Mathematical model calling system and method based on cloud service

Abstract

The invention discloses a mathematical model calling system and method based on cloud service, and the method comprises the steps: a problem analysis module carries out the semantic analysis and feature extraction of a problem description inputted by a user, and generates a standardized problem feature vector; the model feature library stores a predefined mathematical model feature vector set; the dynamic matching module calculates the matching degree of the problem and the model through a hybrid similarity algorithm; the resource scheduling module establishes a multi-objective optimization model according to the model calculation complexity and the current cloud resource state; and the execution engine module automatically generates an adaptive interface, calls the selected mathematical model and monitors the execution state in real time. The method has the advantages that intelligence, elasticity and automation of mathematical model calling are achieved, and the method has remarkable advantages in large-scale complex problem solving in the cloud computing environment.

Description

Technical Field

[0001] The present invention relates to cloud services and model calling, and in particular to a mathematical model calling system and method based on cloud services. Background Art

[0002] With the rapid development of cloud computing technology, cloud services have become an important infrastructure for data processing and computing tasks in all walks of life. In traditional computing models, resources are usually fixed, and the deployment process is complex and lacks flexibility. The mathematical model calling system based on cloud services uses the elastic resources of the cloud platform to provide dynamic computing and efficient resource scheduling capabilities. The high availability and fault tolerance mechanism of the cloud service platform can ensure the stability of model calculations and avoid the risks brought by single point failures. The mathematical model calling system based on cloud services not only improves computing power, but also reduces the user's operation and maintenance costs, making model development, deployment and optimization more convenient, especially suitable for application scenarios that require processing large-scale data and real-time computing.

[0003] The current mathematical model calling systems on the market use cloud computing platforms, API interfaces, and specialized modeling software to call and deploy models. Users can call deployed models through API interfaces to achieve flexible data interaction and automated processing of computing tasks. In addition, traditional mathematical modeling software provides a wealth of mathematical libraries and visualization tools to facilitate users to create, test, and optimize models. With the development of artificial intelligence and machine learning technologies, some advanced model calling systems have also begun to integrate automated learning and adjustment functions. However, compared with cloud-based methods, traditional methods still have gaps in computing resource limitations and scalability, making it difficult to cope with the needs of large-scale data processing and complex models. Summary of the invention

[0004] In order to improve the existing mathematical model calling system and method, a mathematical model calling system and method based on cloud service is provided. The method obtains the best mathematical model through dynamic matching of problem characteristics, optimizes resource scheduling during the model calling process, improves system computing power and reduces operating load.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A mathematical model calling system based on cloud service, comprising:

[0007] Question analysis module: The question analysis module is used to perform semantic analysis and feature extraction on the question description input by the user, and generate a standardized question feature vector Q = (q 1 ,q 2 ,...,q n );

[0008] Model feature library: The model feature library is used to store predefined mathematical model feature vector sets M = {M 1 ,M 2 ,...,M m}, where each model feature vector M j =(m 1 ,m 2 ,...,m n ) contains the computational complexity c j , applicable fields j , input dimension i j , output type o j parameter;

[0009] Dynamic matching module: The dynamic matching module is used to calculate the matching degree between the question and the model through a hybrid similarity algorithm;

[0010] Resource scheduling module: The resource scheduling module is used to calculate the complexity c according to the model j and current cloud resource status, and establish a multi-objective optimization model;

[0011] Execution engine module: The execution engine module is used to automatically generate an adaptation interface and call the selected mathematical model, and monitor the execution status in real time.

[0012] Preferably, the model feature library specifically includes:

[0013] Establish a mathematical model metadata database, including model structure, algorithm complexity, and historical call records;

[0014] A feature encoder is used to map the model parameters to n-dimensional Euclidean space;

[0015] The feature vector is updated through the online learning mechanism, and the formula is:

[0016]

[0017] Among them, η is the learning rate, which controls the step size of the update. is the gradient of the loss function with respect to the feature vector.

