Resource scheduling method and device, electronic equipment, medium and program product

Abstract

This application provides a resource scheduling method, apparatus, electronic device, medium, and program product, which can be applied to the fields of big data technology and artificial intelligence technology. The method includes: acquiring multi-dimensional feature data; constructing an input feature vector based on the multi-dimensional feature data, wherein the multi-dimensional feature data includes at least task information and resource information; inputting the input feature vector into an interpretable model, outputting a scheduling mapping function, and determining a target scheduling factor based on the scheduling mapping function; performing hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multi-dimensional combined classification result; and generating a corresponding first resource scheduling result using the interpretable model based on the multi-dimensional combined classification result.

Description

Technical Field

[0001] This application relates to the fields of big data technology and artificial intelligence technology, and more specifically to a resource scheduling method, apparatus, device, medium and program product. Background Technology

[0002] In modern information systems, cloud computing platforms, and financial scenarios, resource scheduling is a crucial element in ensuring efficient system operation. For example, in financial operations, back-end systems need to handle high-concurrency transaction requests, risk control detection, and model inference tasks under limited computing resources, network bandwidth, and storage space; improper scheduling can lead to increased latency or system congestion.

[0003] However, existing scheduling methods mostly rely on manual rules or black box models. The former is inefficient when facing complex and multidimensional tasks, while the latter lacks interpretability, makes it difficult to track the key factors behind the scheduling results, and is prone to excessive consumption of computing resources in large-scale scenarios. Summary of the Invention

[0004] In view of the above problems, this application provides resource scheduling methods, apparatus, equipment, media and program products.

[0005] According to a first aspect of this application, a resource scheduling method is provided, the method comprising: acquiring multidimensional feature data; constructing an input feature vector based on the multidimensional feature data, wherein the multidimensional feature data includes at least customer information and product information; inputting the input feature vector into an interpretable model, outputting a scheduling mapping function, and determining a target scheduling factor based on the scheduling mapping function; performing hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multidimensional combined classification result; and generating a corresponding first resource scheduling result using the interpretable model based on the multidimensional combined classification result.

[0006] According to an embodiment of this application, the method further includes: obtaining scheduling context information corresponding to the first resource scheduling result; constructing a reinforcement learning state space based on the scheduling context information and the first resource scheduling result, setting the resource adjustment operation based on the first resource scheduling result as the action space, determining the scheduling adjustment strategy based on a preset reward function and using a reinforcement learning strategy optimization algorithm; and adjusting the first resource scheduling result based on the scheduling adjustment strategy to generate a corresponding second resource scheduling result.

[0007] According to an embodiment of this application, the step of inputting the input feature vector into an interpretable model, outputting a scheduling mapping function, and determining a target scheduling factor based on the scheduling mapping function includes: generating a nonlinear mapping relationship between the input feature vector and the resource scheduling result based on the interpretable model; and performing feature contribution analysis on the input feature vector based on the nonlinear mapping relationship, and determining the target scheduling factor based on the feature contribution analysis result.

[0008] According to an embodiment of this application, generating a nonlinear mapping relationship between the input feature vector and the resource scheduling result based on the interpretable model includes: performing a nonlinear transformation on the single variable of the input feature vector based on multiple sets of one-dimensional function units in the interpretable model; and combining the outputs of multiple nonlinearly transformed one-dimensional functions to construct the nonlinear mapping relationship.

[0009] According to an embodiment of this application, the step of hierarchically classifying tasks and resources based on the target scheduling factor to obtain a multi-dimensional combined classification result includes: establishing a multi-level threshold partitioning model with the target scheduling factor as the node, dividing the value range of each target scheduling factor layer by layer to form a hierarchical classification tree; and mapping each combination sample of tasks and resources to a combination category identifier based on the path of the hierarchical classification tree to obtain the multi-dimensional combined classification result.

[0010] According to an embodiment of this application, the training process of the interpretable model includes: inputting the corresponding training samples into a preset interpretable model based on historical multidimensional combined classification results; performing optimization calculations using the preset interpretable model in conjunction with target constraints to obtain a scheduling surface corresponding to each combined classification result; obtaining a target scheduling function based on the scheduling surface; and generating a corresponding first resource scheduling result using the interpretable model based on the multidimensional combined classification results includes: calculating the multidimensional combined classification results based on the target scheduling function to generate the corresponding first resource scheduling result.

[0011] According to an embodiment of this application, the method further includes: generating a classification decision graph based on the multidimensional combined classification result and the corresponding first resource scheduling result, wherein the classification decision graph is used to represent the scheduling boundary conditions corresponding to the multidimensional combined classification result; and / or generating a feature sensitivity curve based on the changing trend of the target scheduling function, wherein the feature sensitivity curve is used to evaluate the marginal response strength of the function input variable to the scheduling result.

[0012] A second aspect of this application provides a resource scheduling apparatus, the apparatus comprising: a data acquisition module, configured to: acquire multidimensional feature data, and construct an input feature vector based on the multidimensional feature data, wherein the multidimensional feature data includes at least task information and resource information; a scheduling factor determination module, configured to: input the input feature vector into an interpretable model, output a scheduling mapping function, and determine a target scheduling factor based on the scheduling mapping function; a hierarchical classification module, configured to: perform hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multidimensional combined classification result; and a scheduling module, configured to: generate a corresponding first resource scheduling result using the interpretable model based on the multidimensional combined classification result.

[0013] According to an embodiment of this application, the scheduling factor determination module can also be used to generate a nonlinear mapping relationship between the input feature vector and the resource scheduling result based on the interpretable model; and to perform feature contribution analysis on the input feature vector based on the nonlinear mapping relationship, and determine the target scheduling factor based on the feature contribution analysis result.

[0014] According to embodiments of this application, the scheduling factor determination module can also be used to perform nonlinear transformation on the single variable of the input feature vector based on multiple sets of one-dimensional function units in the interpretable model; and to combine the outputs of multiple nonlinearly transformed one-dimensional functions to construct the nonlinear mapping relationship.

[0015] According to an embodiment of this application, the hierarchical classification module can also be used to establish a multi-level threshold partitioning model with the target scheduling factor as the node, partition the value range of each target scheduling factor layer by layer to form a hierarchical classification tree; and based on the path of the hierarchical classification tree, map the combined sample of each task and resource to the combined category identifier to obtain the multi-dimensional combined classification result.

