A constrained resource allocation method based on multi-source temporal state perception and associated risk modeling

By using multi-source temporal state perception and associated risk modeling, a global state and object relationship representation is constructed, and risk penalty parameters and resource allocation upper limits are generated. This solves the risk problem in multi-object resource allocation and achieves the rationality and stability of resource allocation.

CN122884711BActive Publication Date: 2026-11-17HUNAN UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611386899.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17
Estimated Expiration
2046-09-08

AI Technical Summary

Technical Problem

Existing technologies struggle to simultaneously characterize multi-scale trends and short-term fluctuations, the relationships between static and dynamic objects, and the associated risks of multiple objects experiencing adverse operational outcomes. This can lead to improper resource allocation, potentially causing conflicts in control direction and system operational risks.

Method used

By using multi-source temporal state perception and associated risk modeling, a global state representation and object relationship representation are constructed, risk penalty parameters and resource allocation upper limits are generated, and combined with utility opportunity parameters and resource constraints, a quadratic programming problem is constructed to allocate resources, thereby reducing system operation risks.

Benefits of technology

It improves the rationality and stability of resource allocation, reduces the risk of task execution anomalies and system operation, and meets resource capacity and object resource constraints.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122884711B_ABST
    Figure CN122884711B_ABST
Patent Text Reader

Abstract

This invention discloses a constrained resource allocation method based on multi-source temporal state perception and associated risk modeling. The method includes: acquiring feature data from multiple decision objects within a preset historical window, processing this data according to the current period to obtain resource allocation input data; modeling based on global state features and historical operating results, fusing them to obtain shared control parameters, and generating risk penalty parameters and resource allocation upper limits; generating utility opportunity parameters through multi-object state and relationship modeling, and generating an associated risk matrix by combining historical adverse results; integrating the above parameters and resource constraints, constructing and solving a constrained quadratic programming problem to obtain the resource allocation results and reserved resources for each object in the current period. This invention can improve the rationality of resource allocation while satisfying resource capacity and object resource constraints. As a constrained resource allocation method based on multi-source temporal state perception and associated risk modeling, this invention can be widely applied in the field of constrained resource allocation technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of constrained resource allocation technology, and in particular to a constrained resource allocation method based on multi-source temporal state perception and associated risk modeling. Background Technology

[0002] Limited resources need to be dynamically allocated among multiple tasks, service instances, or device objects. Object requirements, system load, and service quality change over time, and objects may have dependencies, competition, or common failures. Therefore, resource allocation needs to utilize the temporal state of objects, global state, object relationships, and historical execution results simultaneously.

[0003] Existing prediction, rule-based, reinforcement learning, or explicit optimization methods typically focus on single demand forecasting, node selection, or resource balancing, making it difficult to simultaneously characterize multi-scale trends and short-term fluctuations, static and dynamic object relationships, and the associated risks of multiple objects experiencing adverse operational outcomes. Independently generating risk penalty parameters and resource allocation limits may also cause conflicts in control direction. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a constrained resource allocation method based on multi-source temporal state perception and associated risk modeling, which can improve the rationality and stability of resource allocation while satisfying resource capacity and object resource constraints.

[0005] The first technical solution adopted in this invention is: a constrained resource allocation method based on multi-source temporal state perception and associated risk modeling, comprising the following steps: The feature data of multiple decision objects within a preset historical window are obtained, and the feature data is processed according to the current decision cycle to obtain resource allocation input data. The feature data includes object time-series state data, global state features, historical operation result data, object relationship data, and resource constraint input data. Global state modeling is performed based on the global state characteristics and historical operation results data in the resource allocation input data to obtain a global state representation and a relationship risk summary. The global state representation and the relationship risk summary are then fused to obtain shared control parameters. Risk penalty parameters and resource allocation upper limits are generated based on the shared control parameters. Multi-object state and relationship modeling is performed based on object temporal state data, object relationship data, and global state representation to obtain object temporal representation and object relationship representation, and utility opportunity parameters are generated based on object temporal representation and object relationship representation; Historical adverse operating results are extracted from historical operating results data, and an associated risk matrix is ​​generated based on historical adverse operating results, object relationship representation, and global state representation; Based on the utility opportunity parameter, the associated risk matrix, the risk penalty parameter, the resource allocation upper limit, and the resource constraint input data, a constrained quadratic programming problem is constructed and solved to obtain the resource allocation vector corresponding to the current decision period. Based on the resource allocation vector, the resource allocation result and reserved resources corresponding to each decision object are determined.

[0006] Furthermore, the step of acquiring feature data of multiple decision objects within a preset historical window and processing the feature data according to the current decision cycle to obtain resource allocation input data specifically includes: Based on the current decision cycle, determine the object's time-series status window, global status window, and historical execution result window, and extract the object's time-series status data, global status characteristics, and historical execution result data from the corresponding data sources; Data from different data sources are time-aligned according to a unified time base, and the continuous numerical features after time alignment are scaled uniformly. Based on the data integrity, operational effectiveness, and resource allocability status of each decision object, a valid object mask is generated, and the valid decision objects participating in the resource allocation of the current decision cycle are determined based on the valid object mask; Generate object relationship data based on at least one of object category relationships, topological relationships, dependency relationships, or invocation relationships; Generate resource constraint input data based on at least one of the following: non-negativity constraint, total resource constraint, single-object resource constraint, and numerical stability parameter.

[0007] Furthermore, the step of performing global state modeling based on the global state features and historical operation result data in the resource allocation input data to obtain a global state representation and a relationship risk summary, fusing the global state representation and the relationship risk summary to obtain shared control parameters, and generating risk penalty parameters and resource allocation upper limits based on the shared control parameters specifically includes: The global state features in the resource allocation input data are temporally encoded according to multiple preset time scales to obtain global state representations corresponding to multiple time scales. The global state representations corresponding to different time scales are fused, and the global state representation is obtained through attention pooling. Historical adverse operating results are extracted from historical operating data, and a relationship risk summary is generated based on the historical adverse operating relationships between multiple decision-making objects. By fusing the global state representation and the relational risk summary, shared control parameters are obtained; Risk penalty parameters and resource allocation limits are generated based on shared control parameters, wherein when the shared control parameters increase, the risk penalty parameters increase and the resource allocation limits decrease.

[0008] Furthermore, the step of constructing the temporal representation of the object specifically includes: The time-series state data of each decision object are downsampled according to multiple preset time scales; Trend and fluctuation components are extracted from the time-series state data of objects at different time scales to obtain trend and fluctuation information at the corresponding time scales. By fusing trend and fluctuation information at different time scales across scales, a multi-scale object state representation corresponding to each decision object is obtained. Feature mapping is performed on the multi-scale object state representation to obtain the object temporal representation.

[0009] Furthermore, the step of constructing the object relation representation specifically includes: Static object relationships are generated based on object relationship data, and dynamic object relationships are generated based on the similarity between the temporal representations of each decision object. By fusing static and dynamic object relationships, object relationship data is obtained. Perform cross-object attention processing based on object temporal representations and object relationship data to obtain initial object relationship representations; State gating parameters are generated based on the global state representation, and the degree of injection of the initial object relation representation into the object temporal representation is adjusted based on the state gating parameters. The object-relational representation is obtained by residual fusion of the gated object-relational representation and the object-temporal representation.

[0010] Furthermore, the step of generating utility opportunity parameters based on object temporal representation and object relational representation specifically includes: Residual fusion is performed based on object temporal representation and object relational representation to obtain utility opportunity representation; The original utility opportunity score is generated based on the utility opportunity representation and global state representation corresponding to each decision object. Cross-sectional standardization of the original utility opportunity scores within the current set of effective decision objects; Bounded mapping is performed on the original utility opportunity scores after cross-sectional standardization to obtain the utility opportunity parameters corresponding to each decision object, and the utility opportunity parameters corresponding to ineffective decision objects are set to zero.

[0011] Furthermore, the step of extracting historical unfavorable operating results from historical operating result data and generating an association risk matrix based on historical unfavorable operating results, object relationship representation, and global state representation specifically includes: Based on the one-sided adverse deviation between the historical operating results of each decision-making object and the corresponding operating target, extract the historical adverse operating results; Construct an empirical correlation risk matrix based on historical adverse operating results and non-negative time weights; Based on historical adverse operating results, object relationship representation, and global state representation, potential risk factor loadings and positive idiomatic risk parameters are generated, and a neural association risk matrix is ​​constructed based on the potential risk factor loadings and positive idiomatic risk parameters. Trace normalization was performed on the empirical association risk matrix and the neural association risk matrix respectively, and the two normalized association risk matrices were then weighted and fused. By adding a diagonal stabilizing term to the weighted and fused association risk matrix and performing normalization, the association risk matrix is ​​obtained.

[0012] Furthermore, the step of constructing and solving a constrained quadratic programming problem based on the utility opportunity parameter, the association risk matrix, the risk penalty parameter, the resource allocation ceiling, and the resource constraint input data specifically includes: A resource allocation vector is constructed based on the resource allocation ratio corresponding to each decision object, and the resource allocation vector is used as the decision variable for the quadratic programming problem. A quadratic term is constructed based on the correlation risk matrix, risk penalty parameter, and numerical stability parameter, and a linear term is constructed based on the utility opportunity parameter; Construct non-negativity constraints, total resource constraints, and single-object resource constraints corresponding to the resource allocation vector based on the resource constraint input data; By inputting the quadratic term, the linear term, the nonnegativity constraint, the total resource constraint, and the single-object resource constraint into the quadratic programming solver, the optimal resource allocation vector and the corresponding constraint multipliers are obtained.

[0013] Furthermore, it also includes training the parameter generation process for shared control parameters, utility opportunity parameters, and the associated risk matrix, specifically including: Training samples are constructed based on historical global state characteristics, historical operation result data, historical object time-series state data, and historical object relationship data. Training objectives are constructed based on unfavorable operation results in historical operation result data, corresponding to the system's degree of conservative control, the relative resource allocation value of objects, and the common unfavorable operation relationship of objects. The shared control parameter generation process is trained based on historical global state characteristics and historical operation result data, so that the output shared control parameters can characterize the degree of conservative control of the current system, and risk penalty parameters and resource allocation upper limits are generated based on the shared control parameters. The utility opportunity parameter generation process is trained based on historical object time-series state data, historical object relationship data, and corresponding historical global state features, so that the output utility opportunity parameter represents the relative resource allocation value between different decision objects. The process of generating associated risk parameters is trained based on historical operational results data, historical object relationship data, and corresponding historical global state characteristics, so that the output associated risk matrix represents the degree of association between different decision-making objects that jointly experience adverse historical operational results. After independently training the generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices, the generation processes of these parameters are connected to the solution process of the quadratic programming problem. Based on the running results corresponding to the resource allocation vector obtained from the quadratic programming problem, a resource allocation level loss is constructed. The generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices are then jointly adjusted based on the resource allocation level loss, so that the shared control parameters, utility opportunity parameters, and associated risk matrices can jointly adapt to the constrained resource allocation process.

[0014] Furthermore, after obtaining the resource allocation vector, it also includes: Calculate the proportion and absolute amount of retained resources for the current decision-making cycle based on the resource allocation vector; Based on the resource allocation vector and the resource granularity corresponding to the current resource management system, the resource allocation ratio of each decision object is converted into the corresponding actual resource allocation amount; When resources are allocated using a continuous resource ratio, a corresponding resource quota is generated based on the resource allocation ratio. When resources are allocated using discrete resource units, priority sorting, integer mapping, or discrete optimization are performed according to the resource allocation ratio, and rounding errors are corrected. Resource control instructions are generated based on the actual resource allocation, and then provided to the corresponding resource execution units. After the current decision cycle is completed, collect the object status data, global status data and actual operation results of each decision object, and add them to the historical data sequence. Then, re-execute the constraint resource allocation in the next decision cycle.

[0015] The beneficial effects of the method of this invention are as follows: This invention improves the ability to perceive changes in resource demand and the state of object relationships by uniformly processing multi-source time-series state data and constructing multi-scale state representations, combining static and dynamic relationships between objects for correlation modeling, and adaptively adjusting relationship information according to the global state; furthermore, it generates a correlation risk matrix by using historical unfavorable operating results and object relationships, and constructs a constrained resource allocation problem by combining utility opportunity parameters, risk penalty parameters, and resource allocation upper limits, so that the resource allocation process can simultaneously consider the resource allocation value of objects, the correlation risks between objects, and the current resource constraint state of the system; at the same time, through the joint adjustment of the parameter generation process and the constrained resource allocation process, the generated resource allocation parameters can adapt to the actual resource allocation goals, thereby improving the rationality and stability of resource allocation while meeting resource capacity and object resource constraints, and reducing task execution anomalies and system operation risks caused by improper resource allocation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of a constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to the present invention; Figure 2 This is a schematic diagram of the overall process provided in a specific embodiment of the present invention; Figure 3 This is a schematic diagram of the overall framework provided in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the multi-source resource allocation data construction module structure provided in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the global state control parameter generation module structure provided in a specific embodiment of the present invention; Figure 6 This is a schematic diagram of the multi-object utility opportunity parameter learning module structure provided in a specific embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of the association risk matrix generation module provided in a specific embodiment of the present invention; Figure 8 This is a schematic diagram of the constraint resource allocation solution module structure provided in a specific embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0018] Let the index of the current decision cycle be... The number of decision-makers who can currently participate in resource allocation is The resource allocation vector is composed of the resource allocation ratio corresponding to each decision-making object. ,in, Indicates the first The decision-making object is in the first... The resource allocation ratio obtained in each decision cycle. The utility opportunity parameter is denoted as... , its first element Used to characterize the The resource allocation value of a decision object relative to other decision objects. The global state is represented as […]. Shared control parameters are denoted as The risk penalty parameter is denoted as The upper limit of resource allocation is denoted as Object-relational representation is denoted as The associated risk matrix is ​​denoted as .

