Evolutionary Multi-Objective Bayesian Optimization Based on Multisource Online Transfer Learning

One main challenge in multi-objective Bayesian optimization of expensive problems is that only a very limited number of fitness evaluations can be afforded. To address the above challenge, this article introduces multisource online transfer learning into an evolutionary multi-objective Bayesian optimization algorithm, facilitating the use of knowledge transferred from multiple computationally cheap problems. According to the dominance relationships of the solutions in the cheap and expensive problems, the source selection and style transfer mapping in online transfer learning are adopted to augment the data for training the Gaussian processes for the expensive problems, in which an adaptive online multisource transfer learning method is proposed based on the relationship between the balance factor parameter and the transfer mapping method. Comparative studies on two sets of widely used multi-objective optimization benchmark problems, two sets of multi-task optimization problems, and one real-world expensive optimization problem confirm the effectiveness of the proposed algorithm.