Causal inference based edge cloud device configuration optimization method and system

Abstract

The application discloses an edge cloud device configuration optimization method and system based on causal inference, and comprises the following steps: collecting device state features of all edge cloud devices, wherein the device state features comprise device attribute features, historical bandwidth and historical income; constructing a device set based on any feature in the device attribute features, training other features except the features used for constructing the device set by using a convolutional neural network and an X-Learner method to obtain an income estimation model; constructing an optimization objective function based on the income estimation model with the aim of maximizing the income; and solving the optimization objective function by using a simplex method to obtain an optimal edge cloud device configuration scheme. The application can give the optimal resource configuration of the edge cloud device, finally realizes optimal cost input and output, and realizes maximization of interests of edge cloud merchants.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of edge cloud devices, and particularly relates to an edge cloud device configuration optimization method and system based on causal inference. BACKGROUND

[0002] With the rapid development of 5G, Internet of Things and other technologies, cloud computing services are continuously expanding to the edge, thus generating the concept of edge cloud. The edge cloud is a cloud computing platform built on edge infrastructure, and these computing platforms are composed of node devices. In the edge cloud scenario, the edge cloud node devices may have a huge difference in node bandwidth running due to the heterogeneity of hardware configuration, which ultimately leads to a large difference in the income of each node. In this scenario, how to achieve the optimal hardware configuration scheme of cost investment and income improvement through a certain optimization algorithm is related to the utilization value of the server device and has a non-negligible causal relationship with the final profit of the edge cloud merchant.

[0003] The existing edge cloud device configuration optimization often adopts the method of artificial diagnosis, and has the following shortcomings: 1. It is difficult to fine-tune, often refers to the subjective experience of people, and it is difficult to diagnose each device specifically; 2. It does not evaluate the influence of device heterogeneity on running, but the traffic in the edge scenario is very unstable, and often fluctuates with time and business, which often leads to a large difference in the income of each node. SUMMARY

[0004] In view of the above problems, the application provides an edge cloud device configuration optimization method and system based on causal inference, which can not only be applied to the edge cloud scenario, but also be reused in multiple scenarios such as central cloud, IDC, CDN, terminal device and the like for automatic diagnosis of node devices. To solve the above technical problems, the technical scheme adopted by the application is as follows:

[0005] An edge cloud device configuration optimization method based on causal inference, comprising the following steps:

[0006] S1, collecting device state features of all edge cloud devices, wherein the device state features include device attribute features, historical bandwidth and historical income;

[0007] S2, constructing a device set based on any feature in the device attribute features, and training other features except the features used to construct the device set to obtain a revenue estimation model by using a convolutional neural network and an X-Learner method;

[0008] S3, constructing an optimization objective function based on the revenue estimation model established in step S2 to maximize the revenue;

[0009] S4, the optimal edge cloud device configuration scheme is obtained by solving the optimization objective function in step S3 using the simplex method.

[0010] The device attribute features include the number of CPUs, memory, CPU model, SSD disk model, number of SSD hard disks, SSD hard disk size, HHD disk model, number of HHD disks, and HHD disk size of the edge cloud device.

[0011] The step S2 includes the following steps:

[0012] S2.1, constructing a device set based on one feature in the device attribute features;

[0013] S2.2, training other device attribute features except the feature based on which the device set is constructed using a convolutional neural network to obtain a first revenue model and a second revenue model;

[0014] S2.3, training the first revenue model and the second revenue model based on an X-Learner method to obtain a revenue estimation model.

[0015] The step S2.3 includes the following steps:

[0016] i. training the first revenue model and the second revenue model in turn using the complete data features based on the device set in a permutation and combination manner as experimental and control groups;

[0017] ii. calculating the difference as fitting data using the output of the trained first revenue model and second revenue model respectively;

[0018] iii. creating two fitting models with the fitting data as the target according to the method of step i;

[0019] iv. weighting the two fitting models obtained in step iii to obtain the revenue estimation model.

