Mixed-Initiative Schedule Maintenance: A First Step Toward Plan Steering

When a plan involves hundreds or thousands of events over time it can be di cult or impossible to tell whether those events are unfolding \according to plan" and to assess the impact of dynamic plan modi cations. Pathological states may arise in which goals cannot be attained or are attained too slowly. Plan steering is an agent-based approach to this problem. The agent monitors an unfolding plan, detects and predicts pathological situations, and develops dynamic plan modi cations that will steer the plan around the problem. We present results for a system that performs the related task of schedule maintenance in the transportation planning domain. We evaluate system performance at pathology prediction and pathology avoidance and show that the agent, using limited domain knowledge and simple heuristics, is able to improve throughput signi cantly. We describe experiments in which humans perform the same schedule maintenance task both with and without the aid of the agent, and show that the human and the agent working together achieve better results than either working alone. This work was supported by ARPA/Rome Laboratory under contract #F30602-91-C-0076.