EMaTO-SGN: Evolutionary Many-Task Multi-Objective Optimization with Star-Topology Generative Networks

Abstract

: The evolutionary multitasking algorithm based on generative strategies (EMT-GS) has emerged as a promising approach to evolutionary multitasking. However, when extended to many-task optimization scenarios, EMT-GS may suffer from negative knowledge transfer. Moreover, the fully connected generative adversarial network (GAN) architecture employed in EMT-GS incurs quadratic time complexity with respect to the number of tasks. To address these limitations, we propose a novel algorithm, termed EMaTO-SGN (Evolutionary Many-Task Optimization with Star-Topology Generative Networks). EMaTO-SGN employs a star-topology architecture for centralized knowledge routing and scheduling, in which a central hub evaluates candidate source tasks and determines when knowledge transfer is needed. Each task is assigned a dedicated Wasserstein generative adversarial network with gradient penalty (WGAN-GP), reducing the computational complexity of generative model training and sample generation from quadratic to linear with respect to the number of tasks. To mitigate negative transfer, an inter-task scoring mechanism combines maximum mean discrepancy (MMD) with historical transfer success rates to identify eligible source tasks. In addition, an adaptive triggering strategy monitors evolutionary stagnation to balance cross-task exploration and task-specific exploitation. Experiments on the CEC2019 and WCCI2020 benchmark suites show that EMaTO-SGN achieves the best average rankings among the state-of-the-art algorithms considered. It also requires less computational time than EMT-GS under the tested task configurations. These results demonstrate the effectiveness and scalability of EMaTO-SGN for evolutionary many-task multi-objective optimization.