[0020] The expression of the optimization objective function is:

[0021] maxf(X)=ατ0+(1-α)τ1;

[0022]

[0023] In the formula, p i represents the cost of the i-th edge cloud device, P represents the total cost, f(·) represents the model function of the revenue estimation model, α represents the weight, τ0 represents the first fitting model trained based on the X-Learner method, τ1 represents the second fitting model trained based on the X-Learner method, I represents the total number of edge cloud devices, m i represents the i-th edge cloud device.

[0024] A causal inference-based edge cloud device configuration optimization system, comprising:

[0025] a feature collection module: for collecting device state features of all edge cloud devices, the device state features comprising device attribute features and historical bandwidths;

[0026] a benefit model construction module: for constructing a device set based on one of the device attribute features collected by the feature collection module, and then training other device attribute features of the device set based on the one feature to establish a first benefit model and a second benefit model by using a convolutional neural network;

[0027] a benefit model optimization module: for training the first benefit model and the second benefit model established by the benefit model construction module to obtain a benefit prediction model based on an X-Learner method;

[0028] a configuration scheme generation module: for constructing an optimization objective function based on the benefit prediction model established by the benefit model optimization module to maximize benefits, and solving the optimization objective function by using a simplex method to generate an optimized edge cloud device configuration scheme.

[0029] Advantages of the present application:

[0030] The present application designs an X-Learner and an operational optimization algorithm according to the hardware performance of edge cloud devices, solves the problem of large income difference caused by different cost inputs of edge cloud devices, and can fully exploit the hardware configuration of edge cloud devices; the present application can give the optimal resource configuration of devices, finally realizes the optimization of cost input and output, and maximizes the interests of edge cloud vendors; helps vendors to evaluate the influence of cost input of each heterogeneous device on predicted income, and balances the mismatch between expenditure and benefits. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0032] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0033] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0034] X-Learner: solves the data volume difference problem in T-Learner through cross-training. It is based on T-learner and uses each observation in the training set on a formula similar to the "X" shape. X-learner has two advantages. First, it can prove to be adaptable to structure; second, it is particularly useful when the data volume of one treatment group (usually control) is much larger than that of another treatment, and the reason for developing this situation is that the data of the control group is relatively easy to obtain.

[0035] Linear programming: abbreviated as LP, is an important branch of operations research that is studied early, developed quickly, widely applied, and mature in method. It is a mathematical method to assist people in scientific management, and is a mathematical theory and method for studying the extreme value problem of linear objective function under linear constraint conditions. The basic method for solving linear programming problems is the simplex method, and there is standard software for the simplex method, which can solve linear programming problems with more than 10,000 constraints and decision variables on electronic computers.

[0036] An edge cloud device configuration optimization method based on causal inference includes the following steps:

[0037] S1, collect the device state features of all edge cloud devices, the device state features including device attribute features, historical bandwidth and historical revenue;

[0038] The device attribute features include the number of CPUs, memory, CPU model, SSD disk model, number of SSD hard disks, SSD hard disk size, HHD disk model, number of HHD disks, and HHD disk size of the edge cloud device. The historical bandwidth belongs to the running characteristics of the edge cloud device, that is, the running bandwidth flow of the edge cloud device.

[0039] S2, based on one of the device attribute features in step S1, a device set is constructed, and a revenue estimation model is obtained based on the device set using a convolutional neural network and an X-Learner method, including the following steps:

[0040] S2.1, based on one of the device attribute features, a device set is constructed;

[0041] In this application, M is used to represent the set of all edge cloud devices, M = {m1, m2, ..., m}. i , ..., m I}, m i Let I represent the i-th edge cloud device, and let I represent the total number of edge cloud devices. This represents the number of CPUs owned by the i-th edge cloud device. In this embodiment, a device set is constructed based on the number of CPUs. Since the number of CPUs is a non-continuous value, C is used to represent the set of types of CPU counts owned by edge cloud devices, where C = (c1, c2, ..., c...). n ), where n represents n different types of edge cloud devices with varying numbers of CPUs, when c1 indicates that the i-th edge cloud device has c1 CPUs. Therefore, the relationship between the number of edge cloud devices and the number of CPUs can be represented by the following table:

