Service flow scheduling model generation method and service flow scheduling method

Abstract

The application provides a service flow scheduling model generation method and a service flow scheduling method. The service flow scheduling model generation method comprises the following steps: acquiring multi-source data of service flow scheduling tasks in multiple fields, training a large language model based on the multi-source data to obtain an intermediate model, and the multi-source data comprises a physical mechanism of a service flow; inputting sample scheduling task data into the intermediate model to obtain sample scheduling decision text output by the intermediate model; determining a training reward based on the accuracy of scheduling parameters in the sample scheduling decision text; updating the intermediate model based on the training reward to obtain a service flow scheduling model. The application can reduce the deployment cost of the service flow scheduling model.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for generating a business flow scheduling model and a business flow scheduling method. Background Technology

[0002] In the field of systematic artificial intelligence, the attribute of "network native" is proposed, which aims to incorporate the atomic AI capability service status into the network control plane and management plane. Through capability registration, authentication and cataloging, multi-dimensional information such as AI capability service status (e.g., accuracy, performance, service scenarios, data requirements, cloud / edge / device deployment), network configuration (e.g., bandwidth, latency), and business requirements (e.g., accuracy, efficiency, application scenarios) are linked together, so that AI services can be scheduled and distributed like network slices / microservices, improving the ubiquity and controllability of AI services.

[0003] In existing technologies, intelligent scheduling of business flows for AI services is often achieved by training a scheduling model. However, this approach faces issues of versatility and generalization: a scheduling model trained in one scenario cannot be directly transferred to other scenarios, such as scenarios with different domain knowledge, different resource distributions, or different business combinations. It must be retrained and its parameters tuned, resulting in high deployment costs for scheduling models in existing technologies. Summary of the Invention

[0004] This application provides a service flow scheduling method, apparatus, equipment, medium, and product to address the high deployment cost of existing service flow scheduling models and achieve the effect of reducing the deployment cost of service flow scheduling models.

[0005] This application provides a method for generating a service flow scheduling model, including: Acquire multi-source data of business flow scheduling tasks from multiple domains, train a large language model based on the multi-source data to obtain an intermediate model, wherein the multi-source data includes the physical mechanism of business flow; Input the sample scheduling task data into the intermediate model and obtain the sample scheduling decision text output by the intermediate model; The training reward is determined based on the accuracy of the scheduling parameters in the sample scheduling decision text. The intermediate model is updated based on the training reward to obtain the business flow scheduling model.

[0006] According to the business flow scheduling model generation method provided in this application, the step of determining the training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text includes: Based on the parameter categories and parameter values ​​of the scheduling parameters in the sample scheduling decision text; An accuracy reward is determined based on the parameter category and the accuracy of the parameter value. The training reward is determined based on the accuracy reward.

[0007] According to the service flow scheduling model generation method provided in this application, the step of determining an accuracy reward based on the parameter category and the accuracy of the parameter value includes: Obtain the labeled parameter category and labeled parameter value of each parameter to be scheduled in the labeled scheduling decision text corresponding to the sample scheduling task data; A first accuracy reward is determined based on the degree of overlap between the parameter category and the labeled parameter category; For overlapping parameter categories, a second accuracy bonus is determined based on the difference between the parameter value and the labeled parameter value; The accuracy reward is determined based on the first accuracy reward and the second accuracy reward.

[0008] According to the service flow scheduling model generation method provided in this application, determining the training reward based on the accuracy reward includes: Based on the standardization of the sample scheduling decision text, a standardization reward is determined; The training reward is determined based on the prescriptive reward and the accuracy reward; The normative reward includes a structural normative reward and a grammatical normative reward. The structural normative reward reflects the consistency between the structure of the sample scheduling decision text and the preset structural template, and the grammatical normative reward reflects the consistency between the sample scheduling decision text and the grammatical rules.

[0009] According to the business flow scheduling model generation method provided in this application, the step of training a large language model based on the multi-source data to obtain an intermediate model includes: The multi-source data is classified to obtain knowledge data, spatiotemporal data, and background information. The different categories of multi-source data are converted into natural language. The knowledge data includes knowledge information and physical mechanisms in the multi-source data. The spatiotemporal data includes the values ​​of various business indicators in the multi-source data at different times. The background information is the explanatory information of the business indicators in the multi-source data. The knowledge data, spatiotemporal data, and background information converted into natural language are aligned in the time dimension to form multiple sets of training samples. Each set of training samples is divided into a first part and a second part. The first part of the training sample is input into the large language model to obtain the predicted text output by the large language model. Based on the difference between the predicted text and the second part of the training sample, the training loss is determined. The large language model is updated based on the training loss corresponding to multiple training samples to obtain the intermediate model.

[0010] This application also provides a service flow scheduling method, including: Input the business flow scheduling task data into the business flow scheduling model; Obtain the scheduling decision output by the service flow scheduling model; The service flow scheduling model is generated based on the service flow scheduling model generation method described in any of the above-mentioned methods.

