Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2406.09614 · Machine Learning Science and Technology

Trainability issues in quantum policy gradients

André Sequeira, Luís Paulo Santos, Luís Soares Barbosa

Published Jul 26, 2024• 5 citations
arXivHTMLPDF

Abstract

Abstract This research explores the trainability of Parameterized Quantum Circuit-based policies in Reinforcement Learning, an area that has recently seen a surge in empirical exploration. While some studies suggest improved sample complexity using quantum gradient estimation, the efficient trainability of these policies remains an open question. Our findings reveal significant challenges, including standard Barren Plateaus with exponentially small gradients and gradient explosion. These phenomena depend on the type of basis-state partitioning and the mapping of these partitions onto actions. For a polynomial number of actions, a trainable window can be ensured with a polynomial number of measurements if a contiguous-like partitioning of basis-states is employed. These results are empirically validated in a multi-armed bandit environment.

DOI
10.1088/2632-2153/ad6830

Related papers

Trainability Issues in Quantum Policy Gradients with Softmax Activations

André Sequeira, Luís Paulo Santos, Luís Soares Barbosa

This research addresses the trainability of Parameterized Quantum Circuit-based Softmax policies in Reinforcement Learning. We assess the trainability of these policies by examining the scaling of the expected value of the partial derivative of the log policy objective function. Here, we assume the hardware-efficient ansatz with blocks forming local 2-designs. In this setting, we show that if each expectation value representing the action's numerical preference is composed of a global observable, it leads to exponentially vanishing gradients. In contrast, for n-qubit systems, if the observables are$\log(n)$-local, the gradients vanish polynomially with the number of qubits provided$\mathrm{O}(\log n)$depth. We also show that the expectation of the gradient of the log policy objective depend on the entire action space. Thus, even though global observables lead to concentration, the gradient signal can still be propagated in the presence of at least a single local observable. We validate the theoretical predictions in a series of ansatze and evaluate the performance of local and global observables in a multi-armed bandit setting.

IEEE International Conference on Quantum Computing and Engineering (QCE)•Sep 15, 2024•1 citation

Policy gradients using variational quantum circuits

André Sequeira, Luís Paulo Santos, Luís Soares Barbosa

Abstract Variational quantum circuits are being used as versatile quantum machine learning models. Some empirical results exhibit an advantage in supervised and generative learning tasks. However, when applied to reinforcement learning, less is known. In this work, we considered a variational quantum circuit composed of a low-depth hardware-efficient ansatz as the parameterized policy of a reinforcement learning agent. We show that an 𝜖-approximation of the policy gradient can be obtained using a logarithmic number of samples concerning the total number of parameters. We empirically verify that such quantum models behave similarly to typical classical neural networks used in standard benchmarking environments and quantum control, using only a fraction of the parameters. Moreover, we study the barren plateau phenomenon in quantum policy gradients using the Fisher information matrix spectrum.

arXiv:2203.10591•Quantum Machine Intelligence•Apr 11, 2023•42 citations

A New Framework for Quantum Reinforcement Learning

Alireza Tavakoli, Mohsen Heidari

Quantum reinforcement learning (QRL) has emerged as a promising approach for controlling quantum systems and accelerating learning tasks with quantum resources. However, most existing QRL formulations still retain a classical reinforcement learning interface: states and actions are often encoded in computational bases and rewards are extracted through measurements. This interface is restrictive for genuinely quantum environments: choosing a measurement can discard entanglement information, while repeated intermediate measurements collapse the evolving state and require costly re-preparation. This motivates a fully QRL framework in which the agent, environment, rewards, and return accumulation are all quantum. In this paper, we take a step toward this goal and introduce a new QRL framework in which both the policy and the environment are modeled by conditional quantum channels. We then propose a quantum policy-gradient algorithm for parameterized conditional quantum channels with a trainable control basis. We also derive finite- and infinite-horizon gradient estimators using a new parameter-shift rule. This rule is needed because policy parameters are shared across repeated interaction blocks and the infinite-horizon objective is not a finite circuit; rather than shifting every occurrence and evaluating the whole trajectory, our estimator shifts only one sampled policy block. Sampling this block from a geometric distribution yields an unbiased gradient estimator under bounded rewards. Finally, we establish convergence of the resulting stochastic gradient-ascent procedure to stationary points under standard smoothness and step-size assumptions.