[0042] CPU number f1 f2 ... c n ]]> Device number [b1] [b2] ... b n ]]>

[0043] Where b1 represents the number of edge cloud devices with c1 CPUs, b2 represents the number of edge cloud devices with c2 CPUs, and b n Indicates having c n The number of edge cloud devices with c1 CPUs. Define B1 as the device set of edge cloud devices with c1 CPUs, and B2 as the device set of edge cloud devices with c2 CPUs. n For having c n A set of edge cloud devices with a number of CPUs. and

[0044] When considering the number of CPUs, device attributes other than the number of CPUs do not affect the results and are defined as X. Since there are n device sets with different numbers of CPUs in the edge cloud, each device set alternates between the experimental and control groups during the experiment. Based on the permutations and combinations, a total of [number] tests are required. Group experiment.

[0045] S2.2, use a convolutional neural network to train the device attribute features in the device set other than the features used in step S2 to obtain the first revenue model and the second revenue model;

[0046] The X-Learner method requires two models to be trained on the dataset. In this application, a CNN (Convolutional Neural Network) with one input layer, three hidden layers, and one output layer is selected as the training network model.

[0047] Other device attribute characteristics of the edge cloud devices with different CPU numbers, except the CPU numbers, are taken as inputs to the trained network model for twice training, the other device attribute characteristics adopt representing, obtaining a first revenue model and a second revenue model, the expressions of the outputs of the two models are:

[0048]

[0049] wherein Y represents the model output, f(·) represents the output function, N represents the total number of feature dimensions, W i′l represents the coefficient of the sample feature vector, represents the input sample, i' represents the feature dimension, and l represents the number of hidden layers.

[0050] In the training, the BP algorithm based on gradient descent is used to adjust Θ to make the loss tend to a minimum value, Θ is the parameter set of the MLP (Multi-Layer Perceptron) model, so that the model can more accurately predict the revenue through the characteristics of the edge cloud device, and the loss functions of the two models are both the root mean square error (RMSE), and the expression is:

[0051]

[0052] wherein L RMSE represents the root mean square error value, reflecting the gap between the predicted device revenue and the true label , and n' represents the number of input samples.

[0053] S2.3, training the first revenue model and the second revenue model based on the X-Learner method to obtain a revenue estimation model, including the following steps:

[0054] i. taking the complete data set based on the device set with different CPU numbers as the experimental group and the control group in a permutation and combination manner to train the first revenue model and the second revenue model in turn, and obtaining:

[0055]

[0056]

[0057] wherein Y1 represents the output of the first revenue model after training, Y0 represents the output of the second revenue model after training, f1(·) represents the first revenue model function, f0(·) represents the second revenue model function, E[·] represents the expectation function, and X represents the input.

[0058] ii. calculating the difference value as fitting data using the output of the trained model;

[0059]

[0060]

[0061] In the formula, D1 represents the difference in the first return model, and D0 represents the difference in the second return model. This indicates that when the input dataset is At that time, the predicted profit value output by the trained second profit model, This indicates that when the input dataset is At time t, the predicted profit value output by the first profit model after training, t, q∈[1,…,k], and t≠q.

[0062] iii. Following the method in step i, create two new fitting models with the fitting data as the target;

[0063] τ1=E[D1|X]; (7)

[0064] τ0=E[D0|X]; (8)

[0065] In the formula, τ1 represents the first fitting model established based on the difference of the first profit model, and τ0 represents the second fitting model established based on the difference of the second profit model.

