Information-theoretic transfer learning framework for Bayesian optimisation

Transfer learning in Bayesian optimisation is a popular way to alleviate “cold start” issue. However, most of the existing transfer learning algorithms use overall function space similarity, not a more aligned similarity measure for Bayesian optimisation based on the location of the optima. That makes these algorithms fragile to noisy perturbations, and even simple scaling of function values. In this paper, we propose a robust transfer learning based approach that transfer knowledge of the optima using a consistent probabilistic framework. From the finite samples for both source and target, a distribution on the optima is computed and then divergence between these distributions are used to compute similarities. Based on the similarities a mixture distribution is constructed, which is then used to build a new information-theoretic acquisition function in a manner similar to Predictive Entropy Search (PES). The proposed approach also offers desirable “no bias” transfer learning in the limit. Experiments on both synthetic functions and a set of hyperparameter tuning tests clearly demonstrate the effectiveness of our approach compared to the existing transfer learning methods. Code related to this paper is available at: https://github.com/AnilRamachandran/ITTLBO.git and Data related to this paper is available at: https://doi.org/10.7910/DVN/LRNLZV.