A heterogeneous quantum cloud platform task scheduling method based on reinforcement learning

Abstract

This invention discloses a task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning. This method is implemented using a quantum resource allocation platform integrating a task modeling layer, a task scheduling layer, a quantum hardware layer, and an evaluation module. First, the quantum hardware layer constructs a synchronous real-time hardware feature matrix; the task modeling layer encapsulates heterogeneous quantum task feature vectors; the task scheduling layer merges these two into a state space and constructs a reinforcement learning scheduling decision engine. After filtering illegal operations through an action masking mechanism, legal scheduling is executed. The evaluation module quantifies the evaluation results using a four-dimensional reward function containing penalty terms including task execution time, overdue deadline, resource conflicts, and topology mapping overhead, and feeds the results back to the scheduling decision engine, enabling iterative optimization of the scheduling strategy. This invention achieves explicit hardware feature modeling and standardized task encapsulation, avoids illegal scheduling, dynamically optimizes strategies, supports cross-node distributed scheduling, and significantly improves task execution efficiency, resource utilization, and scheduling robustness.

Description

Technical Field

[0001] This invention relates to the fields of quantum cloud computing and artificial intelligence technology, specifically to a task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning. Background Technology

[0002] As an emerging computing paradigm, quantum computing has shown great potential to trigger paradigm shifts in natural sciences and related fields. In cutting-edge fields such as high-energy physics, quantum chemical simulation, variable quantum algorithms, and quantum machine learning (QML), it has already shown significant acceleration prospects and is expected to achieve quantum superiority over classical computing in specific fields such as numerical simulation.

[0003] Current quantum computing is in the era of Noisy Intermediate-Scale Quantum (NISQ). Limited by the number of qubits, short coherence time, and high gate operation error rate, achieving large-scale, error-free quantum computing remains a significant challenge. Against this backdrop, achieving efficient quantum task scheduling and quantum resource allocation across heterogeneous quantum hardware is crucial for maximizing the throughput, computational fidelity, and scheduling fairness of quantum computing systems.

[0004] Existing quantum task scheduling technologies face the dual challenges of quantum hardware and quantum tasks. At the quantum hardware level, quantum devices exhibit significant heterogeneity, with different qubits exhibiting marked differences in topological connectivity, coherence time, and gate operation error rates, placing extremely high demands on the hardware awareness capabilities of the scheduling framework. At the quantum task level, different quantum applications have varying requirements for the number of qubits, gate operation complexity, and task execution time, and in actual deployment, practical constraints such as task priority, task deadlines, and concurrent execution limits must also be met.

[0005] Despite existing research attempts to address the quantum resource allocation problem, several limitations remain. Firstly, traditional heuristic scheduling methods, such as greedy strategies, while computationally inexpensive, often only achieve local optima when faced with random errors and rapid coherence decay in quantum devices, making it difficult to shift from a throughput-centric to a fidelity-centric scheduling strategy. Secondly, current research suffers from significant shortcomings in low-level feature modeling and system integration. Most scheduling frameworks treat qubits as homogeneous units during scheduling, neglecting key physical constraints such as read fidelity, crosstalk, and specific topologies, leading to significant discrepancies between abstract simulation results and actual quantum hardware performance. Furthermore, much of the research focuses on specific, isolated scheduling strategies, lacking a system-level scheduling platform that organically integrates quantum hardware heterogeneity modeling, quantum task abstraction, and learning-based dynamic scheduling algorithms.

