Networks that learn to predict where the food is and also to eat it

Summary form only given, as follows. Neural networks with recurrent memory units were assumed to be contained in small ´animals´ moving in a 2-D world with ´food´ elements. The networks were first taught to predict the next angle and distance of food given the present angle and distance and the animal´s current move as input. These predicting networks were then taught to approach food, i.e. to select a sequence of moves that was more likely to bring the animal to a food element. The prior learning of a prediction ability greatly increases what the network learns during the successive teaching of food-approaching behavior in comparison with networks that learn to approach food without prior prediction learning. Furthermore, the prediction ability, after learning to approach food, is still present, and it is even enhanced by the intervening learning of another task. The results are discussed in terms of how a map of the environment, as evidenced by the prediction ability, can be useful in acting efficiently in the environment.<>