1

Generative AI-enabled Quantum Computing

Networks and Intelligent Resource Allocation
Minrui Xu, Dusit Niyato, Jiawen Kang, Zehui Xiong, Yuan Cao, Yulan Gao, Chao Ren, and Han Yu

Abstract—Quantum computing networks enable scalable col- infeasible for classical computers to solve in a reasonable
laboration and secure information exchange among multiple clas- period of time. However, quantum computing networks are
sical and quantum computing nodes while executing large-scale still in a developing stage. Currently, quantum channels can
generative AI computation tasks and advanced quantum algo-
rithms. Quantum computing networks overcome limitations such generate several kilobytes of secret keys per second, and Noisy
arXiv:2401.07120v1 [[Link]] 13 Jan 2024

as the number of qubits and coherence time of entangled pairs Intermediate-Scale Quantum (NISQ) devices are limited to
and offer advantages for generative AI infrastructure, including maintaining only hundreds of qubits.
enhanced noise reduction through distributed processing and In quantum computing networks, devising effective resource
improved scalability by connecting multiple quantum devices. allocation schemes is a critical but unresolved issue, which is
However, efficient resource allocation in quantum computing
networks is a critical challenge due to factors including qubit essential for the seamless integration of classical computing
variability and network complexity. In this article, we propose an and networking infrastructure for generative AI with emerging
intelligent resource allocation framework for quantum computing quantum devices [4]. The inherent dynamics of quantum
networks to improve network scalability with minimized resource resources, such as qubit instability and decoherence, pose
costs. To achieve scalability in quantum computing networks, challenges in providing quantum cryptography and quantum
we formulate the resource allocation problem as stochastic
programming, accounting for the uncertain fidelities of qubits computing services, which are crucial for securing and enhanc-
and entangled pairs. Furthermore, we introduce state-of-the-art ing classical infrastructure. Specific challenges in quantum
reinforcement learning (RL) algorithms, from generative learning computing networks include uncertain fidelities of qubits and
to quantum machine learning for optimal quantum resource entangled pairs, the coexistence of multiple communication
allocation to resolve the proposed stochastic resource allocation protocols, and the heterogeneity of networks comprising di-
problem efficiently. Finally, we optimize the resource allocation
in heterogeneous quantum computing networks supporting quan- verse nodes and connections. Addressing these challenges
tum generative learning applications and propose a multi-agent necessitates the efficient allocation of quantum resources.
RL-based algorithm to learn the optimal resource allocation This includes allocating additional qubits for error correction
policies without prior knowledge. and utilizing entangled pairs for quantum state purification.
Index Terms—Quantum key distribution, distributed quantum Furthermore, developing translation policies is vital for in-
computing, resource allocation, reinforcement learning. teroperability among networks utilizing different protocols.
Effective resource allocation frameworks are also crucial to
I. I NTRODUCTION ensure seamless connectivity and functionality across various
types of quantum nodes and links.
Quantum computing networks can integrate existing net-
Fortunately, the inherent uncertainties in quantum comput-
working and computing infrastructure with quantum com-
ing networks can be effectively modeled and addressed by
munication and computing devices to enhance network se-
utilizing reinforcement learning (RL), which is capable of
curity, scalability, and sustainability [1]. Specifically, incor-
modeling the dynamics of quantum links and devices [5].
porating Quantum Key Distribution (QKD) into data center
RL agents learn optimal quantum resource allocation policies
networks [2], especially for generative AI systems, shows
through interactive engagement with the quantum computing
significant promise in bolstering privacy protection during
network environment. To optimize this interaction, state-of-
the training and inference phases of large foundation mod-
the-art RL algorithms employ deep, quantum, or generative
els [3]. Furthermore, the interconnection of quantum devices
neural networks for policy representation. Specifically, deep
in distributed quantum computing architectures enables the
learning enables RL agents to simplify the approximation of
tackling of complex computational tasks, which are currently
the Q-function in extensive state and action spaces. Addition-
Minrui Xu, Dusit Niyato, Yulan Gao, and Han Yu are with the School ally, generative AI enables RL agents to create interaction
of Computer Science and Engineering, Nanyang Technological Univer- trajectories and support long-term decision-making in quantum
sity, Singapore 639798, Singapore (e-mail: minrui001@[Link], dniy-
ato@[Link], [Link]@[Link], [Link]@[Link]). Jiawen Kang is computing networks. Moreover, RL agents using quantum
with the School of Automation, Guangdong University of Technology, China policies, parameterized by quantum neural networks, can expe-
(e-mail: kjwx886@[Link]). Zehui Xiong is with the Pillar of Information dite action selection in centralized quantum resource allocation
Systems Technology and Design, Singapore University of Technology and
Design, Singapore 487372, Singapore (e-mail: zehui xiong@[Link]). scenarios. The synergistic benefits of integrating generative AI
Chao Ren is with the School of Electrical and Electronic Engineering, with quantum computing are manifold:
Nanyang Technological University, Singapore 639798, Singapore (e-mail: • Generative AI for quantum computing networks: Gen-
[Link]@[Link]). Yuan Cao is with the School of Communications and
Information Engineering, Nanjing University of Posts and Telecommunica- erative AI significantly enhances the sample efficiency
tions, Nanjing 210003, China (e-mail: yuancao@[Link]). and training stability of intelligent resource allocation
 2

