EUROPT 2025
Abstract Submission

612. A Novel Stochastic Derivative-Free Trust-Region Algorithm with Adaptive Sampling and Moving Ridge Functions

Invited abstract in session MB-1: Advances in Large-Scale Derivative-Free Optimization , stream Zeroth and first-order optimization methods.

Monday, 10:30-12:30
Room: B100/1001

Authors (first author is the speaker)

1. Benjamin Rees
School of Mathematical Sciences, University of Southampton
2. Christine Currie
School of Mathematics, University of Southampton
3. Vuong Phan
University of Southampton

Abstract

In this talk, we present a novel stochastic, model-based, derivative-free trust-region algorithm, ASTROMoRF. Within simulation optimization (SO) problems, evaluations of the objective function are inherently noisy and often computationally expensive to obtain; therefore, when solving SO problems, it is good practice to limit the number of calls made to the simulation model. ASTROMoRF ensures a stable finite-run performance through the implementation of a trust-region framework. It is also designed with high-dimensional SO problems in mind, using dimensionality-reduction techniques during the model construction stage of the algorithm to remove the model construction’s dependency on the dimension of the problem. ASTROMoRF applies an adaptive sampling scheme when obtaining responses from the model, in order to ensure stochastic sampling error from Monte Carlo runs is kept in lock-step with the model bias. ASTROMoRF employs a variable projection method to construct the surrogate model and active subspace matrix, with certification and geometry improvement on the interpolation set to ensure that the constructed model is fully linear. The talk will present an overview of the algorithm and numerical results showcasing the solvability of ASTROMoRF against other solvers for a range of problems and will demonstrate the superiority of using local active subspace construction over global sensitivity analysis.

Keywords

Status: accepted

Back to the list of papers