Assessing the impact of Learning Techniques in Quantified Boolean Logic Satisfiability

Learning, i.e., the ability to record and exploit some information which is unveiled during the search, proved to be a very effective AI technique for problem solving and, in particular, for constraint satisfaction. In (Giunchiglia et al. 2002) we have introduced learning as a general purpose technique to improve the performances of decision procedures for Quantified Boolean Formulas (QBFs). We have added learning techniques to QUBE, a state-of-the-art QBF solver. Embedding learning techniques in QUBE proved to be a challenging task, which also led to the definition and the implementation of new pruning strategies. In this paper, we report some experimental results that witness the effectiveness of learning and the associated pruning strategies. Noticeably, QUBE augmented with learning is able to solve instances that were previously out if its reach. On the other hand, the additional complexity of learning does impose some overhead on QUBE. We present results that highlight the presence of such overhead, we discuss its causes, and suggest possible remedies in QUBE.