A Meta-Learning - Based Surrogate-Assisted Evolutionary Algorithm for Expensive Multi-Objective Optimization Problems

In recent years, expensive multi-objective optimization problems that involve the costly-computation of multiple competing objectives have gained an increasing concern from the community of evolutionary computation. Surrogate-assisted evolutionary algorithms (SAEA) are often utilized to address such expensive optimization problems. However, in many SAEAs, the initialization of the surrogate model usually requires lots of expensive function evaluations. Due to the limited budget for real function evaluations, it is a great challenge to complete the initialization of the surrogate model using few-shot data. This paper proposes a meta-learning-based SAEA for solving expensive multi-objective optimization problems. This algorithm first uses a gradient-based meta-learning algorithm to learn domain-specific features from related tasks, which are utilized as the common parameters of the surrogate; then, few expensive function evaluations are utilized for fast adaptation to obtain the target task-specific parameters and to complete the initialization of the surrogate; finally, the initialized surrogate is utilized in the MOEA/D-DE for the evolutionary optimization process. Numerical experimental results demonstrate that this algorithm could effectively save the evaluation budget of expensive functions, and the effectiveness is verified on a set of benchmark test problems.