Generative Adversarial Training for Neural Combinatorial Optimization Models

Recent studies show that deep neural networks can be trained to learn good heuristics for various Combinatorial Optimization Problems (COPs). However, it remains a great challenge for the trained deep optimization models to generalize to distributions different from the training one. To address this issue, we propose a general framework, Generative Adversarial Neural Combinatorial Optimization (GANCO) which is equipped with another deep model to generate training instances for the optimization model, so as to improve its generalization ability. The two models are trained alternatively in an adversarial way, where the generation model is trained by reinforcement learning to find instance distributions hard for the optimization model. We apply the GANCO framework to two recent deep combinatorial optimization models, i.e., Attention Model (AM) and Policy Optimization with Multiple Optima (POMO). Extensive experiments on various problems such as Traveling Salesman Problem, Capacitated Vehicle Routing Problem, and 0-1 Knapsack Problem show that GANCO can significantly improve the generalization ability of optimization models on various instance distributions, with little sacrifice of performance on the original training distribution.