[0006] In recent years, reinforcement learning (RL) has become a research hotspot in the field of quantum dynamic scheduling due to its outstanding ability to handle sequential decision-making problems in uncertain environments. By modeling the quantum task scheduling problem as a Markov decision process, reinforcement learning agents can perceive the occupancy state of qubits and the error distribution of quantum hardware, and thus learn adaptive scheduling strategies that balance short-term scheduling feasibility with long-term system efficiency. However, how to achieve efficient optimization of quantum resources under complex quantum hardware topological constraints in a distributed quantum computing environment remains a key technical problem that urgently needs to be solved in the field of quantum computing resource scheduling. Summary of the Invention

[0007] In view of the technical shortcomings of existing technologies, such as incomplete modeling of quantum hardware characteristics, susceptibility to illegal operations in scheduling strategies, lack of a unified integrated scheduling platform, and low efficiency in resource optimization under distributed topology constraints, this invention addresses at least the following technical problems: 1. To address the problem that existing scheduling frameworks lack modeling of the heterogeneous physical characteristics of qubits, leading to discrepancies between abstract simulation and actual hardware execution, resulting in low system computational fidelity and resource utilization; 2. This addresses the problems of traditional scheduling methods, which easily generate invalid scheduling actions that violate the physical constraints of quantum hardware, have large decision search spaces and slow algorithm convergence, and cannot adapt to the complex requirements of heterogeneous quantum tasks. 3. To address the problem that existing research lacks a unified system-level platform that integrates hardware heterogeneity modeling, task abstraction, and learning-driven scheduling, making it difficult to achieve global optimization of scheduling strategies; 4. To address the issues of low efficiency in optimizing quantum resource allocation for complex topological constraints in distributed quantum computing environments, which can easily lead to resource monopolies and processing stagnation, and to realize a shift in scheduling strategies from "throughput-centric" to "fidelity-centric".

[0008] This invention discloses a task scheduling method for a heterogeneous quantum cloud platform based on reinforcement learning. This method is implemented using a quantum resource allocation platform (QRAP) that integrates a task modeling layer, a task scheduling layer, a quantum hardware layer, and an evaluation module. Specifically, it includes the following steps: S1: The quantum hardware layer explicitly models the underlying physical characteristics of the qubit and the topological connectivity features of the quantum hardware, constructs a real-time updated quantum hardware feature matrix based on the modeling results, and synchronizes the quantum hardware feature matrix to the task scheduling layer. S2: The task modeling layer uses a standardized quantum task abstraction model to encapsulate the heterogeneous quantum task with multi-dimensional features, forming a quantum task feature vector, and then synchronizes the quantum task feature vector to the task scheduling layer. S3: A reinforcement learning scheduling decision engine is constructed by the task scheduling layer. The quantum hardware feature matrix and the quantum task feature vector are integrated as the state space for scheduling decisions. Illegal scheduling operations that violate the physical constraints of quantum hardware are filtered through the action masking mechanism. A set of legal scheduling actions is generated and the scheduling is executed. S4: The evaluation module uses a reward function that includes four dimensions of penalties: task execution time, overdue date, resource conflict, and topology mapping overhead, to quantitatively evaluate the actual scheduling execution result after the task scheduling layer performs a legal scheduling action. The evaluation result is then fed back to the reinforcement learning scheduling decision engine to achieve iterative optimization of the scheduling strategy. Furthermore, the underlying physical characteristics of the qubits in the quantum hardware layer include at least coherence time (T1 / T2), gate error rate, and readout error; the quantum hardware feature matrix collects the bit occupancy rate, topological connectivity matrix, and coherence time data of each quantum node in real time; Furthermore, the multidimensional features in the task modeling layer include bit count, gate operation complexity, priority, deadline, and concurrency logic; Furthermore, the action masking mechanism of the task scheduling layer, based on the real-time bit occupancy, topological adjacency, and coherence time constraints of the quantum nodes in the quantum hardware layer, imposes a negative infinite penalty on scheduling actions that violate physical constraints, and dynamically filters and generates the set of legal scheduling actions. Furthermore, the reward function is calculated using the following formula:

[0009] in, This is a penalty item for task execution time; Penalty for tasks exceeding the deadline; Assign penalties for resource overlap; This is a penalty term for topology mapping overhead; Furthermore, the task execution time penalty item ,in, This refers to the actual execution time of a quantum task at the quantum hardware layer. These are the weighting coefficients for quantum tasks; Furthermore, the topology mapping overhead penalty term The numerical value is calculated based on the number of SWAP operations introduced during the execution of the quantum task at the quantum hardware layer; Furthermore, the reinforcement learning algorithm used by the reinforcement learning scheduling decision engine includes one or more of the following: Proximal Policy Optimization (PPO), Deep Q-Network (DQN), Advantageous Actor-Commentator (A2C), and Asynchronous Advantageous Actor-Commentator (A3C). Furthermore, when optimizing the PPO algorithm using a proximal strategy, the range of policy updates is limited by employing a truncated surrogate objective function to ensure training stability. The truncated surrogate objective function is as follows:

