Revised learning based evolutionary assistive paradigm for surrogate selection (LEAPS2v2)

Abstract

Abstract Selecting appropriate surrogate models is crucial. This article upgrades our learning-based surrogate selection paradigm (LEAPS2) to LEAPS2v2. Its key features include: modeling noisy and non-noisy data, avoiding noise-fitting surrogates, powerful data attributes, novel composite metric (Surrogate Quality Score) to assess surrogate accuracy and complexity, and many surrogates and datasets. For any dataset, LEAPS2v2 recommends 3 out of 36 surrogates. A LEAPS2v2 recommendation is successful if it recommends at least one of the top three surrogates based on Surrogate Quality Score. LEAPS2v2 was successful on more than 94 % non-noisy and 89 % noisy datasets. Moreover, it made successful recommendations on 14/16 real industrial datasets. Strong correlation was observed between a surrogate being a true best and being recommended. Our numerical analysis revealed that the best surrogates for noisy data are different from those for non-noisy datasets. No single surrogate outperforms others for all datasets, highlighting the utility of LEAPS2v2. Publication Status: Submitted to Computers & Chemical Engineering