Multi-objective optimization method and device for cross-cloud data transmission scheduling

Abstract

The invention discloses a multi-objective optimization method and device for cross-cloud data transmission scheduling, and relates to the technical field of cloud computing and network optimization. The problems that in the prior art, due to single-target optimization, multiple requirements for transmission efficiency, cost effectiveness and system stability are difficult to meet, and a transmission path and a scheduling strategy cannot be dynamically adjusted are solved. Then, after multi-scale data feature extraction, multi-scale feature decomposition and abnormal data processing are sequentially carried out on the stationary data, an efficient, stable and intelligent cross-cloud data transmission scheduling scheme is generated by combining ant colony algorithm global search, neighborhood search local optimization and asynchronous reinforcement learning technologies, multi-target collaborative optimization can be achieved, and the scheduling efficiency is improved. And meanwhile, a plurality of transmission performance indexes such as transmission delay, bandwidth utilization rate, transmission cost and load balance are balanced, so that the balance between the transmission efficiency and the resource utilization rate is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cloud computing and network optimization, and particularly relates to a multi-objective optimization method and device for cross-cloud data transmission scheduling. BACKGROUND

[0002] With the rapid development of information technology and cloud computing, the demand for large-scale data transmission in cross-cloud environments is increasing. Cloud service providers build multiple heterogeneous cloud platforms to provide users with elastic computing, storage, and network resources, enabling flexible deployment and efficient management of businesses. However, there are problems such as physical geographical location dispersion, complex network topology structure, and differences in resource management mechanisms between different cloud platforms, making cross-cloud data transmission face many challenges. How to efficiently and stably complete cross-cloud data transmission has become a key problem to ensure the quality of cloud services and improve the response speed of businesses.

[0003] Currently, traditional cross-cloud data transmission scheduling methods mainly focus on the optimization of a single performance indicator, such as minimizing transmission delay or maximizing bandwidth. However, these methods often ignore the demand for multi-objective optimization in actual environments, making it difficult to balance transmission efficiency, cost, security, and resource utilization. In addition, the dynamic nature of cross-cloud network environments causes network topology structure and bandwidth resources to change constantly, making it difficult for static scheduling schemes to adapt to real-time fluctuations in network conditions, resulting in transmission bottlenecks, resource waste, and load imbalance. In particular, in a multi-tenant environment, network congestion and resource competition are more intense, and the response speed and adjustment capacity of existing scheduling mechanisms are significantly insufficient.

[0004] Specific problems mainly include:

[0005] 1. Limitations of single-objective optimization. Existing cross-cloud transmission scheduling mostly focuses on minimizing transmission delay or maximizing bandwidth utilization, ignoring the coordinated optimization of multi-dimensional factors such as cost control, load balancing, and security protection, resulting in unreasonable resource allocation and affecting transmission quality and overall system performance.

[0006] 2. Insufficient scheduling adaptability in dynamic network environments. In cross-cloud environments, network status and cloud resources change frequently in real time. Existing methods lack real-time network status perception and intelligent decision-making capabilities, and cannot dynamically adjust transmission paths and scheduling strategies, resulting in large fluctuations in transmission performance and decreased system robustness.

[0007] 3. Compatibility challenges of multi-cloud heterogeneous platforms. Different cloud service providers have significant differences in security policies, access control, and quality of service guarantees. Traditional scheduling algorithms cannot simultaneously meet the security and performance requirements in a multi-cloud environment, affecting the collaborative efficiency of cross-cloud applications.

[0008] 4. Lack of unified multi-objective optimization scheduling framework. The existing technology is mainly designed for single target, and lacks integrated multi-objective trade-off comprehensive scheduling mechanism, which is difficult to meet the multiple requirements of transmission efficiency, cost effectiveness and system stability in complex application scenarios. SUMMARY

[0009] The present application aims to overcome the problems of single target optimization in the prior art, which cannot meet the multiple requirements of transmission efficiency, cost effectiveness and system stability, and cannot dynamically adjust the transmission path and scheduling strategy, and provides a multi-objective optimization method and device for cross-cloud data transmission scheduling.