[0019] like Figure 3 As shown, this embodiment includes a multi-source resource allocation data construction module, a global state control parameter generation module, a multi-object utility opportunity parameter learning module, an association risk matrix generation module, a constrained resource allocation solution module, a resource allocation output module, and a model training module. These modules are connected according to the data processing order of the aforementioned constrained resource allocation method, and are used to perform multi-source data construction, model generation parameter calculation, constrained optimization solution, resource allocation result output, and model training.

[0020] The multi-source resource allocation data construction module is used to acquire object time-series state data, global state features, historical operation result data, object relationship data, and resource constraint input data. Based on the current decision cycle, it performs historical window truncation, time alignment, mean-standard deviation standardization, and effective object filtering to obtain resource allocation input data. The module is also used to generate static object relationship data based on object category, topology, dependency, or call relationship, and to generate effective object masks based on object data integrity, running status, and resource allocability status.

[0021] The global state control parameter generation module is used to perform multi-scale trend-fluctuation encoding and attention pooling on the global state features to obtain a global state representation; generate a relationship risk summary based on historical adverse operating results; fuse the global state representation and the relationship risk summary to obtain shared control parameters, and generate risk penalty parameters and resource allocation limits from the shared control parameters. Specifically, as the shared control parameters increase, the risk penalty parameters increase, and the resource allocation limits decrease.

[0022] The multi-object utility opportunity parameter learning module is used to perform multi-scale trend-fluctuation encoding on object temporal state data to obtain object temporal representation; construct static-dynamic object relationships based on static object relationships and object temporal representation; perform cross-object attention processing in conjunction with effective object masks, and control the degree of injection of object relationship information based on global state representation; and perform cross-sectional normalization and bounded mapping on the fusion result to obtain utility opportunity parameters and object relationship representation.

[0023] The association risk matrix generation module is used to extract historical adverse operation results based on historical operation results and construct an empirical association risk matrix by combining it with non-negative time weights; generate potential risk factor loadings and positive specific risk parameters based on historical adverse operation results, object relationship representation and global state representation, and construct a neural association risk matrix; perform trace normalization, weighted fusion and diagonal stabilization on the empirical association risk matrix and the neural association risk matrix to obtain an association risk matrix that meets the requirements of positive semi-definite or positive definite.

[0024] The constrained resource allocation solution module is used to construct a constrained quadratic programming problem based on utility opportunity parameters, association risk matrix, risk penalty parameters, resource allocation upper limit, and resource constraint input data. It then uses a quadratic programming solver to obtain a resource allocation vector that satisfies both non-negativity and total resource constraints. During model training, this module is also used to construct a smoothing penalty surrogate objective based on the optimal solution of the original quadratic programming problem, and to implicitly pass the gradient of the resource allocation level to each parameter generation module.

[0025] The resource allocation output module is used to output the resource allocation ratio, actual resource allocation amount and reserved resource amount corresponding to each decision object. It can also output shared control parameters, risk penalty parameters, resource allocation upper limit, utility opportunity parameters, associated risk matrix, constraint satisfaction status and solution status for resource management and operation status monitoring.

[0026] The model training module is used to train each parameter generation module according to a two-stage training mechanism that combines parameter role pre-training and decision-oriented joint fine-tuning. In the first stage, the module is trained in the dependency order of the global state control parameter generation module, the multi-object utility opportunity parameter learning module, and the associated risk matrix generation module. In the second stage, the constraint resource allocation solution module is connected, and the parameter generation modules are jointly updated according to the resource allocation level loss and the supervision loss of each parameter role.

[0027] Reference Figure 1 This invention provides a constrained resource allocation method based on multi-source temporal state awareness and associated risk modeling, the method comprising the following steps: S100. Obtain feature data of multiple decision objects within a preset historical window, and process the feature data according to the current decision cycle to obtain resource allocation input data. The feature data includes object time-series state data, global state features, historical running result data, object relationship data, and resource constraint input data. S110. Determine the object time-series status window, global status window, and historical running result window based on the current decision cycle, and extract the object time-series status data, global status characteristics, and historical running result data from the corresponding data sources; In this embodiment, as Figure 4 As shown, multi-source resource allocation data is constructed to form unified resource allocation input data within the current decision-making cycle. This resource allocation input data is respectively input into the global state control parameter generation module, the multi-object utility opportunity parameter learning module, the association risk matrix generation module, and the constraint resource allocation solution module. The goal of this embodiment is to ensure that the data used by different modules has a consistent decision-making time benchmark, avoiding the impact of different data acquisition times, misaligned data windows, or differences in feature numerical scales on the subsequent model parameter generation results.

[0028] In the Within each decision cycle, let the object's temporal state window length be... The object's temporal state feature dimension is The global state window length is The global state feature dimension is The historical results window length is Based on the time index corresponding to the current decision cycle, extract the object's time-series status window, global status window, and historical execution result window respectively.

[0029] The temporal state data of multiple decision objects are represented as follows: .in, Indicates the first Time-series status data of objects in each decision cycle Indicates the number of decision-making objects. Indicates the length of the object's timing state window. The dimension representing the temporal state characteristics of an object. The temporal state sequence of each decision object is represented as follows: .in, Indicates the first The decision-making object at the time location The corresponding object state feature vector.

[0030] Object time-series state characteristics are used to describe the resource requirements, operating load, execution efficiency and service status of each object, and may include demand, utilization, queue length, response latency, arrival rate, throughput or completion status.

[0031] Global state characteristics are used to describe the overall operating environment of the resource system, and are represented as follows: .in, Indicates the first The global state feature matrix corresponding to each decision cycle Indicates the length of the global state window. This represents the dimension of the global state features. The global state feature matrix can be represented as: .in, Indicates time position The corresponding global state feature vector.

[0032] Global state characteristics describe the overall system environment, including total load, overall utilization, average latency, congestion level, energy consumption, available resource ratio, and anomaly level, and together with object-level time-series states, support the generation of subsequent parameters.

[0033] Historical execution result data is used to record the actual execution results of multiple decision-making objects within a preset historical window. The historical execution result data is represented as follows: .in, Indicates the first The historical performance matrix corresponding to each decision cycle Indicates the length of the historical execution results window. The historical sequence of results corresponding to each decision object is represented as follows: .

[0034] Historical operation results record actual results such as completion, timeout, failure, and throughput, which are used to generate a relational risk summary and to extract historical adverse results and construct an empirical association risk matrix.

[0035] Object relationship data is used to describe relatively stable or predetermined structural relationships between multiple decision objects. Object relationship data can be represented as a static object relationship matrix: .in, Indicates the first The decision-making object and the first The static relationship strength between decision objects. The static relationship can be determined based on at least one of the following: object category relationship, topological connection relationship, task dependency relationship, service call relationship, data transmission relationship, or shared device relationship. When two decision objects do not have a preset static relationship, the corresponding matrix element can be set to zero; when two decision objects have a static relationship, a preset weight can be set according to the relationship type, or a continuous weight can be set according to relationship frequency, call count, connection strength, or dependency degree.

[0036] Resource constraint inputs include at least nonnegativity constraints and total resource constraints, and may include single-object upper limits, category constraints, minimum guarantees, capacity constraints, and numerical stability parameters, which are used to limit the feasible domain of constraint resource allocation solutions.

[0037] S120. Time-align the data from different data sources according to a unified time benchmark, and perform scale unification processing on the continuous numerical features after time alignment. In this embodiment, after acquiring the aforementioned data, time alignment processing is first performed on different data sources. Since object state data, global state data, and runtime result data may originate from different acquisition devices, log systems, or monitoring modules, their sampling times and frequencies may differ. The time alignment processing maps each data point to the same time position according to a preset unified time sequence, ensuring that the object's temporal state, global state, and runtime result at the same time position correspond to the same system runtime stage.

[0038] Furthermore, a scaling process is performed on the time-aligned continuous numerical features. In this embodiment, mean-standard deviation standardization is used. For the first... The first decision-making object Each state feature, and its standardized result within a preset historical window, is represented as follows: .in, Indicates the first The decision-making object at the time location The Original state characteristics, and They represent the first The first decision-making object The mean and standard deviation of each feature within the current historical window. This represents the numerical stability constant to prevent the denominator from being zero.

[0039] For global state features, standardization can be performed in the same way within the global state history window. By standardizing the scale, the differences in dimensions between features such as resource utilization, response latency, throughput, and queue length can be reduced, allowing subsequent models to focus more on the relative changes of each feature, rather than letting the absolute value of the feature determine the model output.

[0040] S130. Generate a valid object mask based on the data integrity, operational validity, and resource allocability status of each decision object, and determine the valid decision objects participating in the resource allocation of the current decision cycle based on the valid object mask; S140. Generate object relationship data based on at least one of object category relationship, topological relationship, dependency relationship or invocation relationship; S150. Generate resource constraint input data based on at least one of non-negativity constraints, total resource constraints, single-object resource constraints, and numerical stability parameters.

[0041] In this embodiment, a valid object mask is generated based on the data integrity, operational effectiveness, and resource allocability status of each decision object. The valid object mask is represented as follows: Among them, when the first When a decision-making object has complete data, is in an effective operational state, and allows for resource allocation within the current decision-making cycle, set... When the first When a decision-making object has missing key data, has stopped operating, is temporarily unavailable for resource allocation, or is not within the current resource allocation scope, set The current set of valid objects is represented as follows: .

[0042] After window truncation, time alignment, scale unification, and effective object filtering, the resource allocation input data corresponding to the current decision-making cycle is output as follows: .in, This represents the temporal state data of objects after scaling. This represents the global state features after scaling. This represents historical operation results data. Represents a static object relationship matrix. Indicates the valid object mask. This represents the input data for resource constraints.

[0043] S200. Based on the global state characteristics and historical operation result data in the resource allocation input data, perform global state modeling to obtain a global state representation and a relationship risk summary. Then, fuse the global state representation and the relationship risk summary to obtain shared control parameters. Based on the shared control parameters, generate risk penalty parameters and resource allocation upper limit. S210. Perform time-series encoding on the global state features in the resource allocation input data according to multiple preset time scales to obtain global state representations corresponding to multiple time scales. In this embodiment, as Figure 5As shown, the global state control parameter generation module generates a global state representation, a relationship risk summary, shared control parameters, risk penalty parameters, and resource allocation limits for the current decision cycle based on global state characteristics and the common unfavorable operational relationships among multiple decision objects. This module primarily addresses the question of "to what extent should the current system adopt conservative resource control."

[0044] The global state includes both short-term disturbances and persistent trends. Therefore, this embodiment uses multi-scale trend-fluctuation encoding to describe the long-term operating state while preserving local changes.

[0045] Let the preset time scale set be: .in, Indicates the number of time scales, and The global state feature matrix is ​​downsampled according to each preset time scale to obtain: ; in, Indicates the first The global state sequence corresponding to each time scale. The scale is represented as The time-based pooling operation preserves short-term load changes and local anomalies, while the coarse-scale sequence describes the continuous operating state over a longer time range.

[0046] Perform trend-fluctuation decomposition on the global state sequence corresponding to each time scale. The trend and fluctuation components corresponding to each time scale are represented as follows: ; in, Indicates moving average processing. Indicates the first Trend components corresponding to each time scale Indicates the first The fluctuation components corresponding to each time scale; For the fluctuation component, this embodiment adopts a bottom-up, cross-scale mixing approach, progressively transferring fine-scale fluctuation information to the coarse-scale. One implementation method can be represented as follows: ; in, Indicates the number of times after bottom-up mixing. Fluctuations on a time scale are represented. This indicates downsampling processing to align adjacent scale sequence lengths. This indicates a learnable bottom-up fluctuation mapping.

[0047] For the trend component, this embodiment adopts a top-down cross-scale mixing approach, progressively transmitting coarse-scale trend information to fine-scale components. One implementation method can be represented as follows: ; in, Indicates the first step after top-down mixing. Trends on each time scale are represented. This indicates upsampling or interpolation processing used to align the lengths of adjacent scale sequences. This represents a learnable top-down trend mapping.

[0048] S220. The global state representations corresponding to different time scales are fused, and the global state representation is obtained through attention pooling. In this embodiment, after bidirectional cross-scale mixing of the trend component and the fluctuation component, the two are fused to obtain the state coding sequence corresponding to each time scale: ; in, Indicates the first Feature mappings corresponding to each time scale, residual terms Used to preserve original global state information and improve coding stability; For the Attention pooling is performed on the state encoding sequences corresponding to each time scale. Let... Indicates the first Time position in each time scale The corresponding state encoding is represented as follows: ; in, , and This represents the learnable attention parameters. The corresponding normalized attention weights are expressed as: ; in, Indicates the first The sequence lengths corresponding to each time scale are calculated. Weighted aggregation of each time position based on attention weights yields the scale-level global state representation: ; By concatenating and mapping the scale-level state representations corresponding to each time scale, we obtain the global state representation: ; in, This represents vector concatenation. and Indicates learnable mapping parameters, Representation layer normalization. Global state representation. The overall assessment includes short-term operational disturbances, continuous operational trends, and the importance of different historical time points.

[0049] S230. Extract historical adverse operation results from historical operation result data, and generate a relationship risk summary based on the historical adverse operation relationships between multiple decision objects; S240. The global state representation and the relational risk summary are fused to obtain shared control parameters; In this embodiment, in addition to global state characteristics, the degree of conservative control of the system is also affected by the correlation structure of multiple decision objects experiencing adverse operating results. When different decision objects are highly correlated under adverse conditions, the resource dispersion effect may still be reduced even if the overall system load index does not reach an extreme level. Therefore, this embodiment further constructs a relationship risk summary based on historical adverse operating results.

[0050] Let the first The decision-making object at the time location The corresponding actual operational efficiency is The corresponding preset running target or reference value is The running deviation is defined as: .

[0051] In this embodiment, higher operational utility indicates a more favorable operational result; therefore, the negative portion of the deviation is identified as an unfavorable operational result. .

