Interactions Between Intrinsic and Extrinsic Motivation

In this paper we will discribe a system that motivates robots with a combination of intrinsic and extrinsic sources of motivation. The system is based on reinforcement learning, and uses neural networks to predict the effects of the robot’s actions on the environment and rewarding it for improvement in predictions, while also granting other extrinsic rewards for specific actions. We then test the system in a simulated environment to see whether the combination of different sources of motivation can allow a robot to learn about its environment while also avoiding potentially dangerous actions and keeping its battery charged. We find that robots with extrinsic motivation are able to avoid harm and keep their battery charged successfully, even when intrinsic motivation is also present. However, intrinsically motivated robots are only slightly better at learning about the environment than extrinsically motivated ones, and robots with a combination of both types of motivation do not do any better than those with only extrinsic motivation. Furthermore, robots that choose actions randomly learn about the environment better than those with either intrinsic or extrinsic motivation. From these results, we conclude that the addition of intrinsic motivation does not stop a robot from achieving extrinsic goals, but that flaws in either our system or environment prevent intrinsic motivation from yielding any benefits in the tests performed.