Cumulative Learning Through Intrinsic Reinforcements

摘要

Building artificial agents able to autonomously learn new skills and to easily adapt in different and complex environments is an important goal for robotics and machine learning. We propose that providing reinforcement learning artificial agents with a learning signal that resembles the characteristic of the phasic activations of dopaminergic neurons would be an advancement in the development of more autonomous and versatile systems. In particular, we suggest that the particular composition of such a signal, determined by both extrinsic and intrinsic reinforcements, would be suitable to improve the implementation of cumulative learning in artificial agents. To validate our hypothesis we performed experiments with a simulated robotic system that has to learn different skills to obtain extrinsic rewards. We compare different versions of the system varying the composition of the learning signal and we show that the only system able to reach high performance in the task is the one that implements the learning signal suggested by our hypothesis.