[0010] In a first aspect, a multi-objective optimization method for cross-cloud data transmission scheduling is provided, comprising:

[0011] Obtaining cloud environment original data and preprocessing to obtain stationary data, wherein the preprocessing includes multi-objective feature extraction, multi-scale feature decomposition and abnormal data processing;

[0012] Based on the stationary data, a plurality of scheduling paths are constructed, and an ant colony algorithm is used for global search to obtain a global optimal solution;

[0013] A variable neighborhood search strategy is used to locally optimize the scheduling path in the global optimal solution to obtain a locally optimal path;

[0014] An Actor strategy model for optimizing the locally optimal path is constructed and trained with an asynchronous reinforcement learning mechanism, and a dynamic cross-cloud scheduling scheme is generated using the Actor strategy model;

[0015] According to the cross-cloud scheduling scheme, the transmission scheduling of cross-cloud data is performed.

[0016] In some possible implementation manners, the preprocessing includes:

[0017] A fixed step sliding window is used to time slice the network bandwidth, network delay and resource occupation in the cloud environment original data;

[0018] Statistical features are calculated in each sliding window, wherein the statistical features include mean, variance, extreme value and change rate;

[0019] Wavelet variation is introduced to decompose the bandwidth and delay sequence into low-frequency trend items and high-frequency fluctuation items;

[0020] A method based on Z-score is used to identify bandwidth mutation and delay jitter points, and the bandwidth mutation and delay jitter points are marked as abnormal data, and the abnormal data is removed and filled with sliding median.

[0021] In some possible implementation manners, the multiple scheduling paths are constructed based on the stationary data, and an ant colony algorithm is used for global search, including: generating a multi-path transmission scheme in a virtual topology structure through positive feedback and distributed search mechanism of the ant colony algorithm, iteratively searching for a scheduling path with minimum path cost function through pheromone update and heuristic function adjustment, and converging to a global optimal scheduling path as a global optimal solution through multiple artificial ants selecting routes between different cloud nodes according to path cost in each iteration, wherein a path entropy value index with a penalty is introduced in each search process, and is used to measure information diversity and selection bias of the scheduling path.

[0022] In some possible implementation manners, a variable neighborhood search strategy is used to locally optimize the scheduling path in the global optimal solution, including: constructing a multi-scale neighborhood structure in the space of the global optimal solution, adjusting the initial path through a dynamic switching neighborhood and a disturbance mechanism, performing refined path reconstruction and local scheduling order adjustment in each neighborhood, and dynamically selecting a disturbance operation region based on the path entropy value of the current scheduling path and resource occupation fluctuation in each local search, and preferentially optimizing a bottleneck path segment, while introducing a disturbance guiding mechanism to evaluate the effectiveness of the disturbance strategy according to historical neighborhood search effects, and realizing adaptive selection and evolution of neighborhood operations, so as to obtain a locally optimal path.

[0023] In some possible implementation manners, an Actor policy model for optimizing the locally optimal path is constructed and trained with an asynchronous reinforcement learning mechanism, including: constructing an Actor policy model of a scheduling policy, the Actor policy model adopts a multi-line asynchronous training structure, different threads independently interact with the environment based on shared parameters, in each thread, an Actor network is responsible for outputting scheduling actions, and a Critic network is used to estimate a state value function, a policy execution of Gaussian disturbance is introduced in the training process of the Actor policy model, an expected value of an action in a current state is output by a policy network and a standard Gaussian disturbance is added to generate an actual action, an advantage function is constructed to measure the advantages and disadvantages of the current action, and an Actor network performs policy gradient update based on the advantage function, and a multi-objective trade-off of data transmission delay, bandwidth utilization and path redundancy is performed in a reward function to provide scheduling optimization feedback in a weighted linear combination form.

[0024] In some possible implementation manners, the method further includes:

[0025] A simulation test environment for multiple platforms is constructed to simulate heterogeneous resource configurations of multiple platforms.

[0026] The multi-objective optimization method is deployed in the simulation test environment to process a specified task, where a type of the task includes cross-region big data synchronization, real-time log distribution, and model multi-cloud deployment.

[0027] The cross-cloud scheduling scheme obtained by processing the specified task is embedded into the task scheduling engine, and the cross-cloud transmission path is generated in real time and resource allocation is completed in combination with dynamic task requests and current network status;

[0028] Key performance indicators are collected during operation, including average transmission delay, bandwidth utilization, task completion rate and path redundancy;

[0029] A visual scheduling console is created, which is used for real-time monitoring of scheduling status, transmission bottleneck early warning and resource utilization analysis.

[0030] In some possible implementation manners, the performance data in actual operation is further included as an input of the Actor strategy model, parameters of the Actor strategy model are dynamically adjusted, and the dynamically adjusted data and operation logs are persistently stored.