[0011] This application also provides a service flow scheduling model generation apparatus, comprising: The initial training module is used to acquire multi-source data of business flow scheduling tasks in multiple domains, and to train a large language model based on the multi-source data to obtain an intermediate model. The multi-source data includes the physical mechanism of business flow. The sample scheduling decision text generation module is used to input sample scheduling task data into the intermediate model and obtain the sample scheduling decision text output by the intermediate model. The training reward determination module is used to determine the training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text. The model update module is used to update the intermediate model based on the training reward to obtain the business flow scheduling model.

[0012] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the traffic flow scheduling model generation method as described above and / or the traffic flow scheduling method as described above.

[0013] This application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the traffic flow scheduling model generation method as described above and / or the traffic flow scheduling method as described above.

[0014] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the traffic flow scheduling model generation method as described above and / or the traffic flow scheduling method as described above.

[0015] The business flow scheduling model generation method and business flow scheduling method provided in this application acquire multi-source data of business flow scheduling tasks from multiple domains. This multi-source data includes the physical mechanism of business flow. A large language model is trained based on the multi-source data to obtain an intermediate model. In this way, the intermediate model can learn the inherent physical mechanism of business flow in different domains, rather than being limited to surface observation data. For the intermediate model that has learned the physical mechanism of business flow, it is trained again based on specific sample scheduling task data. During the training process, the training reward is determined based on the accuracy of the scheduling parameters in the sample scheduling decision text output by the intermediate model. The intermediate model is updated based on the training reward to obtain the business flow scheduling model. This allows the trained business flow scheduling model to output accurate scheduling decisions. The business flow scheduling model obtained by combining the two stages of training can not only accurately output scheduling decisions, but also be applicable to different business flow scheduling task scenarios without the need to retrain the model when changing scenarios, thus reducing the deployment cost of the business flow scheduling model. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the business flow scheduling model generation method provided in this application.

[0018] Figure 2 This is a schematic diagram of the model capability transformation process of the service flow scheduling model generation method provided in this application.

[0019] Figure 3 This is a schematic diagram of the model training phase in the business flow scheduling model generation method provided in this application.

[0020] Figure 4 This is a schematic diagram illustrating the composition of training samples in the service flow scheduling model generation method provided in this application.

[0021] Figure 5 This is a schematic diagram illustrating the principle of the training sample combination order in the business flow scheduling model generation method provided in this application.

[0022] Figure 6 This is a flowchart illustrating the service flow scheduling method provided in this application.

[0023] Figure 7 This is a schematic diagram of the business flow scheduling device provided in this application.

[0024] Figure 8 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0029] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0030] The following is combined Figures 1-5 Describe the service flow scheduling method provided in this application. For example... Figure 1 As shown, the service flow scheduling method includes the following steps: S110. Obtain multi-source data of business flow scheduling tasks in multiple fields, train the large language model based on the multi-source data to obtain an intermediate model. The multi-source data includes the physical mechanism of business flow. S120. Input the sample scheduling task data into the intermediate model and obtain the sample scheduling decision text output by the intermediate model. S130. Determine the training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text; S140. Update the intermediate model based on the training reward to obtain the business flow scheduling model.

[0031] The business flow scheduling model generation method provided in this application acquires multi-source data on business flow scheduling tasks from multiple domains. This multi-source data includes the physical mechanism of business flows. Based on the multi-source data, a large language model is trained to obtain an intermediate model. In this way, the intermediate model can learn the inherent physical mechanism of business flows in different domains, rather than being limited to surface observation data. For the intermediate model that has learned the physical mechanism of business flows, it is trained again based on specific sample scheduling task data. During the training process, the accuracy of the scheduling parameters in the sample scheduling decision text output by the intermediate model is used to determine the training reward. The intermediate model is updated based on the training reward to obtain the business flow scheduling model. This allows the trained business flow scheduling model to output accurate scheduling decisions. The business flow scheduling model obtained by combining the two stages of training can not only accurately output scheduling decisions, but also be applicable to different business flow scheduling task scenarios without the need to retrain the model when changing scenarios, thus reducing the deployment cost of the business flow scheduling model.

[0032] In the method provided in this application, such as Figure 2 As shown, firstly, multi-source data related to business flow scheduling tasks are acquired as perceptual information for the large language model, enabling the model to fully explore the real physical mechanism of business operation, avoid model overfitting caused by data deficiency, and improve the model's versatility and generalization. After the model has the ability to accurately understand the physical mechanism, it is trained again using specific business flow scheduling task samples to support the optimization of business flow scheduling decisions, enabling the model to make autonomous decisions on business flow scheduling tasks.

