Computing power network reconstruction strategy generation method and system and storage medium

Abstract

The invention discloses a computing power network reconstruction strategy generation method and system and a storage medium, and the method comprises the steps: responding to a received new task demand, and calculating the adaptation degree of the task demand and a strategy in a historical reconstruction strategy database; judging whether a historical reconstruction strategy of which the adaptation degree meets a preset condition exists or not; if yes, selecting a historical reconstruction strategy meeting the preset condition; if not, inputting the task demand and the current computing power network resource state into a generative model based on artificial intelligence generation content to generate a new reconstruction strategy; and executing computing power network resource reconstruction according to the selected or generated reconstruction strategy.

Description

Technical Field

[0001] This invention relates to the fields of computing power network resource management and artificial intelligence technology, specifically to a method, system, and storage medium for generating computing power network reconstruction strategies. Background Technology

[0002] As a new generation of information infrastructure, computing networks aim to achieve global intelligent scheduling and coordination of computing, storage, network, and algorithm resources. However, with the increasing complexity of application scenarios and the dynamic nature of task requirements, traditional resource reconfiguration methods have shown limitations. These methods are mainly divided into two categories: one is heuristic methods based on fixed rules, whose strategies are rigid and cannot adapt to unknown scenarios; the other is methods based on knowledge graphs and rule-based reasoning, which have high knowledge acquisition and maintenance costs, low real-time reasoning efficiency, and are also limited by preset rules.

[0003] The fundamental flaw of these traditional methods lies in their lack of ability to autonomously learn from massive amounts of historical data and create new strategies. When faced with complex, dynamic, or unprecedented task requirements, they struggle to quickly generate high-quality, adaptive refactoring strategies, resulting in low resource utilization and an inability to meet the ever-increasing demands for high-performance and high-reliability services.

[0004] In summary, the most significant drawback of existing technologies is that current computing power network reconstruction strategies lack adaptability.

[0005] Reasons: Heuristic methods rely on human-defined rules and strategies, which are typically based on past experience or fixed standards. However, these rules may become inapplicable or ineffective in the face of changing environments and diverse needs. Resource reconstruction methods based on knowledge graphs and rule-based reasoning construct knowledge graphs from historical data and domain knowledge, and make decisions through rule-based reasoning. The construction of knowledge graphs and rule-based reasoning depend on existing knowledge and rules; therefore, when encountering new and complex scenarios, they may not be able to quickly generate new solutions.

[0006] Results: Existing methods lack the ability to intelligently learn and generate new reconstruction strategies in response to dynamic changes in adaptive computing power networks.

[0007] Secondary drawbacks: Existing computing power network reconstruction strategies lack real-time performance and computational efficiency.

[0008] Reason: While heuristic methods can provide solutions quickly, they often lack foresight regarding future changes. When faced with unknown scenarios, they struggle to develop optimal resource reconfiguration strategies and cannot guarantee efficient operation over extended periods. Rule-based reasoning-based reconfiguration methods, especially as network size increases, experience significantly higher computational complexity in the reasoning process, impacting real-time performance and decision-making efficiency.

[0009] Results: Existing computing networks are computationally inefficient, and their reconstruction strategies may not be well-suited to the current scenario.

[0010] Secondary drawback: Existing computing power network reconstruction strategies cannot cope with complex and unknown scenarios.

[0011] Reason: Heuristic methods cannot handle complex scenarios, especially when faced with entirely new or unknown requirements, where they encounter unforeseen problems. While knowledge graph-based and rule-based reasoning methods offer high flexibility, their rule sets are typically limited, failing to cover all possible network states or requirements, resulting in a lack of effective decision-making mechanisms when encountering unknown situations.