[0016] According to an embodiment of this application, the scheduling module can also be used to input the corresponding training samples into a preset interpretable model based on historical multidimensional combined classification results, and use the preset interpretable model in combination with target constraints to perform optimization calculations to obtain a scheduling surface corresponding to each combined classification result; obtain a target scheduling function based on the scheduling surface; and calculate the multidimensional combined classification results based on the target scheduling function to generate a corresponding first resource scheduling result.

[0017] According to embodiments of this application, the resource scheduling device may further include an adjustment module. The adjustment module can be used to acquire scheduling context information corresponding to the first resource scheduling result; construct a reinforcement learning state space based on the scheduling context information and the first resource scheduling result; set the resource adjustment operation based on the first resource scheduling result as the action space; determine a scheduling adjustment strategy based on a preset reward function and using a reinforcement learning strategy optimization algorithm; and adjust the first resource scheduling result based on the scheduling adjustment strategy to generate a corresponding second resource scheduling result.

[0018] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0019] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0020] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0021] According to embodiments of this application, by acquiring task and resource information from the computing system to form a multi-dimensional input feature vector, and combining it with an interpretable model to generate a scheduling mapping function and a target scheduling factor, hierarchical classification and combined computation of tasks and resources are achieved, thereby efficiently generating the first resource scheduling result. Compared with existing scheduling methods that rely on manual rules or black-box models, embodiments of this application not only improve the transparency and traceability of scheduling results, but also reduce computational redundancy, reduce system resource consumption, and improve overall scheduling efficiency and system stability, making it suitable for dynamic resource management scenarios under high concurrency and multiple constraints. Attached Figure Description

[0022] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0023] Figure 1 The illustrations depict application scenarios of resource scheduling methods, apparatuses, devices, media, and program products according to embodiments of this application.

[0024] Figure 2 A flowchart illustrating a resource scheduling method according to an embodiment of this application is shown schematically;

[0025] Figure 3A flowchart illustrating a method for generating a second resource scheduling result according to some exemplary embodiments of this application is shown schematically;

[0026] Figure 4 The flowchart illustrates a method for hierarchically classifying tasks and resources according to some exemplary embodiments of this application.

[0027] Figure 5 A schematic diagram illustrating the structure of a resource scheduling apparatus according to an embodiment of this application is shown; and

[0028] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a resource scheduling method according to an embodiment of this application. Detailed Implementation

[0029] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0030] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0031] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0032] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0033] First, the technical terms used in this article are explained and clarified as follows.

[0034] KAN (Kolmogorov–Arnold Network) is a neural network structure based on the Kolmogorov–Arnold representation theorem. By replacing the linear weights and activation functions in traditional neural networks with learnable one-dimensional functions, it can achieve explicit function modeling and nonlinear interaction modeling between features, thus combining prediction accuracy and interpretability.

[0035] Reinforcement learning (RL) is a machine learning method that uses an agent to learn decision-making strategies by continuously trying and learning from feedback (rewards) in an environment in order to maximize long-term benefits.

[0036] Proximal Policy Optimization (PPO) is a commonly used policy gradient optimization algorithm in reinforcement learning to improve training stability and policy update efficiency. This algorithm limits the policy update step size to prevent excessive policy changes, making it suitable for dynamic optimization tasks in complex, continuous action spaces.

[0037] SHAP (SHapley Additive Explanations) is a game-theoretic feature importance analysis method used to evaluate the marginal contribution of each input feature to the model output. By assigning an "attribution value" to each feature, SHAP can provide consistent, locally accurate explanations.

[0038] In modern information systems, cloud computing platforms, and fintech scenarios, resource scheduling is a crucial element in ensuring efficient system operation. Typical resource scheduling tasks include: under conditions of limited computing resources, network bandwidth, storage space, or production equipment, rationally allocating and adjusting resources based on multi-dimensional task characteristics, user needs, and resource constraints, thereby improving overall operational efficiency and service quality.

[0039] Existing resource scheduling methods typically rely on two main types of techniques: rule-based scheduling methods, which depend on pre-defined manual rules or empirical formulas to match and allocate tasks and resources. While simple to implement, these methods are prone to inefficient scheduling and low system utilization in complex scenarios with high dimensionality and multiple constraints, and lack adaptability. The other type is black-box model-based scheduling methods. In recent years, some systems have begun to use machine learning or deep learning models for resource scheduling prediction and allocation. However, these models often suffer from poor interpretability, meaning they struggle to clearly reveal the correlation between task characteristics and scheduling results, leading to difficulties in model adjustment, result verification, and system optimization. Furthermore, black-box models have high computational complexity in large-scale task scenarios, resulting in decreased operating efficiency and excessive resource consumption.

[0040] In addition, existing scheduling methods still have shortcomings in the following aspects: they lack a clear factor contribution analysis mechanism, making it difficult to interpret and track scheduling results; when processing large-scale multidimensional input data, the calculation process is redundant, leading to scheduling delays; they cannot quickly optimize the scheduling scheme according to real-time changes in the task environment, making it difficult to meet the needs of dynamic resource management; and black-box models are difficult to perform incremental optimization, increasing system operation and maintenance costs.

[0041] Based on this, embodiments of this application provide a resource scheduling method, including: acquiring multi-dimensional feature data; constructing an input feature vector based on the multi-dimensional feature data, wherein the multi-dimensional feature data includes at least task information and resource information; inputting the input feature vector into an interpretable model, outputting a scheduling mapping function, and determining a target scheduling factor based on the scheduling mapping function; performing hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multi-dimensional combined classification result; and generating a corresponding first resource scheduling result using the interpretable model based on the multi-dimensional combined classification result. By acquiring task information and resource information in the computing system to form a multi-dimensional input feature vector, and combining it with an interpretable model to generate a scheduling mapping function and a target scheduling factor, hierarchical classification and combined calculation of tasks and resources are realized, thereby efficiently generating a first resource scheduling result. Compared with existing scheduling methods that rely on manual rules or black-box models, embodiments of this application not only improve the transparency and traceability of scheduling results, but also reduce computational redundancy, reduce system resource consumption, and improve overall scheduling efficiency and system stability, making it suitable for dynamic resource management scenarios under high concurrency and multiple constraints.

[0042] It should be noted that the resource scheduling method, apparatus, equipment, medium, and program products defined in this application can be used in the fields of big data technology and artificial intelligence technology, as well as in the field of fintech, and in a variety of other fields besides big data technology, artificial intelligence technology, and fintech. The application fields of the resource scheduling method, apparatus, equipment, medium, and program products provided in the embodiments of this application are not limited.

