Abstract:Many engineering design and industrial manufacturing applications are tasked with finding optimum designs while dealing with uncertainty in the design parameters. The performance or quality of the design may be sensitive to the input variation, making it difficult to optimize. Probabilistic and robust design optimization methods are used in these scenarios to find the designs that will perform best under the presence of known input uncertainty. Robust design optimization algorithms often require a two-level optimization problem (double-loop) to find a solution. The design optimization outer-loop repeatedly calls a series of inner loops that calculate uncertainty measures of the outputs. This nested optimization problem is computationally expensive and can sometimes render the task infeasible for practical engineering robust design problems. This paper details a single-level metamodel-assisted approach for probabilistic and robust design. An enhanced Gaussian Process (GP) metamodel formulation is used to provide exact values of output uncertainty in the presence of uncertain inputs. The GP model utilizes a squared-exponential kernel function and assumes normally distributed input uncertainty. These two factors together allow for an exact calculation of the first and second moments of the marginal predictive distribution. Predictions of output uncertainty are directly calculated, creating an efficient single-level robust optimization problem. We demonstrate the effectiveness of the single-level GP-assisted robust design approach on multiple engineering example problems, including a beam vibration problem, a cantilevered beam with multiple constraints, and a robust autonomous aircraft flight controller design problem. For the optimization problems investigated in this study, the single-level framework found the robust optimum with a 99.9% savings in function evaluations over the standard two-level approach.

GPfit: An R package for Gaussian Process Model Fitting using a New Optimization Algorithm

Leveraging the Nugget Parameter for Efficient Gaussian Process Modeling

Active Design of Dynamic GP Models for Model Predictive Control Using Expected Improvement

A Fast GP Regression Method Using Banded Sparsification of Inverse Covariance

Towards Scalable Gaussian Process Modeling

Fast and robust Bayesian inference using Gaussian processes with GPry

Minibatch Markov chain Monte Carlo Algorithms for Fitting Gaussian Processes

Polynomial Representation of the Gaussian Process

Global Optimization of Gaussian processes

Efficient Gaussian process based on BFGS updating and logdet approximation

Scaling Gaussian Process Optimization by Evaluating a Few Unique Candidates Multiple Times

GP+: A Python Library for Kernel-based learning via Gaussian Processes

ProSpar-GP: scalable Gaussian process modeling with massive non-stationary datasets

Parallel cross-validation: A scalable fitting method for Gaussian process models

GPflowOpt: A Bayesian Optimization Library using TensorFlow

A Solution to the Ill-Conditioning of Gradient-Enhanced Covariance Matrices for Gaussian Processes

Multi-response Robust Optimization Using GP Model with Variance Calibration

Bayesian Modeling with Gaussian Processes using the GPstuff Toolbox

Robust and Conjugate Gaussian Process Regression

A Gaussian Process Modeling Approach for Fast Robust Design with Uncertain Inputs