Improving DPLL Solver Performance with Domain-Specific Heuristics: the ASP Case

In spite of the recent improvements in the performance of the solvers based on the DPLL procedure, it is still possible for the search algorithm to focus on the wrong areas of the search space, preventing the solver from returning a solution in an acceptable amount of time. This prospect is a real concern e.g. in an industrial setting, where users typically expect consistent performance. To overcome this problem, we propose a framework that allows learning and using domain-specific heuristics in solvers based on the DPLL procedure. The learning is done off-line, on representative instances from the target domain, and the learned heuristics are then used for choice-point selection. In this paper we focus on Answer Set Programming (ASP) solvers. In our experiments, the introduction of domain-specific heuristics improved performance on hard instances by up to 3 orders of magnitude (and 2 on average), nearly completely eliminating the cases in which the solver had to be terminated because the wait for an answer had become unacceptable.

💡 Research Summary

The paper addresses a critical weakness of DPLL‑based Answer Set Programming (ASP) solvers: their reliance on generic, one‑size‑fits‑all heuristics can cause the search to wander into unproductive regions of the search space, especially in industrial settings where consistent performance is required. To remedy this, the authors introduce the DORS framework (Domain‑Specific Heuristics for DPLL‑based solvers). The approach consists of an offline learning phase in which a representative set of problem instances from a target domain is solved repeatedly using a modified solver (solvecp) that records every choice point (the literal selected at each branching step) together with its depth (level) in the search tree. For each instance, the algorithm returns either a failure (⊥) or a pair (answer set, ordered list of choice points). By aggregating these decision sequences across many instances, the system builds a statistical policy that maps the current partial assignment (state) to probabilities for each available literal (action). This policy replaces the standard choose_literal routine, biasing the solver toward literals that historically led to solutions in that domain.

Experimental evaluation on industrial ASP problems and benchmarks from the Second ASP Competition shows dramatic improvements: on the hardest instances the runtime is reduced by up to three orders of magnitude, with an average speed‑up of about two times. Moreover, cases that previously timed out almost completely disappear, indicating that the domain‑specific heuristic effectively prunes large portions of the search space before any runtime learning can take effect.

The authors contrast DORS with related techniques such as parameterized generic heuristics, CLASPFOIL’s automatic configuration, and runtime conflict learning. Unlike those methods, DORS retains knowledge across runs, eliminating the “cold‑start” problem of pure runtime learning. However, the paper also acknowledges limitations: the learning phase is offline and may not adapt well to drastically new instances; the current policy is a simple frequency‑based model that may miss complex interactions; and the cost of the offline training is not thoroughly quantified.

Future work suggested includes integrating online updates to the policy, employing reinforcement‑learning methods for more sophisticated state‑action mappings, and extending the framework to other DPLL‑based paradigms such as SAT, SMT, and constraint programming. Overall, the study demonstrates that domain‑specific heuristic learning can substantially enhance the robustness and efficiency of DPLL‑based ASP solvers, offering a practical pathway for deploying these systems in real‑world, time‑critical applications.