arXiv:2611.14731•Nov 10, 2026•0 citations

Data re-uploading and expressivity/trainability trade-off on quantum policy gradients

Bernardo Manuel Barbosa Soares

This thesis investigates the integration of quantum computing into machine learning, specifically focusing on reinforcement learning. The study focuses on the building of a code library to enable the exploration of the use of variational quantum circuits as function approximators within reinforcement learning algorithms, specifically policy gradient methods. A comprehensive framework is developed to analyze the performance and trainability of variational quantum circuit-based reinforcement learning models, supported by this library, to understand how different quantum circuit designs and data re-uploading techniques might affect the performance and trainability. To demonstrate the library’s capabilities, experiments were conducted in classic control environments like CartPole-v1 and Acrobot-v1, evaluating the performance (measured as average return) and trainability of three quantum models. These experiments examined the effects of data re-uploading, classical postprocessing, and inverse temperature parameter scheduling on both metrics. The empirical results show that variational quantum circuits are effective function approximators, with data re-uploading significantly enhancing model performance. Additionally, the globality of the measured observables plays a critical role, with higher globality values leading to better performance. Despite increasing circuit expressiveness, the gradient magnitudes and variances remain substantial throughout training. Furthermore, the tested inverse temperature schedules, used to adjust the policy’s level of exploration, did not show any significant advantage, indicating the challenge in finding optimal settings for these schedules. The study also introduces the universal quantum classifier adapted for reinforcement learning. Contrary to the barren plateau phenomenon, it has been empirically demonstrated that although increasing the number of qubits leads to performance drops in the Acrobot environment, it does not lead to vanishing gradients during the agents’ training process. Furthermore, under specific sets of hyperparameters, all three quantum models demonstrate strong and similar performance levels in both environments.

•Dec 1, 2024•0 citations

Learning parameterized quantum circuits with quantum gradient

Keren Li, Y Wang, Pan Gao +1 more

Parameterized quantum circuits (PQCs) are crucial for quantum machine learning and circuit synthesis, enabling the practical implementation of complex quantum tasks. However, PQC learning has been largely confined to classical optimization methods, which suffer from issues like gradient vanishing. In this work, we introduce a nested optimization model, a hybrid approach that leverages quantum gradients to improve PQC learning for arbitrary polynomial-type cost functions. The proposed approach decomposes the learning problem into multiple subproblems, each aimed at learning the state identified by the quantum gradient using current circuit synthesis methods. Leveraging quantum gradients, our method detects unfavorable local stationary points via an adaptive indicator and resolves them with a guided state. Meanwhile, its warm-started, layer-wise expansion reduces susceptibility to barren-plateau. Numerically, we demonstrate the feasibility of the approach on two tasks: the Max-Cut problem and polynomial optimization. Furthermore, we benchmark against standard fixed and adaptive baselines under matched settings on unfavorable local stationary points prone instances together with reduced BP sensitivity. From the perspective of quantum algorithms, our model improves quantum optimization for polynomial-type cost functions, addressing the challenge of exponential sample complexity growth.

arXiv:2409.20044•npj Quantum Information•Feb 25, 2026•2 citations

Trainability-Simulability Tradeoff in Parameterized IQP-based Reinforcement Learning Policies

André Sequeira

Parameterized instantaneous quantum polynomial-time (IQP) circuits generate distributions conjectured to be hard to sample classically. As reinforcement learning (RL) policies, their $2^n$ measurement outcomes are typically mapped to a smaller action space by a classical decoder. This aggregation can mitigate initialization concentration while making the policy classically simulable. We characterize this tradeoff through the decoder's Fourier--Pauli expansion. Under independent uniform initialization, we express policy variance as an exact sum over groups of nonconstant Pauli components with identical active-generator signatures. Each group contributes its squared aggregate coefficient times $2^{-r}$, where $r$ is the binary rank of its active generators. An inverse-polynomial aggregate coefficient with logarithmic rank prevents exponential policy concentration. Sparse Pauli expansions and contiguous action bins permit efficient classical additive estimation of action probabilities. We introduce a majority decoder combining protected low-weight components with exponentially large exact Fourier support, preserving initialization nonconcentration while excluding polynomial-size enumeration of the exact expansion. Numerical experiments corroborate the variance laws. In ten-qubit CartPole experiments, majority achieves higher returns than recursive parity and the tested local readouts, while higher-weight Pauli sectors contribute an increasing relative share of gradient magnitude during training. A complementary 21-qubit block-chain policy outperforms an extensive pool of fully connected classical neural networks in episode efficiency, even though its bounded-width structure permits efficient classical simulation. These results guide decoder design for quantum and quantum-inspired RL, distinguishing policy nonconcentration, learning performance, and classical simulability.

