Adversarial Instance Generation and Robust Training for Neural Combinatorial Optimization with Multiple Objectives
Researchers have developed a new framework to enhance the robustness of deep reinforcement learning solvers for multi-objective combinatorial optimization problems. This framework includes an adversarial attack method to generate challenging instances that reveal solver weaknesses and an adversarial training strategy to improve performance on unseen data. Experiments on various optimization problems demonstrated that the proposed attack effectively identifies solver vulnerabilities, while the defense mechanism significantly boosts the robustness and generalizability of neural solvers. AI
IMPACT Enhances the reliability of AI solvers for complex optimization tasks, potentially improving efficiency in logistics and operations.