Reinforcement learning when visual sensory signals are directly given as inputs

It is shown that a neural-network based learning system, which obtains visual signals as inputs directly from visual sensors, can modify its outputs by reinforcement learning. Even if each visual cell covered only a local receptive field, the learning system could integrate these visual signals and obtain a smooth evaluation function. It also represented the spatial information smoothly in the hidden layer through the learning, and the area of the state which seemed important for the system was magnified in the hidden neurons´ space. The learning is so adaptive that when a different motion characteristic was employed in the system, the representation became different from the previous one, even if the environment was the same