Efficient analog performance macromodeling via sequential design space decomposition

This paper presents a sequential design space decomposition technique for efficient analog feasibility and performance macromodeling. The performance macromodel is composed of one feasibility classifier and a set of performance regression models. The algorithm sequentially decomposes the design space into smaller partitions. We use a heuristic to identify the regions of significance. Feasibility classifiers of high precision and recall are constructed with low modeling cost within these regions. The final feasibility model is essentially a set of feasibility classifiers, each of which is valid for a subspace of the design space. The performance regression models constructed with much lower training instances have higher accuracy compared to a previous approach.