Machine learning based computer resource intelligent allocation method and terminal

Abstract

The application discloses a computer resource intelligent allocation method and terminal based on machine learning, comprising the following steps: defining a resource pool and a service object, establishing a multi-objective optimization framework and constraint, defining a key performance indicator, collecting original data in multiple aspects and preprocessing to construct a feature set, training and verifying a machine learning model, building a man-machine collaborative decision mechanism, generating a multi-level flexible plan and dynamically binding system indicators, monitoring deviations and collecting feedback, and realizing system full-link iterative optimization. The application provides high-quality input for the model through multi-source data collection and feature construction, combines man-machine collaboration to ensure efficient and controllable decision-making, relies on flexible plans to improve system flexibility, adapts to business changes through closed-loop iteration, effectively improves resource utilization and reduces energy cost, ensures service stability, and enhances system reliability and feasibility.

Description

Technical Field

[0001] This invention relates to the field of intelligent resource allocation technology, specifically to a computer resource intelligent allocation method and terminal based on machine learning. Background Technology

[0002] With the rapid development of cloud computing, edge computing, big data, and artificial intelligence technologies, modern computing infrastructure is supporting increasingly diverse business types and more complex load characteristics. How to efficiently and intelligently allocate and manage computer resources (including computing, storage, network, and energy consumption) has become a core challenge for ensuring service quality, reducing operating costs, and improving overall system performance.

[0003] Traditional computer resource allocation methods primarily rely on static configuration, simple rules based on predefined thresholds, or manual scheduling by operations personnel. Static configuration typically reserves fixed resources based on estimated peak loads, easily leading to low resource utilization and waste. Threshold-based automated scaling (such as setting CPU utilization thresholds to trigger expansion) improves response speed to some extent, but it is still essentially a reactive mechanism that cannot predict future resource demands. Furthermore, fixed thresholds are difficult to adapt to dynamic and non-linear load changes, easily leading to over-allocation of resources or untimely responses in complex and ever-changing business scenarios. While relying entirely on manual scheduling offers high flexibility, it is slow, subjective, and difficult to achieve fine-grained and real-time resource optimization in large-scale, dynamic environments.

[0004] In recent years, machine learning technology has demonstrated enormous potential in the field of resource management due to its powerful data analysis and pattern recognition capabilities. Researchers have begun to explore using time series forecasting models (such as ARIMA and LSTM) to predict future resource demands, or applying reinforcement learning algorithms to learn resource scheduling strategies in simulation environments. These studies have initially verified the feasibility of data-driven methods in improving resource utilization and service quality. However, most existing methods focus on improving the model algorithms themselves, usually embedding them as isolated technical modules into existing systems. They fail to form a complete and systematic engineering framework covering goal definition, data governance, model building, decision execution, effect evaluation, and continuous optimization. In actual production systems, pure machine learning model-driven solutions still face many challenges: the model decision-making process lacks interpretability, making it difficult for operations and maintenance personnel to understand and trust; the model struggles to cope with sudden business patterns or special rule requirements not present in the training data; and there is a lack of effective human-machine collaboration mechanisms, making it impossible to incorporate the experience and domain knowledge of human experts into critical decisions. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based intelligent allocation method and terminal for computer resources to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based intelligent allocation method for computer resources, comprising the following steps: S1. Clearly define the scope of the resource pool, the service targets, and the dimensions of allocable resources; S2. Establish a multi-objective optimization framework and system constraints that balance resource utilization and energy efficiency costs based on service level agreements; S3. Define and set key performance indicators for quantitative evaluation of the achievement of optimization goals; S4. The system collects raw data from multiple levels, including infrastructure, task applications, scheduling and orchestration, and business operations. S5. Clean, align, and standardize the collected multi-source raw data for preprocessing, and construct a feature set containing statistical, temporal, correlation, and semantic information. S6. For prediction, identification, and decision-making tasks, select and train the corresponding machine learning models, and ensure the effectiveness and reliability of the models through time-series verification, loss optimization, and interpretability processing. S7. Construct a model-based automated decision-making review, intervention, and rule injection mechanism to achieve transparent, controllable, and continuously optimized decision-making through interpretable output, multi-solution comparison, and human feedback closed loop. S8. Based on current decisions and future predictions, generate multi-level flexible contingency plans and dynamically bind them to system status indicators to achieve contingency plan early warning, condition triggering and smooth switching. S9. During the execution and switching of the contingency plan, continuously monitor the deviation between the actual effect of the contingency plan and the expected goal, and collect the deviation data as a high-order feedback signal. S10. Through online performance comparison, concept drift monitoring, feedback data-driven model retraining, and rule base maintenance, the system can achieve continuous evaluation, optimization, and full-link iteration.