[0043] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0044] In scenarios involving automated decision-making using personal information, the methods, devices, and systems provided in this application all offer users corresponding entry points for choosing to agree to or reject the automated decision-making results. If the user chooses to reject, the process proceeds to the expert decision-making stage. Here, "automated decision-making" refers to the activity of automatically analyzing and evaluating an individual's behavioral habits, interests, or economic, health, and credit status through computer programs, and then making a decision. Here, "expert decision-making" refers to the activity of making decisions by personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0045] Figure 1 The illustration shows an application scenario diagram of the resource scheduling method, apparatus, device, medium, and program product according to embodiments of this application.

[0046] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0047] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0048] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0049] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0050] It should be noted that the resource scheduling method provided in this application embodiment can generally be executed by server 105. Correspondingly, the resource scheduling device provided in this application embodiment can generally be located in server 105. The resource scheduling method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the resource scheduling device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0051] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0052] The following will be based on Figure 1 The described scene, through Figures 2-4 The resource scheduling method of the disclosed embodiments will be described in detail.

[0053] Figure 2 A flowchart illustrating a resource scheduling method according to an embodiment of this application is shown.

[0054] like Figure 2 As shown, the resource scheduling method 200 of this embodiment includes operations S210 to S240.

[0055] In operation S210, multidimensional feature data is acquired, and an input feature vector is constructed based on the multidimensional feature data, wherein the multidimensional feature data includes at least task information and resource information.

[0056] In the embodiments of this application, the multidimensional feature data may come from the enterprise's internal historical transaction database, resource management system, task information management system, or external data interface, etc.

[0057] Task information may include task identifier, priority level, execution geographical location, business type, task category (e.g., computational task, storage task, transmission task), historical execution status, task sensitivity level, task stability score, task lifecycle stage, resource consumption history, and task execution behavior characteristics. This data is used to characterize the task's resource requirements, execution tendency, and risk level, and is a key input factor in the scheduling model.

[0058] Resource information can include resource classification, resource provider, configuration parameters, specifications, available capacity, cost structure, redundancy status, customization status, inventory / availability, lifecycle stage (e.g., new resources, idle resources), historical utilization rate, and failure rate. Different types of resource information dimensions have significantly different impacts on scheduling strategies. For example, high-performance customized resources may focus more on task profile-driven approaches, while large-scale general-purpose resources may rely more on overall task load fluctuations and comparisons with the external environment.

[0059] In the embodiments of this application, resource information may specifically include basic information of computing, storage, or network resources, such as: computing node type, computing capacity range, allocation duration, task allocation mode (e.g., exclusive, shared), execution load type (fixed load or elastic load), whether it is a policy priority resource (e.g., high priority, latency-sensitive resource), resource protection method (e.g., redundancy, fault tolerance, backup), task purpose (e.g., computing, storage, transmission), resource usage floating ratio, resource source channel, whether it participates in resource sharing strategy, etc.

[0060] It should be noted that all operations involving the collection, use, and modeling of task or resource-related information strictly adhere to relevant laws and regulations regarding data security and privacy protection. When acquiring task or resource information as part of multidimensional feature data, the system ensures that the data source complies with the principle of data processing legality, such as through explicit authorization from the task submitter, service agreement stipulations, or legally authorized enterprise databases obtained through business cooperation channels.

[0061] In some embodiments, the multidimensional feature data may further include multiple dimensions of information such as regional information, channel information, behavioral information, and environmental context information. Regional information can reflect the differences in the geographical location or resource distribution market of the task, and may include the province / city / district where the task is executed, urban / rural level, geographical latitude, economic zone, regional resource supply level index, and regional strategy constraint coefficient. Channel information is used to identify the source path or resource call interface of the task request, and may include online / offline channel type, third-party platform identifier, task request source tag, and access entry code. Different request channels have significant differences in their impact on scheduling strategies and cost structures; incorporating channel features helps to achieve scenario-based differentiated resource allocation strategies.

[0062] In some embodiments, task behavior information can also be introduced as a supplement to dynamic features, such as a sequence of behaviors such as recent task submission, execution, retry, cancellation, migration, expansion or reduction, to construct a behavior embedding vector to reflect the activity level and resource sensitivity level of the task, which is particularly suitable for agile response scenarios such as dynamic scheduling and real-time allocation scheduling.

[0063] Contextual information can also be incorporated as auxiliary features, such as time information (e.g., weekdays / holidays, peak sales periods, quarterly windows), policy information (e.g., resource regulation policies), competition status information (e.g., scheduling trends of adjacent systems, utilization status of competitive resources), and macroeconomic variables (e.g., network traffic, energy supply), to enhance the model's ability to perceive disturbances in external variables and improve the system's robustness in complex economic environments.

[0064] It should be noted that multidimensional feature data can be dynamically configured based on the accuracy requirements of the business model, data availability, and modeling complexity. The system can also integrate a feature selection module to automatically filter, reduce dimensionality, or combine features in high-dimensional data, thereby improving the model's generalization ability and computational efficiency. Regardless of the feature composition method used, the embodiments of this application can support flexible and scalable data input structures, adapting to scheduling modeling needs under various task types, resource forms, and operating environments.

[0065] In the embodiments of this application, multidimensional feature data may simultaneously contain structured fields (such as task priority, node level) and unstructured data (such as task tags, execution log summaries, etc.). The system can complete data cleaning, field completion, outlier handling, and feature encoding operations through the feature engineering module, such as using one-hot encoding for categorical variables and standardizing numerical variables, thereby forming a uniformly formatted input feature vector. The constructed feature vector can be represented as a fixed-length vector structure, suitable for input into subsequent interpretable models.

[0066] In operation S220, the input feature vector is input into the interpretable model, the scheduling mapping function is output, and the target scheduling factor is determined based on the scheduling mapping function.

[0067] In the embodiments of this application, the interpretable model can be a modeling structure with explicit functional expression capabilities, which can output readable and traceable scheduling expressions while maintaining nonlinear fitting capabilities.

[0068] For example, an interpretable model may include a structured neural network based on a combination of one-dimensional functions, such as an explicit modeling network consisting of multiple learnable basis functions superimposed on each variable in the input feature vector, which constructs a complex scheduling function by performing a one-dimensional nonlinear transformation on each variable in the input feature vector and combining the results of the multiple transformations.