Generative AI-enabled Action

network management Agent

DRL GRL QRL

Functions: Agents make decisions
based on network state and their Reward Deep reinforcement learning (DRL)

capabilities Quantum resource allocation

Advantages: Minimize long-term enviroment Understand states and
expected costs and optimize immediate State determine actions using
and future gains​ Reservation and on-demand deep neural networks
Quantum networks
resource allocation
for generative AI Qubits
Quantum channels Generative AI
models Generative reinforcement learning (GRL)
Functions: Utilizes quantum
computing and communication for AI
model training and inference
Synthesized Make actions based on
Advantages: Offers increased security, datasets Entangled pairs Qubits the planning future
speed, energy efficiency, trajectory
Uncertain requests and network
and innovative architectures
Heterogeneous conditions
quantum computing networks Drone node
Quantum reinforcement learning (QRL)
Functions: Integrates existing Satellite
networking and computing with Entangled node
Understand states and
quantum communication and pairs
determine actions using
computing quantum neural
Advantages: Enhances network Quantum Quantum networks
security, scalability, and sustainability channels node

Fig. 1: An overview of intelligent resource allocation in quantum computing networks.

algorithms, which can be achieved through long-horizon generative AI can be enhanced in quantum computing
planning and the synthesis of trajectory experiences [6]. networks. On the other hand, generative AI can enable
• Quantum computing networks for generative AI: By more effective and robust intelligent resource allocation
utilizing quantum computing and communication, the algorithms in quantum computing networks.
training and inference processes of generative AI models • We highlight the potential applications and challenges
can be made more secure, rapid, and energy-efficient. for resource allocation in quantum computing networks,
Moreover, by leveraging the principles of quantum me- where the resource allocation problem is formulated as
chanics, generative AI has the potential to introduce stochastic optimization under the uncertain fidelities of
innovative architectures and models for a wide range of qubits and entangled pairs.
general or specific tasks. • We present a use case of multi-agent reinforcement learn-
Therefore, the synergy of generative AI and quantum comput- ing for quantum resource allocation in heterogeneous
ing networks offers the potential to develop novel algorithms quantum computing networks to illustrate the effective-
for intricate resource allocation problems, as well as expand ness of the proposed framework. The experimental results
the range of applications within quantum computing networks. demonstrate that the proposed framework can reduce the
In this article, we propose an intelligent resource allocation computation latency and energy consumption in provid-
framework for quantum computing networks for allocating ing quantum generative learning applications.
quantum resources with uncertain fidelities. First, we introduce
the quantum computing networks which consist of two main II. Q UANTUM C OMPUTING N ETWORKS
services, i.e., quantum key distribution and distributed quan- Quantum computing networks facilitate collaboration and
tum computing, and three types of quantum resources, i.e., information exchange between multiple classical computers
quantum channels, computation qubits, and entangled pairs. and quantum devices over optical backbone networks and
Furthermore, we propose an intelligent allocation framework space-air-ground integrated networks (SAGINs), as illustrated
for quantum resources in quantum computing networks. We in Fig. 1. Specifically, quantum computing networks enable the
introduce existing state-of-the-art RL algorithms, including, secure collaborative execution of large-scale computation tasks
deep reinforcement learning (DRL), deep generative learning, and advanced quantum algorithms [1]. In quantum computing
and quantum reinforcement learning (QRL). In addition, the networks, tasks like training and inference of foundation mod-
challenges in the intelligent resource allocation framework are els can be securely conducted across multiple data centers us-
highlighted. Finally, we leverage a heterogeneous quantum ing quantum cryptography, such as QKD which ensures secure
computing network that supports quantum generative learning communication by using quantum mechanics principles for
applications to validate the effectiveness of the proposed key generation and distribution. Furthermore, advanced quan-
intelligent resource allocation framework. tum algorithms can be decomposed and compiled into smaller,
Our main contributions can be summarized as: manageable quantum circuits suitable for execution on NISQ
• We propose an intelligent quantum resource allocation for devices. These NISQ devices execute quantum circuits using
quantum computing networks enabled by generative AI. qubits, the fundamental units of quantum information, and
On the one hand, the security and scalability of training quantum gates, which facilitate quantum operations. To enable
and inference of distributed machine learning systems for information exchange in a quantum network, entangled pairs,
 3