[0052] A matrix of historical adverse operating results is formed by the adverse operating results corresponding to each decision-making object. Calculate the adverse correlation matrix between objects based on the historical adverse operating result matrix. . No. The decision-making object and the first The adverse correlation between decision objects is calculated based on the correlation coefficient of their adverse operating results.

[0053] Average adverse correlation is used to characterize the overall degree of association between all effective decision-making objects and their shared adverse operation. One implementation of this concept is as follows: ; in, Indicates the number of currently valid decision-making objects. Indicates the first The decision-making object and the first Adverse correlations among decision-making objects. A higher average adverse correlation indicates that multiple decision-making objects are more likely to experience adverse outcomes simultaneously. The proportion of the largest eigenvalue is used to characterize whether adverse risks among objects are concentrated in a few common risk factors. One implementation of this method is as follows: ; in, Represents the largest eigenvalue of the matrix. Represents the trace of a matrix. This represents the numerical stability constant. The higher the proportion of the largest eigenvalue, the more likely the adverse outcomes of multiple decision-making objects are dominated by a few common factors.

[0054] Object category risk concentration is used to characterize whether adverse operational results are concentrated in a specific object category, topological region, or task group. In one implementation, decision objects are divided based on static object relationship data. For each object group, calculate the mean or corresponding adverse risk energy of the adverse correlation within each object group. Then, perform maximum, mean, or dispersion statistics on the risk values ​​of each object group to obtain the risk concentration of the object category. Therefore, the summary of relationship risk is expressed as follows: Representing the global state Summary of Relationship Risks The data is spliced ​​and merged, and the shared control parameters are obtained by generating a network using shared control parameters. ; in, This indicates a shared control parameter generation network. This represents the Sigmoid activation function. It shares control parameters. Used to characterize the degree of conservative control in the current system. The larger the value, the higher the risk of adverse operation of the current system load, the degree of anomaly, or the combined adverse operation of objects. The smaller the value, the more stable the current system operation is and the lower the risk of adverse operation.

[0055] S250. Generate risk penalty parameters and resource allocation limits based on shared control parameters, wherein when the shared control parameters increase, the risk penalty parameters increase and the resource allocation limits decrease.

[0056] In this embodiment, based on shared control parameters Generate risk penalty parameters respectively and resource allocation limit : ; in, and Let represent the lower and upper bounds of the risk penalty parameter, respectively. ; and Let represent the lower and upper bounds of the resource allocation limit, respectively. .

[0057] As can be seen from the above mapping relationship, the shared control parameters When increased, the risk penalty parameter Strictly increase the upper limit of resource allocation Strictly reduce. By generating two resource control parameters using the same shared control parameters, the system's conservatism can be increased while simultaneously strengthening the penalty for common adverse operational risks and reducing the scale of resource investment. This structurally avoids the contradictory control outcome of increasing risk penalties while simultaneously increasing the upper limit of resource allocation.

[0058] The global state representation output in this embodiment The learning process involves inputting multi-object utility opportunity parameters and generating an association risk matrix. Risk penalty parameters are also included. and resource allocation limit Solving for input constraints and resource allocation. Shared control parameters. It can also serve as an interpretable intermediate state output to characterize the current system resource control tendency, but the shared control parameters themselves are not directly used as the resource allocation ratio.

[0059] S300. Perform multi-object state and relationship modeling based on object temporal state data, object relationship data and global state representation to obtain object temporal representation and object relationship representation, and generate utility opportunity parameters based on object temporal representation and object relationship representation; S310, Constructing object sequential representation; Specifically, the temporal state data of each decision object is downsampled according to multiple preset time scales; trend components and fluctuation components are extracted from the temporal state data of objects at different time scales to obtain trend information and fluctuation information at the corresponding time scales; the trend information and fluctuation information at different time scales are fused across scales to obtain multi-scale object state representations corresponding to each decision object; feature mapping is performed on the multi-scale object state representations to obtain object temporal representations.

[0060] S320. Constructing object-relational representations; Specifically, static object relationships are generated based on object relationship data, and dynamic object relationships are generated based on the similarity between the temporal representations of each decision object. The static and dynamic object relationships are fused to obtain object relationship data. Cross-object attention processing is performed based on the temporal representations of objects and object relationship data to obtain an initial object relationship representation. State gating parameters are generated based on the global state representation, and the degree of injection of the initial object relationship representation into the temporal representation of objects is adjusted based on the state gating parameters. The gated object relationship representation and the temporal representation of objects are residually fused to obtain the final object relationship representation.

[0061] S330. Generate utility opportunity parameters based on the object temporal representation and the object relational representation.

[0062] Specifically, residual fusion is performed based on the temporal representation and relational representation of objects to obtain the utility opportunity representation; the original utility opportunity score is generated based on the utility opportunity representation of each decision object and the global state representation; the original utility opportunity score is cross-sectionally standardized within the current set of valid decision objects; the cross-sectionally standardized original utility opportunity score is boundedly mapped to obtain the utility opportunity parameter corresponding to each decision object, and the utility opportunity parameter corresponding to invalid decision objects is set to zero.

[0063] In this embodiment, as Figure 6 As shown, multi-object utility opportunity parameter learning is used to generate utility opportunity parameters for each decision object within the current decision period based on the object time-series state data, object relationship data, effective object masks, and global state representations of multiple decision objects. Simultaneously, it generates object relationship representations for use in generating the association risk matrix. These utility opportunity parameters characterize the resource allocation value of different decision objects relative to other decision objects within the current decision period. They serve as linear utility parameters in the constrained resource allocation solution and are not directly used as the final resource allocation ratio.

[0064] 1) Module inputs and outputs: No. Within each decision cycle, the multi-object utility opportunity parameter learning module receives scale-unified object temporal state data. Static object relationship matrix and valid object mask and receive global state representation The input and output relationship of the module can be expressed as follows: ; in, Representing object relationships as a matrix, Representing object relationships as dimensions; Represents the original utility chance score vector; This represents the utility opportunity parameter vector after cross-sectional normalization and bounded mapping; Object-relational representation This is used to characterize the cross-object association information obtained by integrating the time-series status, static object relationships, and dynamic object relationships of each decision object, and is further input into the association risk matrix generation module. Original utility chance score. Primarily used for utility opportunity supervision during the parameter role pre-training phase. Final utility opportunity parameters. Used to construct linear utility terms in constrained resource allocation problems; It should be noted that the utility opportunity parameter is not used to predict the absolute resource demand of a decision object, nor does it represent the proportion of resources that the decision object should directly obtain. The utility opportunity parameter mainly describes the relative resource allocation value among multiple effective decision objects within the same decision period. The final resource allocation result also needs to be determined in conjunction with the correlation risk matrix, risk penalty parameters, resource allocation upper limit, and resource constraints.

[0065] 2) Multi-scale trend-fluctuation time series coding of objects: The resource requirements and operational states of different decision-making objects typically contain both short-term disturbances and persistent trends. For example, an object may experience queue growth in a short period due to a sudden surge in requests, or it may maintain high resource requirements for a longer period due to continuous business changes. If only a single fine time scale is used, the model is easily affected by local noise; if only a coarse time scale is used, the model may not be able to respond promptly to recent state changes. Therefore, this embodiment employs multi-scale trend-fluctuation encoding to simultaneously extract local changes and persistent structures in the object's state.

[0066] Let the preset time scale set corresponding to the object temporal coding be . .in, This indicates the number of time scales used in the object's temporal coding, and By performing time pooling or downsampling on the object's time-series state data according to different time scales, we obtain: ; in, Indicates the first The temporal state sequence of objects corresponding to each time scale. This indicates the sequence length at this time scale.

[0067] Fine time scales can preserve short-term changes in features such as task arrival rate, resource utilization, and response latency, while coarse time scales can reduce local noise and describe persistent resource demands and stable operational structures. The temporal state sequences of objects at different scales are converted into a unified hidden dimension through a shared or scale-corresponding input mapping, resulting in an initial scale representation. ; in, Indicates the first Input mapping corresponding to each time scale , Represents the hidden temporal dimension of an object.

[0068] In each multi-scale temporal coding block, trend-fluctuation decomposition is performed on the temporal representation of objects at each scale. For the first... The coded block, its ... The trend and fluctuation components corresponding to each scale are represented as follows: ; in, This indicates a moving average process performed along the time dimension; This represents the trend component, used to characterize the continuous change of an object's state within the current historical window; It represents the fluctuation component, used to characterize local deviations, short-term changes in resource demand, and instantaneous operational disturbances.

[0069] For the fluctuation component, a bottom-up, cross-scale mixing approach is adopted to progressively propagate local changes at the fine scale to the coarse scale. One implementation method is as follows: ; in, This represents the fluctuation component after cross-scale mixing. This indicates the downsampling process used to align the lengths of adjacent scale sequences. This indicates that the learnable wave components are mapped from the bottom up.

[0070] For the trend component, a top-down, cross-scale mixing approach is adopted to progressively propagate coarse-scale persistent changes to fine-scale levels. One implementation method is as follows: ; in, This represents the trend component after cross-scale mixing. This indicates upsampling or interpolation processing used to align the lengths of adjacent scale sequences. This indicates that the learnable trend components are mapped from top to bottom.

[0071] The mixed trend and fluctuation components are fused, and the result is updated via residual connection. Temporal representation of objects at various scales: ; in, Indicates the first In the coded block, the ... Output mappings corresponding to each scale. Residual connections are used to preserve object state information before entering the current coding block and improve the training stability of multi-layer temporal coding processes.

[0072] After a preset number of multi-scale time-series coding blocks, the final trend component and fluctuation component at each scale are converged over time. This embodiment employs both mean convergence and final-time convergence. For the... The average convergence is used to describe the overall operating pattern within the entire historical window, while the last-time convergence is used to retain the object state response closest to the current decision cycle.

[0073] Let the first The final fluctuation component and trend component corresponding to each scale are respectively and Then its mean convergence result is expressed as: ; The final time is represented as follows: and By concatenating and mapping the mean representation and the final time-time representation across all time scales, we obtain the object's temporal representation: ; in, This represents the temporal representation matrix of multiple decision objects. Its first... OK Indicates the first The temporal representation of a decision object obtained by integrating multiple time scales.

[0074] 3) Modeling static and dynamic object relationships: The resource allocation value of a decision-making object depends not only on its own temporal state but may also be affected by the operational state of other decision-making objects. For example, objects with task dependencies may need to coordinate resource allocation; objects sharing the same equipment, network link, or data source may have competitive relationships; and objects belonging to the same category may have similar resource demand changes. Therefore, this embodiment constructs both static and dynamic object relationships.

[0075] Get the static object relationship matrix To unify the neighbor relationship weights of different objects, the static object relationship matrix undergoes effective object filtering and row normalization. The normalized static object relationship matrix is ​​represented as follows: ; in, Indicates the first Within the first decision cycle The decision-making object and the first Static relationship weights between decision objects after effective object screening and row normalization; Indicates the first The decision-making object and the first The weights of the original static relationships between each decision object; and They represent the first The decision-making object and the first The decision-making object is in the first... The effective object mask values ​​corresponding to each decision cycle; Indicates the first The set of effective decision-making objects corresponding to each decision cycle; This represents the index used to traverse the set of valid decision objects; Indicates the first The sum of the static relationship weights between the i-th decision object and all currently valid decision objects is used as the i-th... Normalized denominator for row static relation weights; This represents the numerical stability constant. When the... The first object or the first When an object is invalid, the corresponding relationship weight is set to zero. The static object relationship matrix is ​​used to provide relatively stable structural priors such as categories, topologies, task dependencies, or service calls.

[0076] Static object relationships can describe pre-known object relationships, but they cannot cover time-varying associations that arise as the current running state changes. For example, two objects that originally did not belong to the same category may exhibit similar states within a certain decision cycle because they use the same resource node simultaneously; objects in the same category may also no longer have similar resource requirements due to different running stages. Therefore, this embodiment further constructs a dynamic object relationship matrix based on the object temporal representation.

[0077] First, map the temporal representation of objects to the dynamic relation embedding space: ; in, The temporal representation matrix of the object. The temporal representation dimension of an object; This represents a learnable dynamic relation mapping matrix, used to transform the temporal representation of objects from... 3D feature space mapping to 3D dynamic relationship embedding space; Represents a learnable bias vector; express One-dimensional vector This indicates that the bias vector is copied to all. The row corresponding to each decision object; Represents the dynamic relationship embedding matrix. Indicates the dimension of dynamic relationship embedding. Represents the dynamic relationship embedding matrix The Line, i.e., the first The dynamic relationship embedding corresponding to each decision object. The decision-making object and the first The dynamic relationship weights between decision objects can be represented as: ; in, To preserve non-negative similarity relationships, the denominator only applies to the currently valid set of objects. The objects within are normalized. This yields the dynamic object relation matrix. .

[0078] Static and dynamic object relationships serve different purposes. Static object relationships provide a relatively stable prior structure, which can reduce the random fluctuations that occur when the model learns relationships based solely on short-term states; dynamic object relationships can supplement the time-varying dependencies and state similarities not covered by static relationships, enabling the model to update the scope of cross-object information propagation based on the current running conditions.

[0079] By fusing the object's own temporal representation, the static relation propagation result, and the dynamic relation propagation result, a graph-enhanced object representation is obtained: ; in , and These represent the learnable mapping matrices corresponding to the object's own information, static relationship information, and dynamic relationship information, respectively. The graph-enhanced object representation... It also includes the object's own timing information, stable object relationship information, and dynamic object relationship information.

[0080] 4) Mask cross-object attention processing: Directly summing or averaging the propagation results of static and dynamic relationships may cause each decision-making object to absorb all relationship information indiscriminately. In actual operation, different neighboring objects have different degrees of influence on the target object, and some edges in the same relationship graph may only provide weak information. Therefore, this embodiment further employs masked cross-object attention, enabling each effective decision-making object to selectively aggregate information from other objects based on its current state.