[0007] Preferably, the resource pool ranges from a single cluster to a hybrid cloud across data centers; The service targets include Web services, batch processing jobs, and AI training tasks; The resource dimensions include vCPU core count, memory capacity, GPU count and video memory, network bandwidth limit, disk type, and IOPS quota; S2 specifically includes the following steps: S21. Establish the core optimization objective of maximizing overall resource utilization while satisfying all service level agreements, and set minimizing energy consumption and reducing operating costs as secondary optimization objectives. S22. Define the hard constraints that the system must follow to operate, including the maximum resource requirement limit for a single task, the total capacity limit of physical nodes, and the overall budget limit; S23. Integrate the above objectives and constraints to construct a quantifiable multi-objective optimization problem model.

[0008] Preferably, step S3 specifically includes the following steps: S31. Set average resource utilization rate as a core efficiency indicator; set service level agreement violation rate as a core quality indicator. S32. Set task completion time as the core performance indicator; set energy consumption cost per unit throughput as the core economic indicator.

[0009] Preferably, step S4 specifically includes the following steps: S41. Collect underlying indicators reflecting the status of physical resources from the infrastructure layer. The underlying indicators include continuous collection of CPU utilization, memory usage, I / O throughput, network traffic and energy consumption data of the host machine and physical machine. S42. Collect real-time runtime data of workload instances from the task application layer, including key performance indicators such as resource usage, response time, throughput, and error rate of each instance, as well as application logs. S43. Collect metadata of resource scheduling and task management from the scheduling and orchestration layer, including the waiting status of the global task queue, historical scheduling decision records, task priority tags, and inter-task dependencies. S44. Collect data reflecting external demands and business characteristics from the business layer, including collecting known business characteristics such as external request traffic, user session status, peak business periods, and promotional activity periods.

[0010] Preferably, step S5 specifically includes the following steps: S51. Perform data cleaning operations, including imputation or labeling of missing values, smoothing and filtering of data noise, and identifying and removing outliers. S52. Perform data alignment operation to unify time series data from different acquisition sources and with different acquisition frequencies to the same time base and sampling frequency through timestamp alignment and resampling technology. S53. Perform data standardization or normalization operations, and eliminate the differences in units and numerical ranges of different monitoring indicators through linear transformation or scaling methods. S54. Based on the preprocessed time series data, calculate the statistics of mean, variance, maximum, minimum and quantiles on the rolling time window to construct statistical features; S55. Extract trend components, periodic components, autocorrelation coefficients, and time series pattern features from time series data to construct time series features; S56. Calculate the correlation coefficients between resource usage sequences of different workload instances, the correlation between usage of different resource types, and construct correlation features; S57. Extract task tags, container image names, user groups, and project identifier category fields from task orchestration metadata, encode or embed them to represent semantic features.

[0011] Preferably, step S6 specifically includes the following steps: S61. Select and build a suitable machine learning model structure based on the specific resource prediction task, pattern recognition task, or online decision-making task objective. S62. Following the chronological order of the time series, the dataset containing historical information is divided into training set, validation set and test set in chronological order; S63. Define corresponding loss functions for different types of modeling tasks; for prediction tasks, use mean squared error and mean absolute error loss functions; for decision-making tasks, construct custom loss functions that are directly related to the key performance indicators defined in S3. S64. Employ a time-series cross-validation strategy to evaluate the performance of the machine learning model across multiple time segments. S65. For critical decisions that affect system stability and resource allocation security, introduce interpretability techniques to analyze and explain the reasoning process and decision-making basis of machine learning models.

[0012] Preferably, step S7 specifically includes the following steps: S71. When the underlying machine learning model generates each preliminary decision suggestion, it simultaneously outputs the core basis supporting the decision and the confidence assessment of the decision effect. The decision basis extracts and displays the Top N key features that affect the decision and their contribution. S72. Based on the different optimization tendencies established in S2, the system generates 2 to 3 differentiated alternative solutions, each of which is associated with a set of expected results consisting of key performance indicators defined in S3. S73. Design a decision intervention interface that enables operations and maintenance experts to review the decision suggestions and their basis generated by the system, approve or reject decisions, view detailed comparisons of all alternative solutions, and inject temporary or permanent business rules into the system, including automatic priority approval of resource requests for core service A during major promotional periods. S74. Mark the feedback from operations and maintenance experts on whether the decision results are correct or incorrect, and use it as high-quality labeled data for direct use in the subsequent optimization process of machine learning models.

