Acquiring Criteria for Plan Quality Control

In producing plans, human planners take into account a variety of criteria that guide their decisions. Besides constraints imposed by the domain itself, these criteria often express preferences among alternative plans that achieve the given goals. Human planners can use these criteria for two important purposes: ¯ when asked to generate one plan, human planners are able to discern between an ordinary solution and a better quality one and propose the latter. ¯ when asked to generate several alternative plans, human planners are able to discern between similar alternative solutions and qualitatively different ones relaxing different criteria to explore tradeoffs. Current AI planners are good at generating a solution that satisfies the goals that they are given. Some planners provide facilities to control the quality of the solution to be returned, such as evaluation functions or search control rules. However, they do not usually integrate quality considerations across several plans. In addition, there is not enough data on the adequacy of these representations to reflect the plan quality criteria that are necessary in practice. Often, the quality criteria that human expert planners consider: ¯ are highly dependent on the situation and the scenario at hand. Some criteria may be more important if there is a certain deadline, or new criteria may need to be considered if new consid-erations come up. ¯ include complex factors and tradeoffs that are often not represented by an automatic planner Thus, evaluating plan quality requires both complex reasoning abilities and sophisticated knowledge acqui-sition tools that current planning technology lacks. The goals of this work are twofold: ¯ to provide tools that allow expert planners to define criteria for plan quality and for preferences among alternative plans ¯ to operationalize these criteria to guide generative planners in proposing better quality plans