[0066] iv. Weight the two fitted models obtained in step iii to obtain the profit prediction model;

[0067] In summary, there are currently two CNN models. To make the results more scalable, the models are weighted, and the expression for the profit prediction model is as follows:

[0068] f(x) = ατ0 + (1-α)τ1; (9)

[0069] In the formula, f(·) represents the model function of the revenue prediction model, and α represents the weight. Edge cloud devices can change their revenue by altering device attribute characteristics (such as the number of CPUs) according to the revenue prediction model.

[0070] S3. Based on the profit prediction model established in step S2, construct an optimization objective function with the goal of maximizing profit;

[0071] Since each input set of data yields a difference in revenue, it is necessary to find the attribute change corresponding to the maximum revenue increment. Furthermore, due to cost constraints, the attributes of edge cloud devices are limited; therefore, a linear programming model is considered to find the attribute value that maximizes revenue under these constraints.

[0072] The expression for the optimization objective function is:

[0073] maxf(X) = aτ0 + (1-a)τ1; (10)

[0074]

[0075] where p i represents the cost of the ith edge cloud device, and P represents the total cost.

[0076] The above function is to maximize the output of the model, thereby obtaining the attribute change with the largest increment of revenue. In the linear programming model, the optimization objective function can be written as:

[0077]

[0078] S4, using the simplex method to solve the optimization objective function in step S3 to obtain the optimal edge cloud device configuration scheme;

[0079] First, the original linear programming problem in the present application is standardized to obtain:

[0080] maxf(X) = aτ0 + (1-a)τ1; (12)

[0081]

[0082] m i , p i , s1≥0;

[0083] In the formula, s1 is a slack variable. Some variables are specified as 0, so that the values of other variables can be uniquely solved through the equation constraint. For the above changed constraint condition, the variable coefficient on the left side of the equation is formed into a matrix A, and then all the variable

[0084] coefficients are included in the simplex table:

[0085] f(X) m n ]]> p n ]]> [s1] 0 row 1 0 0 0 1 row 0 1 1 1

[0086] It can be seen from the table that the initial basic variables are (f(X), m i ), and then a variable that can enter the base is found from the table, which requires that the variable can make the increment of the objective function value maximum after entering the base. Since the coefficients of the remaining two variables are the same, and the influence degree on the objective function is the same, any one variable can be selected to enter the base, and the remaining variable is the out-of-base variable. Take the intersection of the rows and columns where the two variables are located as the axis to perform elementary transformation. The updated constraint condition is:

[0087] m1×p1+m2×p2+s1=P;

[0088] m1, m2, p1, p2, s1≥0;

[0089] In the formula, m1 is a promotion variable, and m2 is a demotion variable. The above steps are repeated, and one variable is eliminated each time. When all values in the 0th row are positive, an optimal solution can be obtained.

[0090] The above embodiments are examples of optimal hardware configuration schemes based on CPU quantity cost configuration and benefit improvement. In practice, there are many factors that affect cost configuration, such as CPU model, memory, SSD disk model, SSD hard disk quantity, SSD hard disk size, HHD disk model, HHD disk quantity, and HHD disk size. Specifically, the present application can be applied to configuration selection of various influencing factors, the difference being that different device attribute characteristics are selected when constructing a device set, so as to maximize the expected benefit through optimal hardware configuration selection.

[0091] The embodiment of the present application also provides an edge cloud device configuration optimization system based on causal inference, comprising:

[0092] a feature collection module: configured to collect device state features of all edge cloud devices, wherein the device state features comprise device attribute features and historical bandwidth;

[0093] a benefit model construction module: configured to construct a device set based on one of the device attribute features collected by the feature collection module, and then train other device attribute features of the device set based on the one feature to establish a first benefit model and a second benefit model by using a convolutional neural network;

[0094] a benefit model optimization module: configured to train the first benefit model and the second benefit model established by the benefit model construction module to obtain a benefit estimation model based on an X-Learner method;

[0095] a configuration scheme generation module: configured to construct an optimization objective function based on the benefit estimation model established by the benefit model optimization module to maximize the benefit, and solve the optimization objective function by using a simplex method to generate an optimal edge cloud device configuration scheme.