[0010] in, For strategy parameters, The probability ratio of the strategies. For the dominant function, The truncation coefficient is used to constrain the policy update magnitude. Within the range; Furthermore, the method supports distributed task scheduling across quantum nodes in the quantum hardware layer; when a single quantum node cannot meet the bit quantity requirements of a quantum task, the quantum task is sliced ​​and mapped to multiple adjacent quantum nodes for collaborative execution through a quantum communication link, based on the coupling density attribute of the task.

[0011] This invention deeply integrates explicit modeling of heterogeneous features of quantum hardware, standardized encapsulation of heterogeneous quantum tasks, and dynamic scheduling using reinforcement learning. It constructs a four-layer closed-loop scheduling system based on a quantum resource allocation platform, and combines an action masking mechanism with a four-dimensional penalty-reward function to achieve iterative optimization of the scheduling strategy. Simultaneously, it supports distributed task scheduling across quantum nodes, achieving the following beneficial technical effects: 1. Achieve refined real-time modeling of the underlying physical characteristics and topological connectivity characteristics of quantum hardware, eliminate the deviation between abstract simulation and real hardware execution, and significantly improve the system's computational fidelity and resource utilization; 2. The action masking mechanism accurately filters out illegal scheduling operations, narrows the decision search range, speeds up the algorithm convergence, effectively avoids processing stagnation caused by invalid scheduling, and adapts to the complex requirements of heterogeneous quantum tasks; 3. A system-level platform integrating hardware-aware modeling, task standardization and abstraction, learning-driven scheduling and quantitative evaluation feedback was constructed to achieve global optimization of scheduling strategies and improve the platform's adaptability and versatility to different quantum algorithms and hardware. 4. Cross-node distributed scheduling can reasonably allocate computing power based on task coupling density, solve the resource monopoly problem, and maintain good numerical stability and policy robustness in scenarios with dynamic changes in resource density; 5. A reward function based on four-dimensional penalty terms enables quantitative evaluation of scheduling results and optimization of strategy feedback, promoting the shift of scheduling strategy towards "fidelity-centric" approach. This ensures system throughput while improving task completion rate, maximizing the computing power value of heterogeneous quantum cloud platforms. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning according to the present invention. Figure 2 This is a schematic diagram of the operational framework of the Quantum Resource Allocation Platform (QRAP) of the present invention; Figure 3 A schematic diagram illustrating the structural features of the quantum hardware layer of this invention. Detailed Implementation

[0013] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0014] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0015] This embodiment provides a task scheduling method for a heterogeneous quantum cloud platform based on reinforcement learning. It is implemented based on a quantum resource allocation platform (QRAP) that integrates a task modeling layer, a task scheduling layer, a quantum hardware layer, and an evaluation module. The QRAP platform is a unified architecture that integrates hardware-aware modeling and learning-driven scheduling, specifically designed for distributed noisy intermediate-scale quantum (NISQ) systems. The platform formulates the scheduling problem as a constrained optimization task and instantiates it as a Markov decision process. The system state space explicitly encodes core information such as the occupancy state of qubits, hardware error distribution, topological connectivity matrix, and remaining deadline of the task.

