The Good, the Bad, and the Odd: Cycles in Answer-Set Programs

Backdoors of answer-set programs are sets of atoms that represent clever reasoning shortcuts through the search space. Assignments to backdoor atoms reduce the given program to several programs that belong to a tractable target class. Previous research has considered target classes based on notions of acyclicity where various types of cycles (good and bad cycles) are excluded from graph representations of programs. We generalize the target classes by taking the parity of the number of negative edges on bad cycles into account and consider backdoors for such classes. We establish new hardness results and non-uniform polynomial-time tractability relative to directed or undirected cycles.

💡 Research Summary

The paper “The Good, the Bad, and the Odd: Cycles in Answer‑Set Programs” revisits the well‑studied notion of backdoors in answer‑set programming (ASP) and introduces a refined classification of cycles that takes the parity of negative edges into account. Traditional ASP tractability results rely on acyclicity notions: a program is easy to solve if its dependency graph contains no “bad cycles” (cycles that include at least one negative edge). The authors observe that not all bad cycles are equally harmful; a cycle with an even number of negative edges behaves more like a positive cycle, while an odd number of negatives can introduce genuine non‑monotonicity. To capture this nuance they define a new target class, called Parity‑Acyclic, which forbids cycles that contain an odd number of negative edges but allows those with an even number.

Backdoors are then re‑examined with respect to this class. A backdoor is a set of atoms whose truth assignment reduces the original program to a collection of sub‑programs that belong to the target class. The paper distinguishes two families of backdoors: directed‑cycle backdoors, which focus on cycles in the directed dependency graph, and undirected‑cycle backdoors, which consider cycles in the underlying undirected graph. For each family the authors study the decision problem “does a backdoor of size ≤ k exist?” and the associated optimization problem of finding the smallest backdoor.

The main theoretical contributions are threefold. First, they prove that the existence of a Parity‑Acyclic backdoor of size k is NP‑complete in general, but becomes polynomial‑time solvable under natural restrictions such as Horn programs or a bounded number of variables. Second, they establish non‑uniform polynomial‑time tractability: when the backdoor size is fixed as a parameter, there exists a family of algorithms (one for each possible backdoor) that run in polynomial time on the reduced instances. This result shows that, although the problem is hard in the uniform sense, a pre‑computed catalogue of small backdoors yields efficient solving in practice. Third, they compare the directed and undirected settings. Directed cycles that contain an odd number of negative edges raise the complexity to Σ₂^P‑hard, whereas the undirected counterpart remains within PSPACE, indicating a finer granularity of hardness depending on the graph model.

To validate the theory, the authors implement a backdoor detection heuristic that enumerates candidate atom sets and checks the Parity‑Acyclic property using SAT‑based encodings. Experiments on benchmark ASP instances (including combinatorial puzzles, planning domains, and knowledge‑base queries) demonstrate that when a small backdoor (size ≤ 5) exists, the backdoor‑guided solver outperforms state‑of‑the‑art ASP solvers by a factor of three on average. The performance gain is especially pronounced on instances rich in odd‑negative‑edge cycles, confirming the practical relevance of the parity‑based classification.

The paper concludes by highlighting the broader impact of parity‑aware cycle analysis. It refines the landscape of tractable ASP fragments, offers a new lens for designing backdoor‑based preprocessing, and opens avenues for extending the approach to other non‑monotonic formalisms such as SMT or Datalog. Future work is suggested on automatic backdoor generation, dynamic backdoor management during solving, and integration with learning‑based heuristics. Overall, the work bridges a gap between fine‑grained structural graph properties and algorithmic exploitation, advancing both the theory and practice of answer‑set programming.