arXiv:2611.10793•Nov 7, 2026•0 citations
Abstract This research explores the trainability of Parameterized Quantum Circuit-based policies in Reinforcement Learning, an area that has recently seen a surge in empirical exploration. While some studies suggest improved sample complexity using quantum gradient estimation, the efficient trainability of these policies remains an open question. Our findings reveal significant challenges, including standard Barren Plateaus with exponentially small gradients and gradient explosion. These phenomena depend on the type of basis-state partitioning and the mapping of these partitions onto actions. For a polynomial number of actions, a trainable window can be ensured with a polynomial number of measurements if a contiguous-like partitioning of basis-states is employed. These results are empirically validated in a multi-armed bandit environment.
This research addresses the trainability of Parameterized Quantum Circuit-based Softmax policies in Reinforcement Learning. We assess the trainability of these policies by examining the scaling of the expected value of the partial derivative of the log policy objective function. Here, we assume the hardware-efficient ansatz with blocks forming local 2-designs. In this setting, we show that if each expectation value representing the action's numerical preference is composed of a global observable, it leads to exponentially vanishing gradients. In contrast, for n-qubit systems, if the observables arelog⁡(n)\log(n)log(n)-local, the gradients vanish polynomially with the number of qubits providedO(log⁡n)\mathrm{O}(\log n)O(logn)depth. We also show that the expectation of the gradient of the log policy objective depend on the entire action space. Thus, even though global observables lead to concentration, the gradient signal can still be propagated in the presence of at least a single local observable. We validate the theoretical predictions in a series of ansatze and evaluate the performance of local and global observables in a multi-armed bandit setting.
Abstract Variational quantum circuits are being used as versatile quantum machine learning models. Some empirical results exhibit an advantage in supervised and generative learning tasks. However, when applied to reinforcement learning, less is known. In this work, we considered a variational quantum circuit composed of a low-depth hardware-efficient ansatz as the parameterized policy of a reinforcement learning agent. We show that an 𝜖-approximation of the policy gradient can be obtained using a logarithmic number of samples concerning the total number of parameters. We empirically verify that such quantum models behave similarly to typical classical neural networks used in standard benchmarking environments and quantum control, using only a fraction of the parameters. Moreover, we study the barren plateau phenomenon in quantum policy gradients using the Fisher information matrix spectrum.
Quantum reinforcement learning (QRL) has emerged as a promising approach for controlling quantum systems and accelerating learning tasks with quantum resources. However, most existing QRL formulations still retain a classical reinforcement learning interface: states and actions are often encoded in computational bases and rewards are extracted through measurements. This interface is restrictive for genuinely quantum environments: choosing a measurement can discard entanglement information, while repeated intermediate measurements collapse the evolving state and require costly re-preparation. This motivates a fully QRL framework in which the agent, environment, rewards, and return accumulation are all quantum. In this paper, we take a step toward this goal and introduce a new QRL framework in which both the policy and the environment are modeled by conditional quantum channels. We then propose a quantum policy-gradient algorithm for parameterized conditional quantum channels with a trainable control basis. We also derive finite- and infinite-horizon gradient estimators using a new parameter-shift rule. This rule is needed because policy parameters are shared across repeated interaction blocks and the infinite-horizon objective is not a finite circuit; rather than shifting every occurrence and evaluating the whole trajectory, our estimator shifts only one sampled policy block. Sampling this block from a geometric distribution yields an unbiased gradient estimator under bounded rewards. Finally, we establish convergence of the resulting stochastic gradient-ascent procedure to stationary points under standard smoothness and step-size assumptions.
