Simultaneous knowledge discovery and development of smart neuro-fuzzy surrogates for online optimization of computationally expensive models

This work aims at enabling online optimization and control of computationally expensive models by employing Adaptive Neuro Fuzzy Inference System (ANFIS) as surrogates. ANFIS is governed by several parameters whose estimation based on heuristic assumptions degrade its efficiency. A novel surrogate building algorithm is thus proposed, with the aim of designing optimal ANFIS by balancing the aspects of over-estimation and prediction accuracy. It incorporates Sobol sampling plan and takes physics based model/data as the only input while estimating all other parameters simultaneously. A comparison between robust K-fold based Sample Size Determination (SSD) and an innovative fast Hypercube sampling based SSD technique is presented. Proposed algorithm fine-tunes the human experience which is often biased and thus prone to errors. It also enables the discovery of new knowledge from the existing information. ANFIS is built for industrially validated polymer reaction network model which made its optimization using NSGA-II, 9 times faster, thereby enhancing its scope for online implementation. It is then compared with another state-of-the-art surrogate, Kriging Interpolator, for an unbiased justification of robustness of the proposed algorithm.