Adaptive Planning For Applications With Dynamic Objectives

Planning is commonly viewed as a task to devise a course of action or a plan that conforms as much as possible to a set of goals before acting. The plan will then be used to guide the activities. Most classic planning systems assume a static environment for the planning agents. In a static environment, states remain unchanged between actions, and the outcomes of actions are assumed to be deterministic. In reality, however, most applications are dynamic and stochastic in nature. External events, not caused by controlled actions, may occur; outcomes of actions may differ from expectations; new constraints may be introduced; and a new set of goals may evolve in response to the changes. Recently, we have proposed a multi-modal framework for adaptive planning in a dynamic environment with multiple objectives having the following characteristics: ̄ some of the objectives of the planning process may be conflicting ̄ some objectives may be ill-defined or difficult to measure quantitatively ̄ the objectives may change over time The task domain of production planning and scheduling is a typical example of such an environment. The scheduling objectives typically include the following: meeting due dates; reducing lead times; reducing work-in-process and finished goods inventories; maximizing resource utilization and the throughput of the system; and minimizing the sensitivity of the schedule to random events. These objectives are sometimes in conflict with each other. In our previous work, we developed a real-time distributed scheduling system 1 that observes its environment from different perspectives. These perspectives stem from the different objectives, and the system can react to events as they occur while monitoring the various objectives. This multi-perspective monitoring helps our system achieve better control of the environment. During our study, we discovered that although these global objectives may not change over time, the relevance of each objective is actually a function of time and the state of the system. For example, given a set of N objectives 01, 02, ... On, at time tl, objective 02 may be significantly more important than 01, whereas at another instance 1 For a detailed description of the system, please read the attached paper titled "An Architecture for Real Time Distributed Scheduling" to appear in "Applications of AI in Manufacturing," published by AAAI Press, edited by Dana S. Nau. From: AAAI Technical Report SS-92-01. Compilation copyright © 1992, AAAI (www.aaai.org). All rights reserved.