TABLE I: Comparative Summary of Quantum Computing Network Types

Network
Point-to-point Quantum Networks Multi-protocol Quantum Networks Heterogeneous Quantum Networks
Types
Establish secure connections between Integrate multiple experimental
Extend range and coverage, enhance
Objectives two quantum computing nodes using quantum networks into a global
network robustness
quantum channels quantum network
• Quantum channels
• Quantum channels
Resource • Quantum channels • Computation qubits
• Entangled pairs
Types • Entangled pairs

Fidelity degradation of quantum High complexity in coordination and

Limitations Additional protocol translation costs
signals over long distances management

i.e., pairs of qubits in a quantum-mechanically linked state, relying on the fundamental laws of quantum physics rather
are generated and distributed. This distribution allows for in- than the assumed difficulty of computational problems.
formation exchange through a process known as entanglement
swapping, a key technique for quantum communication [7].
B. Distributed Quantum Computing
By establishing hybrid connections that bridge classical and
quantum devices via quantum channels, quantum computing Distributed quantum computing involves the use of multiple
networks can address and mitigate limitations inherent in both interconnected quantum processors to collaborate to execute
classical and quantum systems, including security vulnerabil- quantum circuits of advanced quantum algorithms [1]. This
ities, quantum noise, and scalability issues. However, due to approach is designed to surpass the inherent limitations of sin-
current hardware constraints in quantum technology, quantum gle quantum processors, particularly in terms of the number of
circuits must be effectively mapped to correspond with the qubits available and overall computational power. The essence
original circuit design of the advanced quantum algorithms, of distributed quantum computing lies in the coordination and
ensuring compatibility and optimal functionality. execution of quantum algorithms across a network of inter-
connected processors. In this setup, each quantum processor
is assigned a specific portion of the overall computational
A. Quantum Key Distribution task. These processors then operate simultaneously, executing
their respective components and communicating through a
QKD is used in quantum computing networks for secure coherent quantum network. Such parallel and coordinated
peer-to-peer communication between computation nodes [2]. execution harnesses the collective computational power of
By leveraging the non-cloning theorem which states that an the network, effectively amplifying the capabilities beyond
unknown quantum state cannot be copied exactly in quantum what individual processors can achieve. Resource requirements
mechanics, QKD focuses on cryptographic key exchange via of distributed quantum computing can be estimated during
quantum channels over fiber optic links for ground-based the compiling of quantum circuits and realized during the
communications and free-space links for aerial and satellite execution through quantum Message Passing Interface (MPI)
communications. During QKD-secured exchanges between via entanglement swapping. Furthermore, the integration of
two computational nodes, both the sender and receiver deploy distributed quantum computing into existing workflow models
their quantum communication infrastructure, establishing a is facilitated by extending workflow languages. This extension
hybrid link composed of classical and quantum channels. The enables the mapping of complex quantum computing tasks to
process begins with the sender preparing and transmitting a native modeling constructs, ensuring both the portability of
sequence of quantum states, typically encoded in photons, these workflows across different quantum computing platforms
through the quantum channel to the receiver. Upon receipt, and ease of orchestration and execution.
the receiver performs quantum measurements on these states
using specific measurement bases that are compatible with
the sender’s preparation method. Afterward, the sender and C. Quantum Resources
receiver engage in a public comparison of a subset of their As listed in TABLE I, there are three main types of quantum
measurement results over a classical channel to estimate the resources in quantum computing networks, including quantum
quantum bit error rate in the transmission. They then apply channels, computation qubits, and entangled pairs.
quantum error correction protocols to rectify any discrepancies 1) Quantum Channels: To establish secure communication
in the key, followed by a process called privacy amplification. links between two computing nodes, quantum channels can
This step involves distilling a shorter, but more secure, final be utilized to exchange secure keys for message transmission.
key from the error-corrected key, eliminating any partial infor- Quantum channels facilitate the distribution of symmetric keys
mation that an eavesdropper might have gained. The distinct encoded as quantum states, enhancing communication security
advantage of QKD lies in its ability to generate secret keys against eavesdropping attacks and quantum attacks. To gener-
over a distance in an information-theoretically secure manner, ate quantum channels over optical transport networks, a suite
 4