[0012] Result: Existing methods are difficult to apply to complex and ever-changing computing networks. Summary of the Invention

[0013] To address the lack of adaptability and creativity in existing computing network resource reconfiguration strategies, this invention provides a method, system, and storage medium for generating computing network reconfiguration strategies. By introducing AIGC technology and utilizing long-term accumulated resource reconfiguration scheme data and real-time dynamic data, a more flexible, intelligent, and adaptive resource reconfiguration strategy is generated. This fundamentally overcomes the limitations of existing computing network resource reconfiguration methods, achieving a flexible, adaptive, efficient, real-time, and future-oriented resource reconfiguration solution. By deeply integrating the generation technology with the computing network, a novel computing network resource management framework with continuous learning and dynamic optimization capabilities is constructed, providing strong support for addressing increasingly complex and dynamic computing demands.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] In a first aspect, the present invention provides a method for generating a computing power network reconfiguration strategy, comprising:

[0016] In response to receiving a new task requirement, calculate the fit between the task requirement and the strategies in the historical reconstruction strategy database;

[0017] Determine if there are any historical reconstruction strategies that meet the preset conditions for fit;

[0018] If so, then the historical reconstruction strategy that meets the preset conditions shall be selected;

[0019] If not, the task requirements and the current computing power network resource status are input into a generative model based on AI-generated content to generate a new reconstruction strategy.

[0020] Based on the selected or generated reconstruction strategy, perform computing network resource reconstruction.

[0021] Secondly, the present invention provides a computing power network reconstruction strategy generation system based on artificial intelligence-generated content, comprising:

[0022] The adaptability calculation and judgment module is configured to, in response to receiving a new task requirement, calculate its adaptability with the strategies in the historical reconstruction strategy database, and determine whether there are any historical reconstruction strategies whose adaptability meets the preset conditions.

[0023] The strategy generation module is configured to: if it exists, output the historical reconstruction strategy that meets the conditions; if it does not exist, input the task requirements and the current computing power network resource status into a generative model based on artificial intelligence-generated content to generate and output a new reconstruction strategy.

[0024] The execution control module is configured to control the computing power network to perform resource reconstruction based on the reconstruction strategy output by the strategy generation module.

[0025] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0026] Preferably, the generative model based on artificial intelligence to generate content is obtained by training a pre-trained generative neural network model using a historical reconstruction scheme dataset, wherein the training aims to optimize the model's generation strategy to meet the constraints of the task requirements.

[0027] Preferably, the calculation of the fit degree includes: extracting the feature representations of the task requirements and the historical strategies respectively, and calculating the similarity between the two feature representations as the fit degree.

[0028] Preferably, the calculation of adaptability includes: evaluating a first performance indicator of the new task requirement under the current resource state; evaluating a second performance indicator of the new task requirement after applying the historical reconstruction strategy; and determining the adaptability based on the difference between the first performance indicator and the second performance indicator.

[0029] Preferably, the method further includes a feedback step: after performing resource restructuring, obtaining task execution effect data; adding the new task requirements, the restructuring strategy adopted, and the task execution effect data to the historical restructuring strategy database and / or for updating the generative model.

[0030] The beneficial effects that the computing power network reconfiguration strategy generation method, system, and storage medium disclosed in this application may bring include, but are not limited to:

[0031] 1. It achieves high adaptability and strong intelligent generation capability of the reconstruction strategy, effectively dealing with unknown and complex scenarios.

[0032] This invention fundamentally changes the strategy generation model, upgrading from relying on "retrieving fixed rules" to "creating adaptive solutions." By introducing a generative model based on AI-generated content, the system is no longer limited to a preset rule base or a finite knowledge graph. When faced with new, complex, or dynamically changing task requirements (such as simultaneously requiring ultra-low latency, high security, and specific algorithms), the system can autonomously generate new and reasonable reconstruction strategies based on its understanding of the requirements and resource status. This breaks through the limitations of traditional methods and significantly improves the responsiveness and service reliability of computing networks when facing unknown challenges.

[0033] 2. An optimal balance between decision-making efficiency and strategy quality has been achieved, ensuring the real-time performance of the system.

