An agent on your shoulder: Human cooperation with algorithmic advice

Abstract

Studies have shown that humans are not quite proficient at playing the Iterated Prisoners’ Dilemma game. Machines on the other hand have been able to learn to play the game quite proficiently. However, their efficacy as advisers for humans has not been studied. In this study, we try to explore how efficient machines can be when acting as advisers. Humans were provided with recommendations from machines while playing against each other in various conditions. However, these humans performed worse than those without advisers. Further analysis showed that humans did not generally adhere to the recommendation. This could be attributed to a couple of factors including the unwillingness to listen to a reputation-less adviser. Thus, this leaves us with another question: How much information is needed for an algorithm to teach a human how to develop cooperative relationships.

Date of Award	Aug 2015
Original language	American English
Supervisor	Iyad Rahwan (Supervisor)