KPI optimization system and cloud-based large-scale work processing system for providing it

Abstract

To provide a KPI optimization system and a cloud-based large-scale job processing system to provide it.SOLUTION: A KPI optimization system is a system that performs prescriptive analysis using an evolutionary algorithm. The KPI optimization system includes: an action value output unit that executes the evolutionary algorithm to update an action value; a digital twin that is a model for predicting a KPI preceding index by receiving the input of an environmental factor value and the action value which is output by the action value output unit. A cloud-based large-scale job processing system is a system allowing division of a large-scale job. The cloud-based large-scale job processing system includes a job manager that when receiving assignment of a large-scale job from a supervisor, divides the large-scale job into a plurality of jobs and stores the plurality of jobs in a queue.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a KPI optimization system and a cloud-based large-scale work processing system for providing the same. The KPI optimization system is based on prescriptive analysis using an evolutionary algorithm (EA), and the cloud-based large-scale work processing system is based in particular on work division. [Background technology]

[0002] The business environment since COVID-19 has become more difficult and uncertain than ever before. Digital transformation is required in all industrial sectors, and the number of variables to consider in business management is continually increasing. It is now more important than ever to make quick and accurate operational strategy decisions that will determine the survival of a company.

[0003] Before deciding on an operating strategy, business managers can use predictive analytics. Predictive analytics is a method of predicting future operating conditions after analyzing current and past data as operating big data and learning a predictive model. When the predictive model predicts future operating conditions, business managers can decide on an operating strategy based on the predicted results. However, business managers have the disadvantage of relying on traditional rule-based or experience-based methods. This is because an operating strategy determined by a traditional method is not necessarily the best operating strategy. For example, if a predictive model predicts that the sales of Company A on the following Monday will decrease by 10% compared to the sales on the previous Monday, the business manager must decide which price change to make to which product in order to reduce the amount of sales decrease or to reverse the sales increase, but traditional methods alone have limitations in determining the best operating strategy in modern society.

[0004] One of the optimization solutions is mathematical optimization, which is a method that has been proven to be globally optimal. However, it has the disadvantage that it requires assumptions such as integerization and normal distribution of variables, resulting in low model accuracy, and the required computing power and calculation time exponentially increase as the number of variables increases.

[0005] Another optimization solution is reinforcement learning-based prescriptive analysis. However, due to the nature of reinforcement learning, it requires a large amount of training data, which has the disadvantage of requiring virtually unlimited computing power and memory to store episodes. In particular, gradient-based learning has the restriction that the solution space must have differentiable smoothness.

[0006] Another optimization solution is large-scale machine learning using the cloud. In conventional machine learning and deep learning-based large-scale data learning, learning is done using one high-performance computing resource. When a certain amount of time has passed in a cloud environment or a certain amount of data has been accumulated, the cloud operator dynamically allocates high-performance computing resources to continue learning and stores the learning results. However, when a service user requests real-time learning, there is a drawback in that the time it takes to dynamically allocate high-performance computing resources may be delayed because high-performance computing resources are limited resources, which delays the time at which learning is completed. On the other hand, there is a drawback in that high-performance computing resources are charged at a high rate. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] U.S. Patent Publication No. 11,514,328 (November 29, 2022) Summary of the Invention [Problem to be solved by the invention]

[0008] The present invention aims to provide a system that can prescribe actions (Prescriptive Analytics) that a user should take to improve the results predicted by a predictive model.

[0009] The present invention aims to provide a system that is versatile in the range of problems that can be solved and is efficient in terms of the size of required data and computing power.

[0010] The present invention aims to provide a system that can progressively search for optimal operating strategies in adverse learning environments.

