A job management method and system based on deep learning

Abstract

The present invention proposes a job management method and system based on deep learning, the method comprising: constructing an initial deep learning model based on a deep belief network; using the deep learning model to predict and optimize resource allocation in a dynamically changing job environment; dynamically revising the prediction results and resource allocation scheme according to changes in scene requirements in the job environment; and continuously training and adjusting the deep learning model for optimization based on the revised prediction and resource allocation scheme. The present invention solves the problem of how to construct a deep learning model based on a deep belief network, and then use the deep learning model to accurately predict and optimize resource allocation in a dynamically changing job environment, while revising the prediction and resource allocation according to the requirements of different scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of job management, and in particular, to a job management method and system based on deep learning. Background Art

[0002] In the existing technology, building a DBN-based deep learning model requires a large amount of data for training and ensuring that the model can learn sufficiently complex and universal features; secondly, ensuring the high accuracy of prediction and resource allocation when resource demand continues to change is a key problem. Especially when the application scenarios are diverse and change frequently, the model not only needs to quickly adapt to new job requirements and environmental changes, but also needs to continuously revise the prediction results and corresponding configurations based on the performance indicator feedback in actual operation.

[0003] In addition, considering the possible human intervention or other unforeseen factors in the actual operation management process, these will also pose a potential threat to the effectiveness of this method.

[0004] Therefore, during implementation, the impact of various variables and the system's ability to respond to these variables must be considered comprehensively. Summary of the invention

[0005] The purpose of this invention is to propose a job management method and system based on deep learning, which can predict and optimize the allocation of job resources according to real-time data, and continuously improve its own performance and adaptability as the working environment changes, thereby ensuring that the system can achieve the best effect in different application scenarios.

[0006] In order to achieve the above object, a first aspect of the present invention provides a job management method based on deep learning, the method comprising:

[0007] S1. Build an initial deep learning model based on a deep belief network;

[0008] S2. Utilizing the deep learning model to predict and optimize resource allocation in a dynamically changing operating environment;

[0009] S3, dynamically revise the prediction results and resource allocation plans according to the changes in scenario requirements in the operating environment;

[0010] S4. Continuously train and adjust the deep learning model for optimization based on the revised prediction and resource allocation plan.

[0011] Preferably, the S1 comprises:

[0012] Preprocess a large amount of historical data, including data cleaning and labeling;

[0013] Initialize the weight values ​​of the multi-layer probabilistic graphical model;

[0014] Adjust the ratio of deep and shallow model structures through comparative analysis to improve computational efficiency;

[0015] Layer-wise unsupervised pre-training using restricted Boltzmann machines;

[0016] Finally, combined with the label data, supervised fine-tuning is completed through the back-propagation algorithm to establish a complete deep belief network model;

[0017] Among them, when initializing the parameters of the multi-layer probabilistic graphical model, the following operations are performed based on specific constraints:

[0018] The number of neurons N in each hidden layer must be greater than or equal to half of the number of nodes D in the input layer, as shown below:

[0019] ;

[0020] Select an appropriate regularization strength λ, and its value must satisfy λ≤0.001 to prevent overfitting;

[0021] The initial learning rate is set to 0.01 to ensure effective update of weights during network training;

[0022] If the learning rate is less than the threshold , then stop updating the weights and consider it to have reached a stable state.

[0023] Preferably, S2 includes:

[0024] Using the constructed deep learning model, input the historical usage of relevant resources in the current operating environment;

[0025] Collect various real-time resource data of the current environment in real time, including resource load, remaining available resources and demand fluctuations;

[0026] By studying the changing trends of the operating environment characteristics, the matching degree of resource supply and demand is predicted;

[0027] Use the forecast results to automatically allocate appropriate resources for tasks in the upcoming time period;

[0028] By maximizing the objective function of the optimization effect, resource allocation is adjusted to ensure maximum resource utilization and system efficiency.

[0029] Preferably, S3 includes:

[0030] Identify new or unusual situations in the operating environment and analyze the potential impact of changes on resource allocation plans;

[0031] Monitor the real-time resource status of the system and evaluate the effectiveness of the current resource distribution;

[0032] Reassess the accuracy of forecasts and the effectiveness of currently allocated resources;

[0033] Timely adjust job priorities and resource allocation plans based on new changes;

[0034] Determine the correction plan and conduct verification and evaluation before implementation to ensure the consistency of resource allocation and changes in the operating environment.

[0035] Preferably, S4 includes:

[0036] Feedback to the deep learning model training set based on the last execution and the newly adjusted results to optimize the model structure;

[0037] Update the learning rate strategy and regularization parameters to avoid overfitting and improve prediction accuracy;

[0038] Further training the model to improve its adaptability and reliability in future job management decisions;

[0039] Use the dynamic correction results as a new round of optimization training data to continuously improve the model's learning of resource scheduling in new environments;

[0040] Iteratively update and test models to maintain predictive capabilities and improve the effectiveness of future resource allocation plans.

[0041] Preferably, the step of constructing an initial deep learning model based on a deep belief network includes:

[0042] Set up the basic architecture of the deep belief network, defining the number of hidden layers and their number of neurons;

[0043] Determine the input feature set and output variables of the model;

[0044] Establish a training sample set and initialize the initial values ​​of network parameters;

[0045] The network is pre-trained according to the given learning rate and number of training rounds, and the optimization algorithm is set. The neural network parameters are updated based on the training samples until the set training accuracy is achieved.

[0046] Preferably, the prediction of the matching degree of resource supply and demand by learning the changing trend of the operating environment characteristics includes:

[0047] To set the number of hidden units N and the learning rate of each layer, the degree of change P of the operating environment characteristics must be considered. The specific operations are as follows:

[0048] If the degree of change of the working environment characteristics is P ≥ 7, each hidden layer will increase N + 50% hidden units;

[0049] If the degree of change in the operating environment characteristics P < 0.3, keep the default configuration unchanged;

[0050] Dynamically adjust the learning rate to reduce the original The ratio of slowing down the change of model learning rate;

[0051] Stop iterative updating of network weights. When the environmental characteristics change significantly, that is, P>0.95 or the predetermined number of times is reached, stop training and updating.

