LATENT VARIABLE SYMBOLIC REGRESSION FOR HIGH-DIMENSIONAL INPUTS

This paper explores symbolic regression when there are hundreds of input variables, and the variables have similar influence which means that variable pruning (a priori, or on-the-fly) will be ineffective. For this problem, traditional genetic programming and many other regression approaches do poorly. We develop a technique based on latent variables, nonlinear sensitivity analysis, and genetic programming designed to manage the challenge. The technique handles 340-input variable problems in minutes, with promise to scale well to even higher dimensions. The technique is successfully verified on 24 real-world circuit modeling problems.