Preferences and Soft Constraints in PDDL 3

In many read-world planning domains, generating good plan quality is a central issue. This is especially true for problems with many solutions, or with many goals that cannot be achieved altogether. We propose an extension to the PDDL language that aims at a better characterization of plan quality by allowing the user to express strong and soft state constraints about the structure of the desired plans, as well as strong and soft problem goals. In the plan quality evaluation, soft goals and constraints are evaluated according the their violation penalty weights, which are expressed by the user in the plan metric. The new language, PDDL3, allows us to distinguish alternative feasible plans (satisfying all strong constraints and goals), preferring plans that minimize the weighted violations for soft goals or constraints, possibly combined with other plan quality criteria. We describe the syntax and semantics of PDDL3.0 and we give several examples, including a domain from the very recent fifth International planning competition, which focused on soft trajectory constraints and goals.