[0052] Preferably, in the process of adjusting the learning rate, based on specific conditions, if the current learning rate reaches or is less than the threshold value T=1e-6, the operation is as follows:

[0053] Calculate the reduction of the loss function in the i-th step of the current iteration as , and the reduction in the previous iteration , the average reduction ratio is defined as , which is expressed as follows:

[0054] ;

[0055] If the loss function changes tends to be stable, that is, ,in, is the ratio calculated last time, the current learning rate is reduced by a fixed coefficient;

[0056] When three times in a row When , the current learning rate is reduced to half of the current learning rate to slow down the update process;

[0057] Once the current learning rate drops to T, the learning update operation is immediately suspended to prevent falling into a training trap or oscillation and non-convergence.

[0058] Preferably, if after three consecutive checks, it is found that the update process still has no obvious progress, an optimized strategy needs to be made to deal with this phenomenon. Under this condition, the operation is performed based on the loss reduction ratio, wherein the optimized strategy includes the following steps:

[0059] The loss in the current iteration step is L, and the initial loss is , the loss is reduced , then the loss reduction ratio ;

[0060] If the loss is reduced by <0.01, that is, the model learning progress is slow, and the strategy of increasing the regularization strength λ is adopted to reduce the risk of overfitting; wherein, the increase in regularization strength get , which is expressed as follows:

[0061] ;

[0062] in, >1, indicating the weight of regularization strength;

[0063] The current loss , and the reduction ratio of the previous iteration is recorded as ,in, , Represents the loss of the previous iteration. The ratio of the two is the current step efficiency E, which is expressed as follows:

[0064] ;

[0065] If E is close to 0, it indicates that the network learning process has entered a stable state;

[0066] If the current step-size efficiency E is stable, the current learning rate is further reduced to half of the current level to promote the convergence process of training and reduce the cost of the number of iterations;

[0067] The detection process is repeated until the loss change no longer satisfies the current step-size efficiency E<0.001 condition to achieve the best learning effect.

[0068] In a second aspect of the present invention, a job management system based on deep learning is provided, the system comprising:

[0069] Initialization module, used to build an initial deep learning model based on deep belief network;

[0070] A resource configuration optimization module, used to use the deep learning model to predict and optimize resource configuration in a dynamically changing operating environment;

[0071] The operating environment correction module is used to dynamically correct the prediction results and resource allocation plans according to the changes in the scene requirements in the operating environment;

[0072] The job management optimization module is used to continuously train and adjust the deep learning model for optimization based on the revised prediction and resource allocation plan.

[0073] The beneficial technical effects of the present invention are at least as follows:

[0074] The present invention can predict and optimize the allocation of job resources based on real-time data, and can continuously improve its performance and adaptability as the working environment changes, thereby ensuring that the system can achieve the best results in different application scenarios. This job management system mainly includes the following four stages: creating a deep learning model based on a deep belief network; using this model for prediction and optimization; adjusting the configuration according to demand changes; and continuing to train and optimize the model based on the revised configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0076] Figure 1 This is a flow chart of a job management method based on deep learning in the present invention.

[0077] Figure 2 This is a framework diagram of a job management system based on deep learning in the present invention. DETAILED DESCRIPTION

[0078] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0079] like Figure 1 As shown, an embodiment of the present invention provides a job management method based on deep learning, and the method includes the following steps.

[0080] The first thing that needs to be built is an initial deep learning model based on a deep belief network (DBN). DBN is a layered neural network structure that performs very well in machine learning tasks, especially unsupervised data representation learning. Building this model usually requires the following steps: training each layer from the bottom up to automatically generate representations of high-level abstract data; then forming a deep network by adjusting the weights in these abstract layers. At the end of this step, a data input and output framework that can handle specific business scenarios will be generated. For example, one can imagine an automated manufacturing assembly line where there are different types of workloads - some are continuous production tasks, while others may be intermittent or emergency maintenance activities - and the system needs to understand these tasks and the dependencies between them.

[0081] Take the smart manufacturing environment as an example, in this environment, the production line may be composed of several robot arms, performing a series of tasks of different complexity such as material picking and placing, welding, assembly, etc. The initial deep learning model first understands and learns the characteristics, priorities, and correlation and complexity of different tasks by training on historical operation data (such as status data of various machines, maintenance records, operating efficiency, etc.).

[0082] After the preliminary model is built, it can be applied in the real environment to complete dynamic prediction tasks and optimize system resource distribution. By collecting new job status information constantly generated in the operating environment, and using the built model for prediction and resource allocation optimization. Continuing with the example of the automated production line environment, the model will predict the tasks that are newly entered into the system - evaluate the time and resources required for each work task, and automatically assign tasks to the most suitable robot workstations to complete them in the most efficient way.

[0083] During this process, if the system detects that the workload of a certain part of the production line suddenly increases beyond the model's estimated range, the model will quickly adjust the relevant work task plan and equipment utilization pattern, such as prioritizing resources to handle urgent or high-value tasks, and postponing less urgent production steps to ensure the smoothness and efficiency of the overall production process.

[0084] Understandably, the environment and resource requirements in actual work are often constantly changing. This requires the system to make necessary changes to the previous resource configuration according to the actual needs of the current scenario in order to better cope with new work requirements. Back to the intelligent manufacturing scenario, assuming that a robot arm on the production line is suspended due to maintenance, the model will immediately recognize this change and automatically re-plan the work order or temporarily replace another robot to perform the task that should have been performed by the damaged robot arm to ensure the smooth progress of the entire process.

[0085] By regularly monitoring the operation process and reallocating resources based on actual operation scenarios and environmental requirements, the operation system can be maintained in a smooth and efficient manner.

[0086] Furthermore, as time goes by and various requirements and constraints in the operating environment continue to evolve, the previously constructed DBN-based learning framework also needs to be adjusted accordingly to ensure its prediction accuracy and adaptability. In the intelligent manufacturing environment, the model may regularly update the actual data from the production line (such as the emergence of new types of materials or equipment wear, etc.), and the feedback and evaluation mechanism can timely reflect the changes within the system and feed back to the algorithm to train the model to improve its adaptability to future situations.