[0031] In a second aspect, a multi-objective optimization device for cross-cloud data transmission scheduling is provided, comprising:

[0032] An acquisition module is configured to acquire cloud environment original data and perform preprocessing to obtain smoothed data, wherein the preprocessing includes multi-objective feature extraction, multi-scale feature decomposition and abnormal data processing;

[0033] A global search module is configured to construct a plurality of scheduling paths based on the smoothed data, and perform global search using an ant colony algorithm to obtain a globally optimal solution;

[0034] A local optimization module is configured to perform local optimization on the scheduling paths in the globally optimal solution using a variable neighborhood search strategy to obtain locally optimal paths;

[0035] A deep optimization module is configured to construct and train an Actor strategy model having an asynchronous reinforcement learning mechanism for optimizing the locally optimal paths, and generate a dynamic cross-cloud scheduling scheme using the Actor strategy model;

[0036] An execution module is configured to perform cross-cloud data transmission scheduling according to the cross-cloud scheduling scheme.

[0037] In a third aspect, a computer readable storage medium is provided, which stores program codes for execution by a device, and the program codes include steps for executing the method in any one of the implementation manners of the first aspect.

[0038] In a fourth aspect, an electronic device is provided, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, and the program or instructions, when executed by the processor, implement the method in any one of the implementation manners of the first aspect.

[0039] The application has the following beneficial effects: the application realizes efficient, stable and intelligent generation of a cross-cloud data transmission scheduling scheme by combining multi-scale data feature extraction, global search of an ant colony algorithm, local optimization of neighborhood search and asynchronous reinforcement learning technology, can realize multi-objective collaborative optimization, balances multiple transmission performance indicators such as transmission delay, bandwidth utilization, transmission cost and load balancing, realizes balance between transmission efficiency and resource utilization, dynamically adjusts the transmission path and resource allocation in combination with real-time network state monitoring, improves the robustness and flexibility of scheduling, in addition, the application also supports security and performance requirements between heterogeneous cloud platforms, and promotes efficient collaboration and intelligent management of cloud resources. BRIEF DESCRIPTION OF DRAWINGS

[0040] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, and are incorporated herein for explanation by illustrating an exemplary embodiment of the present application and to provide further description of the application, but are not intended to limit the present application.

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the accompanying drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0042] Figure 1 is a flowchart of the multi-objective optimization method of the cross-cloud data transmission scheduling of the embodiment 1 of the present application;

[0043] Figure 2 is a flowchart of the multi-objective optimization method of the cross-cloud data transmission scheduling of the embodiment 1 of the present application;

[0044] Figure 3 is a flowchart of the ant colony algorithm in the multi-objective optimization method of the cross-cloud data transmission scheduling of the embodiment 1 of the present application;

[0045] Figure 4 is a flowchart of the variable neighborhood search and strategy in the multi-objective optimization method of the cross-cloud data transmission scheduling of the embodiment 1 of the present application;

[0046] Figure 5 is a principle schematic diagram of the reinforcement learning mechanism in the multi-objective optimization method of the cross-cloud data transmission scheduling of the embodiment 1 of the present application;

[0047] Figure 6 is a structural block diagram of a multi-objective optimization device for cross-cloud data transmission scheduling of embodiment 2 of the application;

[0048] Figure 7 is a schematic diagram of the internal structure of an electronic device of embodiment 4 of the application.

[0049] Reference signs:

[0050] 100, acquisition module; 200, global search module; 300, local optimization module; 400, deep optimization module; 500, execution module. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the application.

[0052] Embodiment 1

[0053] As shown in Figure 1 and Figure 2 , a multi-objective optimization method for cross-cloud data transmission scheduling related to embodiment 1 of the application includes:

[0054] S100, acquiring cloud environment original data and performing preprocessing to obtain smoothed data, wherein the preprocessing includes multi-objective feature extraction, multi-scale feature decomposition, and abnormal data processing.

[0055] In a cross-cloud environment, the transmitted data often comes from multiple heterogeneous cloud platforms, and network bandwidth, time delay, and resource occupation and other indicators have strong time-varying and non-stationary characteristics. In order to guarantee the input data quality of the scheduling algorithm, first, the multi-scale feature extraction and abnormal processing are performed on these time series data.