of specialized components is required, including QKD trans- those used for the actual information exchange.
mitters, QKD receivers, trusted relays, untrusted relays, local
key managers, and the security infrastructure [2]. Specifically, D. Potential Applications
QKD transmitters prepare and send quantum states while QKD
1) Distributed Machine Learning Systems: Distributed ma-
receivers detect these states. Trusted relays are secure nodes
chine learning systems within quantum computing networks
that assist in extending the range of quantum communication,
offer a robust framework for collaborative AI model train-
while untrusted relays facilitate quantum repeater technology
ing and inference across multiple computing nodes. During
without the need for full trust in the relay. As more quantum
the collaborative training process across multiple computing
channels are allocated within these networks, there is a cor-
nodes, the raw data used for training is kept in local databases
responding increase in the secure key rates, which is crucial
for training local models [3]. Upon completion of local model
for effective quantum communication. However, a significant
training, these computing nodes engage in a process of model
challenge arises from the inherent point-to-point nature of
sharing, which involves transmitting the trained models to peer
quantum communication. Unlike classical channels that can
nodes or central servers for model aggregation. In this way,
broadcast to multiple recipients, quantum channels typically
the integrity of the original data can be preserved and the
connect just two nodes directly. To address the increasing
creation of a more robust and comprehensive global model
demands, it is crucial to allocate and reserve quantum channel
can be trained through this aggregation process. However,
resources according to user requests flexibly to improve the
the transmission of these models poses a significant risk,
scalability of quantum computing networks.
particularly in terms of the security of valuable and potentially
2) Computation Qubits: In each quantum processor, a
sensitive AI models. To mitigate the risk of these models being
certain number of qubits are generated to encode and process
intercepted, monitored, or stolen during transmission, QKD
classical information and perform operations using quantum
links are established between computing nodes.
gates [8]. While qubits can be used to run quantum al-
2) Sensor Networks, Smart Grids and Internet of Things:
gorithms for solving complex computational problems more
Sensor networks, smart grids, and the Internet of Things are
efficiently than classical computers, the limited quantity of
constantly collecting diverse types of data from environmen-
physical qubits, errors, and decoherence pose challenges in
tal parameters to user interactions for real-time monitoring,
scaling quantum computing to achieve quantum advantages.
analysis, and decision-making. During the process of data
Consequently, qubit allocation becomes a critical issue in
transmission, some highly sensitive sensing data may be at
quantum computing networks, where logical qubits, which
risk of being intercepted and stolen, which can have an impact
represent the abstract quantum information, can be mapped
on high-risk applications. For example, in secure smart grids,
onto physical qubits. This process is referred to as qubit
where data on power consumption and distribution needs
allocation. Furthermore, the qubit allocation problem can
to be protected, quantum computing networks can provide
be conceptualized as analogous to the quadratic assignment
secure communication channels through quantum key distribu-
problem in classical optimization theory, which is known to
tion [10]. This ensures that the data remains confidential and
be NP-hard, and is computationally challenging to solve as
cannot be tampered with, enhancing the overall security of the
the size of the input increases. The primary objective of qubit
smart grid system. By harnessing the advanced computational
allocation is to find an optimal mapping strategy, which aims
power of quantum computing, the analysis of voluminous
to minimize the impact of quantum noise and errors while
sensing data can be performed with heightened efficiency
simultaneously maximizing the efficiency and scalability of
and precision. This not only leads to more accurate decision-
the quantum computing network.
making but also contributes to substantial resource savings.
3) Entangled Pairs: The exchange of information between
two quantum nodes can be achieved through entanglement
swapping of entangled pairs [9], which are pairs of qubits that III. G ENERATIVE AI- ENABLED I NTELLIGENT R ESOURCE
are quantum-mechanically linked, such that the state of one A LLOCATION FOR Q UANTUM C OMPUTING N ETWORKS
(regardless of the distance separating them) instantaneously in- With the benefits of being adaptable to changing environ-
fluences the state of the other. When entanglement connections ments, RL is used to enhance the efficiency of resource allo-
span long distances, the fidelity of entangled pairs may degrade cation in quantum computing networks. For example, in the
due to imperfect measurements on quantum repeaters during resource allocation problem of QKD networks with multiple
entanglement swapping operations. To ensure the fidelity of tenants, RL-based algorithms are proposed to improve the
entangled pairs, the iterative purification operation is applied blocking probability of tenant requests and the utilization of
to the entangled pairs. This purification process involves the secret-key resources [5]. Specifically, RL agents can model the
use of multiple entangled pairs to distill a smaller number of dynamic environments as a Markov Decision Process (MDP),
pairs with higher fidelity. The fidelity of a purified entangled considering states, actions, rewards, and transition probabili-
pair is quantitatively assessed by comparing its original fidelity ties. In the environment of quantum computing networks, the
with the fidelity post-purification. By iteratively applying this MDP can usually be defined as follows:
purification process, the fidelity of entanglement connections • State space: The state space in quantum computing
can be enhanced to meet the requisite quality standards for networks encompasses key configurations and situations
reliable quantum communication. The purification process has of the network, which includes real-time monitoring
its trade-off between generating additional entangled pairs and of computing nodes, communication links, application
 5