[0087] In order to further improve the overall effect, the system will train a deep belief network on existing data (especially new data streams such as equipment maintenance and fault information) to improve the prediction effect; this method can not only achieve better scheduling plans in the future, but also provide more personalized and accurate support solutions for different types of events - such as quickly responding to emergencies or adjusting the task ratio of different areas in the production line to achieve the goal of maximizing efficiency.

[0088] In summary, this method of managing resources using deep learning technology can maintain a high level of operational and management efficiency in a constantly changing and developing environment. This method solves various problems that may arise in the operation process by applying DBN to the prediction and optimization of resource allocation, such as delayed task completion due to insufficient resources and insufficient response to urgent task changes. It also provides managers with a reasonable decision-making basis based on scientific analysis and simulation, greatly improving management quality and efficiency. This method not only has broad application potential in manufacturing automation scenarios, but also has wide practicality in other industries, such as IT data center operations, traffic scheduling and other fields.

[0089] Next, the contents of step S1 are described. This deep learning-based job management method specifically involves the implementation of multiple technical operation levels, which are described in detail below.

[0090] The first step is to preprocess a large amount of historical data, including data cleaning and labeling. This step refers to the process of cleaning and organizing the original data in order to obtain more accurate results in subsequent training. For example, missing values ​​in the data may need to be supplemented or ignored, and noise data should be removed, so as to ensure that the model can learn in a more standardized and accurate data set environment. In addition, the data is labeled to attach a clear target label or description category to each or each record, so that the model knows what the target result is in subsequent supervised training, further improving the prediction or classification capabilities.

[0091] For example, when applying this job management system in a company, historical order data is the key object that needs to be cleaned. Before cleaning the historical data of millions of orders, it is necessary to first remove duplicate and erroneous information, and then clearly mark the completion status and time consumption of each service. These sorted order history records can be used as the starting point for training models to better optimize job scheduling and management.

[0092] The second step is to initialize the weights of the multi-layer probabilistic graph model. This step involves the parameter setting of the deep model in the early stage of construction. For a deep learning model, randomly initializing the associated weights of each node in the network can start the preparation work before training.

[0093] For example, when building a multi-layer neural network, the standard normal distribution or other distributions can be used to generate initial values ​​for the weights to avoid the problem of all neural units having completely consistent initial states, so that each network layer has random differences for effective learning.

[0094] Then, the ratio of deep and shallow model structures is adjusted through comparative analysis to improve computational efficiency. This indicates that in the actual modeling process, the structure should be repeatedly tested, compared and appropriately modified to ultimately determine the most suitable balance ratio between deep and shallow structures.

[0095] For example, we can compare the model's computational time and final performance by setting different layer depth ratios. In some experiments, we found that using a combination of more shallow layers and an appropriate number of deep layers can achieve the best results. This balance maintains good prediction results while not wasting too much learning and training time, improving the computational cost-effectiveness of the entire model in the job management system.

[0096] The next step is to use the restricted Boltzmann machine (RBM) for layered unsupervised pre-training. In this step, the restricted Boltzmann machine, as a common unsupervised pre-training model, plays the role of preliminary training for each independent layer of the multi-layered neural network. The purpose is to obtain a relatively better initial setting of the weight matrix, provide a better basis before the real labeled data is input for supervised adjustment, and improve the performance in the later stage of training.

[0097] Taking job scheduling as an example, before introducing specific historical order task completion results, we can first build an initial good weight group for the deep learning model by performing restricted Boltzmann machine processing on a large number of unlabeled original log files, so that subsequent adjustments can reach the optimal state faster to assist in the prediction of task allocation.

[0098] The last step is to combine the labeled data and complete the supervised fine-tuning through the backpropagation algorithm to build a complete deep belief network model. This step is to further correct and fine-tune the existing weights based on the labeled historical data after the previous steps are completed, so that the model can eventually form a stable decision boundary and make correct analysis of new data in the future.

[0099] Taking the enterprise scheduling operation as an example, after completing the above steps to obtain the initial version of the operation model, the complete historical order data after annotation is introduced for fine-tuning and optimization. The back propagation algorithm is used to allow the machine to identify whether there are problems with its previous predictions and correct the wrong direction in time. Through continuous iteration, a complete set of operation management solutions suitable for its own specific business scenarios is gradually formed.

[0100] Next, the specific process and examples involved in step S2 are described.

[0101] First, the constructed deep learning model is used to receive statistical data about the past usage of various resources in the current job environment. For example, an enterprise's cloud platform has collected information such as memory usage and CPU utilization over the past year; this information will be used as the basis for model training so that the deep learning model can understand and learn how different types of tasks consume resources.

[0102] Subsequently, in actual use, data on various resources in the current operating environment are collected in real time. This information may include, but is not limited to, data points such as processor load status, idle hardware resources, and changes in business task requirements over time. For example, the system will detect the core processor occupancy and the number of allocable cores in the server cluster once a second over a period of time, while also continuously monitoring new processing requirements initiated by users and counting their changing patterns.

[0103] Based on the models of these historical data and the new information obtained from real-time monitoring, the types and quantities of tasks expected to appear in a specific time period in the future are estimated by studying the trends of environmental characteristics. For example, the system uses the above method to conclude that a large number of data transmission requests are expected to appear in the next hour, and based on this prediction, it decides to reserve part of the network bandwidth to cope with the upcoming peak data throughput.

[0104] Next, decisions are made based on future resource forecasts to automatically allocate resources for each newly generated task that will be run during that period. In the previous example, if a large amount of bandwidth usage is foreseen, corresponding measures may be taken, such as adding some server nodes in advance or limiting the requirements of low-priority tasks for network facilities, so as to ensure the smooth execution of important data transmission and avoid resource bottlenecks as much as possible.

[0105] Finally, make adjustments based on whether the results achieved meet the standards of the established optimal solution, and improve equipment efficiency and utilization as much as possible while meeting all service level agreements (SLAs) and performance indicators. For example, suppose after a period of operation, it is found that although the resource allocation strategy has achieved a good balance overall, it still causes some users to wait in line during a few high-traffic time periods; then you can try to fine-tune the various threshold parameters in those areas with high prediction accuracy until the overall benefit is more ideal, such as further increasing the reserved amount of certain types of resources to alleviate the queuing pressure problem. This continuous cycle of evaluation and adjustment can ensure the long-term stable and efficient operation of the entire job management system.

