Overview
Bayesian Networks for Test Generation is a coverage-directed test generation (CDG) technique used in the functional verification of processors. The approach uses Bayesian networks to model dependencies between test-generation parameters and coverage feedback, steering the generation of instruction streams toward portions of the design space that remain uncovered.
Origin in Functional Verification
The technique is introduced in the context of functional verification of processors, where generating compact, high-coverage test instruction streams is a central challenge. It is documented in the literature as one of the established methods for coverage-directed generation of test programs, alongside other approaches such as constraint-satisfaction formulations, evolutionary generators (e.g., MicroGP), specification-based generators (e.g., MicroTESK), and randomized/SMT-solver-based generation.
The seminal reference for this technique is:
Shai Fine, Avi Ziv. Coverage directed test generation for functional verification using bayesian networks. 2003;286–291. doi:10.1145/775832.775907.
Relation to DIG
The Dynamic Instruction Generator (DIG) paper, Enhancing Functional Verification with Dynamic Instruction Generation by Exploiting Processor Runtime States (Liu et al., 2024), cites the Fine & Ziv Bayesian-network work as part of its related-work survey on coverage-directed and constraint-based test generation. DIG positions itself as a complementary, runtime-state-driven alternative: rather than learning coverage correlations offline, DIG accesses processor runtime architectural information through an instruction set simulator to produce instruction streams with valid semantics, avoiding illegal memory accesses and infinite loops, and incorporates both intra-instruction and inter-instruction test knowledge. In a case study on Western Digital's open-source RISC-V core, VeeR EH2, DIG is reported to reduce the number of test instructions by up to 62.50% and 86.11% relative to the random generators RISC-V DV and RISC-V Torture, respectively, while achieving superior functional coverage.
Limitations of the Available Evidence
The provided evidence consists solely of the DIG paper and its reference list. It does not contain a primary source describing the internal structure of the Bayesian-network technique (e.g., the network topology, the choice of nodes for coverage metrics and generation knobs, or the learning algorithm). Claims about the technique here are restricted to what can be inferred from its role as cited prior art in the DIG paper.