MEGO neural optimizer beats general-purpose solvers with mixture-of-experts
One model trained on classic problems outperforms commercial solvers on real-world binary optimization…
Binary optimization problems — from scheduling to circuit design — are everywhere in engineering and science, but each often requires a custom-crafted solver. Now a team of researchers from China and Singapore has introduced MEGO, a neural optimizer built on a mixture-of-experts (MoE) architecture that tackles a broad range of black-box binary optimization problems with minimal customization. MEGO is trained without any domain-specific knowledge, learning instead to route each new problem instance to the most relevant expert model within its MoE framework.
The results are striking: even when trained only on classic benchmark problems, MEGO generalizes effectively to unseen, complex real-world problem classes. Against widely-used general-purpose optimizers, it delivers both better solution quality and faster runtimes across six distinct problem families. Beyond raw performance, MEGO introduces a computational approach to quantify the similarity between optimization problems — a fundamentally different method from traditional analysis-based classification. This opens the door to automatic problem clustering and could guide future algorithm selection.
- MEGO uses a mixture-of-experts (MoE) architecture with a learned routing policy to dynamically activate the best expert for each binary optimization instance.
- Trained solely on classic problems, it outperforms general-purpose optimizers on real-world applications spanning six disciplines.
- Introduces a computational method to quantify problem similarity, enabling automated problem classification beyond traditional analysis.
Why It Matters
A single, versatile optimizer that replaces custom solvers, saving months of engineering effort across logistics, hardware design, and more.