[0056] Specifically, a fixed step sliding window is used to slice the original transmission indicators in time series, and the window length and step length are set according to the network fluctuation characteristics of the cloud environment, for example: the window length is set to 10 seconds, and the step length is 5 seconds, which ensures capturing dynamic changes while reducing data redundancy. In each window, the mean value μ, the variance σ 2 , the maximum value x max , the minimum value x min and the change rate Δx of the multi-objective (the multi-objective includes network bandwidth, network delay, resource occupation, and link packet loss rate, etc.) are calculated, which comprehensively reflect the fluctuation trend and extreme value of the bandwidth and delay, and provide rich time series information for the subsequent model. The calculation process of the mean value μ of the characteristics can be described as:

[0057]

[0058] where x i represents the i-th sample point in the sliding window, and n is the window size.

[0059] Subsequently, the target feature data such as network bandwidth, network delay and link packet loss rate are decomposed into low-frequency trend items and high-frequency fluctuation items by wavelet transform. The low-frequency component reveals the long-term trend of network resources, and the high-frequency component captures sudden fluctuations and jitter, providing stable trend information and sensitive abnormality prompts for the scheduling model. For the inevitable sudden abnormalities in the network environment, such as bandwidth mutation and delay jitter, an abnormality detection mechanism based on Z-score is used for identification:

[0060]

[0061] When the absolute value of a data point exceeds the set threshold (usually 3), it is considered abnormal data. The abnormal data can be removed and filled by the sliding median method to ensure the continuity and stability of the data and avoid bias caused by abnormal values in the subsequent Actor strategy model training. Through the above steps, the cleaning and multi-scale feature extraction of multi-source transmission data in a heterogeneous cloud environment are completed, which not only guarantees the time sequence integrity of the data, but also enhances the expression ability of the features to dynamic changes, providing stable and high-quality input for subsequent global search and optimization.

[0062] S200, based on the stationary data, a plurality of scheduling paths are constructed, and an ant colony algorithm is used for global search to obtain a global optimal solution.

[0063] After the preprocessing of the original data in the cloud environment is completed, the global search stage is entered, and the goal is to find the scheduling path with the lowest transmission cost in the complex topology of the multi-cloud environment, i.e., the global optimal solution. This stage is based on the multi-path ant colony algorithm (ACO), which simulates the behavior of an ant colony searching for food paths to achieve path optimization in the global range.

[0064] Specifically, the processed data such as network bandwidth, network delay and resource occupation are mapped to a network topology graph, where nodes represent different cloud platforms and edge weights represent transmission costs, considering both delay and resource consumption. As shown in Figure 3 , multiple "ants" start from the starting cloud node in the graph and choose the next node based on the pheromone concentration and path cost probability on the path.

[0065] The probability of ant path selection is given by the following simplified formula:

[0066]

[0067] where, represents the pheromone concentration on path (i,j), represents the historical goodness of the path; is the heuristic information of the path, usually inversely proportional to the path cost (such as delay), helping ants to prefer low-cost paths. k represents all optional next-hop nodes from the current node, ensuring that the path selection probability is normalized between all candidate paths. is the entropy penalty factor of the path, which is used to measure the chaos degree of the path in multiple searches, and its value is obtained by exponential decay calculation of the path entropy. The smaller the value, the more unstable the path structure, thereby reducing the probability of being selected. By comprehensively considering the historical pheromone concentration , the heuristic information of path cost and the stability of the path (entropy penalty) , the path selection probability is finally determined.

[0068] After each round of search, the pheromone is updated: high-quality paths will increase the pheromone concentration, and poor-quality paths will gradually evaporate. This positive feedback mechanism makes the search of the ant colony gradually concentrate on the optimal path. With the increase of the number of iterations, the ant colony algorithm can gradually converge, and output one or more optimal cross-cloud scheduling paths considering bandwidth utilization, delay and resource allocation as the global optimal solution, that is, the global optimal solution is one or more optimal cross-cloud scheduling paths obtained by global search of the ant colony algorithm. The global search step ensures that the initial path selection of cross-cloud scheduling has a global perspective, avoids local traps, and lays a solid foundation for subsequent refinement optimization.

[0069] S300, adopt variable neighborhood search strategy to locally optimize the scheduling path in the global optimal solution to obtain a locally optimal path.

