Researchers have developed a new approach called MEMENTO to improve the performance of neural solvers for routing problems. These problems are challenging due to their hard nature and often require heuristics or reinforcement learning (RL) methods. However, existing learned methods lack adaptability and fail to utilize newly available information within budget constraints. MEMENTO addresses this by leveraging memory to dynamically adjust action distributions based on previous decisions. It has been validated on several routing problems, including the Traveling Salesman and Capacitated Vehicle Routing problems, outperforming other methods such as tree-search and policy-gradient fine-tuning. --- Why it matters: This matters because it can improve the efficiency of neural solvers for complex routing problems, which are crucial in many real-world applications. By leveraging memory to adapt to specific instances and utilize available computational budget, MEMENTO has the potential to enhance the scalability and data-efficiency of these solvers. Source: https://arxiv.org/abs/2406.16424