WeCon neural solver cuts inference time 40% on multi-objective optimization

WeCon, introduced by Xuan Wu and nine co-authors, tackles Multi-Objective Combinatorial Optimization Problems (MOCOPs) – tasks where multiple conflicting objectives must be balanced, such as cost vs. quality in logistics. Existing neural solvers often inject weight vectors only once, limiting context, or suffer from signal dilution during decoding. WeCon addresses this with a custom encoder layer featuring three attention blocks and a Gated Residual Fusion (GRF) block that harmoniously blends instance features and weights. In the decoder, a plug-and-play Residual Fusion (RF) block prevents weight-signal dilution.

To improve training, WeCon introduces Efficient Preference Optimization (EPO), which constructs higher-quality solution pairs instead of relying on random sampling. This yields more informative training signals. Across four MOCOP variants at different problem scales and distributions, WeCon achieves HyperVolume (HV) values comparable to the state-of-the-art POCCO-W solver while slashing inference time by 40%. Ablation studies validate the contributions of the GRF, RF, and EPO components. The work is published on arXiv (2605.22876) and could significantly accelerate real-world multi-objective optimization in supply chain, finance, and engineering.