[0106] Next, the step S3 described refers to a series of steps and technical application details involved in the specific operation of the deep learning-based job management method. First of all, when identifying new or abnormal situations in the operating environment, it means that a mechanism is needed to perceive changes in the environment in real time, such as the launch of new projects, the postponement of key tasks, or sudden problems within the system, such as resource bottlenecks. This mechanism can be implemented with the help of supervised or unsupervised learning methods in AI algorithms. Suppose in a data center, a certain type of service request suddenly surges, causing the CPU occupancy rate to rise to an abnormal level. The combination of preset thresholds and real-time data analysis can identify the occurrence of this situation and evaluate the potential impact on the current resource allocation based on the change-for example, whether it is necessary to temporarily allocate more CPUs from other services to this task to ensure business continuity and service quality.

[0107] Secondly, the process of monitoring the real-time status of system resources and evaluating the effectiveness of existing resource distribution is essentially to continuously monitor the utilization of computing, storage, and network resources used in the job processing process to ensure the efficient operation of the overall architecture, while also taking into account issues such as cost control and energy consumption. In this example, the system will periodically collect data from each node, such as memory usage ratio, CPU core load, etc., and calculate the overall usage based on this data, and then compare it with the preset standards to evaluate whether the configuration is appropriate - for example, if the current memory is sufficient and the CPU is saturated, it is necessary to balance the load distribution of these two elements.

[0108] Next, we use past data learning and simulation prediction to determine future trends, and then combine the latest real-time feedback to continuously calibrate our understanding of the correlation between demand and response and our response strategies, ensuring that each decision is made based on the most accurate information and avoiding premature or delayed configuration adjustments. For example, after predicting that the next wave of request peaks will occur between 9 p.m. and 11 p.m., the system may reserve some additional computing power in advance to wait for that period, and schedule some less critical tasks to be completed before the peak or postpone them for processing.

[0109] This is followed by flexible adjustments to the resource allocation plan, mainly based on the new changes (such as the time of the aforementioned request peak) or potential problems (such as signs of server failure) identified to respond dynamically, and quickly modify the original plan to optimize the entire process. Continuing with the data center case analysis above, before determining the evening peak period, the system may have automatically increased the target CPU utilization index, delayed some of the maintenance operations originally scheduled for this time, and sent notifications and warnings to relevant users or administrators; when executing such a strategy, it is also necessary to comprehensively consider the impact of multiple variables.

[0110] The final stage is the process of confirming and evaluating the change measures. This step is very important. It involves testing and verification before putting the modified workflow implementation plan into practice. The purpose is to prevent the occurrence of unexpected consequences due to inappropriate adjustments, such as hardware damage caused by over-scheduling, and to ensure that the plan can indeed achieve the expected target effect before it is rolled out on a large scale. Taking the resource allocation during the evening peak period mentioned in the present invention as an example, in order to verify whether the modified workload adjustment plan is appropriate, the system first tries the new strategy on a smaller scale and continuously monitors the experimental results, such as paying attention to the response speed performance in actual applications. If it is observed that all indicators are running within a safe range and can significantly improve the job throughput, then this method can be used for global implementation. Otherwise, further iterative modifications are required before it can be safely launched on a full scale.

[0111] Next, the specific steps of step S4 are described.

[0112] The first thing to mention is to use the results of the model's past operations and adjust the training set of the deep learning model based on these execution feedback to optimize the model structure. For example, in a factory automatic scheduling system operation cycle, the model predicted the optimal production configuration based on the existing resources and executed it. After that, by collecting the effect evaluation data after execution, this information was used as a feedback signal and added to the subsequent training set, so that the model can learn from past experience to allocate resources more reasonably. This link helps the model iterate and gradually improve the decision-making process to reduce the occurrence of misjudgments, so that it can complete similar allocation tasks more accurately and effectively in the future.

[0113] Next, a series of measures are taken to optimize the model parameters, including adjusting the learning rate to achieve rapid adjustment without overfitting the training data, and adjusting the regularization coefficient to further prevent overfitting and ensure the stability and accuracy of the model output. For example, in the above example, assuming that the initial learning rate is too fast, the model frequently misjudges when dealing with some particularly complex job plan allocations. To solve this problem, it is necessary to appropriately reduce the learning rate so that it can be more robust and gradually converge to the best solution when updating; at the same time, in order to deal with the overfitting problem, the L1 or L2 norm regularization term is adjusted to limit the situation where some weight parameters are too large and affect the overall accuracy. This ensures that even in the future, when dealing with new types of tasks, the model will not be unstable due to over-adaptation to the characteristics of the samples that have been seen.

[0114] The further model training emphasizes the process of continuous iteration and upgrading on the existing basis to make the overall performance more perfect. The above example continues to advance. After many experiments, it is found that the current system still cannot cope well with special situations such as a large number of emergency processing requests. In this context, it is necessary to strengthen the learning of relevant parts for such situations in order to better support similar situations in the future.

[0115] Then describe the repeated use of dynamic revision results in training for continuous improvement. Taking the above situation as an example, after a period of practice, the knowledge points about effective scheduling strategies under special conditions can be directly used as the basis for the next round of reinforcement, allowing the model to learn the key operating elements in these special scenarios faster and improve the probability and quality of making reasonable decisions when facing new problems. Such a continuous optimization cycle can ensure that the model remains efficient and can cope with more unknown challenges.

[0116] The last part points out the importance of continuous testing of the model. Still using the above scenario as a reference object, as the number of new situations that may appear on the production line increases, the present invention needs to continuously track the performance of the model and test it to ensure that it still exhibits strong processing capabilities and accurate predictive capabilities in the face of new environments and conditions to support the efficient operation of the entire production line; in addition, it may also involve timely adjustment of the algorithm architecture according to new factors to cope with more diverse demand scenarios, thereby improving the overall quality level of the operation and achieving maximum output efficiency.

