Create a Video View Paper

D-VAL: Proving Planning Domains Do the Same Thing

This lightning talk introduces D-VAL, an automated tool that validates whether two planning domain models are functionally equivalent. We explore why this matters for AI planning systems, how D-VAL combines planning and constraint solving to prove equivalence in seconds, and what this means for developing and optimizing planning models with confidence.
Script
When you modify a planning domain model, adding shortcuts or refactoring operators, how do you know it still solves exactly the same problems? The authors of this paper built D-VAL to answer that question automatically, validating functional equivalence in under 43 seconds per domain.
Functional equivalence means two domains have identical reach sets. Every state reachable in one domain must be reachable in the other, even if the operators look completely different. This guarantee is critical when you optimize a domain or learn it from scratch.
D-VAL works in two stages. First, it strips out redundant operators, identifying which are primitives and which are just shortcuts or macros. Then it uses an SMT solver to find a consistent mapping between the predicates of the two domains, proving that one domain's transitions can simulate the other's.
The authors tested D-VAL on 75 validation tasks across 13 competition domains and their modified versions. Every equivalence check completed in under 43 seconds, demonstrating that rigorous formal verification is practical even for complex planning models.
D-VAL handles simple domains where operators have disjoint effects easily, but complex domains with shared effects require careful constraint modeling. The tool assumes operators are deterministic and works within the STRIPS subset of PDDL, so domains with conditional effects fall outside its scope.
D-VAL gives planning researchers confidence to refactor, optimize, and learn domain models without fear of silently breaking functionality. You can explore this work further and create your own research videos at EmergentMind.com.