This paper proposes LocalEscaper, a novel weakly supervised learning framework for the large-scale Traveling Salesman Problem (TSP). Conventional supervised learning-based neural network solutions require a large amount of high-quality labeled data, while reinforcement learning-based solutions, while low-data dependency, suffer from inefficiency. LocalEscaper combines the strengths of supervised and reinforcement learning to enable effective training with low-quality labeled data. Specifically, it introduces a local reconstruction strategy that alleviates the local optimum problem of existing local reconstruction methods, thereby improving solution quality. Experimental results on synthetic and real-world datasets demonstrate that LocalEscaper achieves remarkable results, outperforming existing neural network solutions.