[0117] Next, the steps of constructing the initial deep learning model based on the deep belief network are first to set up the basic architecture and define the number of hidden layers and the number of neurons, which is a basic workflow of deep learning. For example, in the process of building the model, if it is decided to use a deep belief network containing three hidden layers (hidden layer 1, hidden layer 2, and hidden layer 3) to construct the deep belief network of the present invention, then the number of nodes in each hidden layer can be defined as 200, 400, and 200 to form a more typical multi-layer structure.

[0118] After that, it is necessary to determine the specific input features and expected output of the model. For example, in the application of the method for student homework management, the present invention can use factors such as "homework completion time" and "question difficulty level" as input features of the network model, and use the status of students completing homework as the output of the network model (the output result can be in binary form, such as: "1" means completed, 0 means not completed).

[0119] Next, we need to prepare the training data set and randomly initialize the network parameters. Assume that we have collected a thousand job data samples for training the network, and initialize the weight matrix and bias vector in each layer so that we can start the training process.

[0120] The last step involves performing deep belief network pre-training according to a specific learning rate and number of cycles, and adjusting the parameters in the network by means of gradient descent optimization methods, etc., until the performance indicators of the overall network reach the expected effect. For example, when the present invention sets the learning rate of the network to 0.001 and iterates for 30 rounds, the training process is started, and the connection weights and the offset values ​​between neurons are updated in each round using back propagation technology, and the ability to predict the operation data is improved in the process of continuously correcting the network structure, thereby ensuring that the final training accuracy can be controlled at the level of 90%.

[0121] Next, we describe the following operations based on specific constraints when initializing network parameters:

[0122] The number of neurons N in each hidden layer must be greater than or equal to half of the number of nodes D in the input layer, as shown below:

[0123] ;

[0124] Select an appropriate regularization strength λ, and its value must satisfy λ≤0.001 to prevent overfitting;

[0125] Set the initial learning rate to , ensuring effective updating of weights during network training;

[0126] If the learning rate Less than threshold ,For example =1e-8, the weight update stops and is considered to have reached a stable state.

[0127] In order to initialize the parameters in the network to optimize the learning process and model performance, specific steps need to be taken given a series of restrictive constraints:

[0128] For the number of hidden layer neurons N, it is stipulated that the number of units in each layer should be at least half of the number of nodes in the input layer. This rule helps to maintain a certain degree of expressiveness and reduce the risk of over-simplification, while also preventing the network from becoming too large and consuming too many resources.

[0129] For example, assuming that the input layer of the network contains 40 features, the number of neurons in each hidden layer should be no less than 20. This configuration ensures that the model is neither too simple nor too complex to cause overfitting or prolonged training time.

[0130] When selecting the regularization strength (also called λ, regularizationstrength), the value must be in the range of 0<λ<0.01 to adjust the degree of weight update. The role of the regularization term is to suppress the increase of weights and thus reduce the error caused by the random characteristics of the training set; the appropriate degree of regularization is critical to controlling overfitting to specific samples, and this setting can further enhance the reliability of model predictions.

[0131] A possible example might be: choosing λ equal to 0.001 to achieve a good balance between effective penalty intensity and training effect. If λ is set too small, the penalty effect will be weakened, leading to overfitting tendency; on the contrary, if it is set too large, the model will be oversimplified, resulting in impaired generalization ability.

[0132] In order to ensure efficient updating of weights and the normal operation of the entire learning mechanism, the initial learning rate is set to 0.01; this initial value can effectively promote the iteration process and prompt the network to continuously approach the target minimum cost, but the learning curve needs to be monitored to prevent slow convergence or even stagnation.

[0133] Set the learning rate update termination rule. If it drops below 1e-8, it is considered to have reached the "end point" of training and the weights will no longer be changed. This is because an extremely low learning rate may cause the network to be unable to significantly learn new content. This criterion helps to automatically stop the iteration process, while also saving resources to a certain extent and avoiding lengthy and fruitless training.

[0134] Overall, these restrictive measures work together to lay the basic conditions for network training in the initialization phase and provide a good start for the subsequent learning phase.

[0135] Next, the number of hidden units and learning rate of each layer should be set considering the degree of change P of the operating environment characteristics. The specific operations are as follows:

[0136] Set the number of hidden units N and the learning rate for each layer The degree of change P of the working environment characteristics needs to be considered. The specific operations are as follows:

[0137] If the degree of change of the working environment characteristics is P ≥ 7, each hidden layer will increase N + 50% hidden units;

[0138] If the degree of change in the operating environment characteristics P < 0.3, keep the default configuration unchanged;

[0139] Dynamically adjust learning rate Reduced from the original The ratio of slowing down the change of model learning rate;

[0140] Stop iterative updating of network weights. When the environmental characteristics change significantly, that is, P>0.95 or the predetermined number of times is reached, stop training and updating.

[0141] In the first step of this process, we need to assess the complexity of the operating environment, and use the parameter P to represent its value. Assume that this parameter is determined by measuring various factors in the current environment, such as the change in sensor readings or the complexity of the data in the task. If P ≥ 7, it means that the environment is relatively complex; on the contrary, if the environment is relatively simple, then P < 0.3. This division can help determine the number of neurons in the hidden layer.

[0142] Next, consider the specific case of tuning a neural network:

[0143] For example, if the current environment complexity is P=8, and the number of neurons in the hidden layer is 120 units when initially set, this number needs to be adjusted according to the environment. In this environment, P≥7, the present invention should add an additional 50% to accommodate more complex inputs. That is, the newly added number increases the original number of neurons of 120 to 120*150%, which is equal to 180. For a model containing a single hidden layer, its total number of units now becomes this new number of 180.

[0144] In addition, for simple tasks (environmental feature P<0.3), the originally set configuration is maintained without any adjustment; that is, the number of neuron units in the neural network is kept fixed at the original number of 120.

[0145] The present invention then adjusts the learning rate parameters of the neural network to reflect the changes in the complexity of the environment to help the model learn more robustly. When calculating the dynamic learning rate, multiply the base learning rate by a factor of To make adjustments. For example, if an initial learning rate is set to 0.003 and P=5, the learning rate under the new setting is: [0.003*(1-5 / 100)=0.003*0.95], which is 0.00285. This helps avoid overtraining or training too slowly.