[0034] This invention's innovative two-stage decision-making mechanism (first calculating fitness for matching, then deciding whether to reuse or generate) achieves a balance between efficiency and quality. For a large number of common or similar task requirements, the system, through efficient fitness calculation (such as feature similarity comparison), can match and reuse historically optimal strategies in milliseconds, avoiding unnecessary model call overhead. Only when historical experience is insufficient will the computationally intensive generative model be activated. This design significantly reduces the system's average decision latency, makes overall resource scheduling more agile, and meets the stringent real-time requirements of computing networks.

[0035] 3. Improved the utilization efficiency of global computing network resources and task execution performance.

[0036] Because generative models can learn deep optimization patterns from massive amounts of historical data, the strategies they generate can perform more refined and global collaborative scheduling of multi-dimensional resources such as computing power, algorithms, data, and models. Compared with heuristic methods based on local or single-objective optimization, the strategies generated by this invention often lead to better overall performance. Practical applications show that this method can effectively reduce the average task completion latency and improve the overall utilization of key resources such as CPU, GPU, and bandwidth, thereby supporting more efficient task processing with less resource consumption.

[0037] 4. It endows the system with the ability to continuously evolve and optimize itself, thus possessing long-term vitality.

[0038] This invention constructs a complete closed loop of "decision-execution-learning" by introducing a feedback step (corresponding to the system's data feedback module). Each successful strategy execution is recorded as new experience, fed back into the historical reconstructed strategy database, and used for iterative updates of the generative model. This continuously enriches the system's "experience base" and optimizes the generated strategy model. The system thus possesses continuous learning capabilities, its performance improves with runtime, and it can adapt to changes in network environment and task distribution, achieving a leap from "static intelligence" to "dynamic growth," significantly extending the effective lifespan of the technical solution. Attached Figure Description

[0039] Figure 1 This is an exemplary flowchart of a computing power network reconstruction strategy generation method according to some embodiments of this specification;

[0040] Figure 2 This is an exemplary flowchart of a method for generating a computing power network reconstruction strategy in one embodiment; Detailed Implementation

[0041] Various exemplary embodiments, features, and aspects of this application will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0042] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, an indirect connection through an intermediate medium, or the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0043] The technical solution of this application utilizes AIGC technology to train a computing power network reconstruction strategy generation model through transfer learning and a Transformer-based generative model, thereby realizing the automatic generation of computing power network resource reconstruction strategies to cope with the complex and ever-changing needs in computing power networks.

[0044] By collecting, classifying, and cleaning various historical resource reconstruction schemes in the computing power network, we obtain long-term accumulated reconstructable historical data for solving different needs. Through data cleaning, we remove abnormal and missing information and construct a unified dataset for training the reconstruction strategy generation model.

[0045] Please see Figure 1The method of this invention mainly includes two stages: offline preparation and online decision-making. The online decision-making stage embodies the core logic of "calculating fitness - judgment - branch decision".

[0046] Part 1: Offline Preparation Phase

[0047] 1. Construction of the historical reconstruction scheme dataset

[0048] Data is the foundation for model training and matching decisions. The construction process includes:

[0049] (1) Data collection: Collect historical triplet data from the task scheduling logs, resource monitoring system and performance tracing tools of the computing power network, namely: task requirements, execution reconstruction strategy and task execution effect.

[0050] Reconfiguration strategy refers to a scheme or set of instructions that guides a computing network on how to adjust its internal resource allocation and scheduling.

[0051] For example, a refactoring strategy could specify that a particular computing task is offloaded from the cloud to an edge server for execution, a designated data link and algorithm model are allocated to the task, and a corresponding resource reservation ratio is set.

[0052] A dataset is a collection of data used for model training or analysis. For example, a dataset may consist of multiple records, each containing a description of the requirements for a historical task, the text of the reconstruction strategy used, and the final performance metric.

[0053] (2) Data cleaning and standardization: Clean the collected raw data, remove invalid records (such as execution failure or serious missing data fields), and standardize key fields, such as unifying the latency unit to milliseconds and encoding the task type.

