On Using Domain Control Knowledge in Planning

Position Paper

Abstract

Automated planning involves finding a sequence of actions to achieve a given goal. Domain-independent planning decouples a planning task specification from planning engines. Frequently, the planning task specification describes only the physics of the environment, that is, how actions modify the environment. Planning engines are then generic solvers to solve any planning task ``reasonably well''. However, generic planning engines tend to struggle with tasks that domain-specific algorithms can solve easily. Domain Control Knowledge (DCK) narrows the performance gap between domain-dependent and domain-independent solvers by encoding additional information into the planning task specification while keeping the planning engine generic. In this paper, we define the notions of completeness and optimality perseverance of DCK. When DCK has these properties, the generic planner guarantees that it finds a plan (or an optimal plan) if the planning task is solvable and DCK is used. We then define a notion specifying that the use of DCK can eliminate search during plan generation. We discuss the introduced notions in the context of two case studies.