IBMG: interpretable behavioral model generator for nonlinear analog circuits via canonical form functions and genetic programming

The paper presents IBMG, an approach to generate behavioral models of nonlinear analog circuits, with the special distinction that it generates models that are compact and interpretable expressions which are not restricted to any pre-defined functional templates. IBMG outputs a small set of interpretable nonlinear differential equations that approximate the time-domain behavior of the circuit being modeled. The approach uses genetic programming (GP), which evolves functions, but GP has been heavily modified so that the behavioral expressions follow a special "canonical functional form" grammar to remain interpretable. IBMG has explicit error control; it provides a set of models that trade off complexity and accuracy. Experimental results on a strongly nonlinear latch circuit demonstrate the usefulness of IBMG.