[0069] For example, an interpretable model can be a hierarchical decision model based on rule extraction, such as a structure that uses decision paths to express scheduling logic, where each path or node can be traced back to a partitioning rule for a specific feature. Such models can be used to construct scheduling condition boundaries under multi-dimensional feature interactions, and are suitable for application scenarios that require classification-driven scheduling or hierarchical strategy generation.

[0070] For example, the interpretable model can also be combined with a variable-length sequence structure based on local response analysis, which uses feature location embedding and attribution mechanisms to identify key behavioral segments in the input sequence and provide interpretable feedback on the impact of behavior on scheduling results.

[0071] In the embodiments of this application, the output of the scheduling mapping function can be presented in the form of a symbolic function, a graphical structure, or a mathematical expression. Its function structure contains the mathematical correlation between each input variable and the resource scheduling result. This function can not only predict scheduling results but also explain the marginal effect and interactive influence of each input feature through the form, structure, and parameters of its function terms. For example, if the function term of a certain feature variable exhibits a non-monotonic change, it can be identified that the variable has a threshold effect or non-linear interaction on the scheduling result, thereby providing a basis for subsequent classification and strategies.

[0072] In embodiments of this application, the system can evaluate the contribution of each input feature variable based on the scheduling mapping function to determine a set of target scheduling factors. Target scheduling factors refer to the set of variables that have a significant impact on the output of the scheduling function in the overall feature space. Contribution analysis methods may include derivative sensitivity analysis, gradient backpropagation, SHAP value analysis, and feature perturbation impact assessment. Through these methods, the system can quantify the marginal impact of each variable on the scheduling output and rank the features accordingly, selecting the top few factors with significant contributions as target scheduling factors for subsequent classification and rule generation.

[0073] According to embodiments of this application, the number and selection logic of target scheduling factors can be adjusted according to different scenarios. For example, under high-dimensional features, to avoid overfitting, a contribution threshold can be set or the top K variables can be selected; while in scenarios requiring high interpretability, fields with business readability, such as task level, resource category, and resource status, can be preferentially retained as scheduling factors. In addition, the system can also combine the "mandatory factor" list configured by business personnel with the automatic identification results of the model to construct a set of scheduling factors that combines data-driven approaches and expert knowledge.

[0074] In operation S230, tasks and resources are classified hierarchically based on the target scheduling factor to obtain a multi-dimensional combined classification result.

[0075] According to embodiments of this application, the purpose of the hierarchical classification operation is to divide the sample data into several subgroups with similar scheduling characteristic patterns, so as to use more targeted scheduling models or rules in each subgroup and improve scheduling accuracy and controllability.

[0076] In the embodiments of this application, the hierarchical classification operation can adopt a rule-based partitioning structure, such as constructing a hierarchical classification tree. Each layer is divided based on the value range of a target scheduling factor, and the data samples are segmented layer by layer. For example, the first layer can be grouped according to task priority level, the second layer can be subdivided according to resource allocation level, and the third layer can be further divided according to regional resource supply and demand coefficients. Finally, the path represented by each leaf node is a specific combination classification result.

[0077] In this embodiment, the system can construct hierarchical partitioning paths in various ways. Preferably, the partitioning logic can originate from the numerical distribution characteristics of the target scheduling factor (e.g., partitioning based on quantiles), business-defined rule thresholds (e.g., defining high-credit customers as those with a credit score ≥ 700), or data-driven automatic clustering methods (e.g., feature splitting points generated based on decision trees). These partitioning methods can be used individually or in combination to ensure that the generated classification results are both business-readable and possess a certain degree of model discriminability.

[0078] In some embodiments, the target scheduling factor may include both continuous variables (such as task resource consumption and historical average execution time) and categorical variables (such as task type and resource category). In this case, the system can employ interval partitioning and label partitioning strategies respectively to construct classification units based on multi-dimensional feature combinations in the space. For example, the system can define "high-priority tasks / high-configuration resources" as one category and "low-priority tasks / non-standard resources" as another, with each category subsequently corresponding to an independent scheduling model or scheduling strategy function.

[0079] In the embodiments of this application, the hierarchical classification operation can also be implemented using a data-driven automatic clustering method, such as using algorithms like Gaussian mixture models or density clustering to model the similarity of task and resource combination samples in the feature space and automatically generate several combined classification units. The number of clusters can be set according to preset parameters or adaptively optimized by evaluating clustering quality indicators (such as silhouette coefficient).

[0080] In the embodiments of this application, the hierarchical classification operation can be implemented using a hybrid strategy, combining rule setting and data-driven methods. The system can first perform preliminary screening or labeling of samples according to predefined business rules, such as coarse classification based on "whether it is a real-time task" or "whether it is a dedicated resource." Subsequently, within each coarse group, automatic clustering or decision tree methods are applied to further subdivide the internal structure, thereby forming a multi-stage, multi-granularity combined classification structure. This approach achieves a balance between strategy flexibility and interpretability while taking into account both business logic and data distribution.

[0081] The above implementation is merely an example. In other optional implementations, the system can also support ranking-based hierarchical classification based on feature combination scoring functions, community detection-based partitioning based on graph structures, or label-assisted classification trained using historical scheduling response data. The specific hierarchical strategy adopted can be configured based on the stability of the business scenario, data availability, model accuracy requirements, and interpretability requirements. The system can also support dynamic switching or automatic optimization of the hierarchical structure to adapt to the model update process. It should be noted that the hierarchical classification described in this application not only serves the modeling process itself but also constitutes an important unit boundary in the scheduling management system. In subsequent scheduling function calls, dynamic scheduling strategy optimization, and sensitivity analysis, the multi-dimensional combination classification results can serve as the basic unit for controlling granularity, enabling differentiated and refined scheduling from the "task / resource" combination dimension.

[0082] In the embodiments of this application, the number and structure of multi-dimensional combined classification results can be dynamically adjusted according to the complexity of the business. For example, in high-performance computing tasks, a finer-grained classification structure can be used to improve scheduling accuracy; while in standard batch processing tasks, the classification dimensions can be appropriately reduced to improve system computational efficiency and management simplicity. The system can also support automatically merging classification units with too small a sample size or insufficient diversity to enhance model robustness and generalization ability.

[0083] In operation S240, based on the multidimensional combined classification result, the corresponding first resource scheduling result is generated using the interpretable model.

