Double-Strength CAFFEINE: Fast Template-Free Symbolic Modeling of Analog Circuits via Implicit Canonical Form Functions and Explicit Introns

CAFFEINE, introduced previously, automatically generates nonlinear, template-free symbolic performance models of analog circuits from SPICE data. Its key was a directly-interpretable functional form, found via evolutionary search. In application to automated sizing of analog circuits, CAFFEINE was shown to have the best predictive ability from among 10 regression techniques, but was too slow to be used practically in the optimization loop. In this paper, we describe double-strength CAFFEINE, which is designed to be fast enough for automated sizing, yet retain good predictive abilities. We design "smooth, uniform" search operators which have been shown to greatly improve efficiency in other domains. Such operators are not straightforward to design; we achieve them in functions by simultaneously making the grammar-constrained functional form implicit, and embedding explicit \´introns\´ (subfunctions appearing in the candidate that are not expressed). Experimental results on six test problems show that double-strength CAFFEINE achieves an average speedup of 5times on the most challenging problems and 3times overall; thus making the technique fast enough for automated sizing