A GPHH with Surrogate-assisted Knowledge Transfer for Uncertain Capacitated Arc Routing Problem

Abstract

The Uncertain Capacited Arc Routing Problem is an important and challenging problem that has many real-world applications. Genetic Programming is utilised to evolve routing policies for vehicles to make real-time decisions and handle uncertain environments efficiently. However, when the problem scenario changes (e.g. a new vehicle is bought or an existing vehicle breaks down), the previously trained routing policy becomes ineffective and a new routing policy needs to be retrained. The retraining process is time-consuming. On the other hand, by extraction and transfer of some knowledge learned from the previous similar problems, the efficiency and effectiveness of the retraining process can be improved. Previous studies have found that the lack of diversity in the transferred materials (e.g. sub-trees) could hurt the effectiveness of transfer learning. As a result, instead of using the genetic materials from a source domain directly, in this work, we utilise the knowledge from the source domain to create a surrogate model. This surrogate is used on a large number of randomly generated individuals by GP in the target domain to select the promising initial individuals. This way, the diversity of the initial population can be maintained by randomly generated individuals, but also guided by the transferred surrogate model. Our experiments demonstrate that the proposed surrogate-assisted transfer learning method is superior to existing methods and can improve training efficiency and final performance of GP in the target domain.