[0033] To ensure the comprehensiveness of the perceived information, the method provided in this application, such as... Figure 3 As shown, multi-source data includes observational data, historical data, knowledge data, and physical mechanisms of business flows. To improve the versatility of the business flow scheduling model generated by the method provided in this application, the multi-source data is not limited to a single domain but covers multiple domains, such as communications, energy, agriculture, and IT. In practical applications, business flow scheduling processes primarily handle structured spatiotemporal data such as business indicators, scheduling decision schemes are event data, and knowledge is text data. To achieve the fusion of these multi-source data, the method provided in this application processes the multi-source data, merging and aligning it into a natural language corpus for model processing, ensuring the comprehensiveness and versatility of the model's perception.

[0034] After collecting multi-source data, the first step is data cleaning, including anomaly removal and missing data compensation. The multi-source data is then categorized into knowledge data, spatiotemporal data, and background information. Knowledge data includes knowledge information (e.g., domain background knowledge) and physical mechanisms (e.g., explanatory text on the impact of computing resource usage on network latency). Spatiotemporal data includes the values ​​of various business indicators at different times (e.g., real-time network card bandwidth, latency, etc.). Background information provides explanations for the business indicators (e.g., the corresponding business capability type, unit, etc.). Further data preprocessing is then performed, such as data normalization and standardization. For spatiotemporal data normalization, various data characteristics must be fully considered, and expert experience should be incorporated to select the most suitable normalization method.

[0035] After obtaining the knowledge data, spatiotemporal data, and background information converted into natural language, these three elements are aligned along the time dimension to form multiple sets of training samples. In other words, the knowledge data, spatiotemporal data, and background information aligned along the time dimension are grouped into a single training sample. When creating these training samples, special characters can be added to distinguish different types of data. For example, for knowledge data, [a special character can be added].<DATA_KNOWLEDGE> The identifier; for spatiotemporal data, add...<NUM_TABLE> The logo, and can also be used in it. <header>And represent the row and column information of spatiotemporal data; while for background information, you can add<DATA_BACKGROUND> The identifier. In one embodiment, examples of training samples are shown in Table 1.

[0036] Table 1 ###Knowledge Data:<DATA_KNOWLEDGE> 1. Data Domain: Telecommunications Industry. 2. Data Sub-domain: Service Flow Scheduling. 3. Data Purpose: "Network Configuration Data" describes the current state of the network, covering bandwidth, latency, etc.; "Service Data" describes the current state of atomic capabilities, covering accuracy, performance, service scenarios, data requirements, deployment on cloud / edge / device sides, etc.; "Business Requirement Data" mainly describes the business objectives, i.e., the requirements that need to be comprehensively considered in achieving "Network Configuration Data" and "Service Data", covering business accuracy, efficiency, application scenarios, etc. 4. Physical Mechanism: High business accuracy (e.g., 99.9% detection accuracy) often relies on more complex model structures (larger models, deeper networks, more parameters), and model complexity directly affects computational load (FLOPs) and memory usage; large model inference and training require higher data throughput and lower latency. Distributed computation of model parameters or intermediate features will cause mutual constraints among bandwidth, latency, and computing power; network status (bandwidth, latency, packet loss rate) determines whether business requirements can be met. For example, speech recognition tasks have low bandwidth requirements but are sensitive to latency; video generation tasks have extremely high bandwidth and computing power requirements; network congestion or high latency can prevent business objectives from being achieved.< / DATA_KNOWLEDGE> ###Background Information:<DATA_BACKGROUND> 1. Network metrics: 3 dimensions – sampling time, bandwidth, and latency. 2. AI capability metrics: 3 dimensions – accuracy, service scenario, and data requirements.< / DATA_BACKGROUND> ### Spatiotemporal data is as follows: Overall data description: The data expresses a specific description of the network state and AI capabilities.<NUM_TABLE> Sampling Time | Bandwidth | Latency 2025-10-01T00:00 | 310.5 | 1.12 2025-10-01T00:15 | 201.4 | 0.76 2025-10-01T00:30 | 194.4 | 1.33...< / NUM_TABLE> After fusing and aligning multi-source data to form multiple sets of training samples, a dictionary is maintained during model pre-training. Special identifiers are added to the dictionary to ensure that the text is not segmented during word segmentation. Then, natural language processing is used to vectorize the input text (tokenizer) for the dictionary with added special characters, as shown in Table 2.

[0037] Table 2 logo token id <DATA_KNOWLEDGE> token_id: 32008 < / DATA_KNOWLEDGE> token_id: 32009 <DATA_BACKGROUND> token_id: 32010 < / DATA_BACKGROUND> token_id: 32011 In the process of training the large language model to obtain the intermediate model, the "lyrics prediction" method is adopted. That is, a part of the training sample is input into the large language model, and the large language model is required to predict the remaining part. Based on the difference between the predicted text output by the large language model and the remaining part of the training sample, the training loss is determined to update the large language model. This training method can fully learn the various information inherent in the text.

