A t-SNE-based embedding for transfer optimisation with non-overlapping design variables

Abstract The cold start problem is a chief concern in the context of surrogate-based optimisation, as it can slow down or prevent convergence towards a global minimum. Transfer optimisation (TO) has recently emerged as a promising solution, positing the reuse of historical data to improve the quality of the surrogate predictor. However, the requirement for constant design parameters across the source and target tasks severely limits the range of applicability of TO. Several strategies have been proposed to overcome this constraint. However, they typically require either linked samples or linked design variables, and thus only offer a slight extension of the aforementioned scope. This paper proposes a new transfer optimisation method that enables varying design parameters. It removes the link constraint by using 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 the target task’s design variable space. Multiple-output Gaussian processes are used to model the resulting distribution of target and source samples. Results indicate significant improvements of 30-60% in optimisation performance over traditional Kriging-based approaches.