AI-based planning for data analysis

摘要

Over the next 10 years (1994-2004), NASA plans to launch several satellites for studying Earth's ecosystems. This Earth Observing System is expected to provide about a terabyte of data per day. The authors have worked with Earth scientists at NASA Ames to encode the selection and preparation of this data as an AI-based planning task. They extended a domain independent planner called Collage, to meet the requirements of this domain, and are building a prototype application that selects and prepares data for ecosystem models. Collage is a successor to the Gemplan planner. Both systems were designed for large, complex planning domains and have been applied to standard toy problems as well as more realistic domains, such as building-construction planning. Two features distinguish Collage from other domain-independent planners. First, instead of state-based planning algorithms, Collage uses a broad repertoire of domain independent, action-based plan-construction algorithms. It can also be extended to include new planning methods, including domain-specific ones. Second, Collage uses localized search: using information about domain structure, the planner partitions a potentially intractable, global search space into smaller, more manageable ones, each working on a local subproblem of the overall task.