[0011] The present invention aims to provide a system that can shorten learning time, quickly convey learning results to users, and reduce cloud usage costs. [Means for solving the problem]

[0012] One aspect of the present invention is a KPI optimization system, comprising: an analysis unit that sets a target KPI, analyzes a number of first features related to the target KPI, and selects a number of second features; a pre-processing unit that classifies the number of second features into an environmental factor value, an action value, and a KPI leading indicator, and pre-processes the number of second features, the number of second features including at least one feature each corresponding to a first environmental factor value, a first action value, and a first KPI leading indicator; an action value output unit that receives an input of the first environmental factor value and outputs at least one type of second action value; The system includes a digital twin, which is a model that has already been trained to predict a KPI leading indicator or is defined by a mathematical-physical formula, and a KPI calculation unit that calculates at least one first KPI value by calculating a second action value and the predicted first KPI leading indicator, wherein the action value output unit executes an evolutionary algorithm based on the first KPI value to optimize the target KPI, and outputs at least one updated action value, and the digital twin receives an input of the first environmental factor value and the at least one updated action value, and outputs the at least one updated KPI leading indicator.

[0013] The digital twin is a model trained to receive inputs of a first environmental factor value and a first action value and to predict a first KPI leading indicator.

[0014] The first KPI value includes at least one second KPI value that contributes to optimizing the target KPI, and the action value output unit updates at least one action value by applying an evolutionary algorithm to a combination of the action values ​​calculated to calculate the second KPI value.

[0015] The first KPI value includes at least one second KPI value that contributes to optimizing the target KPI, and the action value output unit is an optimizer model, and the optimizer model updates the at least one action value by running an evolutionary algorithm on a combination of hyperparameters or weights that output the action value calculated to calculate the second KPI value, thereby updating the hyperparameters or weights of the optimizer model.

[0016] A number of secondary features preferably have variance inflation factor values ​​of 10 or less, which is a diagnostic for multicollinearity.

[0017] Another aspect of the present invention is a cloud-based large-scale task processing system for providing the above-mentioned KPI optimization system, which includes an API that receives large-scale task processing request information from a user terminal, a supervisor that receives the large-scale task processing request information from the API and assigns the large-scale task to a task manager in charge of the large-scale task, and a task manager that determines the need for dividing the large-scale task, determines division criteria, divides the large-scale task into multiple tasks, and stores the multiple tasks in a queue.

[0018] The present invention further includes container orchestration that pools multiple operations to generate multiple operation containers that perform multiple operations, and preferentially allocates fixed instances out of fixed instances and contract instances to the multiple operation containers.

[0019] A number of work containers store work execution results in a database, the supervisor transmits the work execution results stored in the database to the API, and the API transmits the work execution results to the user terminal. Effect of the Invention

[0020] The present invention overcomes the shortcomings of traditional predictive analytics by prescribing actions for users to take to improve predicted outcomes.

[0021] By utilizing a crowd-based evolutionary algorithm, the present invention can improve accuracy in terms of the range of problems that can be solved and improve efficiency in terms of the size of data required and computing power compared to mathematical optimization and reinforcement learning based prescriptive analysis.

[0022] The present invention can search for an optimal strategy progressively through continuous learning even when initial data is very little or biased. Furthermore, the present invention can track changes in the operating environment due to changes in market trends, policies, etc., and can search for an optimal strategy.

[0023] By dividing large-scale tasks, the present invention first assigns fixed instances to the tasks to start learning, and then, as necessary, proceeds with learning in parallel using spot instances, thereby shortening the learning time, quickly communicating the learning results to users, and reducing cloud usage costs compared to conventional methods by using lower performance instances. [Brief description of the drawings]

[0024] [Figure 1] 1 shows a schematic configuration of a KPI optimization system according to an embodiment of the present invention. [Diagram 2] 2 shows a schematic configuration of an analysis unit according to an embodiment of the present invention. [Diagram 3] 2 shows a schematic configuration of a learning unit according to an embodiment of the present invention. [Figure 4] 1 shows a schematic configuration of a cloud-based large-scale task processing system according to an embodiment of the present invention. [Diagram 5] This shows the configuration of a large-scale work processing system for analyzing big data. [Figure 6] We present the configuration of a large-scale task processing system for learning algorithms and selecting models with excellent performance. [Figure 7] The configuration of a large-scale processing system for learning chromosomes is shown. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0025] The present invention will be described in detail below. However, the present invention is not limited or restricted by the illustrative examples. The objects and effects of the present invention can be naturally understood or become more apparent from the following description, and the objects and effects of the present invention are not limited by the following description alone. In addition, when describing the present invention, if it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description will be omitted.