[0013] Preferably, S8 specifically includes the following steps: S81. After the final decision is made in S7, based on historical forecasting models and predefined business scenarios, a set of tiered flexible contingency plans covering multiple possibilities including optimistic, baseline, and pessimistic scenarios is generated. Each level of the plan specifies the steps, resource amounts, and timelines for subsequent expansion, contraction, or recovery. S82. Bind the flexible contingency plan with quantifiable system status indicators, and pre-verify the resource scheduling feasibility of each contingency plan to form a status contingency plan triggering mapping table. The system status indicators are the number of connections, error rate, and predicted load threshold. S83. Continuously evaluate system status indicators through real-time monitoring and short-term forecasting. When the status is detected to be approaching the trigger condition of a certain plan, issue an early warning to the operation and maintenance personnel. S84. When the triggering conditions are met, the plan is automatically switched according to the preset strategy, or after manual confirmation before the critical operation, the plan smoothly and dynamically transitions from the current execution strategy to a flexible plan that adapts to the new state.

[0014] Preferably, S10 specifically includes the following steps: S101. In an online environment, A / B testing and shadow mode are used to compare the effects of the intelligent allocation strategy with the baseline strategy of static allocation and threshold rules. The key performance indicators defined in S3 are used to evaluate the effectiveness of the strategy, and the operational efficiency of the human-machine collaboration layer in S7 is evaluated, including the average time for manual approval and the decision adoption rate. S102. Establish an online monitoring mechanism to continuously track key indicators of the prediction accuracy and decision execution effect of machine learning models, and identify concept drift. S103. Using the human feedback records, decision effectiveness feedback, and feedback data on deviations in the implementation of contingency plans collected from S7 and S9, construct a high-quality labeled dataset to drive the machine learning model to perform online learning or periodic offline retraining. S104. Analyze the usage frequency and effectiveness of business rules injected through S7, automatically clean up invalid or expired rules, and convert high-frequency effective rules into model features or logic. S105. Based on the evaluation results and monitoring feedback, adjustments to the objectives, constraints and indicators of S1-S3 are triggered, the feature set of S5 is updated, the machine learning model of S6 is retrained, and the business rule base of S7 is maintained, forming a closed-loop iterative mechanism.

[0015] A computer resource intelligent allocation terminal includes: a processor, a memory, and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the processor to implement the machine learning-based computer resource intelligent allocation method as described in any one of claims 1 to 9.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0017] Compared with existing technologies, the beneficial effects of this invention are: by establishing a target-data-model-evaluation closed loop, the system no longer relies on fixed thresholds or rules, and can make dynamic predictions and decisions based on multi-dimensional, real-time system and business data using machine learning models. By generating multi-level flexible contingency plans bound to future states, the system has forward-looking optimization capabilities, realizing a fundamental shift from passive response to proactive planning, and from meeting current needs to adapting to future changes, thereby significantly improving resource utilization, service assurance levels, and energy efficiency and cost-effectiveness. By employing a model-driven automatic decision-making suggestion and human-machine collaborative review and intervention model, the system leverages the advantages of machine learning in processing massive amounts of data and identifying complex patterns. At the same time, it effectively embeds the domain knowledge, business judgment, and ultimate control of human experts. Mechanisms such as interpretable output, multi-solution comparison, rule injection, and human feedback loop ensure the transparency and traceability of the decision-making process, as well as the ability to flexibly respond to sudden business needs and special strategies. This enhances the credibility and practical value of the system in complex and critical production environments. By continuously driving model retraining and rule base updates through online performance monitoring, concept drift detection, and multi-source feedback (including decision aftereffects, contingency plan deviations, and human evaluation), and by adjusting early stages such as optimization objectives and feature sets, the system can continuously adapt to changes in business load, technological evolution, and the accumulation of operational knowledge. This prevents models or rules from becoming outdated and ineffective due to environmental changes, thereby maintaining and continuously improving the long-term performance and lifecycle value of the intelligent allocation system. Attached Figure Description

[0018] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1:

[0022] See Figure 1 As shown in the figure, the machine learning-based intelligent allocation method for computer resources according to an embodiment of the present invention includes the following steps: S1. Clearly define the scope of the resource pool, the service targets, and the dimensions of allocable resources; S2. Establish a multi-objective optimization framework and system constraints that balance resource utilization and energy efficiency costs based on service level agreements; S3. Define and set key performance indicators for quantitative evaluation of the achievement of optimization goals; S4. The system collects raw data from multiple levels, including infrastructure, task applications, scheduling and orchestration, and business operations. S5. Clean, align, and standardize the collected multi-source raw data for preprocessing, and construct a feature set containing statistical, temporal, correlation, and semantic information. S6. For prediction, identification, and decision-making tasks, select and train the corresponding machine learning models, and ensure the effectiveness and reliability of the models through time-series verification, loss optimization, and interpretability processing. S7. Construct a model-based automated decision-making review, intervention, and rule injection mechanism to achieve transparent, controllable, and continuously optimized decision-making through interpretable output, multi-solution comparison, and human feedback closed loop. S8. Based on current decisions and future predictions, generate multi-level flexible contingency plans and dynamically bind them to system status indicators to achieve contingency plan early warning, condition triggering and smooth switching. S9. During the execution and switching of the contingency plan, continuously monitor the deviation between the actual effect of the contingency plan and the expected goal, and collect the deviation data as a high-order feedback signal. S10. Through online performance comparison, concept drift monitoring, feedback data-driven model retraining and rule base maintenance, the system can achieve continuous evaluation, optimization and full-link iteration. Through steps S4-S5, data is systematically collected from multiple levels including infrastructure, applications, scheduling, and business. This data undergoes professional cleaning, alignment, standardization, and feature construction, providing machine learning models with comprehensive, high-quality input that reflects dynamic changes in business and systems. This supports the models in making higher-quality predictions and decisions that better meet the needs of real-world scenarios. Steps S6-S7 utilize validated machine learning models to automate core task decisions. Simultaneously, through interpretable output, multi-solution comparison, human review and intervention, and rule injection mechanisms, human experience and judgment are organically combined with the machine's computing power and pattern recognition capabilities. This ensures both the efficiency of decision-making and the transparency, traceability, and ultimate control of complex and critical decisions by human experts, thereby improving the system's reliability and practical feasibility. Example 2:

[0023] The resource pool includes single clusters and hybrid clouds across data centers; The services offered include web services, batch processing jobs, and AI training tasks. Resource dimensions include vCPU core count, memory capacity, GPU count and VRAM, network bandwidth limit, disk type, and IOPS quota; S2 specifically includes the following steps: S21. Establish the core optimization objective of maximizing overall resource utilization while satisfying all service level agreements, and set minimizing energy consumption and reducing operating costs as secondary optimization objectives. S22. Clearly define the hard constraints that the system must follow to operate. These constraints include the maximum resource requirement limit for a single task, the total capacity limit of physical nodes, and the overall budget limit. S23. Integrate the above objectives and constraints to construct a quantifiable multi-objective optimization problem model; The complex resource allocation problem is structured into a clear multi-objective optimization problem model, which clarifies the trade-offs between core and secondary objectives and sets realistic constraints such as resource limits, node capacity, and budget, providing a precise mathematical description and optimization guidance for machine learning models and decision-making logic.

[0024] Specifically, S3 includes the following steps: S31. Set average resource utilization rate as a core efficiency indicator; set service level agreement violation rate as a core quality indicator. S32. Set task completion time as the core performance indicator; set energy consumption cost per unit throughput as the core economic indicator. A core quantitative indicator system for evaluating system performance was defined, which comprehensively measures performance from four dimensions: efficiency, quality, performance, and economy (utilization rate, SLA violation rate, task completion time, and unit energy cost). This ensures the measurability of optimization objectives and the scientific nature of evaluation, and provides a unified and objective standard for the design of model training loss functions and the final evaluation of system performance.

[0025] Specifically, S4 includes the following steps: S41. Collect underlying metrics reflecting the status of physical resources from the infrastructure layer. These underlying metrics include continuous collection of CPU utilization, memory usage, I / O throughput, network traffic, and energy consumption data of the host machine and physical machines. S42. Collect real-time runtime data of workload instances from the task application layer, including key performance indicators such as resource usage, response time, throughput, and error rate of each instance, as well as application logs. S43. Collect metadata of resource scheduling and task management from the scheduling and orchestration layer, including the waiting status of the global task queue, historical scheduling decision records, task priority tags, and inter-task dependencies. S44. Collect data reflecting external demands and business characteristics from the business layer, including collecting known business characteristics such as external request traffic, user session status, peak business periods, and promotional activity periods. A three-dimensional, multi-source, heterogeneous data acquisition system was built, from underlying physical resources to upper-level business logic, ensuring the comprehensiveness and timeliness of the data on which model analysis and decision-making depend, and providing a solid data foundation for accurately depicting system status, load behavior, business needs and their interrelationships.