[0054] (3) Feature Engineering and Vectorization: To support efficient fit calculation, unstructured task requirement information needs to be converted into machine-processable feature vectors. For example, a task requirement can be quantified into a vector containing dimensions such as [task type encoding, latency requirement, bandwidth requirement, computational complexity rating]. The corresponding reconstruction strategy can also be encoded as a strategy feature vector or summary.

[0055] Matching refers to the process of determining the degree of association or compatibility between two or more objects. For example, matching could involve calculating the cosine similarity between the feature vector of a new task requirement and the feature vector of historical tasks in a database, and then making a judgment based on the similarity score.

[0056] Similarity is a numerical measure used to quantify the degree of similarity between two objects. For example, similarity can be a value between 0 and 1 obtained by methods such as cosine similarity, the reciprocal of Euclidean distance, or Pearson correlation coefficient.

[0057] 2. Training of the generative reconstruction model

[0058] The core intelligent agent of this invention is a generative model based on AIGC, and its training objective is to learn the mapping from "problem description" to "high-quality solution".

[0059] Generative models based on artificial intelligence-generated content (AIGC) refer to machine learning models that can automatically generate new data that meets specific format or content requirements based on input information.

[0060] For example, the model could be a large language model based on the Transformer architecture. After training, it could output a planning text describing how resources should be reconstructed, based on text describing task requirements and resource status.

[0061] (1) Model architecture selection: Transformer-based sequence generation models (such as the GPT architecture) or encoder-decoder-based conditional generation models can be used. These models are good at understanding and generating text with complex structure and semantics.

[0062] (2) Training process: Supervised training is performed using the dataset constructed above. The input is a combination of task requirement description and current resource state, and the output is the reconstructed policy text. The loss function (such as cross-entropy loss) is used to measure the difference between the generated policy and the real historical policy.

[0063] (3) Integration of optimization objectives: In training, reinforcement learning ideas or weighted sampling can be introduced to use the task execution effect as a signal to guide the model to tend to generate those strategies with better historical execution effects, thereby achieving training with the goal of "meeting the requirements and optimizing performance".

[0064] Task execution effect refers to the result or performance index produced after a task is executed in a computing network according to a certain reconstruction strategy.

[0065] For example, task performance can include the actual completion time of the task, the total energy consumption, the average utilization rate of various resources used, and the accuracy of achieving the task objectives.

[0066] Part Two: Online Dynamic Decision-Making and Execution Phase

[0067] Step S101: Receive and parse new task requirements.

[0068] The system interface receives task requests, parses out their specific constraints (such as maximum latency and minimum bandwidth) and resource requirements, and forms a structured requirement object.

[0069] Step S102: Calculate and determine the fitness level.

[0070] This is the key to the first decision-making stage.

[0071] Example 1 of Fit Calculation (Based on Feature Similarity): The new task requirement is transformed into a feature vector V_new. For each strategy in the historical database, its corresponding historical requirement feature vector V_hist is obtained. The similarity (e.g., cosine similarity) between the two feature representations, V_new and V_hist, is calculated, and this similarity is used as the initial fit. This method is fast and suitable for scenarios where the requirement dimensions are clear and quantifiable.

[0072] New task requirements refer to the input information or instructions that trigger the computing network to make resource reconfiguration decisions.

[0073] For example, a new task requirement could be a video stream analysis task that requires low latency, a big data batch processing job that requires high throughput, or an AI model training task that has specific requirements for computational accuracy.

[0074] Adaptability calculation example two (based on performance improvement prediction): This is a more accurate evaluation method. For each candidate historical strategy, the system performs a lightweight simulation:

[0075] a) Estimate the performance metrics P_before of new task requirements under the current network resource conditions (without refactoring) (e.g., estimated latency).

[0076] b) Estimate the performance metric P_after under the expected resource conditions after applying this historical strategy.

