Planning with Goal Preferences and Constraints

In classical planning, the planner is given a concrete goal; it returns a plan for it or a failure message. In the latter case, the user can either quit or modify the goal. For many applications, it is more convenient to let the user provide a more elaborate specification consisting of constraints and preferences over possible goal states. Then, let the system discover a plan for the most desirable among the feasible goal states. To materialize such an approach we require a formalism for specifying preferences and constraints over goals and an algorithm for solving the resulting constrained optimization problem. In this work we motivate the need for planning with preferences and constraints, suggest a rich, yet intuitive formalism for representing goal preferences in the context of a deterministic action model, discuss some of its properties, propose an efficient algorithm for planning with preferences and constraints based on this formalism, and provide extensive experimental analysis in an interesting new domain of configuration planning.