[0146] The last operation is to stop the weight adjustment process of the training model in a specific environment, depending on the value of P or when the number of iterations reaches a predetermined maximum number of times, to reduce the risk of model inaccuracy due to sudden changes in the environment. For example, if P>0.95, that is, the environmental characteristics have changed dramatically, the model training is terminated and the next operation is taken; similarly, if no further improvement is seen after more than 100 iterations, the optimization process should also be terminated.

[0147] Through the comprehensive use of these adjustment means and stopping strategies, the present invention can effectively control the learning process in working environments of different complexities and ensure stable model performance under changing environments.

[0148] Next, in the process of adjusting the learning rate based on specific conditions, if the current learning rate reaches or is less than the threshold T=1e-6, the operation is as follows:

[0149] Calculate the reduction of the loss function in the i-th step of the current iteration as , and the reduction in the previous iteration , the average reduction ratio is defined as , which is expressed as follows:

[0150] ;

[0151] If the loss function changes tends to be stable, that is, ,in, is the ratio calculated last time, the current learning rate is reduced by a fixed coefficient;

[0152] When three times in a row When To slow down the update process;

[0153] Once the learning rate decreases When T=1e-6, the learning update operation is immediately suspended to prevent falling into a training trap or oscillation and non-convergence.

[0154] This algorithm is designed to optimize the dynamic change of learning rate in deep learning model training.

[0155] Furthermore, the learning rate adjustment rules are explained:

[0156] 1. In the initial stage, the algorithm records the rate of decrease of the loss function after each iteration. This decrease is used to calculate the average decrease rate of the loss function in the current step and is expressed as ;

[0157] 2. Monitor stability. As model training progresses, if the loss function changes toward stability, it will be reflected in the current rate of decline. and last time Compared with less than 0.5 (i.e. ), the model is judged to have entered a relatively flat area. At this time, the current learning rate needs to be reduced to maintain a stable training speed. This adjustment is a fixed attenuation coefficient, which aims to smooth the convergence process and reduce the oscillation caused by drastic changes in gradients during training.

[0158] 3. Once continuous detection , which means that the rate of loss decrease has slowed down significantly, so the learning rate is reduced again, this time by half. This operation is intended to more finely control the learning rate when it is found that the model training has entered a very stable state to prevent the model from overfitting the data set.

[0159] 4. If the learning rate of training has dropped to a specific minimum value T=1e-6, it means that even if training continues, the improvement in performance will be minimal. In order to prevent unnecessary resource consumption and waste of computing time, the learning and parameter update process should be stopped in time to avoid unavoidable training instability or falling into a local optimal solution.

[0160] Understandable, specific examples:

[0161] Assume that the convolutional neural network model in the computer vision project starts training from a random initial state. The present invention sets the initial learning rate to = 0.01. After each iteration (e.g., after each epoch) after training starts, statistics are started according to the first step rule described above. and.

[0162] For example, after the fifth round, the model loss dropped from 1 to 0.51, and we can calculate If the loss at the end of the 6th round becomes 0.275 (which is the same as the previous is less than 0.5), according to the second condition, the learning rate will be reduced by a predetermined factor.

[0163] Then after about 10 iterations, we observed Compared with the previous round, the decline rate is significantly slower—less than 0.05, the third condition is performed: the current learning rate Divide by 2, and the learning rate is .

[0164] Finally, if the model training lasts for a certain number of cycles, the final learning rate Once it drops to 1e-6 or less, stop lowering it because the loss will hardly decrease any further.

[0165] Through the above settings and example analysis, it can be seen that this is an adaptive, dynamic adjustment method based on conditional judgment of learning rate, which can effectively help the model learn better and converge to the ideal parameter setting more efficiently.

[0166] Next, if after three consecutive checks, it is found that the update process still has no obvious progress, more specific decisions need to be made to deal with this phenomenon. Under this condition, operations are performed based on the loss reduction ratio. These decisions include the following steps:

[0167] If the loss is reduced by <0.01, which means that the model learning progress is slow, and the strategy of increasing the regularization strength λ is adopted to reduce the risk of overfitting;

[0168] Let the current loss reduction ratio be , and the previous step loss reduction ratio is , the ratio of the two is the effect or efficiency E of each step in the current training process, then

[0169] ;

[0170] If E is close to 0 or very small, it means that the network training progress tends to stagnate or fall into a stable stage;

[0171] If it is found that the value of the efficiency E remains basically unchanged or very small during multiple consecutive inspection cycles, the learning rate is adjusted Reduced to half of the original value to help the training quickly reach a convergence state and save iteration time and resources;

[0172] The above steps are repeated until the efficiency E does not meet the condition of being less than 0.001, ensuring that the ideal model is obtained within a reasonable time frame.

[0173] To illustrate the example: when training a deep learning model, suppose that after three consecutive checks, there is little improvement in the loss.

[0174] Step 1: If the test result is <0.001, this means that the network's learning progress is very slow; in this case, the risk of overfitting can be reduced by increasing the coefficient of λ (for example, from 0.001 to 0.01);

[0175] Step 2: Let =0.0005 in the last iteration =0.001, then the efficiency , such a small value indicates that the network has almost stopped improving, and the parameters should be changed to advance the iteration process.

[0176] Step 3: If similar situations are observed for multiple iterations, the learning rate is reduced from 0.001 to 0.0005 ( =0.0005), which aims to improve the effectiveness of each iteration in reducing the overall error and accelerate the convergence of the model;

[0177] The inspection and adjustment steps are executed repeatedly to ensure that the network can eventually reach the ideal state. This process improves the training efficiency and the accuracy of the results by dynamically adjusting the learning speed, enhancing the data generalization ability, and evaluating the effectiveness of model training.

[0178] A job management method based on deep learning of the present invention includes: first, the technology uses a deep belief network as the basis of the initial deep learning model. This step solves the key problem of building an efficient machine learning architecture based on a specific algorithm. As an unsupervised machine learning framework, the deep belief network can achieve high-level abstract extraction through multi-layer nonlinear processing units, and is suitable for learning and pattern analysis of complex data sets. This method can effectively build a deep learning model that can understand and respond to a variety of situations, laying a good foundation for the management and control of the working environment.

