Selecting optimal source for transfer learning in Bayesian optimisation

Bayesian optimisation offers an efficient solution to optimise black box functions. When coupled with transfer learning methods, Bayesian optimisation can leverage data from other function optimisations. A crucial requirement of transfer learning, however, is to restrict the transfer of knowledge only from related functions. Since the relatedness is not known a priori, selection of useful sources is an important problem. To address this problem, we propose a new method for optimal source selection for transfer learning in Bayesian optimisation. Using multi-armed bandits for source selection, we construct a new technique for identifying the optimal source and then use it for transfer learning in Bayesian optimisation. We show theoretically that the proposed technique is guaranteed to select the most related source and thus helps to improve the optimisation efficiency. We demonstrate the effectiveness of our method for several tasks: synthetic function optimisation, the hyperparameter tuning of support vector machines, and optimisation of short polymer fiber synthesis in an industrial environment.