[0070] After completing the global search based on the ant colony algorithm, a group of initial scheduling paths have been obtained. Although these paths have a relatively optimal overall scheduling cost, due to the complexity of the cross-cloud network topology and the dynamic fluctuations of heterogeneous resources, there may still be problems such as redundant hops, node congestion or low bandwidth utilization in the paths. In order to further improve the scheduling quality, such as Figure 4As shown, the Variable Neighborhood Search (VNS) algorithm is introduced for local optimization. The core idea of VNS is to dynamically switch between multiple neighborhood structures to jump out of local optimal traps and obtain better path adjustment strategies. In cross-cloud data scheduling, different path structures have a significant impact on transmission delay and resource utilization, so it is crucial to construct multiple local disturbance methods to adapt to different scenarios. For example, when there is a bandwidth bottleneck node in the scheduling path, the "node swap" neighborhood can be used for rearrangement; if the path hop count is redundant, the "path segment replacement" neighborhood can be used to simplify the transmission sequence; and in scenarios with high transmission reliability requirements, the "multi-path reconstruction" method is used to enhance redundancy.

[0071] The specific process is as follows: Let the initial path be P = {v1, v2,..., vn}, and VNS defines the following typical neighborhoods in turn:

[0072] Neighborhood one: node swap (Swap) - swap two intermediate nodes in the path to alleviate single-point bottlenecks;

[0073] Neighborhood two: path segment replacement (2-opt) - reverse a segment of the path to reduce hop count and detours;

[0074] Neighborhood three: redundant path insertion / removal - adjust the proportion of multi-path transmission to improve bandwidth utilization.

[0075] After each disturbance, the optimization effect is judged by the scheduling evaluation function. If the current solution is better than the previous solution, accept the solution and start again from the first neighborhood; otherwise, based on the disturbance effect and path state, adaptively adjust the selection of the next neighborhood and the disturbance intensity, continue to switch neighborhood structures until the preset termination condition is reached. When evaluating the disturbance effect, the following two key indicators are considered: transmission delay (the smaller the better) and path average bandwidth utilization (the higher the better), with the target weight assigned to the transmission delay and the target weight assigned to the bandwidth utilization , respectively. The weights and are dynamically adjusted according to the current state of the path, and satisfy . These two weights change in real time with the fluctuations of transmission delay and bandwidth utilization to more accurately reflect the current optimization focus. The objective function with adaptive weights is represented as:

[0076]

[0077] wherein f(P) represents a comprehensive evaluation function of the path P, the smaller the value is, the better the path performance is, Delay(P) represents an end-to-end transmission delay of the path P, the unit is millisecond, BandwidthUtil(P) represents a bandwidth utilization rate of the path P, the value range is 0 to 1. If the current transmission delay is high, the weight a(t) is increased to preferentially reduce the delay; if the bandwidth utilization rate is low, the weight β(t) is increased to preferentially improve the bandwidth use efficiency. Through this mechanism, the algorithm can flexibly adjust the optimization focus, avoid falling into local optimum, and thus more accurately adapt to the dynamic changes of the network state, realizing the dual goals of minimizing the transmission delay and maximizing the bandwidth utilization rate.

[0078] S400, an Actor policy model for optimizing a local optimal path is constructed and trained with an asynchronous reinforcement learning mechanism, and a dynamic cross-cloud scheduling scheme is generated by using the Actor policy model.

[0079] In view of the complex situations such as variable network state, heterogeneous resource distribution, and conflicting scheduling goals in the cross-cloud scheduling process, the traditional ant colony algorithm and the VNS neighborhood search algorithm are prone to fall into local optimum and have weak response ability to environmental changes. Therefore, on the basis of obtaining a local feasible solution by neighborhood search, as shown in Figure 5 , an asynchronous reinforcement learning training method with a Gaussian perturbation mechanism is proposed to improve the global generalization ability and dynamic adaptability of the policy.

[0080] Specifically, the main purpose of introducing the Gaussian perturbation mechanism is to enhance the exploratory nature of the policy and prevent the policy network from falling into a local optimal solution. In actual deployment, the fluctuations of bandwidth and delay often cause the scheduling result to converge too early in the fine-tuning stage, and the Gaussian perturbation can encourage the agent to retain a certain "jumping out ability" in the decision-making process and explore more potential optimal solutions. The Actor policy model adopts a multi-thread asynchronous training structure, and different threads independently interact with the environment based on shared parameters, which accelerates the convergence speed and improves the robustness of the policy. In each thread, the Actor network is responsible for outputting the scheduling action, and the Critic network is used to estimate the state value function. The training process includes the following contents:

[0081] Strategy execution with Gaussian perturbation. The policy network outputs the expected value μ t of the action a t in the current state, and a standard Gaussian perturbation is added to generate the actual action:

[0082]

[0083] wherein, is the standard Gaussian perturbation added.