[0084] It should be noted that the interpretable model used in operation S240 and the model used in operation S220 to construct the scheduling mapping function and identify the target scheduling factor can have the same model structure.

[0085] In the embodiments of this application, the input vector of each combination classification result in the multidimensional combination classification result can be fed into an interpretable model for inference, and the corresponding scheduling prediction value can be output. The first resource scheduling result can be a real number scheduling parameter value or a range of interval scheduling parameters with regular attributes, which can be flexibly adjusted according to the scheduling requirements of the business scenario.

[0086] In other embodiments, the system can utilize the scheduling mapping function (i.e., the explicit mathematical relationship between input variables and scheduling result output) generated in operation S220, and combine it with the feature parameters of the current combined classification, to call the function for evaluation and directly calculate the first resource scheduling result. For example, if the scheduling mapping function is "scheduling value = f(task priority, resource allocation, regional coefficient)", the system can input the values ​​of the current sample on these three variables into the function and output the calculated scheduling result. This process does not require remodeling; instead, it performs the evaluation of the target scheduling function based on the existing function expression.

[0087] In this embodiment, to improve the flexibility and stability of scheduling, the system can perform result constraint processing based on the results obtained from the evaluation of the scheduling mapping function, combined with policy rules. For example, the system can truncate and correct the scheduling results according to the preset upper and lower limits of scheduling for different classification nodes, or smooth the results based on the historical resource fluctuation range to prevent scheduling abrupt changes or policy penetration risks. Such constraint strategies can be embedded in the model calling logic as a post-processing module after function evaluation.

[0088] The resource scheduling method of this application will be specifically described below by way of preferred embodiments.

[0089] In the embodiments of this application, the resource allocation can be further dynamically adjusted based on the first resource scheduling result to generate a second resource scheduling result.

[0090] Figure 3 A flowchart illustrating a method for generating a second resource scheduling result according to some exemplary embodiments of this application is shown schematically.

[0091] like Figure 3 As shown, the method for generating the second resource scheduling result includes operations S310 to S330.

[0092] In operation S310, the scheduling context information corresponding to the first resource scheduling result is obtained.

[0093] In this embodiment, scheduling context information may include current system operating environment data, resource utilization targets, task execution feedback, resource supply or network constraints, scheduling trends of similar systems, global policy rules, time-sensitive factors (such as peak task periods, maintenance windows), compliance or security requirements, etc. This type of information can be integrated and updated through the data warehouse in the resource management system, real-time monitoring interface, external maintenance platform, or historical log system.

[0094] In operation S320, a reinforcement learning state space is constructed based on the scheduling context information and the first resource scheduling result. The resource adjustment operation based on the first resource scheduling result is set as the action space. Based on the preset reward function, the scheduling adjustment strategy is determined using a reinforcement learning strategy optimization algorithm.

[0095] In the embodiments of this application, the reinforcement learning state space can be used to represent the current scheduling situation and can be composed of multiple feature dimensions, such as: the current first resource scheduling result, task execution label (e.g., completion rate prediction), resource utilization target completion degree, resource load score, system demand elasticity score, acceptable scheduling adjustment range, etc. The system inputs this state space into the reinforcement learning agent to determine the current scheduling environment state.

[0096] Simultaneously, an action space can be constructed based on the first resource scheduling result. The action space can be defined as adjustment operations on the first resource scheduling result, and can be a discrete set of actions (e.g., decreasing by 5%, remaining unchanged, increasing by 5%) or a continuous action space (e.g., arbitrarily fluctuating within ±10%). Action design can be configured according to task type, resource attributes, or scheduling strategy complexity. To ensure the stability of model training and the reasonableness of adjustment magnitude, the system can dynamically prune the action space or employ smoothing strategies (such as Gaussian perturbation, action regularization, etc.).

[0097] Reinforcement learning strategy optimization algorithms can preferably employ near-end policy optimization and discrete action reinforcement learning to optimize and update scheduling strategies under different states. The system evaluates the output of each strategy step based on a preset reward function, which can comprehensively consider multiple dimensions such as actual task completion rate, resource utilization efficiency, system load balancing, scheduling fluctuation stability, and policy execution compliance. For example, if a scheduling adjustment action improves task execution efficiency while maintaining it within the resource utilization target range, the system can provide a positive reward for that action.

[0098] In other embodiments, the reward function may also incorporate a time discount factor to optimize for maximizing long-term resource efficiency, or introduce a risk adjustment mechanism to penalize aggressive scheduling strategies that exceed fluctuation tolerance. The system can also continuously iterate and update the state definition and reward function during training, forming a dynamically evolving scheduling strategy library to enhance the system's adaptability and robustness in complex operating environments.

[0099] In operation S330, the first resource scheduling result is adjusted based on the aforementioned scheduling adjustment strategy to generate a corresponding second resource scheduling result. The second resource scheduling result is the final guiding scheduling parameter output by the system after comprehensively considering the initial scheduling suggestions and strategic objectives under the current scheduling context information. The second resource scheduling result retains the static rationality reflected in the original scheduling mapping model while also reflecting strategic adaptation to factors such as real-time system feedback, resource objectives, and task responses, forming a scheduling closed loop that can be dynamically optimized according to the environment.

[0100] After the scheduling adjustment operation is completed, the system can record the execution effect and feedback indicators of each round of scheduling behavior (such as actual task completion rate, changes in resource utilization, changes in load balance, etc.) to further train and update the reinforcement learning model, forming a data-driven long-term strategy iteration path.

[0101] In embodiments of this application, inputting an input feature vector into an interpretable model and outputting a scheduling mapping function, and determining a target scheduling factor based on the scheduling mapping function, may include: generating a nonlinear mapping relationship between the input feature vector and the resource scheduling result based on the interpretable model; and performing feature contribution analysis on the input feature vector based on the nonlinear mapping relationship, and determining the target scheduling factor based on the feature contribution analysis result.

[0102] In the embodiments of this application, the interpretable model can adopt a neural network structure based on the Kolmogorov-Arnold representation theorem, namely the Kolmogorov-Arnold Network (KAN). The KAN network is a neural network structure with explicit function expression capabilities. Its core feature is that it replaces the fixed activation function and linear weight mechanism used in traditional neural networks with a set of learnable one-dimensional function units, thereby more directly approximating the multivariable mapping function.