TABLE II: Comparison of Different Learning-Based Quantum Resource Allocation Schemes (N = number of classic
computing operations, H = length of trajectories, Q = number of quantum computing operations)

Quantum Reinforcement
Schemes Deep Reinforcement Learning Deep Generative Learning
Learning
Efficient management of Offers quantum advantage with
Improved sample quality and
Advantages high-dimensional state and action quantum policy representation and
training stability
spaces optimization
Long-horizon planning via
Target Scenarios Decentralized resource allocation Centralized resource allocation
trajectory synthesizing
Computation and
O(N 2 ) O(HN ) O(QN )
Sample Complexity
Deep Q-network & Deep Decision transformer & Decision Quantum-inspired RL &
Algorithms
Deterministic Policy Gradient diffuser VQC-based Q-learning

requests, network topology, the number of qubits in quan- computational burden and time required for optimal resource
tum nodes, the quantity of both quantum and classical allocation decisions in quantum networks.
channels in quantum links, and the operational protocols 2) Deep Generative Learning: With the inherent capability
of the quantum network. of generating trajectories for extended decision-making peri-
• Action space: The action space is defined by the deci- ods, generative AI plays a pivotal role in simulating prospec-
sions made by RL agents at each time slot, which involves tive states in dynamic environments [6]. Specifically, diffusion
the allocation of available quantum resources, such as models, a type of generative AI, excel in optimizing complete
quantum channels, computation qubits, and entangled trajectories through iterative denoising steps. This process
pairs, for the efficient execution of QKD protocols and allows for the generation of intricate action distributions that
quantum algorithms. surpass the limitations of unimodal distributions. Furthermore,
• Reward function: The reward function is formulated to these models demonstrate significant advantages such as en-
reflect the net utility derived from providing quantum hanced sample quality and improved training stability. In
communication and computing services, which depends scenarios involving long-horizon planning and data synthesis,
on both the benefits of service delivery and the costs generative models are particularly beneficial. For example, the
associated with provisioning quantum resources. Decision Transformer uses a Generative Pre-trained Trans-
• State transition probability: Based on the present state former (GPT) architecture to predict actions autoregressively,
of the quantum network environment, RL agents deter- based on the desired rewards, past states, and actions. There-
mine the allocation of quantum resources. Following this fore, it can generate trajectories that match desired returns on
decision, the environment evolves into a subsequent state, RL tasks based on GPT’s generative capabilities. DRL agents
guided by a state transition probability model. can extrapolate and generate higher returns than the maximum
Subsequently, RL agents learn the policy to optimize the episode return available in the dataset, showcasing their ability
long-term expected efficiency of resource utilization under to generate novel and optimized actions. In addition, the Deci-
changing environmental conditions. A summary of existing RL sion Diffuser is a conditional generative model for sequential
algorithms can be found in TABLE II, which introduces ad- decision-making that simplifies the decision-making pipeline
vantages, target scenarios, complexity, and famous algorithms by framing it as conditional generative modeling, eliminating
of different schemes. the need for reinforcement learning. It outperforms existing
1) Deep Reinforcement Learning: With the expanding scale offline RL approaches on standard benchmarks by sampling
of quantum networks, both the state space and action space for high returns [6]. This allows it to generate future actions
in the MDP formulated by RL agents enlarge correspondingly. that achieve the desired return. Generative AI data synthesizers
This growth presents a challenge in learning an optimal policy significantly augment the quantity and variety of training
due to the “curse of dimensionality.” Fortunately, using deep data available for DRL agents. This expansion is critical
neural networks to represent complex policies, DRL addresses during the policy improvement phase, where agents rely on
this challenge to effectively manage high-dimensional state diverse experiences accumulated during the exploration phase.
and action spaces by integrating deep learning techniques However, it is important to note the heightened computation
with traditional RL. DRL’s use of neural networks empow- and sample complexity associated with these approaches.
ers agents to effectively manage high-dimensional state and 3) Quantum Reinforcement Learning: Beyond using DNNs
action spaces. For instance, the Deep Q-network (DQN) al- to represent the policy of RL agents, quantum reinforcement
gorithm, proficient in outputting discrete actions, leverages learning utilizes quantum computing to calculate the policies
deep neural networks for approximating Q-values in com- represented by quantum neural networks [11]. Leveraging
plex environments. Conversely, the Deep Deterministic Policy fundamental quantum mechanical principles like superposition
Gradient (DDPG) algorithm, suitable for continuous action and entanglement, QNNs exhibit enhanced generalization ca-
spaces, employs deep neural networks in both its actor and pabilities, offering significant improvements in policy explo-
critic components. These capabilities significantly reduce the ration and exploitation tasks. Furthermore, QRL introduces
 6