[0077] c) Calculate the fitness score of the strategy as Adaptation_Score = P_after - P_before. This score directly quantifies the performance improvement brought about by the strategy.

[0078] Judgment Logic: Regardless of the calculation method used, the system generates a suitability score for each candidate strategy. Then, it determines whether there is a strategy with a score exceeding preset conditions (e.g., similarity > 0.85 or performance improvement > 20%). If so, proceed to step S103; otherwise, proceed to step S104.

[0079] Step S103: Select the historical adaptation strategy.

[0080] When a strategy that meets the conditions is determined to exist, the system directly selects the historical reconstruction strategy with the highest adaptability. This approach enables the rapid reuse of millisecond-level decisions and experience.

[0081] Step S104: Invoke the generative model to create a new strategy.

[0082] When no suitable historical strategy is available, the creation process is initiated. A detailed description of the task requirements and a comprehensive snapshot of the current computing network resource status (including node load, link bandwidth, available services, etc.) are fed into the trained generative model. The model infers based on its internal knowledge to generate a completely new text describing a reconstruction strategy specific to the current context.

[0083] The current computing network resource status refers to the description of the availability, load, and performance of various resources within the computing network at a certain moment.

[0084] For example, the current resource status may include: CPU / GPU utilization of each computing node, remaining memory, real-time bandwidth of network links, I / O performance of data storage nodes, and the readiness status of various algorithm models.

[0085] Step S105: Perform computing network resource reconstruction according to the selected or generated reconstruction strategy.

[0086] The execution control module parses the final reconstruction strategy (whether from S103 or S104), decomposes it into specific operation instructions for four dimensions: computing power (scheduling computing instances), algorithms (deploying service images), data (configuring storage and transmission), and models (loading AI models), and executes the reconstruction of computing power network resources by calling the API of the underlying infrastructure.

[0087] Performing computing network resource reconfiguration refers to the process of actually adjusting and configuring resources in the computing network according to the reconfiguration strategy.

[0088] For example, performing a refactoring may include: launching a compute container on a designated edge server, transferring the data to be processed to the container via a selected high-speed network path, and loading a specific AI inference model into the container.

[0089] Step S106: Feedback and optimization (closed loop).

[0090] During and after task execution, the system collects detailed task execution performance data. Then, a feedback mechanism is activated:

[0091] (1) Data entry: The successful case (new task requirements, refactoring strategy adopted, task execution effect) is added to the historical refactoring strategy database as a new case to enrich the experience base.

[0092] A historical reconstruction strategy database refers to a collection that stores past computing power network resource reconstruction schemes and related information. For example, this database can be a relational database table, where each record contains the demand characteristics of a past task, the resource scheduling strategy adopted (such as allocating computing tasks to specific edge nodes), and the performance indicators after the strategy was implemented (such as task completion time and resource utilization).

[0093] (2) Model iteration: Add new successful cases to the training set regularly (e.g., weekly) to incrementally learn or fine-tune the generative model so that its policy generation ability can continue to evolve and better adapt to changes in network environment and task mode.

[0094] Part Three: Examples and Effects

[0095] The solution described in this invention is deployed in a simulated computing network that includes a cloud computing center, multiple edge nodes, and terminal devices to handle mixed workloads.

[0096] Scenario Example 1 (Successful Fit): A "4K video real-time super-resolution" task arrives, requiring a processing latency of <50ms. The system calculates its fit with strategies in the historical database (based on feature similarity) and finds a historical strategy for "high-definition video low-latency processing" with a fit of 0.88 (exceeding the threshold of 0.85). The system directly reuses this strategy, starting a video processing container at a designated edge node, and the task is completed within 45ms.

[0097] Scenario Example 2 (Adaptability Matching Failure, AIGC Generation): A "Cross-Domain Federated Learning Model Aggregation" task arrives, with specific requirements for data security and aggregation latency. No similar strategies are found in the historical database, resulting in a matching failure. The system invokes the generative model, inputting the task requirements, security strategy, and current resource status of each domain. The model generates a new strategy: "Complete model training locally on the participating parties, use a secure multi-party computation protocol to aggregate encrypted model parameters at a trusted relay node, and select the dedicated line with optimal latency for the aggregation path." This strategy is successfully executed, satisfying both security and performance requirements.

