Lazily Reformulating Design Optimization as a Classification Problem

This paper aims to address the unsolved problem of finding truly innovative designs. This is approached by teaching an estimator to directly predict the set of all feasible designs. The estimator is taught using a modified Surrogate Based Optimization (SBO) problem. This modification leads to the discovery of the first ever application of classification surrogates in design optimization. An algorithm is then developed using the new architecture and demonstrates to be on average a superior teacher than Latin Hypercube Sampling (LHS) for all iterations with a confidence greater than 99.7% given that all modes are found. It is envisioned that a practitioner using this approach would find a set of radical feasible designs large enough to survive the uncertainty of all stages of design resulting in an innovative solution.