Learning to Solve the Travelling Salesman Problem via Fourier Representation

Recent studies applying deep learning to solve the travelling salesman problem consistently take city coordinates as input and encode them in the spatial domain. However, relying solely on spatial data may be insufficient to provide comprehensive feature information for model learning. In this paper, we propose a Fourier feature capture method to explore the inherent connections within the input data and enhance the quality of the solution for the travelling salesman prob-lem. The method extracts features from the frequency domain by utilizing Fourier descriptors. Additionally, we propose a multimodal fusion representation scheme to train deep models in a fused space. This scheme enables the models to acquire knowledge from multiple perspectives, thereby enhancing the effectiveness of learning. The experimental results show that our approaches facilitate models acquiring more feature information than baseline neural models.