Using augmented BDI models for intelligent stimuli generation

Constrained random verification (CRV) and formal verification are currently the state-of-the-art approaches to functional verification. CRV's advantage is that does not suffer from size restrictions. Hence it can be applied to large designs. However, CRV is inefficient, with many simulation cycles spent exploring the same state space. Furthermore, guiding the tools into interesting corner cases typically requires considerable manual effort from verification engineers. Conversely, formal verification can find corner cases with little manual steering; but due to complexity limits, it can only be applied exhaustively to relatively small designs. This research aims to employ BDI (Belief-Desire-Intention) intelligence, augmented by machine learning techniques, to investigate the middle ground between CRV and formal verification. A BDI-based approach offers high-level, goal-directed planning with backtracking to achieve set goals. Augmenting a BDI model with a reward/punishment system can facilitate intelligent, agent-based generation of test stimuli which can automatically find and hit the interesting corner cases on a complex design, such as a CPU. This results in less reliance on hand-written constraints and random test generation. The solution would also be applicable to other verification use-cases such as post-silicon debug and test case automation at cluster level.