[0038] Since business flow data contains various structural information, in order for the large language model to learn the inherent relationships between these structures and thus learn a more accurate physical mechanism of the business flow, one possible implementation of the method provided in this application involves rearranging the order of different types of data when assembling training samples, such as... Figure 4 As shown, for example, the original structural order is knowledge data, background information, and spatiotemporal data, and three sets of training samples can be generated by changing the order.

[0039] When dividing training samples, the first one or two structural information items in the training samples can be used as the first part, and the rest as the second part. For example, the first part can be defined as knowledge data and background information, and the spatiotemporal data as the second part; or the first part can be defined as knowledge data and spatiotemporal data, and the background information as the second part; or knowledge data can be the first part, and the background information and spatiotemporal data as the second part, and so on. A certain proportion of training samples with different orders can be set. For example, for 80% of the data, the original structural order can be maintained: knowledge data and physical mechanisms, background information, and spatiotemporal data. In this way, the model can learn information sequentially, learning domain content from "knowledge data" and combining it with the data characteristics learned from "background information" to complete the understanding of specific business trends in "spatiotemporal data." For the remaining 10% of the data, the data structure order can be changed: knowledge data, spatiotemporal information, and background data. The model can learn domain content from "knowledge data and physical mechanisms" and combine it with the understanding of specific business trends in "spatiotemporal data" to complete the generation of "background information." For 10% of the data, by changing the order of the data structure—background information, spatiotemporal data, and knowledge data—the model can generate knowledge data from the background information and spatiotemporal data. By changing the order, a deeper understanding of the intrinsic relationships between the three types of information can be achieved, thereby enabling an understanding of the physical mechanisms of business flows.

[0040] It is worth noting that when dividing training samples, "knowledge data and physical mechanisms" and "background information" cannot be generated from "spatiotemporal data" because the representation of "spatiotemporal data" is not unique. The same "spatiotemporal data" may have completely different semantic representations in different contexts. For example... Figure 5 As shown, the same numerical values ​​[0.581, 0.136] represent completely different information in terms of temperature and electricity.

[0041] In one possible implementation, since the amount of business flow scheduling data may be insufficient, to improve the model's understanding of the physical mechanisms of business flows, domain-specific data other than business flow scheduling data can also be used to generate training samples. Although the domain-specific data does not involve the specific task of business flow scheduling, it still contains knowledge data, spatiotemporal data, and background technology, enabling the model to learn the general physical mechanisms within the domain. In the early stages of training, training samples from the general data can be used to train the large language model, learning the general physical mechanisms within the domain. In the later stages, training samples from the business flow scheduling data can be used to learn the physical mechanisms of more specific business flow scheduling tasks.

[0042] After training the large language model to obtain the intermediate model, the intermediate model is further trained using specific scheduling task data, enabling the model to make autonomous decisions for specific scheduling tasks.

[0043] When training the intermediate model, sample scheduling task data is input into the intermediate model, and the sample scheduling decision text output by the intermediate model is obtained. Based on the accuracy of the scheduling parameters in the sample scheduling decision text, the training reward is determined, and the intermediate model is updated based on the training reward to obtain the business flow scheduling model.

[0044] Sample scheduling task data refers to data used to describe a scheduling task. For example, sample scheduling task data could be in natural language: "You are a professional network business flow decision-making expert, skilled at mining patterns from data to support decision-making. Based on the full volume of business data input by the user, you first comprehensively perceive the data and identify its inherent characteristics; then, combining domain knowledge and physical mechanisms, you accurately understand the perception results and finally form actionable decision recommendations. The full volume of business data is xxx." The full volume of business data describes the scheduling task, such as scenario description text or business metrics. When generating sample scheduling task data, it can be based on a template (prompt). That is, according to a pre-set template, historical scheduling task data is converted into sample scheduling task data as input to the model.

[0045] In one possible implementation, to avoid the impact of insufficient accurate perception of physical mechanisms on model training, the full business data in the sample scheduling task data can also include knowledge data, background information, and spatiotemporal data, which can provide the model with more reference information and help the model learn and output accurate decisions.

[0046] Furthermore, in one possible implementation, the sample scheduling task data can also include a pre-defined structure template for the answer, specifying the template the model should use to output text data. This enables the model to output decisions using a standardized template, improving the readability of the output decision text and facilitating understanding and execution by professionals.

[0047] The training reward is determined based on the accuracy of the scheduling parameters in the sample scheduling decision text, including: Based on the parameter categories and values ​​of scheduling parameters in the sample scheduling decision text; Accuracy rewards are determined based on parameter category and the accuracy of parameter values. The training reward value is determined based on the accuracy reward.

[0048] In this implementation, the accuracy of the scheduling parameters and their values ​​in the scheduling decisions generated by the model is evaluated. Based on the evaluation results, an accuracy reward is generated to update the model. This allows the model to be updated and optimized in the direction of generating accurate scheduling decisions during the training process, and ultimately enables the trained model to output accurate scheduling decisions for business flow scheduling tasks.

