Learning to Learn with Quantum Optimization via Quantum Neural Networks

TL;DR

The paper tackles the challenge of efficiently optimizing QAOA parameters on NISQ devices by introducing a quantum meta-learning approach that uses a Quantum Long Short-Term Memory (QLSTM) network as a parameter optimizer. By training the QLSTM on small problem instances, the method learns transferable parameter-update rules that generalize to larger graphs, reducing the number of iterations and improving solution quality for Max-Cut and Sherrington-Kirkpatrick problems. The authors formalize the meta-learning objective, derive a gradient-based training scheme via backpropagation through time, and demonstrate substantial gains over classical optimizers and standard LSTM, including transfer-learning across problem sizes. This approach offers a scalable, robust pathway to employ variational quantum algorithms more effectively in the NISQ era, with potential applicability to other quantum optimization tasks and quantum-enhanced learning systems.

Abstract

Quantum Approximate Optimization Algorithms (QAOA) promise efficient solutions to classically intractable combinatorial optimization problems by harnessing shallow-depth quantum circuits. Yet, their performance and scalability often hinge on effective parameter optimization, which remains nontrivial due to rugged energy landscapes and hardware noise. In this work, we introduce a quantum meta-learning framework that combines quantum neural networks, specifically Quantum Long Short-Term Memory (QLSTM) architectures, with QAOA. By training the QLSTM optimizer on smaller graph instances, our approach rapidly generalizes to larger, more complex problems, substantially reducing the number of iterations required for convergence. Through comprehensive benchmarks on Max-Cut and Sherrington-Kirkpatrick model instances, we demonstrate that QLSTM-based optimizers converge faster and achieve higher approximation ratios compared to classical baselines, thereby offering a robust pathway toward scalable quantum optimization in the NISQ era.

Figures (8)

• Figure 1: Schematic representation of the QAOA, where a classical optimizer iteratively updates the variational parameters $(\vec{\gamma}, \vec{\beta})$ based on measurement outcomes to maximize the expectation value of the cost Hamiltonian $\langle H_C \rangle$. Each layer $k$ in the quantum circuit consists of two parts: first, an evolution under the cost Hamiltonian, $e^{-i\gamma_k H_C}$, and second, a mixing operation given by $e^{-i\beta_k H_B}$, with $H_B = \sum_i X_i$. In practice, this mixing evolution can be viewed as a product of single-qubit rotations, $\prod_i e^{-i\beta_k X_i}$. After measurement, the resulting bit strings are used to estimate $\langle H_C \rangle$, which serves as the objective function for the classical optimizer to update $\vec{\gamma}$ and $\vec{\beta}$ in subsequent iterations, thereby refining the solution to the optimization problem.
• Figure 2: Schematic of a standard LSTM cell. The input gate, forget gate, and output gate regulate the information flow into, through, and out of the cell state ($C_t$). This gating mechanism enables the network to capture both short-term and long-term dependencies in the data.
• Figure 3: Schematic diagram of the QLSTM architecture. (a) Overview of the QLSTM cell, where classical nonlinear operations in standard LSTM gates are replaced with variational quantum circuits (VQCs), enabling quantum-enhanced memory processing. (b) Detailed structure of the VQC used within each QLSTM gate. The circuit comprises a data encoding layer $U(\vec{x})$, where classical inputs $\vec{x} = (x_1, x_2, \dots, x_i)$ are mapped to quantum states via parameterized rotations (e.g., $R_y(x_i)$), followed by a trainable variational block $V(\vec{\theta})$, consisting of parameterized single-qubit gates (e.g., $R_x, R_y, R_z$) and entangling operations. Measurement outcomes provide nonlinear transformations that act as gate activations within the QLSTM cell.
• Figure 4: Schematic representation of the QAOA framework enhanced with a QLSTM optimizer. (a) The parameterized quantum circuit follows the standard QAOA structure but replaces the classical optimizer with a recurrent quantum-classical model. Following measurement, the estimated expectation value $\langle H_C \rangle$ is provided to a QLSTM-based optimizer, which learns to generate optimized variational parameters $(\vec{\gamma}, \vec{\beta})$ by capturing temporal dependencies across optimization steps. (b) Architecture of the QLSTM optimizer. At each time step $t$, a hybrid QLSTM block receives the current and past QAOA outputs, including previous parameters $\theta_{t-1}$ and measurement outcomes $y_{t-1}$, along with internal hidden states $h_{t-1}$. The hybrid model integrates classical processing with quantum subroutines on a QPU to generate the next parameter set $\theta_t = (\vec{\gamma}_t, \vec{\beta}_t)$, enabling adaptive and history-aware learning dynamics that aim to improve convergence and generalization in variational quantum algorithms.
• Figure 5: Comparison of LSTM (red) and QLSTM (blue) optimization trajectories in the single-layer QAOA parameter space, evaluated on a randomly generated graph with 10 nodes and edge probability $P = 0.4$. The figure highlights the faster convergence and lower final cost achieved by the QLSTM optimizer during training. Both trajectories start from the same initial point (Step 0 at $\gamma = 0, \beta = 0$).