[0103] Specifically, the system can construct multiple one-dimensional function units within the KAN network model structure. Each function unit corresponds to a single variable dimension in the input feature vector, and the system performs a nonlinear transformation operation on this variable. The one-dimensional function unit can adopt trainable spline functions, piecewise polynomials, orthogonal basis functions, or other forms of differentiable function structures. Its function shape parameters are continuously optimized through the backpropagation algorithm during training, thereby adapting to the nonlinear feature distribution of the target scheduling function.

[0104] Furthermore, the outputs of multiple functions after nonlinear transformation can be combined, for example, through weighted summation, function superposition, or nested combination, to construct an overall mapping relationship between resource scheduling results and input features. Compared with traditional multilayer feedforward neural networks, the KAN model not only has stronger functional expressive power but also higher interpretability. Its output can be decomposed into the independent influence of multiple input variables on the scheduling results under different transformation paths, which is helpful for subsequent execution factor contribution analysis, marginal response visualization, and policy rule derivation.

[0105] It should be noted that although the embodiments of this application are not limited to using the KAN structure as the only implementation of an interpretable model, the KAN network has significant advantages in expressing nonlinear relationships, maintaining structural transparency, and supporting symbolic output, making it particularly suitable for scheduling scenarios that require both high-precision prediction and strong interpretability. Therefore, in a preferred embodiment, the system introduces a KAN network to achieve explicit functional modeling from the input feature vector to the scheduling result, constructing a scheduling mapping relationship based on learnable function units.

[0106] In the embodiments of this application, feature contribution analysis can be performed on the input feature vector based on the aforementioned scheduling mapping function. The analysis process can be implemented using various methods, such as local gradient analysis, feature perturbation methods, input-oriented attribution mechanisms, or attribution methods based on game theory. Taking SHAP value analysis as an example, the system can accurately attribute the marginal contribution of each feature variable to the model output, quantifying the influence of each variable on the scheduling result under the current sample or global sample distribution.

[0107] Based on the contribution analysis results, the system can set a weight threshold, the top K contribution rankings, or other business-related criteria, and select several representative key feature variables as the core variable set for subsequent classification modeling and strategy partitioning, denoted as the target scheduling factor. The target scheduling factor will be used as a dimensional input in subsequent hierarchical classification (such as constructing a classification tree or clustering), determining the scheduling path partitioning method for different task / resource combinations.

[0108] Figure 4 The flowchart illustrating a method for hierarchically classifying tasks and resources according to some exemplary embodiments of this application is shown schematically.

[0109] like Figure 4 As shown, the method for hierarchically classifying tasks and resources includes operations S410 to S420.

[0110] In operation S410, a multi-level threshold partitioning model with the target scheduling factor as the node is established, and the value range of each target scheduling factor is partitioned layer by layer to form a hierarchical classification tree.

[0111] In different implementation scenarios, the depth and hierarchical structure of the classification tree can be flexibly set according to the number of target scheduling factors and data distribution. For example, if only three core scheduling factors are identified, the system can construct a three-layer partitioning structure: the first layer is based on task priority level, the second layer on resource allocation level, and the third layer on regional resource supply and demand coefficients, ultimately forming a complete path structure representing the scheduling characteristic path of each task / resource combination.

[0112] In operation S420, based on the path of the hierarchical classification tree, the combined samples of each task and resource are mapped to the combined category identifier to obtain the multidimensional combined classification result.

[0113] After completing the hierarchical classification tree structure, each task-resource combination sample can be mapped to a unique classification path endpoint node according to the path rules of the hierarchical classification tree, and a combination category identifier can be assigned to it. For example, the combination category identifier can be a structured code (such as "T2-R1-Z3" representing task priority level 2, resource type level 1, and region level 3) or a system-defined name, used to refer to the independent scheduling unit of this type of combination in the subsequent scheduling strategy system.

[0114] Furthermore, to achieve customized scheduling functions and efficient scheduling output based on combined classification results, the system can perform a targeted training process on the interpretable model to construct a scheduling function with clear mapping relationships and constraint adaptability, which is then used to generate the first resource scheduling result after model deployment. This training process can include: based on historical multidimensional combined classification results, inputting the training samples corresponding to each combination into a pre-defined interpretable model, performing optimization calculations in conjunction with the target constraints associated with that type of sample, thereby fitting and obtaining the scheduling surface corresponding to each combined classification result.

[0115] The aforementioned scheduling surface is used to characterize the nonlinear mapping relationship between input features and target results under this combined classification. It can preferably be expressed using an interpretable modeling approach with a decomposable structure, such as a KAN network based on the superposition of one-dimensional learnable functions, or a generalized additive model structure, giving the model strong expressive power and structural transparency. By introducing target constraints during training, such as minimum profit requirements, scheduling fluctuation limits, or compliance threshold controls, the system can ensure that the generated scheduling surface conforms to both data patterns and meets the boundary requirements of actual operational strategies.

[0116] After the scheduling surface is fitted, the system can extract the target scheduling function based on the surface. The target scheduling function can be regarded as a local function expression in the input feature space, used to represent the functional mapping relationship between input variables and scheduling results under a specific combination classification. The target scheduling function can be represented by an explicit mathematical expression or a parameterized function structure, and can be directly called after the model is deployed.

[0117] During the model deployment and application phase, the objective function can be evaluated by combining the real-time input customer and product feature vectors and their respective combination classification identifiers, thereby generating the corresponding first resource scheduling result.

[0118] It should be noted that the "scheduling surface" and "target scheduling function" in this implementation can have a one-to-one or one-to-many relationship. The system can select to merge multiple similar classifications and use a unified scheduling function according to classification accuracy requirements, computing resource constraints or model management strategies, so as to improve deployment efficiency and function maintainability.

[0119] In some embodiments, based on the aforementioned multidimensional combined classification results and their corresponding first resource scheduling results, a classification decision graph and a feature sensitivity curve graph can be further constructed to improve the interpretability of the scheduling strategy and the visualization ability of the scheduling logic, thereby assisting business personnel in understanding the model output, adjusting the scheduling strategy, and meeting regulatory transparency requirements.

[0120] Specifically, after completing the combination classification and scheduling function output, the system can generate a classification decision graph based on the classification path structure and the scheduling results corresponding to each category. The classification decision graph can use the target scheduling factor as the horizontal division dimension, presenting each level of nodes in the hierarchical classification structure as a visual path, and marking the corresponding first resource scheduling result or scheduling interval on each terminal classification node. The classification decision graph can adopt a tree structure, rule table, or flowchart format to clearly express the formation rules and scheduling basis of each combination category, thereby constructing an intuitive mapping path from combination classification to scheduling strategy. The classification decision graph can be used in scenarios such as business review, strategy management, and model testing, and is particularly suitable for situations in multi-dimensional scheduling systems where differentiated scheduling control is applied to different task types and resource combinations.

