Human-like Solutions in Combinatorial Optimization via Learning and Search

A recent study explores the near-optimal strategies humans employ to tackle Euclidean traveling salesman problems (TSP) despite their computational constraints. The researchers gathered human-generated solutions from a wide array of TSP scenarios and evaluated them against neural policies derived from Pointer Networks, which utilize recurrent neural networks with attention-driven pointing methods. These networks underwent training through various approaches, including reinforcement learning, supervised learning from optimal routes, supervised learning from human-generated tours, and reinforcement learning fine-tuning. The objective of this research is to gain insights into the characteristics of human-like tours and the learning processes behind such solutions.