[0018] Preferably, the dynamic matching module specifically includes:

[0019] The domain adaptation subunit adjusts the feature weights through transfer learning. The formula is:

[0020] X′ target =T(X source ,X target )

[0021] Among them, X source is the source domain distribution, Xtarget is the target domain distribution, T is the learned mapping function;

[0022] The uncertainty evaluation unit is used to calculate the matching confidence, and the formula is:

[0023]

[0024] Among them, Conf is the matching confidence, proj M(Q) is the projection of the model feature vector space.

[0025] Preferably, the resource scheduler specifically includes:

[0026] Get the model computational complexity c j and cloud resource status R j ;

[0027] Based on the model calculation complexity c j and cloud resource status R j , calculate the maximum value of the objective function, the formula is:

[0028] Minimize computational complexity:

[0029] Maximize cloud resource utilization:

[0030] Model optimization is performed based on multi-objective functions, and the formula is:

[0031]

[0032] Preferably, the execution engine module specifically includes:

[0033] An automatic interface generation unit, wherein the automatic interface generation unit is based on a model input dimension i j , output type o j Generate adaptation code;

[0034] A real-time performance analysis unit that records the actual calculation time t real and the estimated time t pred In contrast, the update complexity parameter is as follows:

[0035]

[0036] Among them, k is the learning rate factor;

[0037] Model effectiveness index unit, the model effectiveness index unit is used to optimize the model selection strategy in the long term, and the formula is:

[0038]

[0039] Among them, PEI j is the model effectiveness index, S j is the performance index of mathematical model j.

[0040] A cloud service-based mathematical model calling method, comprising:

[0041] Receive user questions and generate structured queries through natural language processing;

[0042] Retrieve candidate models from the model feature library to obtain a model set with a high matching degree;

[0043] Based on the obtained high-matching model set, construct a resource optimization problem;

[0044] Obtain the optimal model combination based on the branch and bound algorithm;

[0045] Dynamically allocate computing resources and monitor the execution process, and perform resource reallocation based on response latency.

[0046] Preferably, searching the candidate models in the model feature library to obtain a model set with a high matching degree specifically includes:

[0047] Domain label d based on domain adapter unit j Rough screening to obtain the mathematical model matching the field;

[0048] Fine screening is performed based on the matching confidence of the uncertainty assessment unit to obtain a model set with a high matching degree.

[0049] Preferably, the step of obtaining the optimal model combination based on the branch and bound algorithm specifically includes:

[0050] The model combination problem is modeled as a bipartite graph G = (U, V, E), with edge weights S (Q, M j )-η·c j ;

[0051] Based on the improved Hungarian algorithm, the maximum weight matching is obtained;

[0052] When the resource constraints are not met, the edges with the smallest weights are iteratively removed until it is feasible.

[0053] Preferably, the step of obtaining the maximum weight matching based on the improved Hungarian algorithm specifically includes:

[0054] Transform the weights of all edges of the weight matrix A and construct a new matrix B, where B[i][j] = max(A)-A[i][j];

[0055] For each row, perform B'[i][j]=B[i][j]-minj B[i][j];

[0056] For each column, perform B ‘ '[i][j]=B'[i][j]-min j B'[i][j];

[0057] In the matrix B ‘ 'Get as many zeros as possible so that each zero appears only once in the row and column to establish a match;

[0058] If the number of zeros is less than n, adjust the rows and columns until the condition is met.

[0059] Preferably, when the model fails to execute, an alternative solution is initiated;

[0060] The second best model is selected based on the set of models that obtain the match, and the formula is:

[0061]

[0062] Among them, ξ is the resource difference penalty coefficient, which is used to measure the impact of resource differences on the selection of alternative models;

[0063] Establish a retry decision model, the formula is:

[0064] Retry=σ(β 0 +β 1 S+β 2 ·R remain )

[0065] Among them, σ() is the sigmoid function, R remain is the current remaining resources in the cloud, β 0 , β 1 , β 2 is the model parameter.

