Self-Improved Learning for Scalable Neural Combinatorial Optimization: Overcoming the Limitations of Existing Methods on Large-Scale Problems

This work proposes a novel Self-Improved Learning (SIL) method that enables neural combinatorial optimization (NCO) models to be directly trained on large-scale combinatorial optimization problems with up to 100K nodes, without requiring any labeled data. SIL leverages an efficient self-improved mechanism that iteratively generates better solutions as pseudo-labels to guide model training, significantly boosting the scalability of NCO models.