[0049] In one possible implementation, expert evaluation can be used to assess the accuracy of the scheduling parameters' categories and values ​​in the model-generated sample scheduling decision text, thus obtaining an accuracy reward. However, this approach has low interpretability and requires additional costs. Another possible implementation could employ Supervised Fine-Tuning (SFT), which involves adjusting model parameters using a small amount of labeled data based on a pre-trained intermediate model to adapt it to a specific business flow scheduling task. In this implementation, the accuracy reward is determined based on the accuracy of the parameter categories and values, including: Obtain the labeled parameter category and labeled parameter value of each parameter to be scheduled in the labeled scheduling decision text corresponding to the sample scheduling task data; The first accuracy reward is determined based on the degree of overlap between the parameter category and the labeled parameter category; For overlapping parameter categories, a second accuracy bonus is determined based on the difference between the parameter value and the labeled parameter value; The accuracy reward is determined based on the first accuracy reward and the second accuracy reward.

[0050] In this implementation, for sample scheduling task data, labeled scheduling decision text is pre-generated. This labeled scheduling decision text includes scheduling decisions suitable for the sample scheduling task corresponding to the sample scheduling task data. The labeled scheduling decision text can be generated manually. For example, assuming the labeled scheduling decision text is: "Improve the DNN model with lightweighting, optimizing the inference latency from 0.33s to 0.28s", then the parameter category of the parameter to be scheduled can be extracted as "inference latency of the DNN model", and the parameter value is "0.28s".

[0051] In one embodiment, SFT examples used to train intermediate models are shown in Table 3.

[0052] Table 3 PROMPT_TEMPLATE = Template("""<|im_start|>system You are a professional network data flow analysis and decision-making expert, skilled at identifying patterns and perceiving anomalies from network knowledge data, physical mechanisms, background information, and spatiotemporal data, and providing decision support.<|im_end|><|im_start|>user### Roles and Tasks 1. **Role**: You are a professional network business flow decision-making expert, skilled at mining patterns from data to support decision-making. 2. **Core Task**: Based on the full volume of business data input by the user, first perform "comprehensive perception" of the data to identify its inherent characteristics; then, combining domain knowledge and physical mechanisms, perform "precise cognition" of the perception results, and finally form actionable decision recommendations. ### Work Requirements Please strictly follow the following steps to conduct analysis and output structured results: 1. **Data Understanding** - Analyze the structure of the input data, including row and column definitions, indicator types, and their physical meanings. - Output the analysis process. 2. **Task Understanding** - Clarify the user's task objectives and establish the correlation between data indicators and task requirements. - Output the analysis process. 3. **Comprehensive Perception** - 4. **Precise Understanding** - Combining task objectives and data characteristics, determine which information has a critical impact on decision-making and explain the mechanisms. - Output the analysis process. 5. **Final Conclusions** - After completing the step-by-step analysis, output the overall conclusions, including key findings and configuration recommendations. ### Notes - Strictly adhere to time-series data evidence for analysis; all calculations must be verified to avoid errors. - Conclusions must clearly reflect the constraints between network configuration and business objectives. ### Input Data Description<DATA_KNOWLEDGE> 1. **Data Domain:** Telecommunications Industry 2. **Data Sub-domain:** Service Flow Scheduling 3. **Data Uses:** - "Network Configuration Data": Describes the current network status, including metrics such as bandwidth and latency. - "Service Capability Data": Describes the status of atomic services, including accuracy, performance, service scenarios, data requirements, and deployment location (cloud / edge / device). - "Business Requirement Data": Describes business objectives that need to be achieved by integrating network configuration and service capabilities, including accuracy, efficiency, and scenarios. 4. **Physical Mechanisms:** - High-precision services (e.g., 99.9% detection accuracy) often rely on complex models (large number of parameters, deep structure). Model complexity directly affects computational load (FLOPs) and memory usage. - Large model inference / training requires high data throughput and low latency. Distributed computing may cause constraints between bandwidth, latency, and computing power.- Network conditions (bandwidth, latency, packet loss rate) determine service reachability: voice tasks are sensitive to latency, video tasks have high bandwidth and computing power requirements, and congestion or high latency will cause the failure of business objectives.< / DATA_KNOWLEDGE> ### Data Dimension Explanation<DATA_BACKGROUND> 1. **Network Metrics**: Sampling time, bandwidth (Mbps), latency (s) 2. **AI Capability Metrics**: Accuracy (%), service scenario, data requirements< / DATA_BACKGROUND> ### Spatiotemporal Data Samples<NUM_TABLE> Sampling Time | Bandwidth | Latency 2025-10-01T00:00 | 310.5 | 0.12 2025-10-01T00:15 | 201.4 | 0.16 2025-10-01T00:30 | 194.4 | 0.19...< / NUM_TABLE> ### Task Requirements: Configure the network according to the current network status to meet the classification task objectives: accuracy ≥ 90%, latency ≤ 0.5s. Please output according to the following structure: <think> 1. **Data Understanding:** Analyze the input network status data, including sampling time, bandwidth (in Mbps), and latency (in seconds). The data shows that bandwidth fluctuates between 194.4-310.5 Mbps, and latency is between 0.12-0.19 seconds, indicating good overall network quality. 2. **Task Understanding:** The task objective is to configure the network for categorized services, simultaneously meeting the constraints of accuracy ≥90% and end-to-end latency ≤0.5 seconds. The feasibility of different AI models under these network constraints needs to be evaluated. 3. **Comprehensive Awareness:** Current network latency is below 0.2 seconds, and bandwidth is generally sufficient. However, the additional latency from AI model inference needs to be considered: Lightweight models (such as LightGBM) have an accuracy of approximately 85% and a processing latency of 0.05 seconds. Although the total latency can be controlled within 0.25 seconds, the accuracy is insufficient. Deep models (such as DNN) can achieve an accuracy of 92%, but the single-point inference latency reaches 0.33 seconds, which, combined with network latency, may exceed the 0.5-second limit. 4. **Accurate Recognition**: A single-node deployment cannot simultaneously meet the requirements for accuracy and latency. Based on the spatiotemporal changes in network conditions, a distributed computing strategy can be adopted: data preprocessing (0.05s latency) is performed on node A (0.12s latency) with sufficient bandwidth, and model inference is performed on node B (0.33s latency). Considering the inter-node transmission latency of 0.1s, the total latency = 0.12 + 0.05 + 0.1 + 0.33 = 0.6s, slightly exceeding the limit. However, by optimizing the transmission path (utilizing peak bandwidth periods), the transmission latency can be compressed to 0.05s, thereby reducing the total latency to 0.55s. Further fine-tuning through model lightweighting can reduce the inference latency to 0.28s, thus meeting the ≤0.5s requirement.< / think> <conclusion> A distributed model segmented processing scheme is adopted: 1. **Node Deployment**: Deploy the data preprocessing module on network node A with sufficient bandwidth, and deploy the DNN model inference module on node B with abundant computing resources. 2. **Transmission Optimization**: Utilize peak bandwidth periods (e.g., >300Mbps) for inter-node data transmission to compress transmission latency. 3. **Model Optimization**: Perform lightweight improvements on the DNN model, optimizing the inference latency from 0.33s to 0.28s while maintaining an accuracy of ≥90%. 4. **Latency Budget**: Node A processing (0.12s + 0.05s) + transmission (0.05s) + Node B inference (0.28s) = total latency of 0.5s, with an accuracy of 92%, fully meeting business requirements.< / conclusion> <|im_end|>""" The first accuracy reward is determined by the degree of overlap between the parameter categories in the sample scheduling decision text output by the model and the labeled parameter categories. For the sample scheduling decision text output by the model, the parameter categories and parameter values ​​are extracted. If the parameter categories completely overlap with the labeled parameter categories, then full marks can be given as the first accuracy reward. Otherwise, the first accuracy reward is reduced according to the number or proportion of overlaps.