[0121] In some embodiments, the system can also generate a feature sensitivity curve based on the response characteristics of the target scheduling function, which can be used to measure the marginal response strength of the model input variables to the scheduling results. Specifically, while keeping other input variables constant, the value of a target variable can be gradually adjusted, and the continuous changes in the scheduling function output can be recorded to plot a function curve of "change in input variable to scheduling output result". This curve can be used to identify the trend of the influence of input features on the scheduling output (such as linearity, saturation, critical point, inflection point, etc.), and can be used to evaluate the scheduling leverage, sensitivity, or risk amplification range of a certain feature variable.

[0122] For example, for the input variable "task priority score," the system can identify its critical fluctuation range for scheduling results through feature sensitivity curves (e.g., a significant decrease in scheduling values ​​when the score is within a certain range), thereby assisting in the fine-tuning of resource allocation strategies. Similarly, for the variable "resource allocation level," the system can identify its asymmetric amplification effect on scheduling results and set differentiated scheduling adjustment gradients accordingly. Sensitivity curves can also be used for stability verification during the model testing phase, ensuring that the model's response to key variables is reasonable, controllable, and does not amplify abnormally.

[0123] Optionally, the resource scheduling method can be applied to task scheduling operations on large-scale computing platforms, aiming to achieve fine-grained allocation and differentiated scheduling strategy control at both the task and resource levels in the context of multi-task concurrency and multiple resource constraints. The system first acquires multi-dimensional feature data related to task requests, including task information (such as priority, resource requirements, execution time limit, and historical execution records), resource information (such as resource type, available capacity, performance parameters, redundancy guarantee method, and node source), and network topology parameters and resource utilization coefficients of the task's region. After feature processing, the system constructs an input feature vector in a unified format.

[0124] Subsequently, the input feature vector can be fed into an interpretable model built on the KAN structure to generate a scheduling mapping function between the loan interest rate and the input features. Based on this function, feature contribution analysis can be further performed to identify several target scheduling factors that have a significant impact on scheduling, such as "task priority", "resource availability", "execution time limit" and "node topology level". Based on this, a multi-level threshold partitioning model can be constructed to form a classification tree structure for joint hierarchical classification of tasks and resources.

[0125] Based on the above classification tree structure, each task and resource combination sample can be mapped to a unique combination category. For each combination classification result, the corresponding scheduling function can be directly called using the trained interpretable model structure to complete the automated calculation of resource allocation parameters and output the first resource scheduling result as a basis for personalized scheduling suggestions or decision support.

[0126] To achieve dynamic adaptation to changes in the operating environment, task response, and system objectives, a dynamic scheduling optimization mechanism based on reinforcement learning can be introduced. Specifically, based on the generated first resource scheduling result, further contextual information related to the current scheduling is obtained, including current resource utilization, policy constraint parameters, task response rate, and system load achievement status. This information, along with the first resource scheduling result, is used to construct a reinforcement learning state space. In this state space, the system sets the adjustment range of the first scheduling result as the action space, such as "increase by 5%", "remain unchanged", or "decrease by 10%", and evaluates the resource utilization improvement, task completion rate improvement, and risk control effect brought about by each scheduling adjustment based on a preset reward function. Preferably, the Proximal Policy Optimization (PPO) algorithm can be used to train the reinforcement learning agent, allowing it to learn the optimal dynamic scheduling strategy in multiple rounds of policy iteration. In actual operation, the system can automatically invoke the reinforcement learning strategy to output adjustment decisions based on the classification identifier of the task request sample and the current scheduling context, and adjust the first resource scheduling result based on this strategy to generate the final second resource scheduling result, which serves as a dynamic guidance parameter to cope with rapidly changing system objectives and operational feedback.

[0127] It should be noted that the method proposed in this application is not only applicable to resource scheduling scenarios, but can also be flexibly applied to various fields that require automated decision-making and dynamic optimization based on multi-dimensional features. For example, in product pricing scenarios, this method can be used to generate personalized price suggestions and dynamically adjust them based on market feedback; in cloud computing and data centers, it can be used for intelligent allocation of resources such as computing nodes and storage bandwidth to improve overall utilization and response speed; in logistics and supply chain management, it can be used for vehicle route planning, warehouse allocation, and inventory optimization to reduce operating costs; in financial risk control, it can be used for automated adjustment of loan amounts, interest rate fluctuations, and credit strategies to enhance robustness; in intelligent manufacturing, it can be used for production task scheduling and equipment maintenance cycle optimization to improve production efficiency and equipment utilization. This application does not limit the specific application scenarios.

[0128] Corresponding to the above-described resource scheduling method, embodiments of this application also provide a resource scheduling device.

[0129] Figure 5 A schematic block diagram of a resource scheduling apparatus according to an embodiment of this application is shown.

[0130] like Figure 5 As shown, the resource scheduling device 500 in this embodiment includes a data acquisition module 510, a scheduling factor determination module 520, a hierarchical classification module 530, and a scheduling module 540.

[0131] The data acquisition module 510 can be used to acquire multi-dimensional feature data and construct an input feature vector based on the multi-dimensional feature data, wherein the multi-dimensional feature data includes at least task information and resource information. In one embodiment, the data acquisition module 510 can be used to perform the operation S210 described above, which will not be repeated here.

[0132] The scheduling factor determination module 520 can be used to input the input feature vector into the interpretable model, output a scheduling mapping function, and determine the target scheduling factor based on the scheduling mapping function. In one embodiment, the scheduling factor determination module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0133] The hierarchical classification module 530 can be used to perform hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multi-dimensional combined classification result. In one embodiment, the hierarchical classification module 530 can be used to perform the operation S230 described above, which will not be repeated here.

[0134] The scheduling module 540 can be used to generate a corresponding first resource scheduling result based on the multi-dimensional combined classification result and using the interpretable model. In one embodiment, the scheduling module 540 can be used to execute the operation S240 described above, which will not be repeated here.

