Learn to Collaborate with Humans: Integrating Cheap Talk into Algorithms

Abstract

In human-robot interactions, autonomous robots need to quickly adapt to the behaviors of human partners. Such interactions are challenging because the human and robot may not have identical interests, but still both can benefit from cooperation. In such domains, existing online-learning algorithms do not consistently cooperate with humans. Two primary issues must be addressed to solve this challenge. First, a learning algorithm must be able to learn at time scales that support interactions with a human partner. Second, learning algorithms must be able to overcome the negative effects caused by random exploration, which is a core component of most learning algorithms. This research considers the integration of cheap talk into a learning algorithm (MEGA-S++). Cheap talk is costless signaling which accommodates a broad range of applications. In this thesis, we consider whether such signals help MEGA-S++ to establish cooperative relationships with people. Through an experimental study of 66 subjects, we find that MEGA-S++ integrated with cheap talk learns to cooperate with people as well as people learn to cooperate with each other. The experimental data indicates both human and MEGA-S++ achieve high proportions of mutual cooperation with the condition of cheap talk. Analysis on cheap talk shows substantial differences in the use of speeches by the top performers and low performers. Top performers start games with fair proposals and comply with these proposals. On the other hand, people with low performance fail to carry out the proposals they make. These results demonstrate that communication helps to address deficiencies of the existing online-learning algorithms, allowing them to collaborate with humans even when the goals of human and robot are not fully aligned.

Date of Award	Aug 2015
Original language	American English
Supervisor	Jacob Crandall (Supervisor)