[0053] For overlapping parameter categories, the difference between the parameter values ​​in the sample scheduling decision text and the labeled parameter values ​​is evaluated to determine the second accuracy reward. The difference between the parameter values ​​in the sample scheduling decision text and the labeled parameter values ​​can be measured using various data difference evaluation methods, such as MSE (Mean Squared Error), RMSE (Root Mean Squared Error), and MAE (Mean Absolute Error). Taking MSE as an example, the MSE value between the parameter values ​​in the sample scheduling decision text and the labeled parameter values ​​can be calculated to obtain the second accuracy reward.

[0054] The first and second accuracy rewards can be fused by weighted summation to obtain the total accuracy reward. The intermediate model is then updated based on this accuracy reward, enabling the trained intermediate model to output more accurate business flow scheduling decisions.

[0055] Furthermore, in one possible implementation, in addition to the accuracy reward, other rewards can be set to further improve the quality of the output decision text of the trained business flow scheduling model. In this implementation, training rewards are determined based on the accuracy reward, including: Based on the standardization of the sample scheduling decision text, a standardization reward is determined; The training reward is determined based on the prescriptive reward and the accuracy reward; The normative reward includes a structural normative reward and a grammatical normative reward. The structural normative reward reflects the consistency between the structure of the sample scheduling decision text and the preset structural template, and the grammatical normative reward reflects the consistency between the sample scheduling decision text and the grammatical rules.

[0056] A pre-defined structural template is input into the model along with the business flow scheduling task data. This template specifies that the text output by the model must conform to the pre-defined structural template format. For example, the business flow scheduling task data input into the model can be used to specify that the scheduling decisions output by the model must include the thought process. During training, the higher the consistency between the sample scheduling decision text output by the model and the pre-defined structural objective, the higher the structural regularity reward, and vice versa.