[0098] Quantitative Results: Long-term operational statistics show that after adopting the solution of this invention, the average completion time of the overall system tasks is reduced by about 30%, and the comprehensive utilization rate of resources is improved by about 18%. With the continuous effect of the feedback loop, the success rate of the first application of the generative model generation strategy (without manual intervention) increased from 70% to 92% within six months.

[0099] Part Four: System Architecture Examples

[0100] The system corresponding to the claims can be implemented as a distributed software architecture:

[0101] Adaptability calculation and judgment module: can be deployed as a stateless service, receives task requirements, accesses a vectorized historical policy database, and performs adaptability calculation and threshold judgment.

[0102] The policy generation module includes a policy cache (for returning historical policies) and an AIGC model service client. When generation is needed, it calls the generative model inference service deployed on high-performance computing nodes via RPC.

[0103] Execution control module: Implemented as a resource orchestration engine, containing drivers for different resource types (compute, network, storage), translating policies into specific operations for platforms such as Kubernetes, OpenStack, and SDN controllers.

[0104] Data feedback module: As a background service, it consumes task execution effect data streams from the monitoring system, persists them to the database, and triggers the model training pipeline.

[0105] The modules communicate with each other through message queues and RPC frameworks to ensure high availability and scalability.

[0106] Example 2:

[0107] Figure 2 An exemplary flowchart of the computing power network reconstruction strategy generation method according to an embodiment of the present invention is shown;

[0108] In a cloud-edge-device collaborative computing network, an enterprise needs to process a batch of tasks. These tasks include various types such as real-time video analytics, AI model training, and file storage, and require meeting different performance metrics, such as low latency, high throughput, or low power consumption. Simultaneously, the resource status within the computing network, such as edge node computing power, link bandwidth, and data center load, is dynamically changing. Existing technologies cannot generate resource reconfiguration solutions that meet these diverse needs in a short period of time.

[0109] Step S1: Collect and process historical reconstruction scheme data to form a historical reconstruction scheme dataset.

[0110] Task scheduling and resource reconfiguration schemes are collected from the historical records of the computing power network, including the allocation of computing power nodes for task requirements, task offloading paths, storage usage strategies, and the allocation of computing power, algorithms, data, and models, along with the reconfiguration strategies adopted. After obtaining this data, it is cleaned to remove invalid records, such as outliers and missing values, ensuring data integrity. Finally, the data is categorized into three types based on task type and resource requirements: low-latency, high-throughput, and low-power, forming a multi-objective historical reconfiguration scheme dataset.

[0111] Step S2: Train a generative reconstruction model based on AI-generated content.

[0112] We employ a generative model based on the Transformer architecture, training it to generate reconstruction strategies based on long-term accumulated historical reconstruction strategies. Simultaneously, we utilize transfer learning techniques, pre-training the model on publicly available datasets from similar scenarios (such as cloud computing resource scheduling) to accelerate model learning.

[0113] The generative model is trained using the constructed dataset, and specific optimization objectives are defined: for low-latency tasks, the objective is to optimize task response latency; for high-throughput tasks, the objective is to optimize bandwidth utilization; and for low-power tasks, the objective is to minimize system energy consumption.

[0114] The model was validated using a test dataset to ensure that the generated reconstruction strategy could meet the target requirements in different scenarios, achieving a requirement mismatch rate of less than 10% after strategy execution.

[0115] Step S3: Receive new task requirements and perform matching judgment.

