Bayesian network-based test generation Wiki

Overview

Bayesian network-based test generation is identified in the provided evidence as a form of coverage-guided test generation that uses Bayesian networks. It is discussed among approaches for generating instruction streams for processor verification.

Verification context

The technique appears in a discussion of simulation-based processor verification. The cited source states that extensive processor verification at the Register-Transfer Level (RTL) is important to avoid bugs, and that simulation-based approaches require efficient test generation to achieve thorough verification.

Within that context, the source describes several categories of instruction-stream generation methods:

model-based approaches that separate the test generator from the architecture description;
constraint-solving-based approaches;
optimized frameworks that propagate constraints across multiple instructions;
test generators with coverage models for instruction execution paths;
alternative approaches including Bayesian-network-based coverage-guided test generation, other machine-learning techniques, and fuzzing.

Characterization in the cited comparison

The cited RISC-V verification paper treats Bayesian-network-based test generation as an alternative approach rather than as the paper's proposed method. The paper states that alternative approaches integrating Bayesian networks, other machine-learning techniques, and fuzzing are either not designed for RTL verification or impose restrictions on generated instruction streams. It also states that these approaches do not target the RISC-V ISA.

Scope and limitations from the evidence

The provided evidence supports only a high-level characterization of the technique:

it is coverage-guided;
it is based on Bayesian networks;
it is used or proposed in the area of processor-verification test generation;
in the cited comparison, it is grouped with approaches that do not target RISC-V ISA verification and have limitations for unrestricted RTL instruction-stream testing.

The evidence does not provide algorithmic details such as Bayesian-network structure, training procedure, probability update rules, coverage metrics, or concrete implementation flow.