[0026] 1 shows a schematic configuration of a KPI optimization system according to an embodiment of the present invention. Referring to FIG. 1, a KPI optimization system 10 may include an analysis unit 100, a pre-processing unit 200, and a learning unit 300, and may further include an application unit 400 and a re-learning unit 500.

[0027] 2 shows a schematic configuration of an analysis unit according to an embodiment of the present invention. Referring to FIG. 2, the analysis unit 100 may include a target KPI setting unit 110, a feature analysis unit 120, and a feature selection unit .

[0028] The target KPI setting unit 110 can set a target KPI (Key Performance Indicator). The target KPI can be composed of a KPI to be optimized and an optimization mode. The optimization mode can be any one of maximization, minimization, or convergence to a specific value. The following Examples 1 to 3 can be exemplified as examples of the target KPI. However, the target KPI is not limited to Examples 1 to 3, and any KPI that requires optimization or can be optimized through calculation can be the target KPI. Example 1: Maximizing ticket sales for baseball games. Example 2: Minimization of manufacturing facility power values. Example 3: Flow rate value of a gas production base converges to a specific value.

[0029] According to an embodiment of the present invention, the target KPI setting unit 110 can communicate with an external device, for example, a user terminal, via an API. The user terminal can input target KPI information to the API, and the API can transmit the input target KPI information to the target KPI setting unit 110. The target KPI setting unit 110 can set a target KPI based on the transmitted target KPI information.

[0030] The feature analysis unit 120 can analyze features related to the target KPI. Features related to the target KPI (hereinafter, related features) can mean feature vectors directly or indirectly required for optimizing the KPI. The feature analysis unit 120 preferably analyzes multiple types (at least three types) of related features, and more preferably analyzes at least one each of related features corresponding to environmental factor values, action values, and KPI leading indicators, which will be described later. Each type of related feature preferably includes at least one feature value, and more preferably includes at least two feature values. Here, "at least one" and "at least two" do not necessarily mean a small number such as one or two.

[0031] Next, examples of related features are given in Examples 1 to 3. However, just as the target KPI is not limited to Examples 1 to 3, the related features are not limited to the following examples.

[0032] Example 1: Maximizing baseball game ticket sales - stadium location / climate / weather, playing teams, playing team rankings / selected pitcher / fourth batter, ticket price / discount rate / sales volume, snack price / discount rate / sales volume, etc.

[0033] Example 2: Minimization of manufacturing facility power values ​​- flow rates, concentrations, power measurements, no-load power measurements, etc.

[0034] Example 3: The flow rate value of a gas production base converges to a specific value - pressure value, discharge flow rate value, temperature, atmospheric pressure, etc.

[0035] Categorical data may be included in the related features analyzed by the feature analysis unit 120. The pre-processing unit 200, which will be described later, may convert the categorical data into numerical data by performing one-hot encoding, binary encoding, or label encoding, and provide the converted numerical data to the feature analysis unit 120.

[0036] The feature selection unit 130 can select related features that have a high correlation with the target KPI and low multi-collinearity with other features from among the related features analyzed by the feature analysis unit 120. The feature selection unit 130 preferably selects multiple types (at least three types) of related features, and more preferably selects at least one each of related features corresponding to environmental factor values, action values, and KPI leading indicators, which will be described later. As described above, each type of related feature preferably includes at least one feature value, and more preferably includes at least two feature values. Here, "at least one" and "at least two" do not necessarily mean a small number such as one or two.

[0037] The above-mentioned multicollinearity can be diagnosed through a variance inflation factor (VIF). When there are a number of independent variables, the variance inflation factor is determined by setting a specific independent variable as a dependent variable and setting the remaining independent variables as independent variables, and then performing a regression analysis to measure the relationship between the variables. According to an embodiment of the present invention, if the variance inflation factor value is greater than 10, the multicollinearity can be diagnosed as high, and if the variance inflation factor value is 10 or less, the multicollinearity can be diagnosed as low.