[0081] Using the temporal representation of objects as query information and the graph-augmented object representation as key and value information, construct the query matrix, key matrix, and value matrix respectively: ; in, , and This represents a learnable attention mapping matrix. , This represents the attention dimension.

[0082] Construct an attention mask matrix based on the effective object mask. Its elements can be represented as: ; In actual numerical implementation, Sufficiently small negative numbers can be used as substitutes. The mask matrix ensures that invalid objects do not participate in attention weight normalization, preventing missing data objects, exiting objects, or unassignable objects from affecting the relational representation of valid objects.

[0083] The cross-object attention matrix is ​​represented as follows: ; in, Normalize along the dimensions of the aggregated objects. The row attention weight representation The importance of a decision object when aggregating information from other valid decision objects.

[0084] The object relationship representation is obtained based on the attention matrix: ; in, This represents the attention output mapping matrix. Object-relation representation enables each decision object to selectively absorb information from other objects that are relevant to its current operating state and resource allocation value, rather than performing fixed aggregations based solely on static relation weights or dynamic similarity.

[0085] 5) Global state gating and utility opportunity parameter generation: The effectiveness of object relationship information varies under different global operating states. When the system state is relatively stable, static category relationships and dynamic similarity relationships can provide information on cooperation, competition, and dependency between objects; when the system is under high load, strong disturbance, or abnormal state, the object state may be affected by system-level impacts simultaneously, and excessive propagation of relationship information may amplify noise or cause multiple object representations to converge.

[0086] Therefore, this embodiment generates object-level and feature channel-level state gating matrices based on object temporal representation, object relation representation, and global state representation. First, the global state representation is copied to each decision object, resulting in... .in, express A one-dimensional vector. The state gating matrix is ​​represented as: ; in, Represents the state gating matrix, This represents the gating parameter generation network. This represents the Sigmoid activation function.

[0087] Each element in the gating matrix controls the degree to which the corresponding decision object absorbs object relationship information in the corresponding feature channel. When a certain gating value is close to one, the corresponding relationship features are mostly retained; when the gating value is close to zero, the corresponding relationship features are suppressed.

[0088] The gated object relational representation and the object temporal representation are residually fused to form a utility opportunity representation: ; in, This indicates element-wise multiplication. Presentation layer normalization processing. Residual fusion preserves the object's own temporal information, preventing object relationship information from completely replacing the object's own state; state gating dynamically adjusts the contribution of relationship information based on the current global runtime environment. According to the... The original utility opportunity score is generated from the utility opportunity representation of each decision object and the global state representation: ; in, Indicates the first The utility opportunity representation for each decision-making object. Represents the utility chance parameter generation network, This represents the original utility chance score.

[0089] The absolute scale of the original utility opportunity score may vary with the object state and the number of valid objects in different decision-making cycles. If the original score is directly input into the constraint resource allocation solution, the scale of the linear utility term may change significantly between different decision-making cycles, thereby altering its relative strength with the quadratic associated risk term. Therefore, this embodiment performs cross-sectional standardization on the original utility opportunity score within the current set of valid objects.

[0090] The cross-sectional mean of the original utility chance score of the effective object is expressed as: ; The corresponding standard deviation of the cross section is expressed as: ; in, Indicates the number of currently valid decision-making objects. This represents the numerical stability constant.

[0091] Performing a hyperbolic tangent bounded mapping on the cross-section normalization result yields the first... Utility opportunity parameters corresponding to each effective decision object: ; For ineffective decision-making objects, set This yields the utility opportunity parameter vector: ; Cross-sectional standardization gives the utility opportunity parameter a relatively stable center and scale in different decision-making cycles, and bounded mapping restricts its value to a preset range, thereby improving the numerical stability of the utility opportunity parameter as an input to a linear term in quadratic programming.

[0092] A larger utility opportunity parameter indicates that the corresponding decision object has a relatively higher resource allocation value within the current decision-making cycle; a smaller utility opportunity parameter indicates that the corresponding decision object has a lower resource allocation value relative to other objects. Since the final resource allocation result is also influenced by the association risk matrix, risk penalty parameter, resource allocation ceiling, and other resource constraints, an object with a higher utility opportunity parameter does not necessarily obtain the largest resource allocation ratio.

[0093] In solving constrained resource allocation problems, the utility chance parameter vector is used to construct the linear terms of the quadratic programming problem. Therefore, decision-making objects with higher utility opportunity parameters, all other things being equal, have a stronger tendency to allocate resources.

[0094] S400. Extract historical adverse operation results from historical operation result data, and generate an association risk matrix based on historical adverse operation results, object relationship representation, and global state representation; Specifically, based on the one-sided adverse deviation between the historical operating results of each decision object and its corresponding operating objective, historical adverse operating results are extracted; an empirical correlation risk matrix is ​​constructed based on the historical adverse operating results and non-negative time weights; potential risk factor loadings and positive idiosyncratic risk parameters are generated based on the historical adverse operating results, object relationship representations, and global state representations, and a neural correlation risk matrix is ​​constructed based on the potential risk factor loadings and positive idiosyncratic risk parameters; trace normalization is performed on the empirical correlation risk matrix and the neural correlation risk matrix respectively, and the two normalized correlation risk matrices are weighted and fused; a diagonal stabilizing term is added to the weighted and fused correlation risk matrix and normalized to obtain the correlation risk matrix.

[0095] In this embodiment, as Figure 7 As shown, the association risk matrix generator is used to construct the association risk matrix corresponding to the current decision-making cycle based on the historical adverse operating results, object relationship representations, and global state representations of multiple decision objects. The association risk matrix is ​​used to characterize the strength and association structure of multiple decision objects jointly experiencing adverse operating results, and serves as the matrix parameter for the secondary association risk term in the constraint resource allocation solution.

[0096] Traditional historical association matrices can preserve commonly adverse operational information observed in actual practice, but they rely primarily on past results and may exhibit response lag when the system state changes rapidly. While directly using neural networks to predict each element in the association risk matrix has strong adaptability, the model output does not necessarily satisfy symmetry and positive semi-qualitative requirements, potentially compromising the convexity of subsequent quadratic programming problems. To balance historical stability, state adaptability, and matrix structure effectiveness, this embodiment constructs both an empirical association risk matrix and a neural association risk matrix, and performs trace normalization, weighted fusion, and diagonal stabilization on both to obtain the association risk matrix. .

[0097] 1) Module inputs and outputs: In the Within each decision cycle, the associated risk matrix generation module receives historical operational result data and valid object masks, and receives object relationship representations. and receive global state representation Its input-output relationship can be expressed as: ; in, This represents historical operation results data. Representation of object-relational representation, This represents the global state. Indicates the valid object mask. This represents the associated risk matrix corresponding to the current decision-making cycle.

[0098] In the association risk matrix, the diagonal elements characterize the risk intensity of an adverse outcome occurring within the corresponding decision-making object itself, while the off-diagonal elements characterize the association strength of different decision-making objects jointly experiencing adverse outcomes. The association risk matrix is ​​not a direct estimate of the state covariance of general objects, but rather a one-sided association risk representation constructed for adverse outcomes.

[0099] It should be noted that favorable and unfavorable operational results have different technical implications for resource allocation. Multiple objects achieving favorable operational results simultaneously should not incur the same risk penalty as multiple objects experiencing simultaneous task failures, response timeouts, or service level breaches. Therefore, this embodiment only retains operational results from unfavorable directions to reduce the interference of fluctuations in favorable directions on the modeling of common unfavorable operational risks.

[0100] 2) Extraction and coding of historical unfavorable operating results: Before extracting unfavorable historical operating results, different types of historical operating results are first converted into a unified utility direction, so that the larger the value of the converted operating result, the more favorable the operating state, and the smaller the value, the more unfavorable the operating state.

[0101] For operational metrics such as throughput, task completion rate, and processing efficiency, where higher values ​​are more advantageous, these can be used directly or after scaling to represent utility results. For operational metrics such as response latency, task timeout rate, service level default rate, and energy consumption, where lower values ​​are more advantageous, symbol transformation, target value minus actual value, or other monotonic transformation methods can be used to ensure that the transformed metrics also satisfy the condition that higher values ​​indicate more advantageous operational results.

[0102] Let the first The decision-making object at the time location The corresponding result of running in the same direction is Its preset operating target or reference value is The corresponding deviation in the running result is expressed as: .when When, it indicates that the actual operating results of the decision-making object have reached or exceeded the preset operating targets; when When the actual operating result is lower than the preset operating target, it indicates that the actual operating result is lower than the preset operating target. Therefore, the negative part of the operating result deviation is identified as the historical unfavorable operating result. .

[0103] Through the above processing, the value of the adverse operating result corresponding to the favorable operating result is zero; the higher the degree of adverseness, the better. The larger the absolute value, the better. A historical unfavorable performance matrix is ​​formed by the unfavorable performance outcomes of multiple decision objects within a historical window. .in, Indicates time position The corresponding vector of unfavorable operating results for multiple decision objects. Indicates the length of the historical execution results window.

[0104] When historical execution results are represented using binary adverse events, a preset value can be set to negative for the corresponding element when an adverse event occurs, and to zero when no adverse event occurs. For example, setting a value when a task fails. Set when the task is executed successfully When historical operation results are represented by continuous deviations, continuous negative values ​​can be set based on the degree of timeout, insufficient throughput, task backlog, or service level breach.

[0105] Construct a diagonal mask matrix based on the valid object mask: The mask matrix is ​​used to filter out historically unfavorable performance results: Therefore, the sequence of adverse operating results corresponding to invalid decision objects is set to zero and does not participate in the subsequent object risk representation and the calculation of common adverse operating relationships between objects.

[0106] To extract the unfavorable operating patterns of each decision object within the historical window, a time-series encoding with shared parameters is performed on the sequence of historical unfavorable operating results. This is represented as: ; in, The timing encoder indicates unfavorable operating results. The object representing the adverse operating result is represented.

[0107] In one specific implementation, the adverse running result temporal encoder includes an input mapping layer, a one-dimensional temporal convolutional layer, a nonlinear activation layer, residual connections, and a temporal convergence layer. The input mapping layer maps the adverse running result at each time position to the hidden feature space; the one-dimensional temporal convolutional layer extracts the changes in adverse results between consecutive time positions; the residual connections preserve the original adverse result information; and the temporal convergence layer converges the temporal representations within the historical window into an object-level representation.

[0108] The adverse operational outcome perception object representation primarily describes the adverse operational patterns that have occurred for each decision-making object, including the frequency, duration, recent changes, and severity of adverse events. The object relationship representation, on the other hand, describes cross-object association information under the current object state. Since they focus on different information, they need to be integrated in subsequent steps.

[0109] 3) Object relationship fusion and global state risk channel gating: Representing adverse operating results as objects Relationship between objects By performing feature concatenation and nonlinear mapping, the object risk representation is obtained: ; in, This indicates splicing along the feature dimension. This indicates a risk feature fusion network. Representation of object risk, This represents the risk characteristic dimension.

[0110] Adverse outcome perception object representation provides information on historically occurring adverse outcomes, while object relationship representation provides information on static relationships, dynamic similarity relationships, and current associations obtained from cross-object attention aggregation. Integrating these two allows object risk representation to simultaneously characterize the historical adverse patterns of a single object and the current association structure between multiple objects.

[0111] To enable the associated risk matrix to adjust its risk characteristics based on the current global state, a risk channel gating vector is generated using the global state representation: ; in, This represents the risk channel gating vector. and These represent the lower and upper bounds of the risk channel threshold, respectively. ; This indicates a risk channel gating generation network. This represents the Sigmoid activation function.

[0112] The risk channel gating vector is copied to multiple decision objects and then element-wise multiplied with the object risk representation to obtain the gated object risk representation: ; in, This indicates that the risk channel gating vector will be copied to all decision-making objects.

[0113] Risk channel gating is shared across the object dimension but has different values ​​across the feature channel dimension. Therefore, the same global state can consistently adjust the risk representation of all decision objects across channels, while preserving the specific risk characteristics differences of each decision object in its corresponding channel.

[0114] It should be noted that global state gating controls the degree to which object relationship information is injected into the utility opportunity representation, while risk channel gating controls the importance of different risk characteristic channels in the object risk representation. Although both use global state representation, their corresponding processing objects and technical functions are different.

[0115] 4) Generation of potential risk factors and positive-specific risk parameters: Directly predicting each element in the association risk matrix using a neural network makes it difficult to guarantee that the output matrix will always be symmetric and positive semi-definite. To ensure that the neural association risk matrix meets the subsequent optimization requirements under any network output condition, this embodiment does not directly predict the matrix elements, but instead generates potential risk factor loadings and positive specific risk parameters based on the gated object risk representation.

[0116] Generate a potential risk factor loading matrix based on the gated object risk representation: ; in, This represents the potential risk factor loading generation network. Represents the potential risk factor loading matrix. This indicates the number of potential risk factors.

[0117] Potential risk factors are used to characterize common adverse factors that may simultaneously affect multiple decision-making objects. For example, overload of shared computing devices, congestion of public network links, upstream service anomalies, insufficient system-level energy, or failure of a common data source may all affect multiple decision-making objects simultaneously. Matrix elements Used to characterize the The decision-making object is the first The sensitivity of each potential risk factor.

