learning. Here, non-uniform resolution was also applied. As a result, within a limited number of trials, these functions cannot be achieved. Then, three stages of learning are introduced. The system was trained on “camera motion” task at first and then trained on “pattern recognition” task. After that the system was trained on both “pattern recognition” and “camera motion” simultaneously. It was verified that appropriate camera motion, recognition and recognition timing can be learned only from the reward and punishment through Actor-Q learning.