[0135] According to an embodiment of this application, the scheduling factor determination module 520 can also be used to generate a nonlinear mapping relationship between the input feature vector and the resource scheduling result based on the interpretable model; and to perform feature contribution analysis on the input feature vector based on the nonlinear mapping relationship, and determine the target scheduling factor based on the feature contribution analysis result.

[0136] According to an embodiment of this application, the scheduling factor determination module 520 can also be used to perform nonlinear transformation on the single variable of the input feature vector based on multiple sets of one-dimensional function units in the interpretable model; and to combine the outputs of multiple nonlinearly transformed one-dimensional functions to construct the nonlinear mapping relationship.

[0137] According to an embodiment of this application, the hierarchical classification module 520 can also be used to establish a multi-level threshold partitioning model with the target scheduling factor as the node, partition the value range of each target scheduling factor layer by layer to form a hierarchical classification tree; and based on the path of the hierarchical classification tree, map the combined sample of each task and resource to the combined category identifier to obtain the multi-dimensional combined classification result.

[0138] According to an embodiment of this application, the scheduling module 540 can also be used to input the corresponding training samples into a preset interpretable model based on historical multidimensional combined classification results, and use the preset interpretable model in combination with target constraints to perform optimization calculations to obtain a scheduling surface corresponding to each combined classification result; obtain a target scheduling function based on the scheduling surface; and calculate the multidimensional combined classification results based on the target scheduling function to generate a corresponding first resource scheduling result.

[0139] According to an embodiment of this application, the resource scheduling device 500 may further include an adjustment module. The adjustment module can be used to obtain scheduling context information corresponding to the first resource scheduling result; construct a reinforcement learning state space based on the scheduling context information and the first resource scheduling result; set the resource adjustment operation based on the first resource scheduling result as the action space; determine a scheduling adjustment strategy based on a preset reward function and using a reinforcement learning strategy optimization algorithm; and adjust the first resource scheduling result based on the scheduling adjustment strategy to generate a corresponding second resource scheduling result.

[0140] According to embodiments of this application, any multiple modules among the data acquisition module 510, scheduling factor determination module 520, hierarchical classification module 530, and scheduling module 540 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 510, scheduling factor determination module 520, hierarchical classification module 530, and scheduling module 540 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the data acquisition module 510, the scheduling factor determination module 520, the hierarchical classification module 530, and the scheduling module 540 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0141] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a resource scheduling method according to an embodiment of this application.

[0142] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage portion 608 into a random access memory (RAM) 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0143] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0144] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0145] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0146] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0147] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to enable the computer system to implement the resource scheduling method provided in the embodiments of this application.

[0148] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0149] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0150] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0151] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0152] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0153] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0154] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A resource scheduling method, characterized in that, The method includes: Obtain multidimensional feature data, and construct an input feature vector based on the multidimensional feature data, wherein the multidimensional feature data includes at least task information and resource information; The input feature vector is input into the interpretable model, and the output is a scheduling mapping function. The target scheduling factor is determined based on the scheduling mapping function. Based on the target scheduling factor, tasks and resources are hierarchically classified to obtain multi-dimensional combined classification results; and Based on the multidimensional combined classification results, the corresponding first resource scheduling result is generated using the interpretable model.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the scheduling context information corresponding to the first resource scheduling result; A reinforcement learning state space is constructed based on the scheduling context information and the first resource scheduling result. Resource adjustment operations based on the first resource scheduling result are set as reinforcement learning action spaces. A scheduling adjustment strategy is determined using a reinforcement learning policy optimization algorithm based on a preset reward function. The first resource scheduling result is adjusted based on the aforementioned scheduling adjustment strategy to generate a corresponding second resource scheduling result.

3. The method according to claim 1, characterized in that, The step of inputting the input feature vector into an interpretable model, outputting a scheduling mapping function, and determining the target scheduling factor based on the scheduling mapping function includes: Based on the interpretable model, a nonlinear mapping relationship is generated between the input feature vector and the resource scheduling result; and The input feature vector is analyzed based on the nonlinear mapping relationship, and the target scheduling factor is determined based on the feature contribution analysis results.

4. The method according to claim 3, characterized in that, The process of generating a nonlinear mapping relationship between the input feature vector and the resource scheduling result based on the interpretable model includes: Based on multiple sets of one-dimensional function units in the interpretable model, a nonlinear transformation is performed on the univariate input feature vector; and The nonlinear mapping relationship is constructed by combining the outputs of multiple one-dimensional functions after nonlinear transformation.

5. The method according to any one of claims 1 to 4, characterized in that, The hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multi-dimensional combined classification result includes: A multi-level threshold partitioning model is established with the target scheduling factor as the node, and the value range of each target scheduling factor is partitioned layer by layer to form a hierarchical classification tree; and Based on the path of the hierarchical classification tree, the combined samples of each task and resource are mapped to the combined category identifier to obtain the multidimensional combined classification result.

6. The method according to claim 5, characterized in that, The training process of the interpretable model includes: Based on historical multidimensional combination classification results, the corresponding training samples are input into a preset interpretable model. The preset interpretable model is used in conjunction with the target constraints to perform optimization calculations and obtain the scheduling surface corresponding to each combination classification result. The target scheduling function is obtained based on the aforementioned scheduling surface; The step of generating a corresponding first resource scheduling result based on the multidimensional combined classification result using the interpretable model includes: The multidimensional combined classification results are calculated based on the target scheduling function to generate the corresponding first resource scheduling result.

7. The method according to claim 6, characterized in that, The method further includes: Based on the multidimensional combined classification result and the corresponding first resource scheduling result, a classification decision graph is generated, which is used to represent the scheduling boundary conditions corresponding to the multidimensional combined classification result; and / or Based on the changing trend of the target scheduling function, a feature sensitivity curve is generated. The feature sensitivity curve is used to evaluate the marginal response strength of the function input variables to the scheduling result.

8. A resource scheduling device, characterized in that, The device includes: The data acquisition module is used to: acquire multidimensional feature data and construct an input feature vector based on the multidimensional feature data, wherein the multidimensional feature data includes at least task information and resource information; The scheduling factor determination module is used to: input the input feature vector into the interpretable model, output a scheduling mapping function, and determine the target scheduling factor based on the scheduling mapping function; The hierarchical classification module is used to: perform hierarchical classification of tasks and resources based on the target scheduling factor to obtain a multi-dimensional combined classification result; and The scheduling module is used to: generate a corresponding first resource scheduling result based on the multi-dimensional combined classification result and using the interpretable model.

9. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

11. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images