[0118] Based on the gated object risk representation, the original value of the idiomatic risk is further generated, and the positive idiomatic risk parameter is obtained through the Softplus function and the positive risk lower bound: ; in, This represents a network for generating specific risk parameters. Let represent the vector of positive specific risk parameters, and let its _i_ be the _i_th ... element Indicates the first The specific risk intensity corresponding to each decision object; This indicates a pre-defined positive lower bound for specific risks; express A one-dimensional vector, therefore Indicates by indivual The constant vector is used to add the same positive lower bound element-wise to the Softplus output corresponding to each decision object. Therefore, the specific risk parameters corresponding to each effective decision object all satisfy... This ensures that the diagonal risk term constructed from the positive specific risk parameters has strictly positive diagonal elements. The Softplus function can be expressed as: , Since the Softplus function outputs a positive value, and a positive risk lower bound is further added, for each valid decision object, we have: Positive specific risk parameters are used to characterize the adverse operational risks of an object that cannot be explained by common potential risk factors. For example, an object may have independent adverse risks due to its own task complexity, misconfiguration, or local equipment failure, and these risks need to be reflected in the diagonal elements of the associated risk matrix even if they do not occur together with other objects.

[0119] Construct a neural association risk matrix based on the potential risk factor loading matrix and positive specific risk parameters: ; in, Represents the neural association risk matrix. This represents a diagonal matrix constructed using positive specific risk parameters as diagonal elements; Among them, low-rank matrix A diagonal matrix used to characterize the shared adverse operational relationships among objects caused by potential common risk factors. Used to supplement the independent adverse operational risks of each decision-making entity. Number of factors The parameters can be pre-set based on the number of decision objects, the complexity of their relationships, and the system's computational resources. For invalid decision objects, factor loadings and positive idiomatic risk parameters can be filtered using a valid object mask. ; Then use and Construct the neural association risk matrix corresponding to the currently valid objects. The resource allocation amount corresponding to invalid objects will be restricted to zero in the constraint resource allocation solution.

[0120] 5) Construction of the experience-related risk matrix: The neural association risk matrix can be dynamically adjusted based on object state, object relationships, and global state. However, during model training, especially in decision-oriented joint fine-tuning, the neural network may over-adapt to resource allocation level losses, causing the risk matrix to deviate from the commonly unfavorable operating structure observed in historical data. To preserve risk information from actual operating data, this embodiment further constructs an empirical association risk matrix based on historical unfavorable operating results.

[0121] The historical unfavorable operation results window contains Each time point is assigned a non-negative time weight: .in .

[0122] In one specific implementation, an exponential decay method is used to set the time weight: ; in, This represents the time decay parameter. Historical time points closer to the current decision-making period have higher weights. When... At that time, each historical time position has the same weight. Construct a diagonal weight matrix based on the time weights: The empirical correlation risk matrix is ​​represented as follows: .in, This represents the experience-related risk matrix.

[0123] The first of the empirical correlation risk matrix The diagonal element represents the first... The second-order intensity of adverse outcomes for a decision-making object, weighted by time within a historical window, is as follows: Line 1 Column elements represent the first The decision-making object and the first The time-weighted second-order intensity of adverse operating results occurring together among all decision objects.

[0124] Unlike symmetric covariance matrices that use mean-centered covariance matrices, the empirical correlation risk matrix is ​​constructed directly based on one-sided adverse operating results, without mixing favorable and unfavorable results. Therefore, when multiple decision objects experience adverse operating results simultaneously, the absolute value of the corresponding matrix elements increases; when two objects do not experience adverse operating results simultaneously at the same time, the corresponding empirical correlation risk intensity is lower.

[0125] 6) Trace normalization, empirical-neural matrix fusion, and diagonal stabilization: The empirical association risk matrix and the neural association risk matrix are generated by different computational paths, and their numerical scales may differ. If weighted fusion is performed directly, the matrix with the larger numerical scale may dominate the fusion result, making the preset fusion parameters unable to accurately reflect the relative contributions of the two matrices. Therefore, trace normalization is performed on both types of matrices before fusion.

[0126] Define the trace normalization operator as follows: ; in, Representation matrix traces, This represents the trace-normalized numerical stability constant.

[0127] Trace normalization was performed on the empirical association risk matrix and the neural association risk matrix respectively: ; Let the empirical-neural matrix fusion parameters be... ,and The normalized empirical association risk matrix and the neural association risk matrix are weighted and fused according to the fusion parameters, and a diagonal stabilizing term is added: ; in, Represents the diagonal stability parameter. express 3D identity matrix.

[0128] In this embodiment, the fusion parameters These are pre-defined, fixed parameters determined during model validation. The empirical association risk matrix retains pre-defined non-zero weights, ensuring that actual historical adverse performance information is always preserved during the joint fine-tuning of decision-oriented processes. The neural association risk matrix provides a risk structure that is dynamically adjusted based on object relationships and the current global state.

[0129] when When the fusion result is mainly determined by the empirical correlation risk matrix; when When the fusion result is determined primarily by the neural association risk matrix; when At the same time, the fusion result includes both historical stable risk information and state adaptive risk information.

[0130] Perform trace normalization again on the fusion matrix after adding the diagonal stabilizing term to obtain the final association risk matrix:

[0131] It should be noted that the diagonal stability parameter in the association risk matrix Numerical stability parameters in quadratic programming objectives They have different functions. Diagonal stabilizing parameters Numerical stability parameters are used to prevent the correlation risk matrix from approaching singularity and to improve the numerical stability of the risk matrix itself. This is used to ensure that the objective function of quadratic programming has a uniform lower bound on positive curvature, and to improve the strong convexity and solution stability of the optimization problem.

[0132] 7) Structural properties of the association risk matrix: This embodiment ensures that the empirical correlation risk matrix and the neural correlation risk matrix meet the positive semi-qualitative requirement by constructing a matrix, rather than correcting them by eigenvalue truncation or projection after the model output.

[0133] For any vector The empirical correlation risk matrix satisfies: ; Due to the time weight matrix The diagonal elements of the empirical correlation risk matrix are all non-negative, therefore... It is a positive semi-definite matrix; For the neural association risk matrix, we have: ; because The neural association risk matrix is ​​a positive semi-definite matrix; for all non-zero vectors corresponding to all valid objects, the neural association risk matrix is ​​further a positive definite matrix when all positive specific risk parameters are greater than zero. Trace normalization scales the matrix using only positive numbers, without changing its positive semi-definiteness. The weighted sum of two positive semi-definite matrices using non-negative weights remains a positive semi-definite matrix. Adding... Then, for any non-zero vector ,have:

[0134] Therefore, the fusion matrix after adding the diagonal stabilizing term is a positive definite matrix. After performing trace normalization again, the final association risk matrix is ​​obtained. It remains positive definite. Therefore, the associated risk matrix at least satisfies the positive semi-definite requirement, and is a positive definite matrix under the preferred parameter settings in this embodiment, which can be used as the secondary risk term parameter for subsequent quadratic programming problems.

[0135] With the above structure, the neural network will not cause the associated risk matrix to lose its positive semi-qualitative nature, regardless of the potential risk factor load it outputs, thus avoiding the non-convex quadratic programming problem that may occur when directly predicting matrix elements.

[0136] S500: Based on the utility opportunity parameter, the associated risk matrix, the risk penalty parameter, the resource allocation upper limit, and the resource constraint input data, construct and solve the constrained quadratic programming problem to obtain the resource allocation vector corresponding to the current decision cycle, and determine the resource allocation result and reserved resources for each decision object based on the resource allocation vector.

[0137] Specifically, a resource allocation vector is constructed based on the resource allocation ratio corresponding to each decision object, and this resource allocation vector is used as the decision variable for the quadratic programming problem. A quadratic term is constructed based on the association risk matrix, risk penalty parameter, and numerical stability parameter, and a linear term is constructed based on the utility opportunity parameter. Non-negativity constraints, total resource constraints, and single-object resource constraints corresponding to the resource allocation vector are constructed based on the resource constraint input data. The quadratic term, linear term, non-negativity constraints, total resource constraints, and single-object resource constraints are input into the quadratic programming solver to obtain the optimal resource allocation vector and the corresponding constraint multipliers.

[0138] In this embodiment, as Figure 8 As shown, the constrained resource allocation solution is used to uniformly transform the risk penalty parameter, resource allocation upper limit, utility opportunity parameter, and associated risk matrix into a resource allocation vector that satisfies the preset resource constraints. This step does not directly output the resource allocation ratio for each decision object through a neural network; instead, it constructs a constrained quadratic programming problem based on parameters generated by a model with a clear function, and uses a quadratic programming solver to obtain the optimal solution to the original optimization problem.

[0139] Among them, the utility opportunity parameter describes the relative resource allocation value among the decision objects; the associated risk matrix describes the risk structure of multiple decision objects jointly experiencing adverse operating results; the risk penalty parameter controls the overall penalty intensity of associated risks in the resource allocation objective; and the resource allocation ceiling limits the total amount of resources that can be allocated to all decision objects in the current decision cycle. By uniformly setting the above parameters in the explicit constraint optimization problem, an interpretable resource allocation relationship can be formed among utility opportunities, associated risks, and resource control requirements.

[0140] 1) Module inputs and decision variables: In the Within each decision cycle, the constrained resource allocation solution module receives the following inputs: utility opportunity parameters. Risk Matrix Risk penalty parameters and resource allocation limit , as well as valid object masks and other resource constraint input data; A resource allocation vector is constructed using the normalized resource allocation ratios corresponding to multiple decision objects within the current decision-making cycle. .in, Indicates the first The decision-making object is in the first... Resource allocation ratio obtained in each decision cycle. Resource allocation vector. These are the decision variables in this step; When using a normalized resource representation, the total amount of currently schedulable resources is set to a unit resource, and the resource allocation ratio for each decision-making object is between zero and one. If the actual system uses absolute resource quantities, the normalized resource allocation ratio can be converted to an absolute resource quantity based on the total amount of currently schedulable resources. Let the total amount of currently schedulable resources be... Then the first The actual amount of resources obtained by each decision-making entity can be expressed as: .in, Indicates the first The actual amount of resources obtained by each decision-making entity in the current decision-making cycle; For decision objects marked as invalid by the valid object mask, they can be removed from the current decision variable set before constructing the quadratic programming problem, or the resource allocation limit for the object can be set to zero to satisfy the condition. Therefore, ineffective decision-making entities do not participate in resource competition and allocation during the current decision-making cycle.

[0141] 2) Construction of the quadratic programming objective function: Construct the quadratic term matrix and linear term vector for the quadratic programming problem based on the utility opportunity parameter, the association risk matrix, the risk penalty parameter, and the numerical stability parameter: ; in, Represents a quadratic term matrix. Represents a vector of linear terms. Indicates the numerical stability parameter, express 3D identity matrix.

[0142] Therefore, the original quadratic programming problem constructed by the constrained resource allocation solution module is expressed as: ; It can also be equivalently represented as: ; in, This represents the optimal resource allocation vector for the original quadratic programming problem.

[0143] The first term in the objective function is the linear utility term: ;

[0144] Since this embodiment adopts a minimal form, the utility chance parameter The larger the value, the higher the corresponding resource allocation ratio, all other things being equal. The higher the linear utility term, the more favorable it is for reducing the objective function value. Therefore, a linear utility term makes decision objects with higher utility opportunity parameters have a stronger tendency to allocate resources.

[0145] The utility opportunity parameter only represents the relative resource allocation value among different decision-making objects. Even if an object has a high utility opportunity parameter, its final resource allocation ratio will still be affected by the associated risk matrix, risk penalty parameter, resource allocation ceiling, and other resource constraints. Therefore, this avoids the problem of ignoring the common adverse operational risks among objects when ranking or allocating resources solely based on utility opportunity scores.

[0146] The second term in the objective function is the quadratic association risk term: ;

[0147] Expanding it can be represented as: ;

[0148] In this risk matrix, the diagonal elements penalize the adverse operational risk of an individual decision-making object, while the off-diagonal elements penalize the risk of multiple decision-making objects jointly experiencing adverse operational outcomes. When the shared adverse operational risk between two decision-making objects is high, and both receive a large proportion of resource allocation, the corresponding secondary risk term increases, thereby inhibiting excessive resource concentration on a combination of objects with high shared adverse operational risk.

[0149] Risk penalty parameters This is used to control the overall influence of the associated risk matrix on the objective function. Shared control parameters are used when it is determined that a higher degree of conservative control should be applied to the current system. Increase, thereby generating a larger risk penalty parameter. This increases the associated risk penalty corresponding to the same resource allocation vector.

[0150] The third term in the objective function is the numerical stability term. .

[0151] The numerical stabilizing term provides a uniform lower bound on the positive curvature of the quadratic term matrix, thus ensuring the strong convexity of the objective function. Simultaneously, this term penalizes the square norm of the resource allocation vector, which can, to some extent, reduce the tendency for resources to become excessively concentrated in a few decision objects. It should be noted that the main technical function of the numerical stabilizing term is to ensure the strong convexity of the optimization problem and improve the stability of the numerical solution; its role differs from that of the diagonal stabilizing term added to the associated risk matrix.

[0152] 3) Constraint Construction: In this embodiment, the primal quadratic programming problem includes at least nonnegativity constraints and total resource constraints. The nonnegativity constraint is expressed as... ,Right now Non-negativity constraints are used to ensure that the resource allocation ratio obtained by each decision object is not less than zero, so that the final output conforms to the physical meaning of resource allocation.

[0153] Total resource constraints are expressed as .in, express One-dimensional vector This indicates the upper limit of resource allocation corresponding to the current decision-making cycle. The upper limit of resource allocation is generated by shared control parameters and satisfies... .

[0154] Total resource constraints limit the total proportion of resources allocated to all decision-making objects in the current decision cycle to not exceed the resource allocation ceiling. When the system's conservative control level increases, the resource allocation ceiling decreases, allowing more resources to be reserved and not immediately allocated to the current decision-making objects.

[0155] Based on actual resource management requirements, the primal quadratic programming problem can also include a single-object resource allocation ceiling. Let the... The upper limit of resource allocation for each decision-making object is This can increase An object upper bound vector is composed of multiple object upper bounds. When, it can be represented as .

[0156] The single-object resource allocation limit is used to prevent a single decision-making object from consuming resources exceeding its processing capacity, interface capacity, or safe operating range. For currently invalid decision-making objects, a corresponding limit is set. .