advantages such as improved sample quality and training networks not only significantly extend the coverage range
stability, particularly beneficial in centralized quantum re- of quantum communication but also enhance the practical
source allocation for large-scale and complex networks. With usability and adaptability of quantum applications, such as
these quantum advantages, notably computational acceleration, QKD and entanglement distribution, across varied environ-
QRL policies attain faster convergence rates and superior ments. However, the mobility of satellite and UAV nodes
performance, characterized by reduced inference delay. Ad- introduces greater complexity and uncertainty in resource allo-
ditionally, unlike traditional RL approaches, QRL tends to cation compared to more static terrestrial networks, primarily
exhibit lower computational and sample complexity, making due to dynamic network topologies and variable link stability.
it a more efficient solution for resource allocation in quantum To ensure seamless coverage and integration across space-
computing networks. based, aerial, and terrestrial segments, sophisticated resource
allocation frameworks are essential. These frameworks should
be capable of enhancing real-time data transmission, enabling
A. Challenges in Resource Allocation for Quantum Computing
flexible and adaptive satellite- and UAV-based QKD services,
Networks
and employing a universal provisioning model to optimize
1) Uncertain Fidelities of Qubits and Entangled Pairs: resource utilization and minimize operational costs of the
During the process of allocating quantum resources, e.g., QKD services.
qubits and entangled pairs, to execute quantum services for
users in quantum computing networks, there can be challenges
IV. U SE C ASE : M ULTI - AGENT R EINFORCEMENT
in efficiently provisioning resources due to fluctuations in user
L EARNING - BASED R ESOURCE A LLOCATION FOR
demand and quantum circuit requirements [12]. These fluctua-
Q UANTUM G ENERATIVE L EARNING
tions can be attributed to factors like varying computational re-
quirements, the specifics of quantum algorithms, and dynamic In heterogeneous quantum computing networks, comprising
changes in quantum network topologies. Additionally, natural mobile, edge, and cloud nodes, a collaborative framework
noise inherent in quantum computing and communication sys- is essential to deliver a variety of quantum services. As
tems further complicates these fluctuations. A primary source a notable service, quantum-generative learning can generate
of this uncertainty is the fidelity of qubits and entangled pairs, content with quantum-generative adversarial networks and
as it determines the reliability and accuracy of quantum states compress information through quantum autoencoders. This
during computation and transmission. For remote collaboration technique is a quantum analog of classical autoencoders,
over long distances between computing nodes, maintaining known for compressing and then reconstructing input data.
high-quality entanglement and sustained coherence times is In quantum generative learning, the quantum autoencoder ef-
essential for the successful completion of distributed quantum ficiently processes complex quantum data, offering advantages
computing tasks. To ensure the required fidelities, strategies in speed and dimensionality over classical counterparts. This
such as allocating additional qubits for quantum error cor- process, however, is computationally demanding, necessitating
rection techniques and employing entanglement purification extensive collaboration among the varied computing nodes
protocols with additional entangled pairs are implemented. within the quantum network. Given the diversity of these
2) Multiple-Protocols Quantum Networks: In the current nodes in terms of computational power and functionality,
developmental phase, various small-scale experimental quan- allocating resources optimally poses a significant challenge.
tum networks have been established, primarily focusing on Centralized approaches to resource allocation are impractical
testing and refining quantum communication and computing due to the complexity and scalability issues associated with
algorithms [13]. However, the use of disparate networking managing such a diverse and dynamic network [1]. Therefore,
protocols in these quantum networks presents a significant bar- a decentralized approach as shown in Fig. 2, possibly utilizing
rier to achieving global interconnectedness and interoperabil- multi-agent reinforcement learning (MARL), is better suited.
ity. To facilitate the realization of large-scale, interconnected In this context, each agent (node) learns to make decisions
quantum networks, protocol translation policies have been about resource allocation based on the network’s current state
proposed. These policies aim to serve as a bridge, enabling and its own capabilities, leading to a more efficient and better
communication and coordination among small-scale quantum scalable solution for supporting quantum generative learning
networks operating on different protocols. While protocol applications.
translation policies are instrumental in ensuring secure and In the multi-agent scenario for resource allocation in quan-
successful interoperability between different QKD protocols, tum generative learning, each agent navigates a complex
thus enhancing network scalability, they necessitate additional environment, modeled as a Partially Observable Markov De-
quantum resources. These resources are critical for optimizing cision Process (POMDP), which is modeled for decision-
the probability of successful translation and minimizing the making problems where the agent has incomplete knowledge
risks associated with quantum data relaying. about the environment’s state. In this context, each agent’s
3) Heterogeneous Quantum Networks: The heterogeneous observations include local service statuses and computing ca-
quantum networks comprise a diverse array of quantum nodes, pacities, providing limited but crucial information for decision-
such as terrestrial nodes, satellite nodes, and unmanned aerial making. Each agent employs a policy, i.e., a strategy or set of
vehicle (UAV) nodes, interconnected by various quantum rules, to make resource allocation decisions. These decisions
links, such as fiber links and free-space links [14]. These involve determining whether to offload quantum generative
 7