[0084] The mechanism ensures a certain randomness in the policy execution phase, thereby avoiding the Actor policy model from “overfitting” a certain type of network pattern during training or deployment. Advantage function is calculated: In order to more accurately measure the pros and cons of the current action, a TD error is used to construct an advantage function:

[0085]

[0086] wherein r t is the actual reward corresponding to the action a t , γ is a discount factor, and V(s) represents the expected return of the state S estimated by the Critic network, which is used to guide the policy to converge to the optimal direction. According to the advantage A t of the current time step t, asynchronous parallel policy update is performed. The Actor network performs policy gradient update based on the advantage function, and the optimization objective function of the Actor network is:

[0087]

[0088] wherein, represents the optimization objective value of the Actor network, and policy optimization is achieved by maximizing the objective function; E[·] represents mathematical expectation operation, which embodies the overall optimization of the state-action distribution. π θ (a i | s i ) is a policy function parameterized by the Actor network, representing the probability distribution of selecting action a i under state s i . is a state-dependent reference value function used to stabilize the training process. A t represents the advantage function, which is the difference between the action value function Q and the state value function V, and is used to evaluate the advantage degree of the current action relative to the average level. The objective function realizes policy gradient update through the logπ θ (a i | s i )A t term, and when the advantage function A t is positive, the selection probability of the corresponding action is increased, and when it is negative, the probability is reduced, as a regulating factor to further optimize the training stability. Through the design of the objective function, the present application realizes efficient and stable training of the policy gradient algorithm, and is particularly suitable for control tasks in continuous action space.

[0089] Multiple threads asynchronously execute this process, utilize independent interaction trajectories to jointly improve shared network parameters, effectively improve training efficiency and strengthen the generalization ability of the strategy. In the design of the reward function, multiple target trade-offs such as data transmission delay, bandwidth utilization and path redundancy are considered to provide scheduling optimization feedback in the form of weighted linear combination, ensuring that the generated strategy can stably achieve the scheduling goal in a dynamic environment.

[0090] S500, transmission scheduling of cross-cloud data is performed according to the cross-cloud scheduling scheme.

[0091] After the multi-stage optimization of the scheduling strategy in the above steps S200-S400, the final deployment and verification phase is entered. This step aims to realize the landing of the cross-cloud scheduling strategy from the algorithm model to the actual application system, and to improve the intelligent level and adaptability of the system through a continuous feedback mechanism.

[0092] First, a simulation test environment for multi-cloud platforms is constructed to simulate heterogeneous resource configurations including mainstream cloud platforms such as Alibaba Cloud, Huawei Cloud and AWS. Input includes typical cross-regional big data synchronization, real-time log distribution, model multi-cloud deployment and other task types, covering typical application scenarios such as high concurrency, complex path dependence and multi-objective scheduling.

[0093] During deployment, the optimal scheduling scheme optimized by the Actor strategy model in the aforementioned step S400 is embedded into the task scheduling engine, and the cross-cloud transmission path is generated in real time and the resource allocation is completed in combination with dynamic task requests and the current network state.

[0094] To ensure the actual scheduling ability of the algorithm, key performance indicators are collected during operation, including average transmission delay, bandwidth utilization, task completion rate and path redundancy. For these indicators, a visual scheduling console is provided to support real-time monitoring of scheduling status, transmission bottleneck warning, resource utilization analysis and other functions. At the same time, it supports comparison and evaluation of scheduling effect in task batches, facilitating users and operation and maintenance personnel to analyze the quality of scheduling results and quickly locate problems.

[0095] In further embodiments, an online feedback mechanism is introduced, with actual performance data during operation as input, dynamically adjusting the parameters of the scheduling Actor strategy model (such as Gaussian disturbance amplitude, learning rate of reinforcement learning, neighborhood disturbance structure, etc.), supporting online retraining and adaptive evolution of the Actor strategy model. This feedback closed-loop mechanism ensures that the scheduling strategy always maintains efficient and stable execution capability when facing changing network environment and business demands.

[0096] Finally, the optimized scheduling results and operation logs are persistently stored, supporting long-term strategy evolution and cross-scene migration deployment, forming a cross-cloud intelligent scheduling system with promotional value.

[0097] Embodiment 2

[0098] As Figure 6 shown, the multi-objective optimization device for cross-cloud data transmission scheduling related in Embodiment 2 of the present application comprises:

[0099] An acquisition module 100 is configured to acquire cloud environment original data and perform preprocessing to obtain smoothed data, wherein the preprocessing comprises multi-objective feature extraction, multi-scale feature decomposition and abnormal data processing.

