Curiosity and Boredom Based on Prediction Error as Novel Internal Rewards

Abstract

In this paper, the use of two internal reward models, curiosity and boredom, is proposed. Experiments on a maze navigation task demonstrated that appropriate values of parameters simultaneously improved the performance of the predictor of the environment and increase the external rewards compared with the conventional reinforcement learning. In conclusions, the relation between the proposed method and active learning, diversive curiosity, and specific curiosity is also discussed.