An improved feature ranking method for diagnosis of systematic timing uncertainty

Diagnosis of design-silicon timing mismatch with feature encoding and importance ranking - the methodology explained

“…After applying the learning algorithm ε-SVR [5] to learn the mismatch map, for every path the learned model predicts a mismatch value. If learning is effective, plotting these values should produce a map similar to the original map.…”

“…The learning engine part of the methodology has been studied carefully before in [3,4,5]. As mentioned before, the conclusion is that the SVM algorithm ε-SVR performs the best [5].…”

“…how and what algorithms should be used for learning and how to rank features. The work in [5] improves the SVM classifier algorithm used in [3,4] with the ε-insensitive Support Vector Regression (ε-SVR) algorithm [27]. Although these works were important for developing the overall diagnosis methodology, they also raise questions regarding the applicability of the proposed approach in practice.…”

“…All previous works [3,4,5] use cells and groups of wires as features. In practice, the concerns can go beyond just cells and wires.…”

“…As we allow feature definition to go beyond wires and cells, the feature ranking method in [3,4,5] needs to be generalized as well. This is because the previous ranking method depends on the fact that all feature values can be interpreted in a certain way.…”