Efficient strategies for engineering data re-use

Abstract

One of the central challenges of machine learning is overfitting, whereby algorithms model noise or random fluctuations instead of the true underlying patterns. In engineering design, factors such as noise, uncertainty and complex, nonlinear responses exacerbate the problem. However, the predominant concern remains the prohibitive cost of generating sufficient training data.

This thesis demonstrates the merits of data reuse in addressing the issue of training data sparsity. Leveraging historical data leads to an increase in the effective size and quality of the training set, improving predictive accuracy and enabling a more thorough exploration of the design space. Transfer optimisation, a recently introduced paradigm for engineering data reuse, serves as a mathematical basis, and several surrogate models are demonstrated as its practical realisations. The thesis makes two contributions regarding the effective integration and reuse of historical data.

The first contribution concerns cases where the response values of the historical and new data sources are uncorrelated - a condition shown to be prevalent in engineering design, as even the most subtle topological modifications between data sources can produce pronounced changes in the objective function topography. A novel calibration method is introduced. Results demonstrate that the technique can halve optimisation cost when the relationship between data sources is linear. The second contribution deals with situations where the parameterisations of historical and new data sources differ, with no overlap or known mapping between them. The novel method uses simulation physics, rather than a mapping function, to represent the distribution of source and target samples. Then, it employs a t-SNE-inspired optimisation routine to recreate this distribution in a specified design variable space. Multiple-output Gaussian processes are used to model the resulting distribution of target and source samples. When used as part of the efficient global optimisation algorithm, the proposed approach yields significant reductions in cost - between 30 and 60% - relative to a Kriging-based method with no data reuse.

More information

Identifiers

ORCID for David Toal: orcid.org/0000-0002-2203-0302

ORCID for Andy Keane: orcid.org/0000-0001-7993-1569

Catalogue record

Export record