[0066] Compared with the prior art, the advantages of the present invention are:

[0067] Through the multi-module collaboration of problem analysis, dynamic matching, resource scheduling and execution engine, the hybrid similarity algorithm and transfer learning are used to achieve accurate screening of cross-domain models, and the multi-objective optimization and branch-and-bound algorithm are combined to dynamically balance the computational complexity and resource utilization, and the model features and selection strategies are continuously optimized through online learning and performance index. It supports natural language interaction, automatic interface generation and elastic fault-tolerant mechanism, significantly improving model matching accuracy (20%-30%), resource efficiency (15%-25%) and fault recovery speed, and is suitable for efficient solution of complex problems in the cloud, realizing intelligent, adaptive and low-latency mathematical model calling full-process automation service. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A system schematic diagram of the system and method proposed by the present invention;

[0069] Figure 2 A method schematic diagram of the system and method proposed by the present invention;

[0070] Figure 3 A dynamic matching schematic diagram of the system and method proposed by the present invention;

[0071] Figure 4 A schematic diagram of obtaining the optimal model combination of the system and method proposed by the present invention;

[0072] Figure 5 A schematic diagram of the Hungarian algorithm of the system and method proposed in the present invention;

[0073] Figure 6 A schematic diagram of resource calculation and reallocation of the system and method proposed in the present invention;

[0074] Figure 7 This is a structural diagram of the electronic device proposed by the present invention;

[0075] Figure 8 This is a schematic diagram of the computer-readable storage medium structure proposed by the present invention. DETAILED DESCRIPTION

[0076] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0077] See also Figure 1 As shown, a mathematical model calling system based on cloud service includes:

[0078] Question analysis module: The question analysis module is used to perform semantic analysis and feature extraction on the question description input by the user, and generate a standardized question feature vector Q = (q 1 ,q 2 ,...,q n );

[0079] Model feature library: The model feature library is used to store predefined mathematical model feature vector sets M = {M 1 ,M 2 ,...,M m}, where each model feature vector M j =(m 1 ,m 2 ,...,m n ) contains the computational complexity c j , applicable fields j , input dimension ij , output type o j parameter;

[0080] Dynamic matching module: The dynamic matching module is used to calculate the matching degree between the question and the model through a hybrid similarity algorithm;

[0081] Resource scheduling module: The resource scheduling module is used to calculate the complexity c according to the model j and current cloud resource status, and establish a multi-objective optimization model;

[0082] Execution engine module: The execution engine module is used to automatically generate an adaptation interface and call the selected mathematical model, and monitor the execution status in real time.

[0083] See also Figure 1 As shown in the figure, the model feature library specifically includes:

[0084] Establish a mathematical model metadata database, including model structure, algorithm complexity, and historical call records;

[0085] A feature encoder is used to map the model parameters to n-dimensional Euclidean space;

[0086] The feature vector is updated through the online learning mechanism, and the formula is:

[0087]

[0088] Among them, η is the learning rate, which controls the step size of the update. is the gradient of the loss function with respect to the feature vector.

[0089] Specifically, in order to establish a mathematical model metadata database, a unified structure can be first designed to store the basic information of the model, including the model structure (such as the number of layers and nodes of the neural network), algorithm complexity (time complexity and space complexity), and historical call records (model application scenarios, execution time, performance evaluation, etc.). On this basis, a feature encoder (such as a deep encoder) is used to map the parameters of each model into an n-dimensional Euclidean space, and the characteristics of the model are expressed through high-dimensional vectors. An online learning mechanism is introduced to continuously optimize the model. The feature vector is continuously updated according to new data. In each iteration, the learning rate is used to control the step size of the parameter update, and the learning rate can be fixed or dynamically adjusted. The error between the feature vector and the target value is calculated through the loss function, and the gradient of the loss function to the feature vector is solved using optimization algorithms such as the gradient descent method, thereby adjusting the value of the feature vector.

[0090] See also Figure 1 As shown, the dynamic matching module specifically includes:

[0091] The domain adaptation subunit adjusts the feature weights through transfer learning. The formula is:

[0092] X′ target =T(X source ,X target )

[0093] Among them, X source is the source domain distribution, X tar get is the target domain distribution, T is the learned mapping function;

[0094] The uncertainty evaluation unit is used to calculate the matching confidence, and the formula is:

[0095]

[0096] Among them, Conf is the matching confidence, proj M(Q) is the projection of the model feature vector space.

