Virtual knowledge architecture for intelligent robot planning

The authors present an architecture for a system which will learn to anticipate and avoid problems, including both problems in accomplishing the overall goal of the system as well as problems normally associated with cooperation. The authors envision a future of autonomous robots which will be asked to accomplish a common goal as a team. The coach agent is a case based planner whose purpose is to learn employing cases of very successful teams and to anticipate and avoid unsuccessful teams. The coach is allowed to trap in situations such as deadlock and repetition of effort among agents so long as these events are defined as failures by the user. The user may also define failures as instances of poor performance. In these cases, the system will learn to avoid employing agents which lead to poor performance. The architecture will also avoid plans that involve agents which are malfunctioning and will attempt to achieve the goal using those resources made available by the rest of the team.<<ETX>>