Evolutionary Knowledge Transfer via Graph Learning for Many-Task Heterogeneous Routing Problems
提出基于图学习的进化知识迁移方法，用在线更新的神经网络替代传统基于解的迁移载体，实现异构任务间的知识共享，在多达100个任务的路径规划问题上表现优于五种现有算法。
Evolutionary multitasking optimization (EMTO) is widely applied to solve combinatorial optimization, notably routing problems. While existing handcrafted-rule-based knowledge transfer operators facilitate inter-task information exchange and overall convergence in EMTO, they face key limitations: (1) reliance on excellent solutions as carriers hinders knowledge transfer between dimensionally heterogeneous tasks; (2) limited rule generalization impedes modeling complex mappings in many-task environments. To address these, this work proposes evolutionary knowledge transfer via graph learning (EKT-GL). Its core module is an online-updated neural network. EKT-GL introduces a model-based knowledge carrier to transfer common knowledge, replacing solution-based carriers. It employs end-to-end learning to seamlessly integrate similarity measurement, knowledge extraction, and transfer, overcoming manual rule limitations and ensuring superior generalization through data-driven training. In comparative experiments, two routing problems are used to construct many-task heterogeneous data sets with different scale (the maximum number of tasks is 100). Experimental results indicate that EKT-GL have competitive performance compared to other five state-of-the-art EMTO algorithms.