Multi-Source Knowledge Fusion Based on Graph Attention Networks for Many-Task Optimization

Although knowledge transfer methods are developed for many-task optimization problems, they tend to utilize solutions from a single task for knowledge transfer. Indeed, there are usually multiple relevant source tasks with commonality. Multi-source data fusion can capture complementary knowledge of distinct source tasks to better assist the optimization of target tasks. However, biases potentially flow with the interaction between tasks during multi-source fusion, resulting in performance degeneration. Thus, how to select multiple relevant source tasks and perform multi-source knowledge transfer is challenging. To address these issues, this paper proposes a multi-source knowledge fusion (MKF) method based on graph attention networks. In MKF, tasks are structured using a relational graph, in which each vertex represents a task and each directed edge from vertex u to vertex v represents that u is a source task of v. Particularly, for each task, multiple source tasks are selected based on the distribution similarity and evolutionary performance. In the graph, local message is passed from source tasks to target tasks using graph attention networks, which automatically learn the adjacency weight of each directed edge and aggregate solutions from multiple source tasks to obtain fused representations for target tasks. These fused representations are adopted to generate new solutions through mutation. In this way, multi-source knowledge is fused and transferred according to their importance to the target task. Integrating MKF into differential evolution, a new algorithm named MKF-DE is put forward. Experimental results on GECCO2020MaTOP and CEC2022MaTOP show that MKF-DE outperforms state-of-the-art algorithms on most instances.