IEEE Taskforce on Evolutionary Scheduling and Combinatorial Optimisation Webinar #18, by Dr. Zhenkun Wang from Southern University of Science and Technology. More information can be found at https://homepages.ecs.vuw.ac.nz/~yimei/iee...

Neural Combinatorial Optimization (NCO) aims to learn directly from data a neural network that can solve complex combinatorial optimization problems like the travel salesman problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP). Existing NCO models achieve good performance on small-scale problem instances but fail to generalize to solve large-scale ones. In this talk, I will systematically review these existing NCO methods and introduce their basic principles. Thereafter, I will reveal some possible reasons for their poor large-scale generalization capability. Beyond that, I will introduce a new model structure, Light Encoder and Heavy Decoder (LEHD), and illustrate why it can achieve excellent large-scale generalization performance. A data-efficient training scheme and a flexible solution construction mechanism will be presented along with the LEHD model. Finally, some experimental results on TSP and CVRP will be provided to indicate the superiority of the LEHD.