[0038] The preprocessing unit 200 can classify the related features selected by the feature selection unit 130 into an environmental factor value, an action value, and a KPI leading indicator, and can preprocess each value. The environmental factor value, the action value, and the KPI leading indicator are defined as follows.

[0039] Environmental factor value: The environmental factor value is defined as a value that cannot be intentionally changed by the user of the KPI optimization system 10. Depending on the target KPI, various related features may correspond to the environmental factor value. In the first embodiment, the environmental factor value may be the location / climate / weather of the stadium, the competing team, the ranking of the competing team, etc.

[0040] -Action value: The action value is defined as a value that the user of the KPI optimization system 10 can intentionally change in order to optimize the KPI, and the KPI can be optimized by optimizing the action value. In the first embodiment, the action value can be the price / discount rate of a ticket, the price / discount rate of a snack, etc. Meanwhile, the action value can be said to be a value that numerically expresses an action that the user of the KPI optimization system 10 should take in order to optimize the KPI. When the user of the KPI optimization system 10 is the operator of a baseball stadium and the operator wants to change the ticket price in order to maximize the sales of baseball game tickets, the action value can be adopted as the changed ticket price.

[0041] -KPI Leading Indicators: KPI leading indicators are defined as indicators that have a leading effect on KPI optimization and should be predicted or derived in advance for KPI optimization. In the first embodiment, the KPI leading indicator can be the number of spectators. This is because the number of spectators can be used to determine whether or not sales of baseball game tickets can be maximized. Meanwhile, the KPI leading indicator corresponds to a value that must be maintained for the stability of the KPI optimization system 10 among the variables required for KPI calculation.

[0042] The preprocessing unit 200 can perform preprocessing in the following order.

[0043] 1. Missing value processing (common to environmental factor values, action values, and KPI leading indicators): Processing missing values ​​that occur during data collection. The preprocessing unit 200 can replace missing values ​​with 0 or replace missing values ​​with the value at the time before the missing value occurred.

[0044] 2. Removal of duplicate data (common to environmental factor values, action values, and KPI leading indicators): Remove duplicate data that was entered incorrectly during data collection.

[0045] 3. Embedding Cyclical Information (Environmental Factor Values): This is a pre-processing method that uses the cyclical characteristics of information that appears over time, and converts categorical variables or time values ​​into two-dimensional vectors. With reference to Equation 1 below, the first dimension is converted into Sin values, and the second dimension is converted into Cosine values.

[0046]

number

[0047] 4. Generation of differential data (action value, KPI leading indicator): Due to the nature of time-series data, the amount of change from a previous point in time to the current point in time can be important information when generating the KPI optimization system 10. Therefore, the differential value of the action value and the differential value of the KPI leading indicator are used as one of the learning data.

[0048] 5. Resampling (common to environmental factor values, action values, and KPI leading indicators): When inputting learning data, the data input pattern may not be constant. For example, data should be input in 10-second intervals, but data may be input randomly after 30 or 40 seconds. To deal with this, resampling is performed.

[0049] 6. Data Scaling (Common to Environmental Factor Values, Action Values, and KPI Leading Indicators): Adjusts the range of data values. If the data values ​​are too large or small, they may converge to 0 or diverge infinitely during the algorithm learning process. For classification tasks, a Standard Scaler is applied - which makes the feature values ​​have a normal distribution with a mean of 0 and a variance of 1 - and for regression tasks, a MinMax Scaler is applied - which makes the maximum and minimum feature values ​​1 and 0, respectively.

[0050] 3 shows a schematic configuration of a learning unit according to an embodiment of the present invention. Referring to FIG. 3, the learning unit 300 may include an action value output unit 310, a digital twin model 320, and a KPI calculation unit 330.

