Emergence of Prediction by Reinforcement Learning Using a Recurrent Neural Network

Figure 1

Object catching task and a recurrent neural network. An agent moves up or down and catches a moving object. The initial direction of motion and velocity of the object are chosen randomly for every episode. The invisibility area is also chosen randomly in the range of $x>3.0\text{.}$$x,y$ coordinates of the object and y coordinate of the agent are input to an Elman-type recurrent neural network. Each input signal represents local information, as shown in Figure 2.