[0026] Specifically, S5 includes the following steps: S51. Perform data cleaning operations, including imputation or labeling of missing values, smoothing and filtering of data noise, and identifying and removing outliers. S52. Perform data alignment operation to unify time series data from different acquisition sources and with different acquisition frequencies to the same time base and sampling frequency through timestamp alignment and resampling technology. S53. Perform data standardization or normalization operations, and eliminate the differences in units and numerical ranges of different monitoring indicators through linear transformation or scaling methods. S54. Based on the preprocessed time series data, calculate the statistics of mean, variance, maximum, minimum and quantiles on the rolling time window to construct statistical features; S55. Extract trend components, periodic components, autocorrelation coefficients, and time series pattern features from time series data to construct time series features; S56. Calculate the correlation coefficients between resource usage sequences of different workload instances, the correlation between usage of different resource types, and construct correlation features; S57. Extract task tags, container image names, user groups, and project identifier category fields from task orchestration metadata, encode or embed them to represent them, and construct semantic features. By employing standardized data processing procedures and diverse feature construction methods, multi-dimensional information, including statistical, temporal, relational, and semantic information, is extracted from raw data. This process transforms the raw data into a feature representation that machines can learn efficiently, which is a key step in improving the accuracy and generalization ability of machine learning models.

[0027] Specifically, S6 includes the following steps: S61. Select and build a suitable machine learning model structure based on the specific resource prediction task, pattern recognition task, or online decision-making task objective. S62. Following the chronological order of the time series, the dataset containing historical information is divided into training set, validation set and test set in chronological order; S63. Define corresponding loss functions for different types of modeling tasks; for prediction tasks, use mean squared error and mean absolute error loss functions; for decision-making tasks, construct custom loss functions that are directly related to the key performance indicators defined in S3. S64. Employ a time-series cross-validation strategy to evaluate the performance of the machine learning model across multiple time segments. S65. For critical decisions that affect system stability and resource allocation security, introduce interpretability techniques to analyze and explain the reasoning process and decision-making basis of machine learning models; It provides specific technical solutions for building and evaluating machine learning models suitable for resource allocation scenarios, emphasizing temporal logic, loss functions linked to business KPIs, temporal verification, and interpretability processing, ensuring that the trained models not only perform well on historical data, but can also be safely, reliably, and interpretably applied to online decision-making environments.

[0028] Specifically, S7 includes the following steps: S71. When the underlying machine learning model generates each preliminary decision suggestion, it simultaneously outputs the core basis supporting the decision and the confidence assessment of the decision effect. The decision basis extracts and displays the Top N key features that affect the decision and their contribution. S72. Based on the different optimization tendencies established in S2, the system generates 2 to 3 differentiated alternative solutions, each of which is associated with a set of expected results consisting of key performance indicators defined in S3. S73. Design a decision intervention interface that enables operations and maintenance experts to review the decision suggestions and their basis generated by the system, approve or reject decisions, view detailed comparisons of all alternative solutions, and inject temporary or permanent business rules into the system, including automatic priority approval of resource requests for core service A during major promotional periods. S74. Mark the feedback from operations and maintenance experts on whether the decision results are correct or incorrect, and use it as high-quality labeled data for direct use in the subsequent optimization process of machine learning models. A specific human-machine collaborative workflow was designed. By displaying decision-making basis, providing alternative solutions, opening up review and rule injection, and integrating human feedback, the domain knowledge and experience of human experts are seamlessly integrated into the automated decision-making process. This effectively solves the problems of insufficient flexibility and untimely response to new business rules that may exist in the actual operation of purely automated models.

[0029] Specifically, S8 includes the following steps: S81. After the final decision is made in S7, based on historical forecasting models and predefined business scenarios, a set of tiered flexible contingency plans covering multiple possibilities including optimistic, baseline, and pessimistic scenarios is generated. Each level of the plan specifies the steps, resource amounts, and timelines for subsequent expansion, contraction, or recovery. S82. Bind the flexible contingency plan with quantifiable system status indicators, and pre-verify the resource scheduling feasibility of each plan to form a status contingency plan trigger mapping table. The system status indicators are the number of connections, error rate, and predicted load threshold. S83. Continuously evaluate system status indicators through real-time monitoring and short-term forecasting. When the status is detected to be approaching the trigger condition of a certain plan, issue an early warning to the operation and maintenance personnel. S84. When the triggering conditions are met, the plan switch is automatically executed according to the preset strategy, or the plan is manually confirmed again before the critical operation, so as to smoothly and dynamically transition from the current execution strategy to a flexible plan that adapts to the new state. A forward-looking contingency management mechanism has been introduced, shifting from passive response to proactive preparation. By generating multi-level contingency plans based on predictions and binding them to measurable status indicators, automatic early warning and smooth switching of resource allocation have been achieved. This enhances the system's resilience, agility, and fault tolerance in the face of future load fluctuations, and better ensures service stability and resource response speed.

