Effective Knowledge Transfer Strategy by Promising Predictive Solutions for Evolutionary Multitasking Optimization

Traditional evolutionary algorithms require solving complex optimization problems from scratch every time, resulting in time-consuming and inefficient processes. However, many optimization problems share common characteristics, and knowledge gained from solving similar optimization tasks can enhance their performance. In light of this, multitasking optimization (MTO) are proposed, which can solve multiple tasks simultaneously. Recently, many MTO algorithms aim to improve the overall performance by knowledge transfer across tasks. However, a significant challenge is identifying promising knowledge to help tasks achieve positive knowledge transfer. To address this challenge, this paper introduces an evolutionary multitasking algorithm based on effective knowledge transfer strategy by promising predictive solutions (EMT-EKTS). Specifically, a logistic regression classifier is utilized to identify the valuable solutions. Subsequently, the historical evolutionary direction of the target task population and the mean difference direction between different tasks are employed to calculate the promising direction. Additionally, the valuable solutions are clustered to generate multiple classes, and the promising regions are determined based on the promising direction and these classes. Lastly, the proposed effective knowledge transfer strategy generates promising predictive solutions with good diversity in the identified promising regions. The experimental results demonstrate that EMT-EKTS outperforms several competitive evolutionary multitasking algorithms and achieves better performance.