Chebyshev inequality–based inflated convex hull for uncertainty quantification and optimization with scarce samples

Author Type

Outside Researcher

Co-Author Type 1

Outside Researcher

Co-Author Type 2

Faculty

Co-Author Type 3

Outside Researcher

College

Engineering and Computer Science

Department

Ocean and Mechanical Engineering

Document Type

Article

Publication/Event/Conference Title

Structural and Multidisciplinary Optimization

Publication Status

Version of Record

Abstract

Dealing with the unknown-but-bounded uncertainty and insufficient or scarce data is an often-encountered challenge in real life engineering. In this study, we develop an efficient convexity approach to construct the uncertainty model in the form of bounds to address such a challenge. Approaches that use probabilistic models require comprehensive information about the uncertainty, which is often difficult or expensive to obtain, whereas the convex models that approximate the uncertain region using different bounding geometries can operate with scarce data or parameter bounds. The novelty of the current work is to use (i) convex hull as the bounding geometry in the uncertain design space (ii) Chebyshev inequality to inflate the convex hull that can include future data points which will then be used in design. Convex hull provides the least volume compared to other geometries reported in literature, such as interval, ellipse, super ellipse, and parallelepiped that are used in convexity approaches. In addition, to obtain the bounds of linear limit states, it is sufficient to evaluate the limit states only at the vertices of the convex hull, thereby saving on expensive simulations at nonvertex points. The proposed method is demonstrated on engineering examples with different coefficient of variation and random variables following different distributions. Results reveal the superiority of the proposed approach over the existing convexity approaches for uncertainty quantification. Also, the optima obtained by proposed method are always conservative to ones from a large sample Monte Carlo simulation and thus avoids under- or over-design associated with safety factor approach.

First Page

2267

Last Page

2285

DOI

10.1007/s00158-021-02981-5

Publication Date

10-1-2021

This document is currently not available here.

Share

COinS