[0016] The specific execution steps of the scheduling method of the present invention are as follows: Figure 1 As shown, the entire scheduling process is completed using a closed-loop pipeline design based on the QRAP platform. Figure 2 The QRAP operational framework shown sequentially generates standardized quantum tasks through a task modeling layer, makes resource allocation decisions through a task scheduling layer, completes program mapping and execution through a quantum hardware layer, and implements quantitative evaluation and policy feedback through an evaluation module, forming a continuous optimization iterative process, specifically including the following steps: S1: The quantum hardware layer explicitly models the underlying physical properties of qubits and the topological connectivity features of quantum hardware. Based on the modeling results, it constructs a real-time updated quantum hardware feature matrix and synchronizes the quantum hardware feature matrix to the task scheduling layer. like Figure 3 As shown, the quantum hardware layer consists of multiple interconnected quantum nodes, each of which integrates multiple qubits. To achieve hardware awareness, this embodiment defines multi-dimensional physical characteristics for each qubit, including at least coherence time (T1 / T2), gate error rate, readout error, and readout fidelity. At the same time, the quantum hardware layer collects the bit occupancy rate, topological connectivity matrix, and coherence time data of each quantum node in real time, summarizes the above heterogeneous characteristics into a real-time updated quantum hardware feature matrix, and synchronizes it to the task scheduling layer as an important hardware basis for scheduling decisions.

[0017] S2: The task modeling layer uses a standardized quantum task abstraction model to encapsulate the multi-dimensional features of heterogeneous quantum tasks, forming a quantum task feature vector, and then synchronizes the quantum task feature vector to the task scheduling layer. The multidimensional features of the task modeling layer include the number of bits, gate operation complexity, priority, deadline, remaining deadline, and concurrency logic. By using a standardized quantum task abstraction model, heterogeneous quantum tasks from different fields and with different needs are encapsulated into a unified six-tuple feature vector, achieving standardized abstraction of quantum tasks. This allows the scheduling layer to quickly identify the core requirements of the task and achieve accurate matching with the characteristics of quantum hardware.

[0018] S3: The task scheduling layer constructs a reinforcement learning scheduling decision engine, which integrates the quantum hardware feature matrix and the quantum task feature vector as the state space for scheduling decisions. It filters illegal scheduling operations that violate the physical constraints of quantum hardware through an action masking mechanism, generates a set of legal scheduling actions, and executes the scheduling. The reinforcement learning scheduling decision engine has a built-in reinforcement learning algorithm library, which can employ one or more of the following algorithms: Proximal Policy Optimization (PPO), Deep Q-Network (DQN), Advantageous Actor-Commentator (A2C), and Asynchronous Advantageous Actor-Commentator (A3C). Algorithm selection can be flexibly chosen to match the feature requirements of the quantum task. During the scheduling decision process, the scheduling engine first integrates the hardware feature matrix and the task feature vector to form a complete state space. Then, it activates the core action masking mechanism. This mechanism transforms the underlying hard physical constraints into feasible solution space constraints for higher-level algorithms. Based on the real-time bit occupancy, node capacity limitations, topological adjacency relationships, and coherence time constraints of quantum nodes in the quantum hardware layer, this mechanism imposes a negative infinite penalty on scheduling actions that violate physical constraints, dynamically filtering and generating a set of legal scheduling actions. If a scheduling action violates a physical constraint, such as the available bit topology within the target node not matching the task's coupling requirements, the masking mechanism will impose a negative infinite penalty on that action, ensuring that the reinforcement learning agent only searches for policies within the physically feasible range, effectively avoiding pipeline stalls caused by illegal operations.

[0019] S4: The evaluation module uses a reward function that includes four dimensions of penalties: task execution time, overdue date, resource conflict, and topology mapping overhead. It quantitatively evaluates the actual scheduling results after the task scheduling layer performs a legal scheduling action. The evaluation results are fed back to the reinforcement learning scheduling decision engine to achieve iterative optimization of the scheduling strategy.

[0020] The formula for calculating the reward function is:

[0021] in, This is a penalty item for task execution time; This is a penalty for tasks exceeding the deadline; tasks that exceed the deadline will be subject to strict penalties. Penalties for resource overlap are assigned to reflect the cost of violating resource overlap rules; The topology mapping overhead penalty terms are all negative penalty values, and are quantified to penalize the corresponding scheduling problems.

