Multitask differential evolution with adaptive dual knowledge transfer

The goal of multitasking optimization (MTO) is to handle multiple tasks simultaneously. In MTO, effective knowledge transfer (KT) techniques significantly influence the performance of multitasking evolutionary algorithms (MTEAs). These techniques vary in their impact, and by assigning suitable techniques to individuals, algorithms can leverage them to enhance overall performance. With this purpose, we propose MTDE-ADKT, a novel MTEA integrating adaptive dual knowledge transfer and improved differential evolution. The MTDEADKT introduces several key innovations: Firstly, a novel domain adaptation (DA)-based KT technique rooted in transfer learning is proposed. Secondly, the DA-based KT technique is integrated with the traditional unified search space-based KT technique. This integration dynamically adjusts the probability allocation for each KT technique, tailoring it to suit the specific needs of each task. Thirdly, a new mutation strategy for offspring generation is presented, facilitating genetic material exchange among different tasks. The experimental results show that MTDE-ADKT outperforms 18 state-of-the-art algorithms on two MTO benchmark suites, a many-task optimization benchmark suite, and two real-world applications.