A Model of Human Teamwork for Agent-Assisted Search Operations

Abstract

Coalition forces are engaged in distributed collaborative decision making in time-pressured, high-stakes situations. Providing automated decision support for such environments is a very challenging problem, due to shortening decision cycles, the changing nature of threats, opponent tactics, and environmental unpredictability. Intelligent agents have the promise to provide timely assistance in various areas of decentralized, collaborative decision making, such as information gathering, information dissemination, monitoring of team progress and alerting the team to various unexpected events. In order to fulfill the promise of agent technology in providing effective team assistance, better understanding of robust human-agent teamwork is crucial. The goal of our research project is to develop a theoretically grounded and empirically tested framework to allow for effective agent support for human teams that are engaged in adaptive teamwork in dynamic environments. In order to (a) establish an experimental baseline of the performance of human-only teams and (b) understand where agents can provide the best utility in supporting human teamwork, we designed a scenario and experimentally evaluated team work where human teams performed a time-stressed, collaborative search task in a multi-player gaming environment. The collaborative search task recreates some of the challenges faced by human teams during search and rescue operations in the real world. In our experiments, we analyze (1) verbal communication between team members and (2) team coverage patterns. By ascertaining the information processing and coordination requirements of this team task, we can identify ``insertion points'' for agent assistance to human teams. The search patterns demonstrated by the experimental subjects exhibited similar problems to the behavior of actual search and rescue teams: (1) the creation of accidental holes in the search pattern due to poor execution of the search plan, and (2) poor priority assignments in the search plan due to false clues and hunches. We have identified that this is a promising area for agent assistance. By having agents monitor and track individual team members' coverage, gaps in the team coverage are exposed earlier in the search process allowing repairs to be made in a more timely fashion. Our model predicts that aiding the state of coordination between team members will result in task performance improvement.