Transfer Learning-Based Evolutionary Multi-task Optimization

Multi-task optimization (MTO) is an emerging research topic to optimize multiple related tasks simultaneously. It aims to enhance task interrelationships by leveraging shared information and features, thereby improving model performance. Evolutionary transfer optimization (ETO), applied to address multi-task problems using evolutionary algorithms, incorporates the principles of transfer learning. It utilizes knowledge and experience from source tasks to expedite the optimization process of target tasks. We introduce a transfer learning-based strategy where valuable information from one task is transferred as comprehensively as possible to another task. This article proposes an idea that is based on joint distribution adaptation (JDA) and employs population individual replacement methods as knowledge transfer, differential evolution as the underlying optimizer, called transfer learning-based evolutionary multi-task optimization algorithm (TLEMTO). To validate the effectiveness of the proposed algorithm, the experiment is conducted on CEC17 multi-task optimization problem benchmarks, the results show that TLEMTO is superior to the compared state-of-the-art algorithms.