A Cyclic Layerwise QAOA Training

TL;DR

This paper tackles the training bottleneck of Multi-Angle QAOA by introducing Orbit-QAOA, a round-robin, layerwise optimization with selective freezing that repeatedly revisits all layers to adapt to evolving circuit expressibility. By updating only one complete layer per step and dynamically freezing stabilized layers via an activeness threshold, Orbit-QAOA achieves substantial reductions in training steps (up to 81.8% in benchmarks) and per-step overhead while maintaining MA-like final ACR (≈1) across diverse graphs. The authors demonstrate Orbit-QAOA’s effectiveness not only for MA-QAOA but also its extensions to the Quantum Alternating Operator Ansatz and single-angle QAOA, underscoring broad applicability for scalable hybrid quantum-classical optimization. Overall, Orbit-QAOA offers a practical, robust strategy to balance expressibility and training efficiency, enabling faster convergence with comparable quality for large-scale QAOA deployments.

Abstract

The quantum approximate optimization algorithm (QAOA) is a hybrid quantum-classical algorithm for solving combinatorial optimization problems. Multi-angle QAOA (MA-QAOA), which assigns independent parameters to each Hamiltonian operator term, achieves superior approximation performance even with fewer layers than standard QAOA. Unfortunately, this increased expressibility can raise the classical computational cost due to a greater number of parameters. The recently proposed Layerwise MA-QAOA (LMA-QAOA) reduces this overhead by training one layer at a time, but it may suffer from obtaining the precise solution due to the previously fixed parameters. This work addresses two questions for efficient MA-QAOA training: (i) What is the optimal granularity for parameter updates per epoch, and (ii) How can we get precise final cost function results while only partially updating the parameters per epoch? Despite the benefit of reducing the parameters that update per epoch can reduce the classical computation overhead, too fine or coarse a granularity of Hamiltonian update can degrade the MA-QAOA training efficiency. We find that optimizing one complete layer per epoch is an efficient granularity. Moreover, selectively retraining each layer by tracking gradient variations can achieve a final cost function equivalent to the standard MA-QAOA while lowering the parameter update overhead. Based on these insights, we propose Orbit-QAOA, which cyclically revisits layers and selectively freezes stabilized parameters. Across diverse graph benchmarks, Orbit-QAOA reduces training steps by up to 81.8%, reduces approximation ratio error by up to 72x compared to the unified stop condition-applied enhanced LMA-QAOA, and achieves equivalent approximation performance compared to the standard MA-QAOA.

Figures (12)

• Figure 1: Comparison of MA-QAOA training strategies: (a) Standard MA-QAOA trains all parameters simultaneously across all layers in each epoch. (b) LMA-QAOA incrementally adds layers (and freezes previously optimized layers), reducing per-epoch optimization cost. (c) The proposed Round-Robin training revisits each layer at a time in a cyclic fashion.
• Figure 2: Training performance of 9-qubit 5-layer MA-QAOA circuits on graphs for Power-Law and Sherrington–Kirkpatrick models under sublayer-partitioned optimization by round-robin training approach. The value $k$ denotes the number of sublayers into which each QAOA layer is partitioned. In other words, 1/$k$ of the single-layer's parameters are updated per step.
• Figure 3: Training performance comparison of 11-qubit Orbit-QAOA circuits under the two update strategies: Sequential vs. Random selection. Each subfigure shows the evolution of the ACR over training steps for QAOA circuits targeting different graph models (Path, Power-Law, and Sherrington-Kirkpatrick) ordered from left to right. The top row corresponds to circuits with $p=5$ layers, and the bottom row corresponds to $p=10$ layers.
• Figure 4: Evaluations of training performance of shallow Orbit-QAOA circuits across different numbers of layers (p = 1, 2, 3) and types of target graphs. Each subfigure shows the approximated cut ratio over training steps for graph instances with 5, 10, 15, and 20 nodes (left to right), under three target graph models: Path (top row), Power-Law (middle row), and Sherrington-Kirkpatrick (bottom row).
• Figure 5: Comparison of 6-qubit 5-layer QAOA circuits' training processes across different optimization strategies (MA, LMA, LMA+, RR, and Orbit) on various graph models. Each graph shows the evolution of the ACR over training steps for a specific network type: Power-Law (PL), Erdős–Rényi (ER), Barabási–Albert (BA), Bianconi–Barabási (BB), Watts–Strogatz (WS), and Sherrington–Kirkpatrick (SK). The graphs are arranged from left to right in order of increasing average degree (i.e., number of edges divided by number of nodes).