Mining High Quality Assertions Using Best-Gain Decision Forests

We introduce the Best-Gain Decision Forest algorithm, an assertion mining methodology that generates high quality assertions. Our methodology uses static analysis and a novel machine learning technique to mine assertions from register-transfer level (RTL) simulation traces. Our machine learning technique is inspired by decision tree algorithms and generates concise, high coverage RTL assertions. The Best-Gain Decision Forest algorithm induces assertions from all decision trees optimized for maximum gain and uses a set containment algorithm to minimize redundancy. We show that our methodology generates assertions with up to 2 fewer propositions and 10% greater functional coverage than those generated by existing methodologies.