Designing for multi-agent collaboration: a shared mental model perspective

© 2018 Dr. Ronal Rajneshwar SinghWith the growing use of intelligent agents, it is important to know how to design artificial agents that would be capable of working in mixed teams of humans and artificial agents. In this thesis, formal models of teamwork and concepts successful in explaining effective teamwork in human teams are used to design artificial agents to enable the agents to be effective team members. For artificial agents to be effective teammates, they need to reason explicitly with concepts that achieve effective teamwork. To give artificial agents this reasoning ability, it is intuitive to explore teamwork concepts that work well with human teams. After all, humans will be an integral part of the human-agent system, and considering concepts that work well with human teams can provide useful insights when designing artificial agents. Two important factors that bind human teams together are: 1) a shared understanding of the team, tasks and goals between members; and 2) the interdependence relationships between members. This thesis seeks to explore ways of representing the shared understanding, designing a computational model of the shared understanding, identifying links between the elements of the shared understanding and the interdependence relationships, and enabling artificial agents to make communication decisions using the shared understanding in the context of artificial agent teams. This thesis uses shared mental models (SMM), which have been used to explain effective teamwork in human teams, as a way of establishing and maintaining the shared understanding between the team members, and proposes a computational model of SMM. The computational SMM consists of five components that represent the SMM knowledge and three types of processes that are required when designing and implementing artificial agents capable of leveraging SMM in their decision-making processes. Empirical studies demonstrate the success of the proposed SMM in capturing the required knowledge. A fine-grained analysis of various types of interdependence relationships results in the definitions of task, goal, and agent interdependence. These definitions clearly distinguish between the different forms of interdependence, and examples are used to show that such an analysis can identify the SMM related communication requirements. The results gathered using simulation-based experiments point out the sub-components of the SMM impacted by increasing levels of agent interdependence and provide guidance to agents on what to communicate given particular levels of interdependence. Examples also highlight that there are benefits to system designers as these definitions provide a formal framework in which to consider the different forms of interdependence relationships and resulting design options. Agents need to know how to determine what, when, and to whom to communicate to in order to establish SMM. This thesis proposes two novel and generalisable communication planning approaches, named Temporal-Projection based (TP) and Team-Model based (TM). These communication planning approaches enable agents to make their communication decisions using SMM. Empirical studies demonstrate the success of the two approaches in enabling agents to communicate the different SMM components, and illustrate that the two approaches perform better than hand-crafted communication protocols