This thesis investigates the integration of quantum computing into machine learning, specifically focusing on reinforcement learning. The study focuses on the building of a code library to enable the exploration of the use of variational quantum circuits as function approximators within reinforcement learning algorithms, specifically policy gradient methods. A comprehensive framework is developed to analyze the performance and trainability of variational quantum circuit-based reinforcement learning models, supported by this library, to understand how different quantum circuit designs and data re-uploading techniques might affect the performance and trainability. To demonstrate the library’s capabilities, experiments were conducted in classic control environments like CartPole-v1 and Acrobot-v1, evaluating the performance (measured as average return) and trainability of three quantum models. These experiments examined the effects of data re-uploading, classical postprocessing, and inverse temperature parameter scheduling on both metrics. The empirical results show that variational quantum circuits are effective function approximators, with data re-uploading significantly enhancing model performance. Additionally, the globality of the measured observables plays a critical role, with higher globality values leading to better performance. Despite increasing circuit expressiveness, the gradient magnitudes and variances remain substantial throughout training. Furthermore, the tested inverse temperature schedules, used to adjust the policy’s level of exploration, did not show any significant advantage, indicating the challenge in finding optimal settings for these schedules. The study also introduces the universal quantum classifier adapted for reinforcement learning. Contrary to the barren plateau phenomenon, it has been empirically demonstrated that although increasing the number of qubits leads to performance drops in the Acrobot environment, it does not lead to vanishing gradients during the agents’ training process. Furthermore, under specific sets of hyperparameters, all three quantum models demonstrate strong and similar performance levels in both environments.
Parameterized quantum circuits (PQCs) are crucial for quantum machine learning and circuit synthesis, enabling the practical implementation of complex quantum tasks. However, PQC learning has been largely confined to classical optimization methods, which suffer from issues like gradient vanishing. In this work, we introduce a nested optimization model, a hybrid approach that leverages quantum gradients to improve PQC learning for arbitrary polynomial-type cost functions. The proposed approach decomposes the learning problem into multiple subproblems, each aimed at learning the state identified by the quantum gradient using current circuit synthesis methods. Leveraging quantum gradients, our method detects unfavorable local stationary points via an adaptive indicator and resolves them with a guided state. Meanwhile, its warm-started, layer-wise expansion reduces susceptibility to barren-plateau. Numerically, we demonstrate the feasibility of the approach on two tasks: the Max-Cut problem and polynomial optimization. Furthermore, we benchmark against standard fixed and adaptive baselines under matched settings on unfavorable local stationary points prone instances together with reduced BP sensitivity. From the perspective of quantum algorithms, our model improves quantum optimization for polynomial-type cost functions, addressing the challenge of exponential sample complexity growth.
Parameterized instantaneous quantum polynomial-time (IQP) circuits generate distributions conjectured to be hard to sample classically. As reinforcement learning (RL) policies, their 2n2^n2n measurement outcomes are typically mapped to a smaller action space by a classical decoder. This aggregation can mitigate initialization concentration while making the policy classically simulable. We characterize this tradeoff through the decoder's Fourier--Pauli expansion. Under independent uniform initialization, we express policy variance as an exact sum over groups of nonconstant Pauli components with identical active-generator signatures. Each group contributes its squared aggregate coefficient times 2−r2^{-r}2−r, where rrr is the binary rank of its active generators. An inverse-polynomial aggregate coefficient with logarithmic rank prevents exponential policy concentration. Sparse Pauli expansions and contiguous action bins permit efficient classical additive estimation of action probabilities. We introduce a majority decoder combining protected low-weight components with exponentially large exact Fourier support, preserving initialization nonconcentration while excluding polynomial-size enumeration of the exact expansion. Numerical experiments corroborate the variance laws. In ten-qubit CartPole experiments, majority achieves higher returns than recursive parity and the tested local readouts, while higher-weight Pauli sectors contribute an increasing relative share of gradient magnitude during training. A complementary 21-qubit block-chain policy outperforms an extensive pool of fully connected classical neural networks in episode efficiency, even though its bounded-width structure permits efficient classical simulation. These results guide decoder design for quantum and quantum-inspired RL, distinguishing policy nonconcentration, learning performance, and classical simulability.