[0097] See also Figure 1 As shown, the resource scheduler specifically includes:

[0098] Get the model computational complexity c j and cloud resource status R j ;

[0099] Based on the model calculation complexity c j and cloud resource status R j , calculate the maximum value of the objective function, the formula is:

[0100] Minimize computational complexity:

[0101] Maximize cloud resource utilization:

[0102] Model optimization is performed based on multi-objective functions, and the formula is:

[0103]

[0104] It is understandable that in multi-objective optimization, minimizing computational complexity and maximizing resource utilization often conflict. For example, reducing computational complexity may require sacrificing some resource utilization, and vice versa. How to balance these two goals and avoid extreme solutions is a difficult problem. We can make trade-offs by assigning different weights to the objective function, or use the Pareto optimal solution to find a compromise between different objectives. At the same time, we can use the method of dynamically adjusting the weight of the objective function to adjust the weight of the optimization goal according to the current resource status or demand changes of the system to avoid over-optimization of a single goal.

[0105] See also Figure 1 As shown, the execution engine module specifically includes:

[0106] An automatic interface generation unit, wherein the automatic interface generation unit is based on a model input dimension i j , output type o j Generate adaptation code;

[0107] A real-time performance analysis unit that records the actual calculation time t real and the estimated time t pred In contrast, the update complexity parameter is as follows:

[0108]

[0109] Among them, k is the learning rate factor;

[0110] Model effectiveness index unit, the model effectiveness index unit is used to optimize the model selection strategy in the long term, and the formula is:

[0111]

[0112] Among them, PEI j is the model effectiveness index, S j is the performance index of mathematical model j.

[0113] Specifically, when generating the adaptation code, the adaptation interface is automatically generated according to the model input and output requirements. For example, if the input is a two-dimensional matrix, the generated code will ensure that the input data conforms to the matrix structure. If the model is a classification model, the generated code will automatically adapt the output to the category label. If it is a regression model, the generated code will ensure that the output is a continuous value.

[0114] See also Figure 2 As shown, a cloud service-based mathematical model calling method includes:

[0115] Step 1: Receive user questions and generate structured queries through natural language processing;

[0116] Step 2: Retrieve candidate models from the model feature library to obtain a model set with a high matching degree;

[0117] Step 3: Based on the obtained high-matching model set, construct a resource optimization problem;

[0118] Step 4: Obtain the optimal model combination based on the branch and bound algorithm;

[0119] Step 5: Dynamically allocate computing resources and monitor the execution process, and perform resource reallocation based on response delay.

[0120] See also Figure 3As shown, searching for candidate models in the model feature library and obtaining a model set with a high matching degree specifically includes:

[0121] Domain label d based on domain adapter unit j Rough screening to obtain the mathematical model matching the field;

[0122] Fine screening is performed based on the matching confidence of the uncertainty assessment unit to obtain a model set with a high matching degree.

[0123] See also Figure 4 As shown in the figure, obtaining the optimal model combination based on the branch and bound algorithm specifically includes:

[0124] The model combination problem is modeled as a bipartite graph G = (U, V, E), with edge weights S (Q, M j )-η·c j ;

[0125] Based on the improved Hungarian algorithm, the maximum weight matching is obtained;

[0126] When the resource constraints are not met, the edges with the smallest weights are iteratively removed until it is feasible.

[0127] See also Figure 5 As shown in the figure, based on the improved Hungarian algorithm, obtaining the maximum weight matching specifically includes:

[0128] Transform the weights of all edges of the weight matrix A and construct a new matrix B, where B[i][j] = max(A)-A[i][j];

[0129] For each row, perform B'[i][j]=B[i][j]-min j B[i][j];

[0130] For each column, perform B ‘ '[i][j]=B'[i][j]-min j B'[i][j];

[0131] In the matrix B ‘ 'Get as many zeros as possible so that each zero appears only once in the row and column to establish a match;

[0132] If the number of zeros is less than n, adjust the rows and columns until the condition is met.