[0022] Task execution time penalty ,in, This refers to the actual execution time of a quantum task at the quantum hardware layer. The weight coefficient is the weight of the quantum task. The longer the execution time, the larger the weight coefficient, and the larger the corresponding penalty value, so as to effectively punish quantum tasks with excessively long execution times. Topology mapping overhead penalty term The value is calculated based on the number of SWAP operations introduced during the execution of the quantum task at the quantum hardware layer. The more SWAP operations there are, the larger the corresponding penalty value, thereby quantifying the additional overhead in the topology mapping process.

[0023] The evaluation module calculates the comprehensive reward value of the scheduling execution result based on the above reward function, and transmits the reward value as feedback information to the reinforcement learning scheduling decision engine. The scheduling engine adjusts and optimizes the scheduling strategy based on the feedback result, so as to realize the continuous iterative upgrade of the scheduling strategy.

[0024] In a preferred embodiment, for the high-penalty training environment formed by the above four-dimensional penalty terms, when the reinforcement learning scheduling decision engine selects the Proximal Policy Optimization (PPO) algorithm as the core scheduling algorithm, in order to ensure training stability, the algorithm adopts a truncated agent objective function:

[0025] in, For strategy parameters, The probability ratio of the strategies. For the dominant function, The cutoff coefficient is used to constrain the policy update magnitude within a certain range. Within the interval, it effectively prevents the agent from policy collapse when encountering severe resource allocation failures or timeout penalties, ensuring the robustness of model convergence and adapting to the scheduling policy learning needs in high-penalty environments.

[0026] In another preferred embodiment, the method of the present invention supports distributed task scheduling across quantum nodes in the quantum hardware layer; when a single quantum node cannot meet the bit quantity requirements of a large quantum task, the quantum task is partitioned and mapped to multiple adjacent quantum nodes for collaborative execution through high-speed quantum links based on the coupling density attribute of the task, thereby maximizing the utilization of computing power at the distributed hardware level and effectively solving the resource monopoly problem in distributed quantum computing.

[0027] The quantum resource allocation platform (QRAP) of this invention adopts a highly decoupled modular design, achieving independence between the task layer, scheduling layer, and hardware layer. It enables deep topology awareness of the underlying hardware characteristics of distributed NISQ systems, fundamentally solving the problem of traditional scheduling algorithms being disconnected from physical constraints. This allows it to seamlessly adapt to various mainstream quantum chip topologies, including Grid, Ring, and Heavy-hex. The platform provides standardized interfaces and pre-built algorithm libraries, suitable not only for the needs of various cutting-edge quantum algorithms such as Variational Quantum Algorithm (VQA) and Quantum Machine Learning (QML), but also for quantum hardware systems from various vendors. It possesses strong versatility and scalability, significantly lowering the R&D and migration threshold for quantum resource management systems and supporting the rapid deployment of large-scale quantum computing tasks in real heterogeneous hardware environments.

[0028] The scheduling method of this invention can effectively cope with qubit topological constraints. In the QRAP platform environment, various reinforcement learning scheduling algorithms can achieve extremely high task completion rates, demonstrating the platform's universality and efficiency in resource allocation. Among them, the PPO algorithm stands out for its reward value, achieving fast convergence and optimal policy acquisition under various reward configurations, significantly outperforming traditional heuristic benchmarks. Furthermore, the scheduling method of this invention exhibits strong numerical stability and policy robustness in complex environments with varying resource density (e.g., 5-10 bits / node). Even in distributed environments with drastic changes in resource density (e.g., 2-10 bits / node), it can effectively cope with various stringent qubit topological constraints. While ensuring task real-time performance, it significantly optimizes system throughput, effectively solving the problems of resource monopoly and processing stagnation in distributed computing, demonstrating its core competitiveness in optimizing resource allocation.

[0029] This invention has broad industrial application prospects in fields such as quantum cloud services, high-energy physics simulation, and quantum artificial intelligence. It can serve as the underlying resource scheduling engine for quantum cloud platforms, significantly improving job fidelity and system operating efficiency in multi-tenant environments through fine-grained management of heterogeneous computing resources. This technical solution exhibits good compatibility with mainstream algorithms, demonstrates excellent performance in dynamic task flow processing, and possesses extremely high commercialization value, providing crucial technical support for the efficient operation and maintenance of future distributed quantum computing clusters.

