Routine Computing Tasks: Planning as Understanding

Abstract

A system called NETWORK is described which implements the construction-integration model of Kintsch (1988) in a routine computing-task domain. This system builds a plan of action on-line for a given task from a set of plan elements. These plan elements are simple overlearned production rules that are put together by NETWORK to produce plans for novel tasks. This approach is contrasted with other types of planners as NETWORK is shown to plan solutions to a variety of tasks. Discussions focussing on the use of long-term memory, case-based reasoning, and planning and acting are presented. NETWORK takes as input a task description, uses this information to select related knowledge from its long-term memory, and constructs a network representation of the task. This network is then integrated through a spreading-activation procedure where irrelevant items in the network become deactivated, and things that appear related sustain each other's higher activation. Subsequently, a decision process chooses a plan element for firing, depending upon its level of activation with those more highly activated being considered for action first. When a plan element is found that can fire, its outcomes are added to the state of the world. The process repeats until a selection of plan elements is produced to complete the task.