[0057] The grammatical regularity reward is used to optimize the fluency and readability of the text output by the model. The grammatical regularity reward can be output by another language evaluation model, that is, the generated sample scheduling decision text is input into the language evaluation model, and the language evaluation model outputs a grammatical regularity reward that reflects the consistency between the sample scheduling decision text and the grammatical rules.

[0058] The structural normativity reward, grammatical normativity reward, and accuracy reward are combined to obtain the final reward. This combination can be achieved using a weighted summation method, where different weights can be assigned based on importance. Since structural normativity and accuracy are more important, they can be given higher weights.

[0059] Updating intermediate models based on training rewards can be achieved using the GRPO (Group Relative Policy Optimization) algorithm. GRPO is a policy optimization algorithm specifically designed for reinforcement learning training of large language models (LLMs), and it is a simplification and improvement of the PPO (Proximal Policy Optimization) algorithm. Using the GRPO algorithm to train intermediate models for reinforcement learning can balance performance and efficiency.

[0060] Based on the above-mentioned service flow scheduling model generation method, this application further provides a service flow scheduling method, such as... Figure 6 As shown, the steps include: S610. Input the service flow scheduling task data into the service flow scheduling model; S620. Obtain the scheduling decision output by the service flow scheduling model.

[0061] The service flow scheduling model is generated based on the aforementioned service flow scheduling model generation method.

[0062] In the service flow scheduling method provided in this application, service flow scheduling task data can be constructed based on the aforementioned sample service flow scheduling task data. The service flow scheduling task data is input into the service flow scheduling model generated based on the method described above, and the scheduling decision output by the model is obtained.

[0063] The business flow scheduling model generation method provided in this application acquires multi-source data on business flow scheduling tasks from multiple domains. This multi-source data includes the physical mechanisms of business flows. A large language model is trained based on this multi-source data to obtain an intermediate model. This intermediate model can learn the inherent physical mechanisms of business flows in different domains, rather than being limited to surface observation data. The intermediate model, having learned the physical mechanisms of business flows, is then trained again based on specific sample scheduling task data. During training, a training reward is determined based on the accuracy of the scheduling parameters in the sample scheduling decision text output by the intermediate model. This reward is used to update the intermediate model, resulting in a business flow scheduling model that can output accurate scheduling decisions. Therefore, the business flow scheduling model generated by this method not only accurately outputs scheduling decisions but is also applicable to different business flow scheduling task scenarios, eliminating the need for retraining when changing scenarios and reducing the deployment cost of the business flow scheduling model.

[0064] The following describes the service flow scheduling model generation apparatus provided in this application. The service flow scheduling apparatus described below can be referred to in correspondence with the service flow scheduling method described above. Figure 7 As shown, the service flow scheduling device provided in this application includes: The initial training module 710 is used to acquire multi-source data of business flow scheduling tasks in multiple domains, and to train a large language model based on the multi-source data to obtain an intermediate model. The multi-source data includes the physical mechanism of business flow. The sample scheduling decision text generation module 720 is used to input sample scheduling task data into the intermediate model and obtain the sample scheduling decision text output by the intermediate model. The training reward determination module 730 is used to determine the training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text. The model update module 740 is used to update the intermediate model based on training rewards to obtain the business flow scheduling model.

[0065] Figure 8 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 8 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute a business flow scheduling model generation method and / or a business flow scheduling method. The business flow scheduling model generation method includes: acquiring multi-source data of business flow scheduling tasks from multiple domains; training a large language model based on the multi-source data to obtain an intermediate model; the multi-source data includes the physical mechanism of business flows; inputting sample scheduling task data into the intermediate model to obtain sample scheduling decision text output by the intermediate model; determining a training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text; and updating the intermediate model based on the training reward to obtain the business flow scheduling model. The service flow scheduling method includes: inputting service flow scheduling task data into a service flow scheduling model; obtaining the scheduling decision output by the service flow scheduling model; wherein the service flow scheduling model is generated based on the above-mentioned service flow scheduling model generation method.

[0066] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0067] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the business flow scheduling model generation method and / or business flow scheduling method provided by the above methods. The business flow scheduling model generation method includes: acquiring multi-source data of business flow scheduling tasks from multiple domains; training a large language model based on the multi-source data to obtain an intermediate model; the multi-source data includes the physical mechanism of business flows; inputting sample scheduling task data into the intermediate model to obtain sample scheduling decision text output by the intermediate model; determining a training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text; and updating the intermediate model based on the training reward to obtain a business flow scheduling model. The business flow scheduling method includes: inputting business flow scheduling task data into the business flow scheduling model; and obtaining the scheduling decision output by the business flow scheduling model; wherein the business flow scheduling model is generated based on the above-described business flow scheduling model generation method.

