Biasing symbolic search by means of dynamic activity profiles

We address BDD based reachability analysis, which is the core technique of symbolic sequential verification and Model Checking. Within this framework, non purely breadth-first and guided traversals have shown their value to improve efficiency by reducing memory consumption for BDD representation. We propose a guided search strategy exploiting performance statistics. These activity figures are gathered through a continuous and dynamic learning process on a variable-by-variable basis. This technique is currently integrated with the reachability analysis routine, as it is fully compatible with dynamic reordering and allows multiple partial traversal phases. We thus move away from the static and manual schemes, which are one of the main limitations of previous approaches. Experiments are given to demonstrate the efficiency and robustness of the approach.