[0030] Specifically, S10 includes the following steps: S101. In an online environment, A / B testing and shadow mode are used to compare the effects of the intelligent allocation strategy with the baseline strategy of static allocation and threshold rules. The key performance indicators defined in S3 are used to evaluate the effectiveness of the strategy, and the operational efficiency of the human-machine collaboration layer in S7 is evaluated, including the average time for manual approval and the decision adoption rate. S102. Establish an online monitoring mechanism to continuously track key indicators of the prediction accuracy and decision execution effect of machine learning models, and identify concept drift. S103. Using the human feedback records, decision effectiveness feedback, and feedback data on deviations in the implementation of contingency plans collected from S7 and S9, construct a high-quality labeled dataset to drive the machine learning model to perform online learning or periodic offline retraining. S104. Analyze the usage frequency and effectiveness of business rules injected through S7, automatically clean up invalid or expired rules, and convert high-frequency effective rules into model features or logic. S105. Based on the evaluation results and monitoring feedback, the adjustment of the objectives, constraints and indicators of S1-S3 is triggered, the feature set of S5 is updated, the machine learning model of S6 is retrained, and the business rule base of S7 is maintained, forming a closed-loop iterative mechanism. By comparing results, monitoring concept drift, and using multi-source feedback (human evaluation, contingency plan deviation, etc.) to drive model retraining and rule base optimization, and triggering full-link iteration of preceding steps (objectives, features, models, rules), the system can continuously adapt to business changes, correct deviations, and absorb experience, achieving self-evolution and long-term value maintenance, and preventing models and rules from becoming outdated due to environmental changes. Example 3:

[0031] A computer resource intelligent allocation terminal includes: a processor, a memory, and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the processor to implement the machine learning-based computer resource intelligent allocation method as described in any one of claims 1 to 9. Example 4:

[0032] The system's functional boundaries (resource pool scope, service targets) and operational objects (allocatable resource dimensions) are clearly defined to delineate the scope of all subsequent operations. Core business constraints based on Service Level Agreements (SLAs) are established, and a multi-objective optimization framework balancing resource utilization, energy consumption, and cost is constructed. The physical and budgetary constraints of system operation are also clearly defined, transforming the resource allocation problem into a structured and quantifiable optimization model. To ensure measurable optimization objectives, a set of multi-dimensional key performance indicators (KPIs) is defined, including average resource utilization reflecting efficiency, SLA violation rate reflecting service quality, task completion time reflecting performance, and energy cost per unit throughput reflecting economy. To achieve data-driven intelligent decision-making, data reflecting resource status, load behavior, and other factors are systematically collected from four layers: infrastructure, task application, scheduling and orchestration, and business. The system utilizes raw data from historical data and business needs to construct a comprehensive data view. It preprocesses the heterogeneous data from multiple sources, performing cleaning, alignment, and standardization to eliminate noise and inconsistencies. By calculating statistics, extracting temporal patterns, analyzing correlations, and encoding semantic information, it constructs a feature set containing statistical, temporal, correlation, and semantic features to provide high-quality input for the model. For specific tasks such as resource prediction, pattern recognition, and online decision-making, it selects appropriate machine learning model structures for training. The training process emphasizes temporal verification to prevent information leakage, defines a loss function linked to the aforementioned key performance indicators to drive business goal optimization, and introduces interpretability analysis for key decisions to ensure the model's effectiveness, reliability, and transparency. After the model automatically generates preliminary decision suggestions (with supporting evidence and confidence levels) and multiple alternative solutions, it enters the human-machine collaboration phase. Through the decision intervention interface, operations and maintenance experts can review the basis, compare solutions, approve / reject decisions, and inject temporary or permanent business rules. Expert feedback (approval / rejection flags) is used as high-quality labels for continuous model optimization, forming a closed loop of human feedback. Based on the final decision and future load forecasts, a tiered set of flexible contingency plans covering different scenarios (optimistic, baseline, pessimistic) is generated. These contingency plans are dynamically bound to quantifiable system status indicators such as connection count and error rate, forming a status-contingency plan trigger mapping. The system continuously monitors the status, provides early warnings when approaching trigger conditions, and automatically or smoothly switches contingency plans when the conditions are met. During the execution and switching of the above flexible contingency plans, the system synchronously monitors their actual effects and continuously records deviations from expected goals. These deviation data reveal the effectiveness of the contingency plans and, as a type of high-level feedback data, are collected along with regular post-decision effect data to provide deeper insights for system optimization. The system then enters a continuous self-optimization phase.By comparing the effectiveness of strategies with A / B testing and shadow mode, concept drift is monitored; multi-source feedback data, such as human feedback and contingency plan deviations collected from human-machine collaboration and contingency plan monitoring, is used to drive the periodic retraining of machine learning models and the dynamic cleaning and optimization of the business rule base; based on the evaluation results, adjustments and updates to previous steps (such as objectives and indicators, feature sets, models, and rule bases) are ultimately triggered, thus forming a closed-loop iterative mechanism covering the entire link to ensure that the system can adapt to changes and continuously improve.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0034] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope of this disclosure is indicated by the following claims.