[0068] Furthermore, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, this computer program implements the business flow scheduling model generation method and / or business flow scheduling method provided by the above-described methods. The business flow scheduling model generation method includes: acquiring multi-source data of business flow scheduling tasks from multiple domains; training a large language model based on the multi-source data to obtain an intermediate model; the multi-source data including the physical mechanism of business flows; inputting sample scheduling task data into the intermediate model to obtain sample scheduling decision text output by the intermediate model; determining a training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text; and updating the intermediate model based on the training reward to obtain a business flow scheduling model. The business flow scheduling method includes: inputting business flow scheduling task data into the business flow scheduling model; and obtaining the scheduling decision output by the business flow scheduling model; wherein the business flow scheduling model is generated based on the above-described business flow scheduling model generation method.

[0069] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0070] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.< / header>

Claims

1. A method for generating a service flow scheduling model, characterized in that, include: Acquire multi-source data of business flow scheduling tasks from multiple domains, train a large language model based on the multi-source data to obtain an intermediate model, wherein the multi-source data includes the physical mechanism of business flow; Input the sample scheduling task data into the intermediate model and obtain the sample scheduling decision text output by the intermediate model; The training reward is determined based on the accuracy of the scheduling parameters in the sample scheduling decision text. The intermediate model is updated based on the training reward to obtain the business flow scheduling model.

2. The service flow scheduling model generation method according to claim 1, characterized in that, The determination of training rewards based on the accuracy of the scheduling parameters in the sample scheduling decision text includes: Based on the parameter categories and parameter values ​​of the scheduling parameters in the sample scheduling decision text; An accuracy reward is determined based on the parameter category and the accuracy of the parameter value. The training reward is determined based on the accuracy reward.

3. The service flow scheduling model generation method according to claim 2, characterized in that, The determination of accuracy reward based on the parameter category and the accuracy of the parameter value includes: Obtain the labeled parameter category and labeled parameter value of each parameter to be scheduled in the labeled scheduling decision text corresponding to the sample scheduling task data; A first accuracy bonus is determined based on the degree of overlap between the parameter category and the labeled parameter category; For overlapping parameter categories, a second accuracy bonus is determined based on the difference between the parameter value and the labeled parameter value; The accuracy reward is determined based on the first accuracy reward and the second accuracy reward.

4. The service flow scheduling model generation method according to claim 2, characterized in that, The determination of the training reward based on the accuracy reward includes: Based on the standardization of the sample scheduling decision text, a standardization reward is determined; The training reward is determined based on the prescriptive reward and the accuracy reward; The normative reward includes a structural normative reward and a grammatical normative reward. The structural normative reward reflects the consistency between the structure of the sample scheduling decision text and the preset structural template, and the grammatical normative reward reflects the consistency between the sample scheduling decision text and the grammatical rules.

5. The service flow scheduling model generation method according to claim 1, characterized in that, The process of training the large language model based on the multi-source data to obtain an intermediate model includes: The multi-source data is classified to obtain knowledge data, spatiotemporal data, and background information. The different categories of multi-source data are converted into natural language. The knowledge data includes knowledge information and physical mechanisms in the multi-source data. The spatiotemporal data includes the values ​​of various business indicators in the multi-source data at different times. The background information is the explanatory information of the business indicators in the multi-source data. The knowledge data, spatiotemporal data, and background information converted into natural language are aligned in the time dimension to form multiple sets of training samples. Each set of training samples is divided into a first part and a second part. The first part of the training sample is input into the large language model to obtain the predicted text output by the large language model. Based on the difference between the predicted text and the second part of the training sample, the training loss is determined. The large language model is updated based on the training loss corresponding to multiple training samples to obtain the intermediate model.

6. A service flow scheduling method, characterized in that, The method includes: Input the business flow scheduling task data into the business flow scheduling model; Obtain the scheduling decision output by the service flow scheduling model; The service flow scheduling model is generated based on the service flow scheduling model generation method as described in any one of claims 1-5.

7. A service flow scheduling model generation device, characterized in that, include: The initial training module is used to acquire multi-source data of business flow scheduling tasks in multiple domains, and to train a large language model based on the multi-source data to obtain an intermediate model. The multi-source data includes the physical mechanism of business flow. The sample scheduling decision text generation module is used to input sample scheduling task data into the intermediate model and obtain the sample scheduling decision text output by the intermediate model. The training reward determination module is used to determine the training reward based on the accuracy of the scheduling parameters in the sample scheduling decision text. The model update module is used to update the intermediate model based on the training reward to obtain the business flow scheduling model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the service flow scheduling model generation method as described in any one of claims 1 to 5 and / or the service flow scheduling method as described in claim 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the service flow scheduling model generation method as described in any one of claims 1 to 5 and / or the service flow scheduling method as described in claim 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the service flow scheduling model generation method as described in any one of claims 1 to 5 and / or the service flow scheduling method as described in claim 6.

Images