[0179] Then, after the initial model is built, it will be used to predict the best resource allocation in a dynamic environment. In the operating environment, there are often problems such as uncertainty and demand fluctuations. This method can make the best choice by simulating and predicting the current resource usage in a dynamic environment. For example, applying this method in areas such as task scheduling or service provision can significantly improve the effectiveness and efficiency of resource allocation, thereby avoiding the expiration and redundancy caused by static planning.

[0180] In addition, in order to adapt to the changing environment and needs, the system needs to revise the original prediction results and configuration plans according to the real-time situation. This revision makes the resource allocation strategy more flexible and efficient. Specifically, when the system detects new data points or environmental changes, it will adjust the previous decision based on the updated information and recalculate the resource allocation plan to keep the policy relevance synchronized with the actual working conditions.

[0181] Finally, the system uses the latest feedback to continuously train and adjust the model. In this way, as the model experiences more different instances and obtains new information, it becomes more powerful and accurate, so that it can make more accurate predictions and optimization suggestions when facing future uncertain challenges, and finally form a self-improvement, iterative cycle of work. This method ensures that the deep learning model can continue to improve as practical data accumulates, so as to more accurately solve the problem of job management in complex scenarios.

[0182] The entire technical process revolves around the characteristics of the dynamic operating environment, and improves the performance of resource allocation and task planning through the application of deep learning, especially deep belief networks. It not only improves the response flexibility in the operating environment and reduces operation and maintenance costs, but also improves key indicators such as resource utilization and service level.

[0183] To better implement the above method, Figure 2 As shown, the present invention also proposes a job management system based on deep learning, the system comprising:

[0184] Initialization module 101, used to construct an initial deep learning model based on a deep belief network;

[0185] A resource configuration optimization module 102, for using the deep learning model to predict and optimize resource configuration in a dynamically changing operating environment;

[0186] The operating environment correction module 103 is used to dynamically correct the prediction results and resource allocation scheme according to the changes in the scene requirements in the operating environment;

[0187] The job management optimization module 104 is used to continuously train and adjust the deep learning model for optimization based on the revised prediction and resource allocation plan.

[0188] Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0189] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0190] In the description of the present application, it should be noted that the terms "upper" and "lower" etc. indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the inventive product is usually placed when in use. These are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0191] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A job management method based on deep learning, characterized in that: The method comprises: S1. Build an initial deep learning model based on a deep belief network; S2. Utilizing the deep learning model to predict and optimize resource allocation in a dynamically changing operating environment; S3, dynamically revise the prediction results and resource allocation plans according to the changes in scenario requirements in the operating environment; S4. Continuously train and adjust the deep learning model for optimization based on the revised prediction and resource allocation plan; The S1 includes: Preprocess a large amount of historical data, including data cleaning and labeling; Initialize the weight values ​​of the multi-layer probabilistic graphical model; Adjust the ratio of deep and shallow model structures through comparative analysis to improve computational efficiency; Layer-wise unsupervised pre-training using restricted Boltzmann machines; Finally, combined with the label data, supervised fine-tuning is completed through the back-propagation algorithm to establish a complete deep belief network model; Among them, when initializing the parameters of the multi-layer probabilistic graphical model, the following operations are performed based on specific constraints: The number of neurons N in each hidden layer must be greater than or equal to half of the number of nodes D in the input layer, as shown below: ； Select an appropriate regularization strength λ, and its value must satisfy 0≤λ≤0.001 to prevent overfitting; The initial learning rate is set to 0.01 to ensure effective update of weights during network training; If the learning rate is less than the threshold , then stop updating the weights and regard it as reaching a stable state; The S4 includes: Feedback to the deep learning model training set based on the last execution and the newly adjusted results to optimize the model structure; Update the learning rate strategy and regularization parameters to avoid overfitting and improve prediction accuracy; Further training the model to improve its adaptability and reliability in future job management decisions; Use the dynamic correction results as a new round of optimization training data to continuously improve the model's learning of resource scheduling in new environments; Iteratively update and test models to maintain predictive capabilities and improve the effectiveness of future resource allocation plans; In the process of adjusting the learning rate, based on specific conditions, if the current learning rate reaches or is less than the threshold T=1e-6, the operation is as follows: Calculate the reduction of the loss function in the i-th step of the current iteration as , and the reduction in the previous iteration , the average reduction ratio is defined as , which is expressed as follows: ； If the loss function changes tends to be stable, that is, ,in, is the ratio calculated last time, the current learning rate is reduced by a fixed coefficient; When three times in a row When , the current learning rate is reduced to half of the current learning rate to slow down the update process; Once the current learning rate drops to T, the learning update operation is immediately suspended to prevent falling into a training trap or oscillation and non-convergence; If after three consecutive checks, it is found that the update process still has no obvious progress, an optimized strategy is needed to deal with this phenomenon. Under this condition, the operation is performed based on the loss reduction ratio. The optimized strategy includes the following steps: The loss in the current iteration step is L, and the initial loss is , the loss is reduced , then the loss reduction ratio ; If the loss is reduced by , that is, the model learning progress is slow, and the strategy of increasing the regularization strength λ is adopted to reduce the risk of overfitting; wherein, the increase in regularization strength get , which is expressed as follows: ； in, , represents the weight of regularization strength; The current loss , and the reduction ratio of the previous iteration is recorded as ,in, , Represents the loss of the previous iteration. The ratio of the two is the current step efficiency E, which is expressed as follows: ； If E is close to 0, it indicates that the network learning process has entered a stable state; If the current step-size efficiency E is stable, the current learning rate is further reduced to half of the current level to promote the convergence process of training and reduce the cost of the number of iterations; Repeat the detection process until the loss change no longer meets the current step-size efficiency E < 0.001 condition to achieve the best learning effect; The S2 includes: Using the constructed deep learning model, input the historical usage of relevant resources in the current operating environment; Collect various real-time resource data of the current environment in real time, including processor load status, idle hardware resources, and demand changes of business tasks over time; By studying the changing trends of the operating environment characteristics, the matching degree of resource supply and demand is predicted; Use the forecast results to automatically allocate appropriate resources for tasks in the upcoming time period; By maximizing the objective function of the optimization effect, resource allocation is adjusted to ensure maximum resource utilization and system efficiency.

