Date of this Version
Human users form coalitions to solve complex tasks and earn rewards. Examples of such coalition formation can be found in the military, education, and business domains. Multiagent coalition formation techniques cannot be readily used to form human coalitions due to the unique aspects of the human coalition formation problem, e.g., uncertainty in human user behavior and changes in human user behaviors due to human learning. Thus, a multiagent system designed to form human coalitions has to solve a learning problem, that is further made difficult by the limited learning opportunities and usability issues (i.e., actions or decisions being perceived as not useful due to loss of immediate re-wards while the agents are learning or exploring) intrinsic to the human coalition formation process. We propose and design a multiagent framework that distinguishes the impact of the model of a human user from that of the agent support for that model. This novelty allows an agent to (1) better compute the types of support it should provide to its assigned user and (2) more accurately estimate the value of a coalition by its ability of (a) solving the current task and (b) improving the coalition members’ behavior due to learning. In our design, each agent models its environment using a Bayesian network and forms human coalitions for its assigned user using a negotiation-based protocol. Each coalition balances the immediate and future rewards by analyzing the benefits of solving tasks and facilitating human learning. To evaluate the proposed framework, we have built a comprehensive simulation where agents support students to form coalitions in a collaborative learning environment. Our results show that the framework is able to form successful coalitions that facilitate student learning while solving tasks, leading to overall better re-wards for student coalitions.