[0157] In other implementations, minimum resource guarantee constraints, object category resource constraints, device capacity constraints, or object dependency constraints can be added. For example, for objects that require minimum guaranteed resources, a minimum resource guarantee constraint can be set. .in, Indicates the first The minimum guaranteed resource ratio corresponding to each decision-making object.

[0158] For those belonging to the same object group For multiple decision-making objects, you can set the resource limit for the object group:

[0159] in, Indicates the first The upper limit of resource allocation corresponding to each object group.

[0160] The aforementioned extended constraints can all be incorporated into the quadratic programming problem in the form of linear equations or linear inequalities. Under the premise that the utility opportunity parameter, the associated risk matrix, the risk penalty parameter, and the resource allocation ceiling all participate in the constraint optimization, the specific type of resource constraint is adjusted according to the actual resource system.

[0161] The basic feasible region in this embodiment is represented as follows:

[0162] Therefore, the complete primal quadratic programming problem can be expressed as:

[0163] 4) Existence and uniqueness of the primal quadratic programming problem: Association Risk Matrix For a positive semi-definite matrix or a positive definite matrix, the risk penalty parameter satisfies The numerical stability parameters satisfy Therefore, for any non-zero vector ,have:

[0164] Therefore, the quadratic term matrix satisfies:

[0165] The Hessian matrix representation of the objective function with respect to the resource allocation vector is as follows:

[0166] Since the Hessian matrix is ​​a positive definite matrix, the objective function with respect to the resource allocation vector is: - Strongly convex functions.

[0167] Furthermore, the zero vector Satisfies the nonnegativity constraint and the total resource constraint, therefore the feasible region is... Non-empty. Both the non-negativity constraint and the total resource constraint are closed half-space constraints, and their intersection is a closed set. For any feasible resource allocation vector, we know from the non-negativity constraint and the total resource constraint that:

[0168] Therefore, the feasible region is a bounded set. According to the Heine-Borel theorem in finite-dimensional Euclidean space, a closed and bounded feasible region... For compactness.

[0169] Since the objective function is a continuous function, according to the extremum existence theorem, a continuous function can attain its minimum on a non-empty compact set. Therefore, the primal quadratic programming problem has at least one optimal solution. Furthermore, because the objective function is strongly convex, and a strongly convex function has at most one optimal solution on a convex feasible region, the primal quadratic programming problem has a unique optimal solution. .

[0170] The above properties demonstrate that, under any model output condition that satisfies the parameter value range, the present invention can obtain a definite and unique resource allocation result, avoiding the uncertainty of the execution result caused by multiple different optimal resource allocation vectors corresponding to the same set of model generation parameters.

[0171] 5) Forward solution of the original quadratic programming problem: The constraint resource allocation solution module directly solves the original quadratic programming problem during the forward computation process and uses the unique optimal solution of the original problem as the resource allocation vector for the current decision cycle.

[0172] The non-negativity constraint and the total resource constraint are represented as follows:

[0173] In the implementation of the basic constraint:

[0174] The primal quadratic programming problem can be represented as:

[0175] Let the Lagrange multipliers corresponding to the inequality constraints be . The Karush-Kuhn-Tucker conditions for the primal quadratic programming problem are expressed as:

[0176] Among them, the first term is the stationarity condition, the second term is the original feasibility condition, the third term is the dual feasibility condition, and the fourth term is the complementary relaxation condition.

[0177] Since the objective function is a strongly convex function and the constraints are linearly convex constraints, the resource allocation vector that satisfies the above conditions is the unique global optimal solution to the original quadratic programming problem.

[0178] In practical implementation, interior-point methods, effective set methods, operator splitting methods, or other numerical solvers capable of solving convex quadratic programming problems can be used. The solver receives the quadratic term matrix, the linear term vector, and the linear constraints, and outputs the optimal resource allocation vector and the corresponding constraint multipliers.

[0179] The resource allocation vector output from the forward solution of the original quadratic programming problem already satisfies the non-negativity constraint and the total resource constraint, eliminating the need for post-solution normalization, truncation, projection, or heuristic correction. Therefore, the actual resource allocation result remains consistent with the original constraints of the model, avoiding post-solution processing that alters the resource allocation objective or undermines the pre-defined role of parameters in the optimization problem.

[0180] When the constraint residuals of the quadratic programming solver fall within a preset tolerance range due to numerical accuracy, the result can be considered as satisfying the constraints. The preset tolerance is set based on the resource allocation accuracy and the solver's numerical accuracy. If the solver fails to meet the preset convergence condition, it can use the feasible resource allocation vector from the previous decision cycle, adopt a zero resource allocation vector, or activate a preset safe resource allocation strategy to ensure the continuous operation of the resource system.

[0181] 6) Reserve resources and resource allocation output: When the total amount of currently schedulable resources is normalized to one, the proportion of reserved resources is calculated based on the optimal resource allocation vector:

[0182] in, Indicates the first The proportion of resources reserved for each decision cycle.

[0183] because:

[0184] therefore:

[0185] Reserved resources include resources that are forcibly reserved due to the resource allocation cap being less than one, and resources that are proactively not allocated due to the combined effects of utility opportunities and associated risks in the optimization problem. Reserved resources can be used for unforeseen tasks, minimum resource guarantees for critical objects, system failure recovery, system fault tolerance, or resource reserves for the next decision-making cycle.

[0186] When the total amount of currently schedulable resources is At that time, the absolute amount of reserved resources is expressed as .

[0187] The resource allocation output module generates resource allocation results based on the optimal resource allocation vector. The resource allocation results may include the resource allocation ratio, absolute amount of resource allocation, proportion of retained resources, absolute amount of retained resources, constraint satisfaction status, and solution status for each decision object.

[0188] Depending on the specific resource management system, the resource allocation vector can be further converted into resource control instructions. For example, the resource allocation ratio can be converted into task resource quotas, object processing time ratios, bandwidth percentages, storage capacity limits, or energy usage limits. The above conversion only changes the execution unit of the resource allocation result and does not change the model parameter generation and constraint optimization process of this invention.

[0189] 7) Forward solution of the original problem and backward differentiation of the smooth surrogate: During the model training phase, the gradient generated by the resource allocation level loss needs to be passed through the constrained resource allocation solution module to the utility opportunity parameter, the associated risk matrix, the risk penalty parameter, and the resource allocation upper limit, and then the model parameters are updated.

[0190] Directly expanding the internal iterative process of the quadratic programming solver increases storage and computational overhead. Furthermore, when the set of active constraints in the original quadratic programming problem changes, the local derivatives of the optimal solution with respect to the input parameters may abruptly change, thus affecting the stability of model training.

[0191] Therefore, this embodiment adopts a separate processing method for the original quadratic programming forward solution and the smoothed surrogate backward differentiation. During the forward computation, the quadratic programming solver still directly solves the original constrained problem, and the actual output always uses the optimal solution of the original problem. During backpropagation, a smoothing penalty surrogate problem is constructed based on the optimal solution of the original problem and the corresponding constraint multipliers, and implicit differentiation is performed through this surrogate problem.

[0192] Suppose the original quadratic programming problem includes The inequality constraint, the first The inequality constraints are expressed as follows:

[0193] in, and They represent the first The coefficient vector and right-hand side terms corresponding to each linear constraint.

[0194] For nonnegativity constraints It can be equivalently represented as The total resource constraint can be expressed as... Define the smooth positive part function. .in, This represents the smoothed temperature parameter. The positive part function of the smoothing is... The function is a continuously differentiable approximation. When the constraints are satisfied and there is a large relaxation, the function value is close to zero; when the constraints are violated, the function value increases with the degree of violation.

[0195] The penalty weights for each constraint are set according to the constraint multipliers corresponding to the original quadratic programming problem. In one implementation, the first... The penalty weights corresponding to each constraint are represented as follows:

[0196] in, Indicates the base penalty weight. This represents the constraint multiplier scaling factor. In the original quadratic programming problem, the first... The optimal multiplier corresponding to each constraint.

[0197] The larger the constraint multiplier, the stronger the impact of the corresponding constraint on the current optimal resource allocation result. Therefore, a higher smoothing penalty weight can be set for the constraint. The basic penalty weight is used to ensure that the surrogate problem can still perceive the corresponding constraint boundary when the constraint multiplier is small or zero.

[0198] The objective function of the smooth penalty agent is expressed as:

[0199] During backpropagation, the optimal solution of the original quadratic programming problem is used. As a local expansion point or iteration initial point for the smooth surrogate problem, a differentiable mapping from the parameters to the resource allocation result is constructed using the local first-order conditions of the smooth surrogate objective. It is important to emphasize that the smooth surrogate problem is only used for gradient calculation; its output does not replace the forward resource allocation result of the original quadratic programming problem.

[0200] The first derivative of the smoothed positive part function is expressed as:

[0201] Its second derivative is expressed as:

[0202] Therefore, the gradient of the smooth agent objective function with respect to the resource allocation vector is expressed as:

[0203] The Hessian matrix representation of the smoothing agent objective function with respect to the resource allocation vector is as follows:

[0204] Because of each item All are positive semi-definite matrices, and their corresponding coefficients are non-negative. Therefore:

[0205] Thus, the smoothed surrogate objective function maintains strong convexity, and its Hessian matrix is ​​invertible, enabling the stable computation of the gradient of the resource allocation result relative to the model generation parameters through implicit differentiation.

[0206] Let all optimized input parameters be denoted as:

[0207] The local first-order conditions of the smooth surrogate problem are expressed as:

[0208] in, This represents the locally differentiable resource allocation mapping corresponding to the smooth proxy problem.

[0209] According to the implicit function theorem, the local derivative of the resource allocation result with respect to the optimization input parameters is expressed as:

[0210] in:

[0211] because Since it is a positive definite matrix, the aforementioned implicit derivative exists.

[0212] Let the resource allocation level loss be Then its gradient with respect to the optimized input parameters can be expressed as:

[0213] In actual calculations, it is not necessary to explicitly calculate the inverse of the Hessian matrix. The vector-Hessian inverse product can be completed by solving a system of linear equations, thereby reducing computational overhead and improving numerical stability.

[0214] The utility opportunity parameter enters the objective function through a linear term, the risk penalty parameter and the associated risk matrix enter the objective function through a quadratic term, and the resource allocation upper bound enters the surrogate penalty term through the total resource constraint. Therefore, the resource allocation level gradient can be passed to:

[0215] Specifically, the gradient corresponding to the utility opportunity parameter is further passed to the multi-object utility opportunity parameter learning module; the gradient corresponding to the association risk matrix is ​​further passed to the association risk matrix generation module; and the gradients corresponding to the risk penalty parameter and the resource allocation upper limit are further passed to the global state control parameter generation module.

[0216] The separation of the original quadratic programming forward solution and the smooth surrogate backward differentiation has the following advantages: the original quadratic programming forward solution ensures that the actual output resource allocation vector corresponds completely with the original objective function and the original resource constraints; the smooth surrogate backward differentiation reduces gradient discontinuities caused by activity constraint switching and avoids expanding the entire iteration process of the quadratic programming solver; through the resource allocation level gradient, the generated parameters can be jointly adjusted according to the final resource allocation objective.

[0217] In addition, the training process for generating shared control parameters, utility opportunity parameters, and associated risk matrices is also included. First, training samples are constructed based on historical global state characteristics, historical operational result data, historical object time-series state data, and historical object relationship data. Then, training objectives are constructed based on unfavorable operational results in the historical operational result data, corresponding to the system's conservative control degree, the relative resource allocation value of objects, and the common unfavorable operational relationships of objects. The shared control parameter generation process is trained based on historical global state characteristics and historical operational result data, ensuring that the output shared control parameters characterize the current system's conservative control degree. Risk penalty parameters and resource allocation limits are generated based on the shared control parameters. Finally, the utility opportunity parameter generation process is trained based on historical object time-series state data, historical object relationship data, and corresponding historical global state characteristics, ensuring that the output utility opportunity parameters... The system represents the relative resource allocation value among different decision-making objects. Based on historical operational data, historical object relationship data, and corresponding historical global state characteristics, the system trains the associated risk parameter generation process, so that the output associated risk matrix represents the degree of correlation between different decision-making objects in the occurrence of historically unfavorable operational results. After independently training the generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices, the system connects these processes with the solution process of the quadratic programming problem. Based on the operational results corresponding to the resource allocation vector obtained from the quadratic programming problem, a resource allocation level loss is constructed. The generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices are jointly adjusted based on the resource allocation level loss, so that the shared control parameters, utility opportunity parameters, and associated risk matrices jointly adapt to the constrained resource allocation process.

[0218] In this embodiment, the present invention adopts a two-stage training mechanism that combines parameter role pre-training and decision-oriented joint fine-tuning, so that the model generation parameters output by the global state control parameter generation module, the multi-object utility opportunity parameter learning module, and the associated risk matrix generation module can maintain their respective preset functions while adapting to the final resource allocation goal.

[0219] When end-to-end training is performed solely based on the final resource allocation result, different models may generate parameters that correspond to the same or approximately the same resource allocation result. For example, simultaneously scaling the utility opportunity parameter and the strength of the quadratic term, or changing the utility opportunity parameter corresponding to a decision object that did not receive a resource allocation, may not significantly alter the current resource allocation vector. In this case, relying solely on the resource allocation level loss is insufficient to uniquely determine the specific role that each model-generated parameter should represent, potentially causing the utility opportunity parameter, shared control parameter, and associated risk matrix to drift during training.

[0220] To address the aforementioned issues, this embodiment employs a first-stage pre-training method that utilizes supervisory objectives corresponding to the preset roles of parameters generated by each model. This method establishes correspondences between shared control parameters, utility opportunity parameters, and associated risk matrices, and between these parameters and the degree of conservative control, the relative resource allocation value of the object, and the common adverse operational relationships. In the second stage, the three parameter generation modules are connected to the constraint resource allocation solution module. Based on the final resource allocation results, a decision-oriented joint fine-tuning is performed, while retaining the parameter role supervisory loss from the first stage.

