Direct-Vision-Based Reinforcement Learning Using a Layered Neural Network

摘要

In this paper, Direct-Vision-Based Reinforcement Learning is proposed not only for the learning of motion but for the learning of the whole process, which includes recognition, from sensors to motors in robots. In this learning, raw visual sensory signals are put into a layered neural network directly, and the network is trained by Back Propagation using the training signal generated based on reinforcement learning. By employing neural network, whole the process becomes seamless and is trained purposively and harmonically.Two simulations of the mobile robot with visual sensors are performed as examples. One task requires stereo-vision, and the other requires obstacle avoidance. By observing the hidden representation in the neural network, it is shown that some abstract representation of spatial recognition is formed without any advance knowledge.Finally it is shown that each visual sensory cell makes a role of localization of the global information of the space and it helps the fast and stable learning.