[0116] The new task requirement is: real-time video analysis task, requiring low latency of less than 20ms and high throughput (>500MB / s). Upon task arrival, a matching judgment is performed: the historical reconstruction strategy database is searched to see if a historical strategy that meets the requirements already exists. The fit is evaluated by calculating the cosine similarity between feature representations. If the fit exceeds a preset threshold, it is considered a successful match, and the historical strategy is directly selected. If the fit does not reach the threshold (i.e., the latency or throughput requirements cannot be fully met by the historical strategy), then proceed to step S4.

[0117] Step S4: Invoke the generative model to create a new reconstruction strategy.

[0118] The trained generative reconstruction model is invoked, inputting the current computing network resource status (including the specific availability of multiple resource elements such as computing power, algorithms, data, and models) and the aforementioned task requirements. Based on the current task requirements and the existing resource structure, the model generates a completely new reconstruction strategy. This strategy specifies how to adjust each resource element: to meet low latency requirements, the strategy schedules the task to an edge server geographically close to the data source and employs an efficient distributed algorithm; to meet high throughput requirements, the strategy specifies the use of high-throughput data links for transmission.

[0119] Step S5: Perform resource restructuring and provide feedback for optimization.

[0120] Based on the generated reconstruction strategy, the computing network resources are reconstructed, specifically adjusting the resource allocation of the computing network. The task execution effect is monitored, recording metrics such as task completion time, resource utilization, and energy consumption. The monitored task execution effect data, along with the task requirements and the adopted reconstruction strategy, are fed back into the historical reconstruction strategy database and used for subsequent optimization and training of the generative reconstruction model.

[0121] Technical Results: For unmatched task requirements, the generative model can generate reconstruction strategies that meet the current needs in real time. The system can adapt to dynamically changing network states and task requirements, improving the flexibility of task execution. The two-stage decision-making mechanism ensures rapid response for matched tasks, while unmatched tasks generate strategies in real time through the generative model, with a generation time of less than 500ms. The generated strategies can handle complex scenarios with multi-objective optimization. The strategies generated through AIGC technology optimize the allocation of heterogeneous resources, significantly improving the overall performance of the computing network, resulting in an average reduction of 32% in task completion time and an increase of 18% in resource utilization.

[0122] This embodiment has at least the following beneficial effects:

[0123] 1. Addressing the main drawback: lack of adaptability.

[0124] It achieves adaptive resource reconfiguration strategies, dynamically generating new optimization strategies based on real-time demands and environmental changes. Freed from the constraints of fixed rules, it possesses intelligent learning and self-evolution capabilities, significantly enhancing the flexibility and responsiveness of computing networks.

[0125] This invention introduces AIGC technology, utilizing a Transformer-based generative model to construct a generative reconstruction model by training it on long-term accumulated reconstruction scheme data. The generative model possesses continuous learning capabilities, dynamically optimizing through a feedback mechanism and automatically generating new reconstruction strategies when real-time demands arise.

[0126] 2. Regarding secondary drawback 1: Insufficient real-time performance and computational efficiency.

[0127] This significantly improves the efficiency of generating reconstruction strategies for computing networks, ensuring real-time decision-making. The two-stage mechanism optimizes the model invocation frequency, reduces computational overhead, and improves the accuracy of strategy matching. The generative model can respond to changing needs in real time, providing highly compatible reconstruction strategies for the current scenario.

[0128] This invention employs a two-stage policy generation mechanism. First, it searches for matching historical solutions in the reconstructed policy database to quickly respond to simple needs. For complex or unmatched needs, it invokes a generative model to generate policies in real time, ensuring real-time performance. It utilizes an efficient Transformer-based generative model for training, thereby improving the speed of policy generation.

[0129] 3. Regarding secondary drawback 2: Inability to handle complex and unknown scenarios.

[0130] This invention possesses the ability to handle complex and unknown scenarios, and the strategies generated through AIGC technology can cover a wider range of applications. When faced with diverse, dynamic, and complex computing network demands, it can generate high-quality reconstruction schemes, unrestricted by fixed rules and limited knowledge. By extending the generative model's capabilities through transfer learning, it enables rapid adaptation from known domains to new domains, solving resource scheduling problems in unknown scenarios.

