A Multi-objective Bayesian Optimization Algorithm with Online Knowledge Transfer

Since transfer learning can learn effective knowledge from the source task to help solve the target task, it offers a powerful tool for assisting multi-objective evolutionary Bayesian optimization in solving expensive multi-objective problems in the presence of data paucity using. Most existing optimization algorithms use genetic operators or construct Gaussian process models to achieve knowledge transfer and much less work have been reported that make use of transfer learning methods for knowledge transfer. To fill the gap, this paper proposes an online knowledge transfer-based multi-objective Bayesian optimization algorithm. The algorithm first classifies the data of cheap and expensive problems based on the dominance relationship and trains a classifier between cheap and expensive problems using the balanced distribution adaptation method. Then, this classifier is employed to generate synthetic data for training the surrogates for expensive optimization. In addition, to enhance the prediction accuracy for the classifier, an online transfer approach is adopted to retrain the classifier using the newly evaluated data in each iteration of optimization. Finally, the proposed algorithm is examined in comparison with two popular algorithms on two benchmark test problems to verify its performance.