[0100] A global search module 200 is configured to construct a plurality of scheduling paths based on the smoothed data, and perform global search by using an ant colony algorithm to obtain a globally optimal solution.

[0101] A local optimization module 300 is configured to perform local optimization on the scheduling paths in the globally optimal solution by using a variable neighborhood search strategy to obtain a locally optimal path.

[0102] A deep optimization module 400 is configured to construct and train an Actor strategy model for optimizing the locally optimal path, which has an asynchronous reinforcement learning mechanism, and generate a dynamic cross-cloud scheduling scheme by using the Actor strategy model.

[0103] An execution module 500 is configured to perform transmission scheduling of cross-cloud data according to the cross-cloud scheduling scheme.

[0104] It should be noted that other specific embodiments of the multi-objective optimization device for cross-cloud data transmission scheduling in the present embodiment can refer to the specific embodiments of the multi-objective optimization method for cross-cloud data transmission scheduling described above, and to avoid redundancy, details are not described here.

[0105] Embodiment 3

[0106] The computer readable storage medium related in Embodiment 3 of the present application stores program code for execution by a device, and the program code comprises steps for executing the method in any one of the implementation manners in Embodiment 1 of the present application.

[0107] The computer readable storage medium can be a read only memory (ROM), a static storage device, a dynamic storage device or a random access memory (RAM). The computer readable storage medium can store program code, and when the program stored in the computer readable storage medium is executed by a processor, the processor is configured to execute the steps of the method in any one of the implementation manners in Embodiment 1 of the present application.

[0108] Embodiment 4

[0109] As Figure 7 Embodiment 4 of the present application relates to an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, and the program or instructions, when executed by the processor, implement the method in any one of the implementation manners in Embodiment 1 of the present application.

[0110] The processor can be a general-purpose central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), a graphics processing unit (GPU), or one or more integrated circuits, for executing related programs to implement the method in any one of the implementation manners in Embodiment 1 of the present application.

[0111] The processor can also be an integrated circuit electronic device with signal processing capability. In the implementation process, each step of the method in any one of the implementation manners in Embodiment 1 of the present application can be completed by the integrated logic circuit of hardware in the processor or the instruction in the form of software.

[0112] The processor can also be a general-purpose processor, a digital signal processor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block diagram disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory, and combines the hardware to complete the functions required by the units included in the data processing device of the embodiments of the present application, or executes the method in any one of the implementation manners in Embodiment 1 of the present application.

[0113] The above is only the preferred specific implementation manner of the present application; however, the protection scope of the present application is not limited to this. Any skilled person in the art can make equivalent replacements or changes to the technical solution and the improved concept of the present application within the technical scope disclosed in the present application, which should be covered in the protection scope of the present application.

Claims

1. A multi-objective optimization method for cross-cloud data transmission scheduling, characterized in that, include: The raw data of the cloud environment is acquired and preprocessed to obtain stable data, wherein the preprocessing includes multi-target feature extraction, multi-scale feature decomposition and outlier data processing; Multiple scheduling paths are constructed based on the stabilized data, and an ant colony algorithm is used for global search to obtain the global optimal solution. A variable neighborhood search strategy is used to locally optimize the scheduling path in the global optimal solution to obtain the local optimal path; Construct and train an Actor policy model with an asynchronous reinforcement learning mechanism for optimizing local optimal paths, and use the Actor policy model to generate dynamic cross-cloud scheduling schemes. Cross-cloud data transmission scheduling is performed according to the aforementioned cross-cloud scheduling scheme.

2. The multi-objective optimization method for cross-cloud data transmission scheduling according to claim 1, characterized in that, The preprocessing includes: A time-series slice of network bandwidth, network latency, and resource usage in the raw data of the cloud environment is performed using a sliding window with a fixed step size. Statistical features are calculated in each sliding window, wherein the statistical features include mean, variance, extreme values, and rate of change; Wavelet transform is introduced to perform multi-scale feature decomposition on bandwidth and delay sequences to extract low-frequency trend terms and high-frequency fluctuation terms; A Z-score-based method is used to identify bandwidth mutations and latency jitter points. These points are marked as anomalous data, which are then removed and replaced with the moving median.

