Using Refinement Search to Unify and Synthesize Classical Planners

Though domain-dependent planners can be more efficient than general-purpose planners, no flexible and straight-forward framework is available to generate them. The trend in implementing planners has been to write general-purpose planners through traditional human-intensive programming effort. This thesis presents a general method for automatically generating efficient domain-specific classical planners from formal specifications using a semi-automated software synthesis system called Kestrel Interactive Development System (KIDS, for short). To do this, a unifying theory for general-purpose classical planning was needed so that one tackles core planning issues rather than get tied up with the shallow distinctions between state-space and plan-space models of search. Main contributions of this thesis are: (i) Combining classical planning in a unified model called Universal Classical Planner (UCP, for short) which casts plan-space and state-space planning as special cases of primitive refinements in a generalized planning algorithmic template. (ii) Using the UCP notion, developing CLAY, a general architecture for writing declarative theory of domain-specific planning. The planning theory is then used by KIDS to synthesize efficient domain-specific planners in concert with domain knowledge. It will be empirically shown that such planners perform well.