Neural QAOA²: Differentiable Graph Partitioning and Parameter Initialization for Quantum Combinatorial Optimization
A recent study introduces Neural QAOA², a fully differentiable framework designed for quantum combinatorial optimization. This approach tackles the scalability challenges associated with the quantum approximate optimization algorithm (QAOA) by simultaneously creating graph partitions and setting initial parameters. Utilizing a generative evaluative network (GEN) alongside a differentiable quantum evaluator, it serves as a reliable performance surrogate, facilitating direct gradient guidance to learn the relationship between graph topology and the configurations of partitions and parameters.
Key facts
- Neural QAOA² is an end-to-end differentiable framework for quantum combinatorial optimization.
- It jointly generates graph partitions and initial parameters for QAOA.
- The method uses a generative evaluative network (GEN).
- A differentiable quantum evaluator serves as a high-fidelity performance surrogate.
- The framework provides direct gradient guidance for learning.
- It addresses limitations of existing divide-and-conquer frameworks like QAOA².
- Existing methods suffer from misalignment between heuristic partitioning metrics and quantum optimization goals.
- Topology-blind parameter initialization leads to optimization cold starts.
Entities
—