[0133] It is understandable that in the process of converting the weight matrix A to B, the maximum value of each column needs to be calculated, which may lead to high computational complexity, especially when processing large-scale data, so this process can be accelerated by preprocessing or parallel computing. Pre-calculate the maximum value of each column in the matrix and store it in an auxiliary array so that it can be quickly used in subsequent conversions. At the same time, use a distributed computing framework to parallelize the calculation of these maximum values ​​to further reduce time overhead.

[0134] See also Figure 6 As shown, dynamically allocating computing resources and monitoring the execution process, and performing resource reallocation based on response delay specifically include:

[0135] When the model fails to execute, the alternative plan is launched;

[0136] The second best model is selected based on the set of models that obtain the match, and the formula is:

[0137]

[0138] Among them, ξ is the resource difference penalty coefficient, which is used to measure the impact of resource differences on the selection of alternative models;

[0139] Establish a retry decision model, the formula is:

[0140] Retry=σ(β 0 +β 1 S+β 2 ·R remain )

[0141] Among them, σ() is the sigmoid function, R remain is the current remaining resources in the cloud, β 0 , β 1 , β 2 is the model parameter.

[0142] Specifically, The value of this function is between [0, 1]. α is a model parameter used to adjust the slope of the sigmoid function and control the sensitivity of the remaining resources to the retry decision. Before scheduling a task, we predict the changing trend of cloud resources and reserve certain resources to reduce retries caused by resource fluctuations. By accumulating historical task scheduling data, we continuously optimize the parameter β. 0 , β 1 , β 2 , making the retry decision model gradually tend to the optimal.

[0143] Furthermore, the method according to the embodiment of the present application can also be performed by Figure 7 The electronic device architecture shown in FIG. Figure 7As shown, the electronic device 500 may include a bus 501, one or more CPUs 502, a read-only memory (ROM) 503, a random access memory (RAM) 504, a communication port 505 connected to a network, an input / output component 506, a hard disk 507, etc. The storage device in the electronic device 500, such as the ROM 503 or the hard disk 507, may store a mathematical model calling system and method based on cloud services provided by the present application. The electronic device 500 may also include a terminal interface 508. Of course, Figure 7 The architecture shown is only exemplary and can be omitted according to actual needs when implementing different devices. Figure 7 One or more components of an electronic device are shown.

[0144] Figure 8 Schematic diagram of a computer-readable storage medium structure provided by an embodiment of the present application. Figure 8 As shown, a computer-readable storage medium 600 according to an embodiment of the present application is shown. Computer-readable instructions are stored on the computer-readable storage medium 600. When the computer-readable instructions are executed by the processor, a cloud-based mathematical model calling system and method according to an embodiment of the present application described with reference to the above figures can be executed. The storage medium 600 includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory (cache), etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0145] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The above is a description of a specific embodiment of this specification. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0146] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0147] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A mathematical model calling system based on cloud service, characterized in that: include: Question analysis module: The question analysis module is used to perform semantic analysis and feature extraction on the question description input by the user, and generate a standardized question feature vector Q = (q1, q2, ..., q n ); Model feature library: The model feature library is used to store a predefined mathematical model feature vector set M = {M1, M2, ..., M m }, where each model feature vector M j =(m1,m2,...,m n ) contains the computational complexity c j , applicable fields j , input dimension i j , output type o j parameter; Dynamic matching module: The dynamic matching module is used to calculate the matching degree between the question and the model through a hybrid similarity algorithm; Resource scheduling module: The resource scheduling module is used to calculate the complexity c according to the model j and current cloud resource status, and establish a multi-objective optimization model; Execution engine module: The execution engine module is used to automatically generate an adaptation interface and call the selected mathematical model, and monitor the execution status in real time.

2. A cloud service-based mathematical model calling system according to claim 1, characterized in that: The model feature library specifically includes: Establish a mathematical model metadata database, including model structure, algorithm complexity, and historical call records; A feature encoder is used to map the model parameters to n-dimensional Euclidean space; The feature vector is updated through the online learning mechanism, and the formula is: Among them, η is the learning rate, which controls the step size of the update. is the gradient of the loss function with respect to the feature vector.