[0131] This invention constructs a generative model using AIGC technology, learns latent patterns in complex scenarios by combining long-term accumulated data, and utilizes transfer learning to transfer knowledge from related domains, enhancing the model's adaptability to new scenarios. The generative model can not only handle complex requirements in existing scenarios but also generate reasonable resource reconfiguration strategies based on input requirements when encountering unknown scenarios.

[0132] Those skilled in the art should understand that the above embodiments are merely examples illustrating the principles and implementation of the present invention and are not intended to limit the present invention. For example, the calculation method of fitness is not limited to the two described above; any algorithm that can measure the correlation or fit between task requirements and historical strategies falls within the scope of the "fit" concept of the present invention. The specific architecture of the generative model can also adopt other advanced generative technologies such as diffusion models. The timing of feedback updates can be real-time streaming or periodic batch processing. Any modifications, equivalent substitutions, or improvements made within the core framework of "two-stage (reuse / generation) decision based on fitness judgment" and "execution feedback closed loop" proposed in this invention should be included within the protection scope of this invention.

Claims

1. A method for generating a computing power network reconfiguration strategy, characterized in that, include: In response to receiving a new task requirement, calculate the fit between the task requirement and the strategies in the historical reconstruction strategy database; Determine if there are any historical reconstruction strategies that meet the preset conditions for fit; If so, then the historical reconstruction strategy that meets the preset conditions shall be selected; If not, the task requirements and the current computing power network resource status are input into a generative model based on AI-generated content to generate a new reconstruction strategy. Based on the selected or generated reconstruction strategy, perform computing network resource reconstruction.

2. The method according to claim 1, characterized in that, The generative model based on artificial intelligence is obtained by training a pre-trained generative neural network model using a historical reconstruction scheme dataset. The training aims to optimize the model's generation strategy to meet the constraints of the task requirements.

3. The method according to claim 2, characterized in that, The steps for constructing the historical reconstruction scheme dataset include: Collect data including historical task requirements, the refactoring strategies adopted, and the task execution results; The collected data is cleaned and standardized.

4. The method according to claim 2, characterized in that, The generative neural network model is based on the Transformer architecture.

5. The method according to claim 1, characterized in that, The steps for calculating the fitness degree include: Extract the feature representations of the task requirements and the historical strategies respectively; Calculate the similarity between two feature representations as the fit.

6. The method according to claim 1, characterized in that, The steps for calculating the fitness degree include: Evaluate the first performance metric of the new task requirements under the current resource conditions; Evaluate the second performance metric of the new task requirements after applying the history reconstruction strategy; The fit is determined based on the difference between the first performance index and the second performance index.

7. The method according to any one of claims 1 to 6, characterized in that, The method also includes a feedback step: After performing resource restructuring, obtain task execution effect data; The new task requirements, the refactoring strategy adopted, and the task execution effect data are added to the historical refactoring strategy database and / or used to update the generative model.

8. A computing power network reconstruction strategy generation system based on artificial intelligence-generated content, characterized in that, include: The adaptability calculation and judgment module is configured to, in response to receiving a new task requirement, calculate its adaptability with the strategies in the historical reconstruction strategy database, and determine whether there are any historical reconstruction strategies whose adaptability meets the preset conditions. The strategy generation module is configured to: if it exists, output the historical reconstruction strategy that meets the conditions; if it does not exist, input the task requirements and the current computing power network resource status into a generative model based on artificial intelligence-generated content to generate and output a new reconstruction strategy. The execution control module is configured to control the computing power network to perform resource reconstruction based on the reconstruction strategy output by the strategy generation module.

9. The system according to claim 8, characterized in that, Also includes: The data feedback module is configured to collect task execution effect data after resource reconstruction is executed, and add the new task requirements, the reconstruction strategy adopted, and the execution effect data to the historical reconstruction strategy database and / or to update the generative model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images