Claims

1. A machine learning-based intelligent allocation method for computer resources, characterized in that, Includes the following steps: S1. Clearly define the scope of the resource pool, the service targets, and the dimensions of allocable resources; S2. Establish a multi-objective optimization framework and system constraints that balance resource utilization and energy efficiency costs based on service level agreements; S3. Define and set key performance indicators for quantitative evaluation of the achievement of optimization goals; S4. The system collects raw data from multiple levels, including infrastructure, task applications, scheduling and orchestration, and business operations. S5. Clean, align, and standardize the collected multi-source raw data for preprocessing, and construct a feature set containing statistical, temporal, correlation, and semantic information. S6. For prediction, identification, and decision-making tasks, select and train the corresponding machine learning models, and ensure the effectiveness and reliability of the models through time-series verification, loss optimization, and interpretability processing. S7. Construct a model-based automated decision-making review, intervention, and rule injection mechanism to achieve transparent, controllable, and continuously optimized decision-making through interpretable output, multi-solution comparison, and human feedback closed loop. S8. Based on current decisions and future predictions, generate multi-level flexible contingency plans and dynamically bind them to system status indicators to achieve contingency plan early warning, condition triggering and smooth switching. S9. During the execution and switching of the contingency plan, continuously monitor the deviation between the actual effect of the contingency plan and the expected goal, and collect the deviation data as a high-order feedback signal. S10. Through online performance comparison, concept drift monitoring, feedback data-driven model retraining, and rule base maintenance, the system can achieve continuous evaluation, optimization, and full-link iteration.

2. The intelligent allocation method for computer resources based on machine learning according to claim 1, characterized in that, The resource pool scope includes a single cluster and a hybrid cloud across data centers; The service targets include Web services, batch processing jobs, and AI training tasks; The resource dimensions include vCPU core count, memory capacity, GPU count and video memory, network bandwidth limit, disk type, and IOPS quota; S2 specifically includes the following steps: S21. Establish the core optimization objective of maximizing overall resource utilization while satisfying all service level agreements, and set minimizing energy consumption and reducing operating costs as secondary optimization objectives. S22. Define the hard constraints that the system must follow to operate, including the maximum resource requirement limit for a single task, the total capacity limit of physical nodes, and the overall budget limit; S23. Integrate the above objectives and constraints to construct a quantifiable multi-objective optimization problem model.

3. The intelligent allocation method for computer resources based on machine learning according to claim 2, characterized in that, S3 specifically includes the following steps: S31. Set average resource utilization rate as a core efficiency indicator; set service level agreement violation rate as a core quality indicator. S32. Set task completion time as the core performance indicator; set energy consumption cost per unit throughput as the core economic indicator.

4. The intelligent allocation method for computer resources based on machine learning according to claim 3, characterized in that, S4 specifically includes the following steps: S41. Collect underlying indicators reflecting the status of physical resources from the infrastructure layer. The underlying indicators include continuous collection of CPU utilization, memory usage, I / O throughput, network traffic and energy consumption data of the host machine and physical machine. S42. Collect real-time runtime data of workload instances from the task application layer, including key performance indicators such as resource usage, response time, throughput, and error rate of each instance, as well as application logs. S43. Collect metadata of resource scheduling and task management from the scheduling and orchestration layer, including the waiting status of the global task queue, historical scheduling decision records, task priority tags, and inter-task dependencies. S44. Collect data reflecting external demands and business characteristics from the business layer, including collecting known business characteristics such as external request traffic, user session status, peak business periods, and promotional activity periods.

5. The intelligent allocation method for computer resources based on machine learning according to claim 4, characterized in that, S5 specifically includes the following steps: S51. Perform data cleaning operations, including imputation or labeling of missing values, smoothing and filtering of data noise, and identifying and removing outliers. S52. Perform data alignment operation to unify time series data from different acquisition sources and with different acquisition frequencies to the same time base and sampling frequency through timestamp alignment and resampling technology. S53. Perform data standardization or normalization operations, and eliminate the differences in units and numerical ranges of different monitoring indicators through linear transformation or scaling methods. S54. Based on the preprocessed time series data, calculate the statistics of mean, variance, maximum, minimum and quantiles on the rolling time window to construct statistical features; S55. Extract trend components, periodic components, autocorrelation coefficients, and time series pattern features from time series data to construct time series features; S56. Calculate the correlation coefficients between resource usage sequences of different workload instances, the correlation between usage of different resource types, and construct correlation features; S57. Extract task tags, container image names, user groups, and project identifier category fields from task orchestration metadata, encode or embed them to represent semantic features.