Continuous Agent Agent 1 Discrete Agent

Actor-Critic Network Continuous Agent Actor-Critic Network
Continuous Discrete
Action Discrete Agent Action

Value Value
Latent Agent u Latent
Continuous Agent
Update Parameters Update Parameters
Discrete Agent
Continuous Calculate Calculate Discrete
Optimizer Loss Loss Optimizer

Sample a batch of Sample a batch of

experiences Discrete Actions Continuous Actions experiences
Action, Reward, Next Action, Reward, Next
Quantum Computing Networks
Replay Buffer Observation Observation Replay Buffer

Observation Observation

Fig. 2: The intelligent resource allocation framework with discrete and continuous action spaces.

learning tasks to either edge or cloud nodes and deciding

50 MARL, d=0, e=1
the proportion of tasks to offload. This offloading strategy MARL, d=0.5, e=0.5
is critical as it affects the computational load distribution MARL, d=1, e=0
40 Random, d=0, e=1
and overall network efficiency. The reward signal for each Random, d=0.5, e=0.5
agent is the global resource cost, incentivizing agents to find Random, d=1, e=0
30
cost-effective resource allocation strategies. The experiences
Cost

from these interactions are stored in replay buffers, which

20
are used to refine and improve the agents’ policies over time.
Furthermore, DRL agents learn from past experiences in replay 10
buffers, aiding in the optimization of future decisions. The
primary objective of each agent is to minimize the long-term 0
expected cost within the quantum computing networks which 0 250 500 750 1000 1250 1500 1750 2000
involves balancing immediate rewards with future gains. By Epochs

optimizing their policies, the agents aim to achieve efficient Fig. 3: The convergence of proposed multi-agent RL
resource allocation in the dynamic environment of quantum algorithms under different QoS settings, where d is the
computing networks. weight of latency and e is the weight of energy consumption.

In quantum computing networks, each computing node

can run quantum generative learning applications, such as for resource allocation in heterogeneous networks. Compared
quantum autoencoders [15]. Quantum autoencoders are a type to the random policy, the cost can be reduced by at least
of quantum circuit designed to compress quantum states onto 80% under different QoS settings. Therefore, the proposed
fewer qubits while preserving the information from the initial algorithm can effectively reduce the system cost by allocating
state. In this way, the dimensionality of the information can quantum resources properly for running quantum generative
be reduced while the fidelity is maintained. For each quantum learning applications.
autoencoder, there is an input layer, a bottleneck layer for
compression, and an output layer for reconstruction. For V. C ONCLUSIONS
instance, to compress n qubits into k qubits using a quantum In this article, we have proposed an intelligent resource
autoencoder, there are n qubits for the input state, n-k qubits allocation framework for quantum computing networks with
for the reference state, and one auxiliary qubit. Therefore, the uncertain quantum resources and service demands. We have
total number of qubits required would be 2n-k+1 qubits. In discussed several services in quantum computing networks for
the experiment, each computing node executes the quantum improving the security and sustainability of existing network-
autoencoders to compress information from [6,9] qubits to ing and computing infrastructure. Furthermore, to improve
[3,5] qubits. In the networks, the number of mobile nodes the scalability of quantum computing networks, we have
is set to 10, the number of edge nodes is set to 5, and there is proposed an intelligent resource allocation framework, which
one cloud node. The cost is calculated based on the execution can leverage RL agents, from generative learning to quantum
latency and energy consumption according to the quality of learning, to model the environments and learn the optimal
service (QoS) settings. We leverage multi-agent reinforcement allocation policy without prior knowledge. Finally, we have
learning algorithms to allow decentralized decision-making designed a use case of quantum generative learning service
among computing nodes. As we can observe from Fig. 3, the provisioning in heterogeneous quantum computing networks
proposed MARL algorithm can converge the optimal policy and proposed a MARL algorithm to learn the optimal policy.
 8