[0221] 1) Training sample construction: The training samples are constructed according to the decision-making cycle. For the first... The training samples, and the input data include data up to the [number]th [number]. The observable object time-series state data, global state characteristics, historical operation result data, object relationship data, effective object mask, and resource constraint input data for each decision cycle.

[0222] The training samples can be represented as:

[0223] in, Indicates as of the date Resource allocation input data for each decision cycle Indicates object-level future utility label, This represents a matrix indicating a common unfavorable operational target in the future. This indicates a label indicating potential adverse future performance at the system level.

[0224] Object-level future utility labels are used to train the multi-object utility opportunity parameter learning module; the future common adverse run objective matrix is ​​used to train the associated risk matrix generation module; and system-level future adverse run labels are used to train the global state control parameter generation module. These future labels are only used to construct the supervised loss during the training phase and are not used as model input during model inference and resource allocation execution.

[0225] To prevent future information from entering the model input, construct the first When training the nth sample, only the nth sample will be used. Using the data from the first decision cycle and previous periods as input, the data from the first decision cycle will be used as input. The results of the operation within a decision cycle or a preset future supervision window are used as supervision labels. The training set, validation set, and test set are divided in chronological order, and samples from different time intervals are not randomly crossed to avoid information from later time intervals from entering the previous model training process.

[0226] 2) Training of the first-stage global state control parameter generation module: The first phase involves training a global state control parameter generation module. This module generates shared control parameters based on the global state characteristics and relational risk summary up to the current decision period. Risk penalty parameters are generated from shared control parameters. and resource allocation limit .

[0227] To enable shared control parameters to characterize the future degree of conservative control in the system, a system-level future adverse operation label is constructed based on the adverse operating results of multiple decision objects in the future. Let the future monitoring scope be... The decision cycle, the first The vector of adverse operating results corresponding to each future decision cycle is: Then the system-level future adverse operational intensity can be expressed as:

[0228] in, Represents the weight of future time, and satisfies:

[0229] The system-level future adverse operation intensity comprehensively reflects the degree of adverse operation of multiple decision-making objects within the future monitoring scope. When future objects experience timeouts, execution failures, insufficient throughput, or increased service level violations... Increase.

[0230] By unifying the scale of training samples and applying bounded mapping to the system-level future adverse operational intensity, shared control parameter supervision labels are obtained:

[0231] in, and Let represent the mean and standard deviation of the system-level future adverse operational intensity in the training data, respectively. Represents the numerical stability constant. This represents the Sigmoid activation function.

[0232] The parameter role loss of the global state control parameter generation module can be expressed as:

[0233] In other implementations, binary cross-entropy loss, quantile loss, rank correlation loss, or other loss functions that can monitor the degree of conservative control of the system may be used.

[0234] By minimizing the loss of the aforementioned parameter roles, the shared control parameters are made to take larger values ​​when the future system is operating under adverse conditions, and smaller values ​​when the future system is operating in a relatively stable state. Since the risk penalty parameter and the resource allocation limit are generated from the shared control parameters through monotonically increasing and monotonically decreasing mappings, respectively, the role supervision of the shared control parameters simultaneously establishes a correspondence between the intensity of risk penalty and the scale of resource investment and the degree of conservative control of the system.

[0235] After the global state control parameter generation module reaches the preset training termination condition, its model parameters are saved, and these model parameters are fixed when training the multi-object utility opportunity parameter learning module. The preset training termination condition may include the validation set loss no longer decreasing within a preset number of rounds, the number of training rounds reaching a preset upper limit, or the model performance reaching a preset threshold.

[0236] 3) Training of the first-stage multi-object utility chance parameter learning module: After fixing the model parameters of the global state control parameter generation module, use this module to generate the global state representation. The global state representation, along with object temporal state data, object relationship data, and effective object mask, is then input into the multi-object utility opportunity parameter learning module.

[0237] The multi-object utility opportunity parameter learning module outputs the raw utility opportunity score. Utility opportunity parameter Relationship between objects The role loss is constructed based on the object-level future utility labels in the training samples, using utility opportunity parameters.

[0238] In one implementation, cross-sectional correlation loss is used:

[0239] in, This represents the original utility opportunity score corresponding to the currently valid object. This indicates the future utility label corresponding to the currently valid object.

[0240] In another implementation, an amplitude-weighted pairwise ranking loss can be used. For any two decision objects in the current set of valid objects... and Define the difference in future utility as The corresponding difference in original utility opportunity scores is .

[0241] The magnitude-weighted pairwise ranking loss can be expressed as: ; Among them, objects with greater differences in future utility have higher loss weights, thus enabling the model to prioritize and correctly distinguish objects with more obvious differences in resource allocation value.

[0242] The first-stage training loss for the multi-object utility chance parameter learning module can be cross-sectional correlation loss, magnitude-weighted pairwise ranking loss, or a weighted combination of the two: ; in, and This represents the loss weight.

[0243] Through the above training, the utility opportunity parameter is able to characterize the relative resource allocation value among multiple decision objects within the same decision-making period, rather than simply fitting the absolute value of the object's future utility. After training is complete, the model parameters of the multi-object utility opportunity parameter learning module are saved, and the model parameters of the global state control parameter generation module and the multi-object utility opportunity parameter learning module are fixed when training the association risk matrix generation module.

[0244] 4) Training of the first-stage correlation risk matrix generation module: After fixing the aforementioned two parameter generation modules, the association risk matrix generation module is trained based on the global state representation output by the global state control parameter generation module and the object relationship representation output by the multi-object utility opportunity parameter learning module, combined with historical adverse operating results.

[0245] The correlation risk matrix generation module generates a correlation risk matrix based on historical adverse operating results, object relationship representation, and global state representation. Construct a future common adverse operation target matrix based on the pre-set adverse operation results of multiple decision-making objects within the scope of future supervision. .

[0246] The parameter role loss of the association risk matrix generation module is represented as follows: ; in, The effective object pair mask matrix is ​​represented by the following elements: ; In the above formula, Describing the Frobenius norm, This represents the sum of the absolute values ​​of the matrix elements.

[0247] The loss is used to enable the diagonal elements of the associated risk matrix to characterize the adverse operating risk of each decision object in the future, and the off-diagonal elements to characterize the intensity of the adverse operating results that may occur together in the future for different decision objects.

[0248] The three parameter generation modules in the first phase are trained sequentially in the order of global state control parameter generation module, multi-object utility opportunity parameter learning module, and association risk matrix generation module. This order is determined by the data dependencies between the modules: the multi-object utility opportunity parameter learning module needs to use the global state representation output by the global state control parameter generation module; the association risk matrix generation module needs to use both the global state representation and the object relationship representation output by the multi-object utility opportunity parameter learning module.

[0249] Using a sequential training approach can prevent downstream modules from being trained before the upstream representation has established a stable preset role, thereby reducing the interference caused by the simultaneous random changes of multiple modules on parameter role learning.

[0250] 5) Second-stage model connection and resource allocation level loss construction: After completing the first stage of parameter role pre-training, the model parameters of the three parameter generation modules are loaded and connected to the constraint resource allocation solution module to form a complete processing link from multi-source resource allocation input data to resource allocation vector.

[0251] For the For each training sample, the global state control parameters are first generated, the multi-object utility opportunity parameters are learned, and the association risk matrix is ​​generated to obtain the risk penalty parameters. Resource allocation limit Utility opportunity parameter and associated risk matrix Then, by solving the original quadratic programming problem, the resource allocation vector is obtained. .

[0252] To adapt the model's generated parameters to the final resource allocation target, a resource allocation-level loss is constructed based on the resource allocation vector and the future object's running results.

[0253] Let the actual utility vector of objects within the next decision-making cycle or the pre-defined future supervision scope be: ; The utility resulting from resource allocation can be expressed as: ; To ensure that resource allocation results mitigate future common adverse operational risks while pursuing object utility, an adverse risk item can be constructed based on the future common adverse operational objective matrix: ; In one implementation, the resource allocation level loss is expressed as: ; in, This represents the loss weight corresponding to the common adverse operating risks in the future.

[0254] When the primary goal of resource allocation is to improve the future utility of an object, it can be set as follows: Resource allocation level losses are constructed solely based on realized utility. When it is necessary to consider both realized utility and future jointly adverse operating outcomes, a setting can be made. .

[0255] In other implementations, resource allocation level losses can also be constructed based on task completion rate, system throughput, average response latency, service level default rate, energy consumption, or other resource system performance indicators. The specific form of resource allocation level losses can be determined based on the technical objectives of the resource system, but they all transmit the final decision feedback to the three parameter generation modules through the resource allocation vector.

[0256] 6) Second-stage parameter role loss preservation and joint fine-tuning: Using only resource allocation level loss for the second stage of training may allow different parameter generation modules to achieve the same or similar resource allocation results through mutual compensation. For example, the overall increase in utility opportunity parameter and the simultaneous increase in risk penalty parameter may partially offset each other; when the resource allocation upper limit is not an activity constraint in some samples, its value may change without immediately altering the resource allocation vector; the utility opportunity parameter corresponding to an object that does not receive resource allocation may also drift without affecting the current decision.

[0257] Therefore, this embodiment continues to calculate and retain the parameter role loss from the first stage in the second stage. The joint training loss in the second stage is expressed as: ; in, Indicates resource allocation level loss. This indicates the loss of the shared control parameter role. This represents the role loss, a parameter representing the utility opportunity. This indicates the loss of a role in the associated risk matrix; , and These represent the non-negative weights corresponding to the loss of each parameter role.

[0258] For a training batch, the joint training loss can be expressed as: ; in, This represents the set of samples included in the current training batch. This indicates the training batch size.

[0259] During the second stage of backpropagation, the resource allocation level loss is calculated using a smoothed surrogate back-differentiation mechanism, with respect to the utility opportunity parameter, the association risk matrix, the risk penalty parameter, and the resource allocation upper limit, and these gradients are then passed to the corresponding parameter generation modules. The parameter role loss directly provides supervisory gradients to each parameter generation module.

[0260] Let the trainable parameters of the global state control parameter generation module, the multi-object utility opportunity parameter learning module, and the association risk matrix generation module be respectively... , and Phase Two Implementation: ; in, This indicates the step size for updating model parameters. In actual training, an adaptive gradient optimizer can be used, with settings for learning rate decay, weight decay, gradient clipping, and early stopping conditions.

[0261] Resource allocation level loss causes the three parameter generation modules to coordinate and adjust around the final resource allocation goal; parameter role loss limits each module from deviating from its preset role. Therefore, the second stage does not completely replace the parameter supervision of the first stage, but rather conducts joint fine-tuning based on the parameter roles established in the first stage, guided by decision-making.

[0262] After obtaining the resource allocation vector, the retention ratio and absolute amount of retained resources corresponding to the current decision cycle are calculated based on the resource allocation vector. According to the resource allocation vector and the resource granularity corresponding to the current resource management system, the resource allocation ratio for each decision object is converted into the corresponding actual resource allocation amount. When resources are allocated using continuous resource ratios, corresponding resource quotas are generated based on the resource allocation ratios. When resources are allocated using discrete resource units, priority sorting, integer mapping, or discrete optimization are performed based on the resource allocation ratios, and rounding errors are corrected. Resource control instructions are generated based on the actual resource allocation amount and provided to the corresponding resource execution unit. After the current decision cycle is completed, object status data, global status data, and actual operating results for each decision object are collected and added to the historical data sequence. Constrained resource allocation is then re-executed in the next decision cycle.

[0263] In this embodiment, after completing the two-stage training, the model parameters of the global state control parameter generation module, the multi-object utility opportunity parameter learning module, and the associated risk matrix generation module are fixed, and rolling resource allocation is performed according to the preset decision cycle.

[0264] 1) Data boundaries during the inference phase: In the When performing model inference in the decision cycle, only the first decision cycle is used. The data observable for each decision cycle and prior to that cycle. Object-level future utility labels, future common adverse run target matrices, and system-level future adverse run labels are used only for model training and do not enter the model inference process; The input for the inference phase is represented as The end point of each historical window shall not be later than the current decision-making cycle. This processing method avoids the model using unobservable future data to generate current resource allocation results.

[0265] 2) Reasoning process in a single decision cycle: In the Within each decision cycle, the first step is to construct multi-source resource allocation data. Based on the current time index, the object time-series status window, global status window, and historical running result window are extracted. Time alignment, scale unification, and effective object filtering are then performed on various types of data. Subsequently, the global state control parameter generation is performed. The global state characteristics are input into the global state control parameter generation module to obtain the global state representation. Construct a risk summary of relationships based on historical adverse operating results. The two are then combined to obtain shared control parameters. Risk penalty parameters are generated according to the monotonic mapping. and resource allocation limit ; Then, multi-object utility opportunity parameter learning is performed. Object temporal state data, object relation data, effective object mask, and global state representation are input into the multi-object utility opportunity parameter learning module to obtain object relation representation. And utility opportunity parameters ; Further, the association risk matrix generation is performed by inputting historical adverse operating results, object relationship representations, and global state representations into the association risk matrix generation module to obtain empirical association risk matrices and neural association risk matrices. These are then processed through trace normalization, weighted fusion, and diagonal stabilization to obtain the final association risk matrix. ; Finally, the constraint resource allocation problem is solved. Based on the utility opportunity parameter, the association risk matrix, the risk penalty parameter, and the resource allocation upper limit, the original quadratic programming problem is constructed, and the quadratic programming solver is called to obtain the resource allocation vector. and the proportion of reserved resources ; During the inference phase, there is no need to perform back-derivative calculation of the smooth agent, nor is it necessary to calculate the parameter role supervision loss and resource allocation level loss. The smooth agent and various supervision labels are only used during the model training phase.