2. A job management method based on deep learning according to claim 1, characterized in that: The S3 includes: Identify new or unusual situations in the operating environment and analyze the potential impact of changes on resource allocation plans; Monitor the real-time resource status of the system and evaluate the effectiveness of the current resource distribution; Reassess the accuracy of forecasts and the effectiveness of currently allocated resources; Timely adjust job priorities and resource allocation plans based on new changes; Determine the correction plan and conduct verification and evaluation before implementation to ensure the consistency of resource allocation and changes in the operating environment.

3. The job management method based on deep learning according to claim 1, characterized in that: The step of constructing an initial deep learning model based on a deep belief network includes: Set up the basic architecture of the deep belief network, defining the number of hidden layers and their number of neurons; Determine the input feature set and output variables of the model; Establish a training sample set and initialize the initial values ​​of network parameters; The network is pre-trained according to the given learning rate and number of training rounds, and the optimization algorithm is set. The neural network parameters are updated based on the training samples until the set training accuracy is achieved.

4. The job management method based on deep learning according to claim 1, characterized in that: The matching degree prediction of resource supply and demand is performed by learning the changing trend of the operating environment characteristics, including: To set the number of hidden units N and the learning rate of each layer, the degree of change V of the operating environment characteristics must be considered. The specific operations are as follows: If the degree of change of the working environment characteristics V ≥ 7, each hidden layer adds N + 50% hidden units; If the degree of change in the operating environment characteristics V is less than 0.3, keep the default configuration unchanged; Dynamically adjust the learning rate to reduce the original The ratio of slowing down the change of model learning rate; Stop iterative updating of network weights. Stop training and updating when the environmental characteristics change significantly, that is, when V>0.95 or when the predetermined number of times is reached.

5. A job management system based on deep learning, characterized in that: The system comprises: Initialization module, used to build an initial deep learning model based on deep belief network; A resource configuration optimization module, used to use the deep learning model to predict and optimize resource configuration in a dynamically changing operating environment; The operating environment correction module is used to dynamically correct the prediction results and resource allocation plans according to the changes in the scene requirements in the operating environment; A job management optimization module, for continuously training and adjusting the deep learning model for optimization based on the revised forecast and resource allocation plan; The initialization module execution steps include: Preprocess a large amount of historical data, including data cleaning and labeling; Initialize the weight values ​​of the multi-layer probabilistic graphical model; Adjust the ratio of deep and shallow model structures through comparative analysis to improve computational efficiency; Layer-wise unsupervised pre-training using restricted Boltzmann machines; Finally, combined with the label data, supervised fine-tuning is completed through the back-propagation algorithm to establish a complete deep belief network model; Among them, when initializing the parameters of the multi-layer probabilistic graphical model, the following operations are performed based on specific constraints: The number of neurons N in each hidden layer must be greater than or equal to half of the number of nodes D in the input layer, as shown below: ； Select an appropriate regularization strength λ, and its value must satisfy 0≤λ≤0.001 to prevent overfitting; The initial learning rate is set to 0.01 to ensure effective update of weights during network training; If the learning rate is less than the threshold , then stop updating the weights and regard it as reaching a stable state; The operation management optimization module execution steps include: Feedback to the deep learning model training set based on the last execution and the newly adjusted results to optimize the model structure; Update the learning rate strategy and regularization parameters to avoid overfitting and improve prediction accuracy; Further training the model to improve its adaptability and reliability in future job management decisions; Use the dynamic correction results as a new round of optimization training data to continuously improve the model's learning of resource scheduling in new environments; Iteratively update and test models to maintain predictive capabilities and improve the effectiveness of future resource allocation plans; In the process of adjusting the learning rate, based on specific conditions, if the current learning rate reaches or is less than the threshold T=1e-6, the operation is as follows: Calculate the reduction of the loss function in the i-th step of the current iteration as , and the reduction in the previous iteration , the average reduction ratio is defined as , which is expressed as follows: ； If the loss function changes tends to be stable, that is, ,in, is the ratio calculated last time, the current learning rate is reduced by a fixed coefficient; When three times in a row When , the current learning rate is reduced to half of the current learning rate to slow down the update process; Once the current learning rate drops to T, the learning update operation is immediately suspended to prevent falling into a training trap or oscillation and non-convergence; If after three consecutive checks, it is found that the update process still has no obvious progress, an optimized strategy is needed to deal with this phenomenon. Under this condition, the operation is performed based on the loss reduction ratio. The optimized strategy includes the following steps: The loss in the current iteration step is L, and the initial loss is , the loss is reduced , then the loss reduction ratio ; If the loss is reduced by , that is, the model learning progress is slow, and the strategy of increasing the regularization strength λ is adopted to reduce the risk of overfitting; wherein, the increase in regularization strength get , which is expressed as follows: ； in, , represents the weight of regularization strength; The current loss , and the reduction ratio of the previous iteration is recorded as ,in, , Represents the loss of the previous iteration. The ratio of the two is the current step efficiency E, which is expressed as follows: ； If E is close to 0, it indicates that the network learning process has entered a stable state; If the current step-size efficiency E is stable, the current learning rate is further reduced to half of the current level to promote the convergence process of training and reduce the cost of the number of iterations; Repeat the detection process until the loss change no longer meets the current step-size efficiency E < 0.001 condition to achieve the best learning effect; The resource configuration optimization module execution includes: Using the constructed deep learning model, input the historical usage of relevant resources in the current operating environment; Collect various real-time resource data of the current environment in real time, including processor load status, idle hardware resources, and demand changes of business tasks over time; By studying the changing trends of the operating environment characteristics, the matching degree of resource supply and demand is predicted; Use the forecast results to automatically allocate appropriate resources for tasks in the upcoming time period; By maximizing the objective function of the optimization effect, resource allocation is adjusted to ensure maximum resource utilization and system efficiency.

Images