UC Berkeley

Even with impressive advances in formal methods, certain major challenges remain. Chief among these are environment modeling, incompleteness in specifications, and the hardness of underlying decision problems. In this paper, we characterize two trends that show great promise in meeting these challenges. The first trend is to perform verification by reduction to synthesis. The second is to solve the resulting synthesis problem by integrating traditional, deductive methods with inductive inference (learning from examples) using hypotheses about system structure. We present a formalization of such an integration, show how it can tackle hard problems in verification and synthesis, and outline directions for future work.