[0266] 3) Execution of resource allocation results: The resource allocation output module converts the normalized resource allocation vector into control instructions that the resource system can execute. Let the total amount of currently schedulable resources be... Then the first The actual resource allocation for each decision-making object is: The absolute amount of reserved resources is expressed as... ; The resource allocation output module can discretize the actual resource allocation amount according to the resource granularity. For example, when resources can only be allocated in fixed units, the actual resource amount can be rounded according to a preset discrete unit, and the rounding error can be corrected without exceeding the total resource capacity. When the execution system supports the proportion of continuous resources, it can be directly based on... Adjust object resource quotas. When the execution system uses discrete resource units, priority sorting, integer mapping, or discrete optimization post-processing can be performed according to the resource allocation ratio. The post-processing must not cause the total amount of actually allocated resources to exceed the currently available resource capacity.

[0267] 4) Rolling update process: In the After resource allocation is completed in each decision cycle, the system collects the object status data, global status data, and actual operating results of each decision object within that cycle, and adds this data to the historical data sequence. (Entering the next...) At each decision cycle, the historical window is scrolled forward one decision cycle, the oldest data outside the historical window range is deleted, and new observation data is added. The updated object time series status window is represented as follows: The global status window and the historical execution results window are updated in the same way. The process is re-executed based on new input data to obtain the resource allocation vector for the next decision cycle.

[0268] Therefore, this invention does not calculate a long-term resource allocation scheme based on fixed parameters in a one-time calculation. Instead, it regenerates optimization parameters and resolves the resource allocation problem in each decision cycle based on the latest object state, global state, and historical running results, so that the resource allocation results are updated as the system running state changes.

[0269] In summary, such as Figure 2 As shown, Figure 2 The upper part illustrates the forward processing flow of resource allocation within a single decision-making cycle according to this invention. Figure 2The lower part illustrates the two-stage training mechanism corresponding to the aforementioned forward processing flow. In the first stage, based on the data dependencies between the global state control parameter generation module, the multi-object utility opportunity parameter learning module, and the associated risk matrix generation module, parameter role pre-training is performed on S2, S3, and S4 sequentially. In the second stage, S2, S3, and S4 are connected to the constraint resource allocation solution in S5. A resource allocation-level loss is constructed based on the resource allocation results output by S5, and the gradient of the resource allocation level is backpropagated to S2, S3, and S4 through the constraint resource allocation solution process, while retaining the parameter role supervision loss corresponding to each parameter generation module. Therefore, Figure 2 The two-stage training mechanism shown in the lower part and Figure 2 The resource allocation processing flow shown above corresponds to each other, where solid lines represent forward data processing relationships and dashed lines represent training and gradient propagation relationships. This embodiment of the invention acquires object temporal states, global states, historical running results, object relationships, and resource constraint data; generates utility opportunity parameters through multi-scale temporal coding, static-dynamic relationship learning, and state gating; generates risk penalty parameters and a total resource allocation upper limit based on the global state and relationship risk summary, and fuses the empirical risk matrix and the neural risk matrix to obtain a positive semidefinite association risk matrix; constructs and solves a quadratic programming problem with non-negative constraints and total resource constraints based on the above parameters to obtain the resource allocation vector and reserved resources; and employs a two-stage training mechanism to maintain the preset role of the model-generated parameters, improving the synergy and stability of utility opportunity identification, association risk control, and resource constraint execution.

[0270] Therefore, the embodiments of the present invention represent relative utility through multi-scale state encoding and static-dynamic relationship learning, synchronously shrink the upper limit of resources when risk penalty is enhanced by sharing control parameters, construct a stable associated risk matrix through experience-neural risk fusion and structure, and take into account constraint executability, gradient stability and parameter effect preservation through forward solving of the original problem, smooth surrogate back-derivative and two-stage training.

[0271] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this is not intended to limit the scope of the embodiments of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the embodiments of the present application.

Claims

1. A constrained resource allocation method based on multi-source temporal state perception and associated risk modeling, characterized in that, Includes the following steps: The feature data of multiple decision objects within a preset historical window are obtained, and the feature data is processed according to the current decision cycle to obtain resource allocation input data. The feature data includes object time-series state data, global state features, historical operation result data, object relationship data, and resource constraint input data. Global state modeling is performed based on the global state characteristics and historical operation results data in the resource allocation input data to obtain a global state representation and a relationship risk summary. The global state representation and the relationship risk summary are then fused to obtain shared control parameters. Risk penalty parameters and resource allocation upper limits are generated based on the shared control parameters. The shared control parameters are used to characterize the degree of conservative control corresponding to the current system. When the shared control parameters increase, the risk penalty parameters increase and the resource allocation upper limit decreases. Multi-object state and relationship modeling is performed based on object temporal state data, object relationship data, and global state representation to obtain object temporal representation and object relationship representation, and utility opportunity parameters are generated based on object temporal representation and object relationship representation; Historical adverse operating results are extracted from historical operating results data, and an associated risk matrix is ​​generated based on historical adverse operating results, object relationship representation, and global state representation; Based on the utility opportunity parameter, the associated risk matrix, the risk penalty parameter, the resource allocation upper limit, and the resource constraint input data, a constrained quadratic programming problem is constructed and solved to obtain the resource allocation vector corresponding to the current decision period. Based on the resource allocation vector, the resource allocation result and reserved resources corresponding to each decision object are determined. The step of generating utility opportunity parameters includes: performing multi-scale trend-fluctuation encoding on object time-series state data to obtain object time-series representations; constructing static-dynamic object relationships based on static object relationships and object time-series representations; performing cross-object attention processing in conjunction with effective object masks and controlling the degree of injection of object relationship information based on global state representations to obtain object relationship representations; and performing cross-sectional standardization and bounded mapping on the fusion results to obtain utility opportunity parameters, which are used to describe the relative resource allocation value among various decision objects. The step of extracting historical adverse operating results from historical operating result data and generating an association risk matrix based on historical adverse operating results, object relationship representation, and global state representation specifically includes: extracting historical adverse operating results based on the one-sided adverse deviation between the historical operating results of each decision object and the corresponding operating objective; constructing an empirical association risk matrix based on historical adverse operating results and non-negative time weights; generating potential risk factor loadings and positive idiosyncratic risk parameters based on historical adverse operating results, object relationship representation, and global state representation, and constructing a neural association risk matrix based on potential risk factor loadings and positive idiosyncratic risk parameters; performing trace normalization on the empirical association risk matrix and the neural association risk matrix respectively, and then performing weighted fusion on the two normalized association risk matrices; adding a diagonal stable term to the weighted fused association risk matrix and performing normalization to obtain the association risk matrix.

2. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 1, characterized in that, The step of acquiring feature data of multiple decision objects within a preset historical window and processing the feature data according to the current decision cycle to obtain resource allocation input data specifically includes: Based on the current decision cycle, determine the object's time-series status window, global status window, and historical execution result window, and extract the object's time-series status data, global status characteristics, and historical execution result data from the corresponding data sources; Data from different data sources are time-aligned according to a unified time base, and the continuous numerical features after time alignment are scaled uniformly. Based on the data integrity, operational effectiveness, and resource allocability status of each decision object, a valid object mask is generated, and the valid decision objects participating in the resource allocation of the current decision cycle are determined based on the valid object mask; Generate object relationship data based on at least one of object category relationships, topological relationships, dependency relationships, or invocation relationships; Generate resource constraint input data based on at least one of the following: non-negativity constraint, total resource constraint, single-object resource constraint, and numerical stability parameter.

3. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 2, characterized in that, The step of performing global state modeling based on global state features and historical operation result data in resource allocation input data to obtain a global state representation and a relationship risk summary, fusing the global state representation and the relationship risk summary to obtain shared control parameters, and generating risk penalty parameters and resource allocation upper limits based on the shared control parameters specifically includes: The global state features in the resource allocation input data are temporally encoded according to multiple preset time scales to obtain global state representations corresponding to multiple time scales. The global state representations corresponding to different time scales are fused, and the global state representation is obtained through attention pooling. Historical adverse operating results are extracted from historical operating data, and a relationship risk summary is generated based on the historical adverse operating relationships between multiple decision-making objects. By fusing the global state representation and the relational risk summary, shared control parameters are obtained; Risk penalty parameters and resource allocation limits are generated based on shared control parameters, wherein when the shared control parameters increase, the risk penalty parameters increase and the resource allocation limits decrease.

4. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 3, characterized in that, The step of constructing the temporal representation of the object specifically includes: The time-series state data of each decision object are downsampled according to multiple preset time scales; Trend and fluctuation components are extracted from the time-series state data of objects at different time scales to obtain trend and fluctuation information at the corresponding time scales. By fusing trend and fluctuation information at different time scales across scales, a multi-scale object state representation corresponding to each decision object is obtained. Feature mapping is performed on the multi-scale object state representation to obtain the object temporal representation.

5. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 4, characterized in that, The step of constructing the object relation representation specifically includes: Static object relationships are generated based on object relationship data, and dynamic object relationships are generated based on the similarity between the temporal representations of each decision object. By fusing static and dynamic object relationships, object relationship data is obtained. Perform cross-object attention processing based on object temporal representations and object relationship data to obtain initial object relationship representations; State gating parameters are generated based on the global state representation, and the degree of injection of the initial object relation representation into the object temporal representation is adjusted based on the state gating parameters. The object-relational representation is obtained by residual fusion of the gated object-relational representation and the object-temporal representation.

6. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 5, characterized in that, The step of generating utility opportunity parameters based on object temporal representation and object relational representation specifically includes: Residual fusion is performed based on object temporal representation and object relational representation to obtain utility opportunity representation; The original utility opportunity score is generated based on the utility opportunity representation and global state representation corresponding to each decision object. Cross-sectional standardization of the original utility opportunity scores within the current set of effective decision objects; Bounded mapping is performed on the original utility opportunity scores after cross-sectional standardization to obtain the utility opportunity parameters corresponding to each decision object, and the utility opportunity parameters corresponding to ineffective decision objects are set to zero.

7. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 6, characterized in that, The step of constructing and solving a constrained quadratic programming problem based on the utility opportunity parameter, the association risk matrix, the risk penalty parameter, the resource allocation ceiling, and the resource constraint input data specifically includes: A resource allocation vector is constructed based on the resource allocation ratio corresponding to each decision object, and the resource allocation vector is used as the decision variable for the quadratic programming problem. A quadratic term is constructed based on the correlation risk matrix, risk penalty parameter, and numerical stability parameter, and a linear term is constructed based on the utility opportunity parameter; Construct non-negativity constraints, total resource constraints, and single-object resource constraints corresponding to the resource allocation vector based on the resource constraint input data; By inputting the quadratic term, the linear term, the nonnegativity constraint, the total resource constraint, and the single-object resource constraint into the quadratic programming solver, the optimal resource allocation vector and the corresponding constraint multipliers are obtained.

8. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 7, characterized in that, It also includes training the parameter generation process for shared control parameters, utility opportunity parameters, and association risk matrices, specifically including: Training samples are constructed based on historical global state characteristics, historical operation result data, historical object time-series state data, and historical object relationship data. Training objectives are constructed based on unfavorable operation results in historical operation result data, corresponding to the system's degree of conservative control, the relative resource allocation value of objects, and the common unfavorable operation relationships of objects. The shared control parameter generation process is trained based on historical global state characteristics and historical operation result data, so that the output shared control parameters can characterize the degree of conservative control of the current system, and risk penalty parameters and resource allocation upper limits are generated based on the shared control parameters. The utility opportunity parameter generation process is trained based on historical object time-series state data, historical object relationship data, and corresponding historical global state features, so that the output utility opportunity parameter represents the relative resource allocation value between different decision objects. The process of generating associated risk parameters is trained based on historical operational results data, historical object relationship data, and corresponding historical global state characteristics, so that the output associated risk matrix represents the degree of association between different decision-making objects that jointly experience adverse historical operational results. After independently training the generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices, the generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices are connected with the solution process of the quadratic programming problem. Based on the running results corresponding to the resource allocation vector obtained from the quadratic programming problem, a resource allocation level loss is constructed. The generation processes of shared control parameters, utility opportunity parameters, and associated risk matrices are jointly adjusted based on the resource allocation level loss, so that the shared control parameters, utility opportunity parameters, and associated risk matrices can jointly adapt to the constrained resource allocation process. The joint adjustment includes the following steps: adopting a separate processing method of forward solving of the original quadratic programming and backward differentiation of the smoothing surrogate; in the forward calculation process, the original constraint problem is still solved directly by the quadratic programming solver, and the actual output always adopts the optimal solution of the original problem; in the backward propagation process, based on the optimal solution of the original problem and the corresponding constraint multipliers, a smoothing penalty surrogate problem is constructed, and implicit differentiation is performed through the surrogate problem.

9. The constrained resource allocation method based on multi-source temporal state perception and associated risk modeling according to claim 8, characterized in that, After obtaining the resource allocation vector, the following is also included: Calculate the proportion and absolute amount of retained resources for the current decision-making cycle based on the resource allocation vector; Based on the resource allocation vector and the resource granularity corresponding to the current resource management system, the resource allocation ratio of each decision object is converted into the corresponding actual resource allocation amount; When resources are allocated using a continuous resource ratio, a corresponding resource quota is generated based on the resource allocation ratio. When resources are allocated using discrete resource units, priority sorting, integer mapping, or discrete optimization are performed according to the resource allocation ratio, and rounding errors are corrected. Resource control instructions are generated based on the actual resource allocation, and then provided to the corresponding resource execution units. After the current decision cycle is completed, collect the object status data, global status data and actual operation results of each decision object, and add them to the historical data sequence. Then, re-execute the constraint resource allocation in the next decision cycle.

Citation Information

Patent Citations

  • Deep learning agent collaborative decision-making method and system for operator risk control

    CN122549929A

  • Dynamic scheduling method and system based on double-domain trust optimization

    CN122593201A