ITR: Modeling Domain Independent Planning to Advance Application

Colorado State University, Fort Collins CO

Investigators

Abstract

This project will contribute to the field of automatic planning by developing methods to determine what planner works well under various sets of conditions. It will construct a metaplanner that will be able to test performance of existing and hybridized planners on a suite of domains, especially more realistic domains than have been used commonly before. Based on the results of initial studies, predictive models of planner performance will be developed that translate problem and domain features to estimates of how much time would be needed by a given planner to solve the problem. The models, in turn, will be operationalized in the metaplanner. There are three steps to the work: 1) collect and develop applications and realistic domains to assemble a challenging benchmark set of problems; 2) modeling the performance of several state of the art planners, and 3) developing a metaplanner which is directed by the models to select a planner to solve problems in a particular domain. Planning underlies any complex human endeavor, from organizing a busy life to operating a large airline or testing large software systems. In each case, people need to track an enormous number of details, marshal resources and ensure that actions necessary to achieve their goals can be taken in a timely matter. However, automated planning has had little effect on how planning is actually accomplished. This research seeks to improve that situation through new methods to decide which planner is appropriate for a given task. The metaplanner can be used by both people learning about planning and new application developers who may not know which planner would be best for their application.