Weekly review of Reinforcement Learning papers #11

Every Monday, I present 4 publications from my research area. Let’s discuss them!

Image by the author

[← Previous review][Next review →]

Paper 1: Reward is enough

Silver, D., Singh, S., Precup, D., & Sutton, R. S. (2021). Reward Is Enough. Artificial Intelligence, 103535.

The hypothesis that is formulated in this paper, is that reward maximization in a sufficiently complex environment, is a sufficient condition for the emergence of intelligence. The example they take in this paper is that of a squirrel, who wants to get as many nuts as possible. To achieve his goal, he must perform a number of subsequent tasks: perceiving his environment, moving around, climbing trees, communicating with other squirrels, understanding the cycles of the seasons… In their hypothesis, all these subsequent tasks will be learned implicitly thanks to the maximization of a single reward, the maximization of the number of nuts obtained.

Figure from the article: A squirrel learns complex behaviors that are required to maximize its consumption of food. Also relevant for the example of a kitchen robot.

They take the example of their previous work concerning the game of chess and the game of Go. The agent was only trained by the…