Training Agents by Crowds

On-line learning algorithms are particularly suitable for developing interactive computational agents. These algorithm can be used to teach the agents the abilities needed for engaging in social interactions with humans. If humans are used as teachers in the context of on-line learning algorithms a serious challenge arises: their lack of commitment and availability during the required extensive training. In this work we address this challenge by showing how ”crowds of human workers” rather than ”single users” can be recruited as teachers for training each learning agent. This paper proposes a framework for training agents by the crowds. The focus of this proposal is narrowed by using Reinforcement Learning as the human guidance method for teaching agents how to engage in simple negotiation games (such as the Ultimatum Bargaining Game and the Dictator Game).