OS-BiTP: Objective sorting-informed bidomain-information transfer prediction for dynamic multiobjective optimization

Abstract

Prediction response mechanisms based on transfer learning are extensively prevalent in dynamic multiobjective optimization algorithms (DMOAs), which transform historical information into a new environment for tracking the Pareto set (PS) or front (PF). However, many existing methods learn information of overall changes from old to new populations for prediction. Due to the different characteristics of individual variation within the population, this inevitably causes the valid information of more relevant individuals to be partially weakened during the training process, thus reducing transfer prediction-based accuracy. Therefore, this paper proposes an objective sorting-informed bidomain-information transfer prediction (OS-BiTP) for the DMOA based on individual objective variation, with the aim of transferring individuals within the same characteristics. The three core components in OS-BiTP are variation-based objective sorting (VOS), bidomain-information transfer within objective space (BiTOS), and bidomain-information transfer within decision space (BiTDS). Specifically, VOS divides the current PF into high- and low-objective variation classes and designs a modified linear prediction mechanism to forecast new environmental objective vectors. Afterward, VOS trains an easy transfer learning model to match old and new environmental individuals with the same objective variation classes to increase the transfer efficiency of individuals. To accurately track dynamic PFs and PSs, BiTOS and BiTDS perform intraclass correlation alignment for the same class of objective vectors and nondominated solutions and fine-tune the predicted objective vectors and solutions based on their variation differences. The numerical results demonstrate the superior performance and application of OS-BiTP via a systematic comparison with seven state-of-the-art DMOAs.