6. The intelligent allocation method for computer resources based on machine learning according to claim 5, characterized in that, S6 specifically includes the following steps: S61. Select and build a suitable machine learning model structure based on the specific resource prediction task, pattern recognition task, or online decision-making task objective. S62. Following the chronological order of the time series, the dataset containing historical information is divided into training set, validation set and test set in chronological order; S63. Define corresponding loss functions for different types of modeling tasks; for prediction tasks, use mean squared error and mean absolute error loss functions; for decision-making tasks, construct custom loss functions that are directly related to the key performance indicators defined in S3. S64. Employ a time-series cross-validation strategy to evaluate the performance of the machine learning model across multiple time segments. S65. For critical decisions that affect system stability and resource allocation security, introduce interpretability techniques to analyze and explain the reasoning process and decision-making basis of machine learning models.

7. The intelligent allocation method for computer resources based on machine learning according to claim 6, characterized in that, S7 specifically includes the following steps: S71. When the underlying machine learning model generates each preliminary decision suggestion, it simultaneously outputs the core basis supporting the decision and the confidence assessment of the decision effect. The decision basis extracts and displays the Top N key features that affect the decision and their contribution. S72. Based on the different optimization tendencies established in S2, the system generates 2 to 3 differentiated alternative solutions, each of which is associated with a set of expected results consisting of key performance indicators defined in S3. S73. Design a decision intervention interface that enables operations and maintenance experts to review the decision suggestions and their basis generated by the system, approve or reject decisions, view detailed comparisons of all alternative solutions, and inject temporary or permanent business rules into the system, including automatic priority approval of resource requests for core service A during major promotional periods. S74. Mark the feedback from operations and maintenance experts on whether the decision results are correct or incorrect, and use it as high-quality labeled data for direct use in the subsequent optimization process of machine learning models.

8. The intelligent allocation method for computer resources based on machine learning according to claim 7, characterized in that, S8 specifically includes the following steps: S81. After the final decision is made in S7, based on historical forecasting models and predefined business scenarios, a set of tiered flexible contingency plans covering multiple possibilities including optimistic, baseline, and pessimistic scenarios is generated. Each level of the plan specifies the steps, resource amounts, and timelines for subsequent expansion, contraction, or recovery. S82. Bind the flexible contingency plan with quantifiable system status indicators, and pre-verify the resource scheduling feasibility of each contingency plan to form a status contingency plan triggering mapping table. The system status indicators are the number of connections, error rate, and predicted load threshold. S83. Continuously evaluate system status indicators through real-time monitoring and short-term forecasting. When the status is detected to be approaching the trigger condition of a certain plan, issue an early warning to the operation and maintenance personnel. S84. When the triggering conditions are met, the plan is automatically switched according to the preset strategy, or after manual confirmation before the critical operation, the plan smoothly and dynamically transitions from the current execution strategy to a flexible plan that adapts to the new state.

9. The intelligent allocation method for computer resources based on machine learning according to claim 8, characterized in that, S10 specifically includes the following steps: S101. In an online environment, A / B testing and shadow mode are used to compare the effects of the intelligent allocation strategy with the baseline strategy of static allocation and threshold rules. The key performance indicators defined in S3 are used to evaluate the effectiveness of the strategy, and the operational efficiency of the human-machine collaboration layer in S7 is evaluated, including the average time for manual approval and the decision adoption rate. S102. Establish an online monitoring mechanism to continuously track key indicators of the prediction accuracy and decision execution effect of machine learning models, and identify concept drift. S103. Using the human feedback records, decision effectiveness feedback, and feedback data on deviations in the implementation of contingency plans collected from S7 and S9, construct a high-quality labeled dataset to drive the machine learning model to perform online learning or periodic offline retraining. S104. Analyze the usage frequency and effectiveness of business rules injected through S7, automatically clean up invalid or expired rules, and convert high-frequency effective rules into model features or logic. S105. Based on the evaluation results and monitoring feedback, adjustments to the objectives, constraints and indicators of S1-S3 are triggered, the feature set of S5 is updated, the machine learning model of S6 is retrained, and the business rule base of S7 is maintained, forming a closed-loop iterative mechanism.

10. A computer resource intelligent allocation terminal, characterized in that, include: A processor, a memory, and one or more computer programs, said one or more computer programs being stored in said memory and configured to be executed by said processor to implement the machine learning-based intelligent allocation method for computer resources as described in any one of claims 1 to 9.

Images