R EFERENCES
[1] M. Caleffi, M. Amoretti, D. Ferrari, D. Cuomo, J. Illiano, A. Manzalini,
and A. S. Cacciapuoti, “Distributed quantum computing: a survey,” arXiv
preprint arXiv:2212.10609, 2022.
[2] Y. Cao, Y. Zhao, Q. Wang, J. Zhang, S. X. Ng, and L. Hanzo, “The
evolution of quantum key distribution networks: On the road to the
qinternet,” IEEE Communications Surveys & Tutorials, vol. 24, no. 2,
pp. 839–894, 2022.
[3] C. Ren, H. Yu, R. Yan, M. Xu, Y. Shen, H. Zhu, D. Niyato, Z. Y. Dong,
and L. C. Kwek, “Towards quantum federated learning,” arXiv preprint
arXiv:2306.09912, 2023.
[4] L. Chen, K. Xue, J. Li, N. Yu, R. Li, Q. Sun, and J. Lu, “Simqn: A
network-layer simulator for the quantum network investigation,” IEEE
Network, 2023.
[5] Y. Cao, Y. Zhao, J. Li, R. Lin, J. Zhang, and J. Chen, “Multi-tenant
provisioning for quantum key distribution networks with heuristics and
reinforcement learning: A comparative study,” IEEE Transactions on
Network and Service Management, vol. 17, no. 2, pp. 946–957, 2020.
[6] Z. Zhu, H. Zhao, H. He, Y. Zhong, S. Zhang, Y. Yu, and W. Zhang,
“Diffusion models for reinforcement learning: A survey,” arXiv preprint
arXiv:2311.01223, 2023.
[7] Z. Li, K. Xue, J. Li, L. Chen, R. Li, Z. Wang, N. Yu, D. S. Wei,
Q. Sun, and J. Lu, “Entanglement-assisted quantum networks: Mechan-
ics, enabling technologies, challenges, and research directions,” IEEE
Communications Surveys & Tutorials, 2023.
[8] Y. Mao, Y. Liu, and Y. Yang, “Qubit allocation for distributed quantum
computing,” in IEEE INFOCOM 2023-IEEE Conference on Computer
Communications, pp. 1–10, IEEE, 2023.
[9] J. Li, M. Wang, K. Xue, R. Li, N. Yu, Q. Sun, and J. Lu, “Fidelity-
guaranteed entanglement routing in quantum networks,” IEEE Transac-
tions on Communications, vol. 70, no. 10, pp. 6748–6763, 2022.
[10] P.-Y. Kong, “A review of quantum key distribution protocols in the per-
spective of smart grid communication security,” IEEE Systems Journal,
vol. 16, no. 1, pp. 41–54, 2020.
[11] Y. Ren, R. Xie, F. R. Yu, T. Huang, and Y. Liu, “Nft-based intelligence
networking for connected and autonomous vehicles: A quantum rein-
forcement learning approach,” IEEE Network, vol. 36, no. 6, pp. 116–
124, 2022.
[12] R. Kaewpuang, M. Xu, D. T. Hoang, D. Niyato, H. Yu, R. Li, Z. Xiong,
and J. Kang, “Elastic entangled pair and qubit resource management in
quantum cloud computing,” arXiv preprint arXiv:2307.13185, 2023.
[13] Y. Cao, Y. Zhao, J. Zhang, Q. Wang, D. Niyato, and L. Hanzo, “From
single-protocol to large-scale multi-protocol quantum networks,” IEEE
Network, vol. 36, no. 5, pp. 14–22, 2022.
[14] M. Xu, D. Niyato, Z. Xiong, J. Kang, X. Cao, X. S. Shen, and
C. Miao, “Quantum-secured space-air-ground integrated networks: Con-
cept, framework, and case study,” IEEE Wireless Communications, 2022.
[15] J. Romero, J. P. Olson, and A. Aspuru-Guzik, “Quantum autoencoders
for efficient compression of quantum data,” Quantum Science and
Technology, vol. 2, no. 4, p. 045001, 2017.