Date of this Version
Ahuja, P. (2017). Investigating Diversity in Open Multiagent Team Formation (Master's thesis, University of Nebraska, Lincoln). Retrieved from http://digitalcommons.unl.edu/computerscidiss/
Team formation is the most rudimentary form of interactions in distributed AI and multiagent systems as it allows coherent collections of agents to work together in a beneficial manner towards a common goal of interest. Basically, individual expertise are assembled together in an additive fashion for accomplishing tasks together. A plethora of the related studies found in the literature often make several unrealistic assumptions such as coordination amongst the agents, or agents having knowledge of the whole environment, or agents and/or tasks are of the same kind, or a static environment setting. Against this background, we argue that there are real-world characteristics that make team formation more challenging: (1) There is no or minimal pre-coordination since storage and retrieval is a costly affair, (2) There is diversity amongst types of agents (Apprentices, Generalists, and Specialists) and tasks (Low, Medium, and High), (3) The environment is open i.e., agents and tasks can leave and enter the environment, and (4) Agents are continuously learning and improving their capabilities.
The main contribution of this research is to study in great depths the impacts of various permutations of open and diverse environments on team formation and how agents learn to form these teams. Based on the findings of these studies, we demonstrate that both diversity and openness have impacts on the team formation. Having evaluated the results of the impacts of openness and diversity on the environment we, to strengthen the robustness of the original model, we introduce an enhanced version of this model. The next contribution of this thesis is putting forth an enhanced probabilistic modelling solution. To be able to carry out new investigations and introduce the new model, we have restructured and cleaned up the simulation software used for building the original model. Having implemented the enhanced model, we then show how this new model performs better than the original model. The final contribution of this thesis was to show why the new model performed better than the original model.
Advisor: Leen-Kiat Soh