3. The multi-objective optimization method for cross-cloud data transmission scheduling according to claim 1 or 2, characterized in that, Multiple scheduling paths are constructed based on the stabilized data, and a global search is performed using an ant colony algorithm, including: By leveraging the positive feedback and distributed search mechanism of the ant colony algorithm, a multi-path transmission scheme is generated in the virtual topology. Through pheromone updates and heuristic function adjustments, the scheduling path with the minimum path cost function is iteratively searched. In each iteration, multiple artificial ants select routes between different cloud nodes based on path costs, gradually converging to the globally optimal scheduling path as the global optimal solution. In each search process, a path entropy index with penalties is introduced to measure the information diversity and selection bias of the scheduling path.

4. The multi-objective optimization method for cross-cloud data transmission scheduling according to claim 3, characterized in that, A variable neighborhood search strategy is employed to locally optimize the scheduling path in the global optimal solution, including: A multi-scale neighborhood structure is constructed in the space of the global optimal solution. The initial path is adjusted by dynamically switching neighborhoods and perturbation mechanisms. Fine-grained path reconstruction and local scheduling order adjustment are performed in each neighborhood. In each round of local search, the area of ​​action of perturbation operation is dynamically selected based on the path entropy value and resource consumption fluctuation of the current scheduling path, and bottleneck path segments are optimized first. At the same time, a perturbation guidance mechanism is introduced to evaluate the effectiveness of the perturbation strategy based on the historical neighborhood search effect, so as to realize the adaptive selection and evolution of neighborhood operation, thereby obtaining the local optimal path.

5. The multi-objective optimization method for cross-cloud data transmission scheduling according to claim 1 or 4, characterized in that, Construct and train an Actor policy model with an asynchronous reinforcement learning mechanism for optimizing locally optimal paths, including: An Actor policy model for scheduling is constructed, which adopts a multi-threaded asynchronous training structure. Different threads interact independently of the environment based on shared parameters. In each thread, the Actor network is responsible for outputting scheduling actions, and the Critic network is used to estimate the state value function. During the training process of the Actor policy model, Gaussian perturbation is introduced into the policy execution. The policy network outputs the expected value of the action in the current state and adds standard Gaussian perturbation to generate the actual action. An advantage function is constructed to measure the merits of the current action. The Actor network performs policy gradient updates based on the advantage function. In the reward function, multiple objectives such as data transmission delay, bandwidth utilization, and path redundancy are weighed, and scheduling optimization feedback is provided in the form of a weighted linear combination.

6. The multi-objective optimization method for cross-cloud data transmission scheduling according to claim 5, characterized in that, Also includes: Build a simulation testing environment for multiple platforms to simulate heterogeneous resource configurations across multiple platforms; The multi-objective optimization method is deployed in a simulation test environment to handle specified tasks, wherein the types of tasks include cross-regional big data synchronization, real-time log distribution, and multi-cloud model deployment. The cross-cloud scheduling scheme obtained from processing a specified task is embedded into the task scheduling engine. By combining dynamic task requests and the current network status, cross-cloud transmission paths are generated in real time and resource allocation is completed. Key performance indicators are collected during operation, including average transmission delay, bandwidth utilization, task completion rate, and path redundancy. Create a visual scheduling console, which is used for real-time monitoring of scheduling status, early warning of transmission bottlenecks, and analysis of resource utilization.

7. The multi-objective optimization method for cross-cloud data transmission scheduling according to claim 1, characterized in that, Also includes: The actual performance data is used as input to the Actor strategy model, the parameters of the Actor strategy model are dynamically adjusted, and the dynamically adjusted data and operation logs are persistently stored.

8. A multi-objective optimization device for cross-cloud data transmission scheduling, characterized in that, include: The acquisition module is used to acquire raw data from the cloud environment and preprocess it to obtain stable data. The preprocessing includes multi-target feature extraction, multi-scale feature decomposition, and outlier data processing. The global search module is used to construct multiple scheduling paths based on the stabilized data and use the ant colony algorithm to perform a global search to obtain the global optimal solution. The local optimization module is used to perform local optimization on the scheduling path in the global optimal solution using a variable neighborhood search strategy to obtain the local optimal path; The deep optimization module is used to build and train an Actor policy model with an asynchronous reinforcement learning mechanism to optimize local optimal paths, and to use the Actor policy model to generate dynamic cross-cloud scheduling schemes. The execution module is used to schedule the transmission of cross-cloud data according to the cross-cloud scheduling scheme.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code for execution by the device, the program code including steps for performing the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in any one of claims 1-7.

Images