A coordination technique in a highly automated train rescheduling system

Abstract

Realistic problems are often not only complex and dynamic, but also involve various kinds of interaction among software processes, with human beings, and with a lot of distributing equipment and computers connected through communication networks. Train rescheduling is such a problem in plan revision. In this kind of problem, the AI system's output can not be used without human acceptance. And a person such as the train dispatcher has to be responsible for the acceptance of the output. Further, situations dynamically change in real-time. Thus, the central problem here is to let him quickly grasp and accept the AI system's suggestions. Focussing on such problems, this paper presents an AI coordination technique especially for plan revision consisting of: a flexible and hierarchical representation mechanism for goal decomposition and its alternative attainment knowledge, for solving complex problems through divide and conquer; dynamic goal generation, goal prioritizing, and its concurrent achievement through preemption, for solving dynamic problems; a fine-grained goal coordination mechanism for a real time complex system; and a man-machine coordination mechanism for quick user acceptance (rule-type knowledge representation for encapsulating multiple suggestions such as for proposing alternative suggestions together and/or for proposing subordinate suggestions together to solve side effects). This technique was used and evaluated in a train rescheduling AI system, which succeeded in its practical use for the first time in one of Japanese largest cities where trains run every 2-3 minutes. This proved that the technique could contribute to solving a real-world dynamic and complex man-in-the-loop problem.