Towards Automated Knowledge Transfer in Evolutionary Multitasking via Large Language Models

A research team led by Xuebin Lyu, Yuxiao Huang, and colleagues has published a groundbreaking paper introducing the Self-guided Knowledge Transfer Design (SKTD) framework. This system leverages large language models (LLMs) like GPT-4 to autonomously generate knowledge transfer methods (KTMs) for evolutionary multi-task optimization (EMTO) algorithms. EMTO is an advanced optimization paradigm that improves search efficiency by enabling knowledge transfer across multiple tasks solved in parallel. Previously, designing effective KTMs required substantial domain expertise and careful manual customization, as different EMTO scenarios required distinct transfer strategies.

The SKTD framework represents the first attempt to automate KTM generation for EMTO. Through extensive experiments on well-established EMTO benchmarks with varying degrees of task similarity, the researchers demonstrated that SKTD consistently achieves superior or highly competitive performance compared to both state-of-the-art program search approaches and manually designed EMTO methods. The framework enables data-driven and self-adaptive construction of transfer strategies, facilitating effective knowledge reuse across heterogeneous tasks and diverse EMTO scenarios. This breakthrough opens new possibilities for automating the design of optimization solvers using LLMs' demonstrated capabilities in autonomous programming and algorithm synthesis.