[0096] The embodiment of the present application also provides an electronic device, comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is executed by the processor to implement the edge cloud device configuration optimization method based on causal inference.

[0097] The embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by the processor to implement the edge cloud device configuration optimization method based on causal inference.

[0098] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for optimizing the configuration of edge cloud devices based on causal inference, characterized in that, Comprising the following steps: S1, collecting device state features of all edge cloud devices, the device state features comprising device attribute features, historical bandwidth and historical revenue; S2, constructing a device set based on one of the device attribute features, training other features of the device set based on the constructed device set by using a convolutional neural network and an X-Learner method to obtain a revenue estimation model; S3, constructing an optimization objective function based on the revenue estimation model established in step S2 with the goal of maximizing revenue; S4, solving the optimization objective function in step S3 by using a simplex method to obtain an optimized edge cloud device configuration scheme; The expression of the optimization objective function is: ； ； wherein, represents the cost of the th edge cloud device, represents the total cost, represents a model function of the revenue estimation model, represents a weight, represents a first fitted model trained based on an X-Learner method, represents a second fitted model trained based on an X-Learner method, represents the total number of edge cloud devices, represents the th edge cloud device, represents an input sample; The original linear programming problem corresponding to the optimization objective function is standardized, and the constraint condition is updated as: ； ； In the formula, is the relaxation variable.

2. The method of claim 1, wherein, The device attribute features comprise the number of CPUs, memory, CPU model, SSD disk model, number of SSD hard disks, size of SSD hard disks, HHD disk model, number of HHD hard disks and size of HHD hard disks of the edge cloud device.

3. The method of claim 1, wherein, The step S2 comprises the following steps: S2.1, constructing a device set based on one of the device attribute features; S2.2, training other device attribute features of the device set based on the features by using a convolutional neural network to obtain a first revenue model and a second revenue model; S2.3, training the first revenue model and the second revenue model by using an X-Learner method to obtain a revenue estimation model.

4. The method of claim 3, wherein, The step S2.3 comprises the following steps: i, training the first revenue model and the second revenue model in turn by using the complete data features based on the device set in a permutation and combination manner as experimental groups and control groups respectively; ii, calculating the difference as fitting data by using the outputs of the trained first revenue model and the second revenue model respectively; iii, creating two fitting models with the fitting data as the target according to the method of step i; iv, weighting the two fitting models obtained in step iii to obtain a revenue estimation model.

5. A system for optimizing configuration of edge cloud devices based on causal inference, the system comprising: Comprise: A feature collection module for collecting device state features of all edge cloud devices, the device state features comprising device attribute features and historical bandwidth; A revenue model construction module for constructing a device set based on one of the device attribute features collected by the feature collection module, and then training other device attribute features of the device set based on the features by using a convolutional neural network to establish a first revenue model and a second revenue model; A revenue model optimization module for training the first revenue model and the second revenue model established by the revenue model construction module by using an X-Learner method to obtain a revenue estimation model; A configuration scheme generation module for constructing an optimization objective function based on the revenue estimation model established by the revenue model optimization module with the goal of maximizing revenue, and solving the optimization objective function by using a simplex method to generate an optimized edge cloud device configuration scheme; The expression of the optimization objective function is: ； ； wherein, denotes the cost of the edge cloud device, denotes the total cost, denotes a model function of the revenue estimation model, denotes a weight, denotes a first fitted model trained based on the X-Learner method, denotes a second fitted model trained based on the X-Learner method, denotes the total number of edge cloud devices, denotes the edge cloud device, denotes an input sample; The original linear programming problem corresponding to the optimization objective function is standardized, and the constraint condition is updated as: ； ； In the formula, is a slack variable.

Images