Novel solution strategies for multiparametric nonlinear optimization problems with convex objective function and linear constraints

• Published in: Optimization and engineering
• Main Authors: Narciso, Diogo A. C., Pistikopoulos, Efstratios N.
• Format: Journal Article
• Language: English
• Published: 22-04-2024
• ISSN: 1389-4420; 1573-2924
• DOI: 10.1007/s11081-024-09888-2

This paper expands the multiparametric quadratic programming (mp-QP) framework presented in Narciso et al. (Comput Chem Eng 164:107882, 2022.
https://doi.org/10.1016/j.compchemeng.2022.107882
) to the more general multiparametric nonlinear programming (mp-NLP) case. First, the vector of parameters in mp-NLP problems is recast so that a unique transformed parameter is implicitly assigned to each of the inequality constraints. Maps of critical regions in this transformed space of parameters feature a set of 1-dimensional parametric edges (two per inequality constraint), which then greatly facilitate solution calculation. In the mp-NLP case, however, parametric edges define nonlinear semi-infinite lines; this requires an adaptation to the mp-QP algorithm (deals with linear parametric edges only), to enable a suitable calculation path to the more general nonlinear case. Three routes are proposed to mp-NLPs: the first route delivers solutions in compact form (same format as in mp-QP) using a single reference point per edge; the second route delivers explicit solutions using a hybrid approach for critical region construction, where all active sets not detected in the parameters space are excluded from the solution (equivalent to first route concerning accuracy); the third route builds on the initial explicit solution and further partitions the parameters space until all solution fragments satisfy an error check. Five algorithms were coded for these routes, and tested in a large range of mp-NLP problems. These strategies enable significant improvements in terms of solution accuracy, algorithm efficiency, and interpretability when compared to the state-of-the-art mp-NLP algorithms.