[0030] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A task scheduling method for a heterogeneous quantum cloud platform based on reinforcement learning, characterized in that, The method is based on a quantum resource allocation platform (QRAP) that integrates a task modeling layer, a task scheduling layer, a quantum hardware layer, and an evaluation module, and includes the following steps: S1: The quantum hardware layer explicitly models the underlying physical characteristics of the qubit and the topological connectivity features of the quantum hardware, constructs a real-time updated quantum hardware feature matrix based on the modeling results, and synchronizes the quantum hardware feature matrix to the task scheduling layer. S2: The task modeling layer uses a standardized quantum task abstraction model to encapsulate the heterogeneous quantum task with multi-dimensional features, forming a quantum task feature vector, and then synchronizes the quantum task feature vector to the task scheduling layer. S3: A reinforcement learning scheduling decision engine is constructed by the task scheduling layer. The quantum hardware feature matrix and the quantum task feature vector are integrated as the state space for scheduling decisions. Illegal scheduling operations that violate the physical constraints of quantum hardware are filtered through the action masking mechanism. A set of legal scheduling actions is generated and the scheduling is executed. S4: The evaluation module uses a reward function that includes four dimensions of penalties: task execution time, overdue date, resource conflict, and topology mapping overhead, to quantitatively evaluate the actual scheduling execution result after the task scheduling layer performs a legal scheduling action. The evaluation result is then fed back to the reinforcement learning scheduling decision engine to achieve iterative optimization of the scheduling strategy.

2. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 1, characterized in that, The underlying physical characteristics of the qubits in the quantum hardware layer include at least coherence time (T1 / T2), gate error rate, and readout error; the quantum hardware feature matrix collects the bit occupancy rate, topological connectivity matrix, and coherence time data of each quantum node in real time.

3. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 1, characterized in that, The multidimensional features in the task modeling layer include bit count, gate operation complexity, priority, deadline, and concurrency logic.

4. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 1, characterized in that, The action masking mechanism of the task scheduling layer applies a negative infinite penalty to scheduling actions that violate physical constraints based on the real-time bit occupancy, topological adjacency, and coherence time constraints of quantum nodes in the quantum hardware layer, and dynamically filters and generates the set of legal scheduling actions.

5. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 1, characterized in that, The formula for calculating the reward function is as follows: in, This is a penalty item for task execution time; Penalty for tasks exceeding the deadline; Assign penalties for resource overlap; This is a penalty term for topology mapping overhead.

6. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 5, characterized in that, The task execution time penalty item ,in, This refers to the actual execution time of a quantum task at the quantum hardware layer. These are the weighting coefficients for quantum tasks.

7. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 5, characterized in that, The topology mapping overhead penalty term The numerical value is calculated based on the number of SWAP operations introduced during the execution of the quantum task at the quantum hardware layer.

8. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 1, characterized in that, The reinforcement learning algorithm used by the reinforcement learning scheduling decision engine includes one or more of the following: Proximal Policy Optimization (PPO), Deep Q-Network (DQN), Advantageous Actor-Commentator (A2C), and Asynchronous Advantageous Actor-Commentator (A3C).

9. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 8, characterized in that, When optimizing the PPO algorithm using a proximal strategy, a truncated surrogate objective function is employed to limit the range of policy updates in order to ensure training stability. The truncated surrogate objective function is as follows: in, For strategy parameters, The probability ratio of the strategies. For the dominant function, The truncation coefficient is used to constrain the policy update magnitude. Within the range.

10. The task scheduling method for heterogeneous quantum cloud platforms based on reinforcement learning as described in claim 1, characterized in that, The method supports distributed task scheduling across quantum nodes in the quantum hardware layer. When a single quantum node cannot meet the bit quantity requirements of a quantum task, the quantum task is sliced ​​and mapped to multiple adjacent quantum nodes for collaborative execution through a quantum communication link, based on the coupling density attribute of the task.

Images