3. The cloud service-based mathematical model calling system according to claim 1, characterized in that: The dynamic matching module specifically includes: The domain adaptation subunit adjusts the feature weights through transfer learning. The formula is: X′ target =T(X source ,X target ) Among them, X source is the source domain distribution, X target is the target domain distribution, T is the learned mapping function; The uncertainty evaluation unit is used to calculate the matching confidence, and the formula is: Among them, Conf is the matching confidence, proj M(Q) is the projection of the model feature vector space.

4. The cloud service-based mathematical model calling system according to claim 1, characterized in that: The resource scheduler specifically includes: Get the model computational complexity c j and cloud resource status R j ; Based on the model calculation complexity c j and cloud resource status R j , calculate the maximum value of the objective function, the formula is: Minimize computational complexity: Maximize cloud resource utilization: Model optimization is performed based on multi-objective functions, and the formula is:

5. The cloud service-based mathematical model calling system according to claim 1, characterized in that: The execution engine module specifically includes: An automatic interface generation unit, the automatic interface generation unit generating adaptation code based on the model input and output mode; A real-time performance analysis unit that records the actual calculation time t real and the estimated time t pred In contrast, the update complexity parameter is as follows: Where k is the learning rate factor; Model effectiveness index unit, the model effectiveness index unit is used to optimize the model selection strategy in the long term, and the formula is: Among them, PEI j is the model effectiveness index, S j is the performance index of mathematical model j.

6. A cloud service-based mathematical model calling method, applicable to the cloud service-based mathematical model calling system according to any one of claims 1 to 5, characterized in that: include: Receive user questions and generate structured queries through natural language processing; Retrieve candidate models from the model feature library to obtain a model set with a high matching degree; Based on the obtained high-matching model set, construct a resource optimization problem; Obtain the optimal model combination based on the branch and bound algorithm; Dynamically allocate computing resources and monitor the execution process, and perform resource reallocation based on response latency.

7. The cloud service-based mathematical model calling method according to claim 6, characterized in that: The step of retrieving candidate models from the model feature library to obtain a model set with a high matching degree specifically includes: Domain label d based on domain adapter unit j Rough screening to obtain the mathematical model matching the field; Fine screening is performed based on the matching confidence of the uncertainty assessment unit to obtain a model set with a high matching degree.

8. The cloud service-based mathematical model calling method according to claim 6, characterized in that: The method of obtaining the optimal model combination based on the branch and bound algorithm specifically includes: The model combination problem is modeled as a bipartite graph G = (U, V, E), with edge weights S (Q, M j )-η·c j ; Based on the improved Hungarian algorithm, the maximum weight matching is obtained; When the resource constraints are not met, the edges with the smallest weights are iteratively removed until it is feasible.

9. The cloud service-based mathematical model calling method according to claim 8, characterized in that: The method of obtaining the maximum weight matching based on the improved Hungarian algorithm specifically includes: Transform the weights of all edges of the weight matrix A and construct a new matrix B, where B[i][j] = max(A)-A[i][j]; For each row, perform B'[i][j]=B[i][j]-min j B[i][j]; For each column, perform B''[i][j]=B'[i][j]-min j B'[i][j]; Get as many zeros as possible in the matrix B'' so that each zero appears only once in the row and column, and establish a match; If the number of zeros is less than n, adjust the rows and columns until the condition is met.

10. The cloud service-based mathematical model calling method according to claim 6, characterized in that: The dynamically allocating computing resources and monitoring the execution process, and performing resource reallocation based on response delay specifically include: When the model fails to execute, the alternative plan is launched; The second best model is selected based on the set of models that obtain the match, and the formula is: Among them, ξ is the resource difference penalty coefficient, which is used to measure the impact of resource differences on the selection of alternative models; Establish a retry decision model, the formula is: Retry=σ(β0+β1·S+β2·R remain ) Among them, σ() is the sigmoid function, R remain is the current remaining resources in the cloud, and β0, β1, and β2 are model parameters.

Images