Adaptive Bias Estimation Multitask Optimization

Evolutionary multitask optimization (EMTO) seeks to improve optimization efficiency by utilizing knowledge transfer (KT) between different tasks. However, existing KT strategies typically migrate individuals directly from source population to target population without fully considering the adaptability of the target population. This direct KT strategy may result in a mismatch between the transferred individual and the target population, thereby influencing KT efficiency. Moreover, while some KT strategies take bias into account, they often employ fixed adjustment methods, neglecting the dynamic variations in population distributions, which may lead to negative KT. To address this limitation, this paper proposes a novel adaptive bias estimation multitask optimization (ABEMTO). Specifically, ABEMTO first estimates the bias between two populations and adaptively adjusts it based on population distribution similarity, thereby effectively eliminating bias during the KT process, reducing the impact of negative KT, and ultimately improving KT efficiency. The experimental results reveal that ABEMTO significantly outperforms other state-of-the-art EMTO algorithms on two widely used multitask benchmarks, CEC2017 and CEC2022, demonstrating the effectiveness of ABEMTO.