Evaluation of a Mixed-Initiative Approach to Schedule Maintenance

Plans and schedules formulated to run in the real world will often fail due to the complexity and unpredictability of the environment. An unexpected, high priority order may arrive at the job shop or a piece of equipment used to process orders may break down. Existing methods to deal with this problem include real time recovery from plan failures [1] [2] [7] and post-hoc plan repair based on failures observed while executing the plan [6]. Failure recovery mechanisms, such as replanning, can be expensive, and it may not be feasible to repair a plan by letting it repeatedly fail. An alternative strategy is to monitor the execution of the plan, attempting to predict pathological states that make it di cult or impossible to achieve goals [5]. Doing so admits the possibility of e ecting plan modi cations in real time to avoid pathological states. Plan steering is a mixed-initiative approach to real time prediction and avoidance of plan failures [3]. A plan steering system comprises a pathology demon that monitors the execution environment to detect and predict pathological states, a plan steering agent that evaluates the demon's predictions and formulates plan modi cations to avoid predicted pathologies, and a human user who monitors the environment, the demon, and the agent. The human and the agent work together to steer the plan away from potential problems by intervening before they develop. The bene ts of keeping computers in the loop are clear. For large, complex plans, involving hundreds or thousands of events over time, determining whether events are unfolding according to plan and assessing the impact of dynamic plan modi cations are impossible for humans. As a rst step toward plan steering, we built a demon and an agent for the related task of schedule maintenance in the transportation planning domain. The problems that we address in this domain are closely related to common problems in scheduling of manufacturing processes. Ships and cargo (orders) must ow through a series of ports (processing points) that have limited capacity in a speci c order and in a timely manner. The primary di erence is that the \schedule" for a ship orders the ports that must be visited but only speci es the time at which the ship should begin its journey. We experimentally assessed the performance of the system at its two primary tasks: predicting schedule pathologies and formulating schedule modi cations to avoid those pathologies. We then assessed the performance of humans at the same task of schedule maintenance. We looked at performance in three conditions: the human acting alone, the agent acting alone, and the human and the agent working together.