[0051] The digital twin model 320 can receive the input of the environmental factor value and the action value as shown in the following Equation 2, and can predict the KPI leading indicator. The environmental factor value and the action value received as input to the digital twin model 320 are related features preprocessed by the preprocessing unit 200. The digital twin model 320 may be a machine learning model or a deep learning model, and may be a rule-based model in some cases. When predicting the KPI leading indicator, in the case of a KPI leading indicator with a relatively high degree of complexity, the digital twin model 320 is configured as a learning model, and in the case of a KPI leading indicator with a relatively low degree of complexity, the digital twin model 320 is configured as a mathematical physical model.

[0052]

number

[0053] The method for constructing the digital twin model 320 will be described in detail.

[0054] According to one embodiment of the present invention, the digital twin model 320 may be a model that learns to predict KPI leading indicators by receiving inputs of environmental factor values ​​and action values ​​as shown in Equation 2.

[0055] According to another embodiment of the present invention, the digital twin model 320 may be derived through a mathematical-physical formula, so that the digital twin model 320 may be a mathematical-physical model. Simply put, the digital twin model 320 may be called a mathematical model, and may be derived as a formula. The digital twin model 320, which is a mathematical-physical model, may receive an environmental factor value and an action value as input, as shown in Equation 2 above, and predict a KPI leading indicator.

[0056] The digital twin model 320 can receive at least one type of environmental factor value from the pre-processing unit 200. In the first embodiment, the digital twin model 320 can receive only the environmental factor value corresponding to the stadium position, or can receive both the environmental factor value corresponding to the stadium position and the environmental factor value corresponding to the playing team.

[0057] The digital twin model 320 can receive at least one type of action value input from the preprocessing unit 200. In the first embodiment, the digital twin model 320 may receive only an action value corresponding to the ticket price as input, or may receive both an action value corresponding to the ticket price and an action value corresponding to the snack price as input.

[0058] The digital twin model 320 can predict (derive) at least one type of KPI leading indicator, and each type of KPI leading indicator can include at least one value. In the first embodiment, the digital twin model 320 can predict the number of spectators per day and the number of spectators by seat. It is preferable that the digital twin model 320 outputs a KPI leading indicator that matches the feature value of the related feature that the preprocessing unit 200 has classified into a KPI leading indicator and preprocessed.

[0059] When the digital twin model 320 is a learning model, it is preferable to select the most reliable model from among a large number of learned digital twin models to configure the learning unit 300. Since at least two digital twin models can be generated for each task, the learning unit 300 can be configured by selecting the digital twin model with the best performance from among the generated digital twin models, thereby improving the interaction between the action value output unit 310 and the KPI calculation unit 330.

[0060] The action value output unit 310 may receive at least one type of environmental factor value--associated features preprocessed by the preprocessing unit 200--and output at least one type of action value, where each type of action value may include at least one value. The action value output unit 310 has an ultimate purpose of outputting an action value capable of optimizing the KPI when inputted to the already configured digital twin model 320. The action value output unit 310 may execute an evolutionary algorithm and may update and output an action value based on the KPI value calculated by the KPI calculation unit 330. The action value output unit 310 may output an action value that is necessarily consistent with the feature value of the associated features preprocessed by the preprocessing unit 200 after classifying it into an action value, but may also output an inconsistent action value.

[0061] The action value output unit 310 can be configured with an optimal value search unit 310A or an optimizer model 310B. When optimizing a KPI, in the case of a KPI with a relatively low degree of complexity, the action value output unit 310 is configured with the optimal value search unit 310A, and in the case of a KPI with a relatively high degree of complexity, the action value output unit 310 is configured with the optimizer model 310B.

[0062] First, a case where the action value output unit 310 is configured with an optimum value search unit 310A will be described. The optimum value search unit 310A can use an evolutionary algorithm to search for and output an optimum action value that can optimize the KPI when input to the digital twin model 320. A genetic algorithm (GA), which is an example of an evolutionary algorithm, is driven in the following order.

[0063] 1. Generation of a set of initial chromosomes: A set of initial chromosomes is generated with an arbitrary value. Here, as the arbitrary value, the feature value of the related feature that is classified and preprocessed by the preprocessing unit 200 as an action value is used.

[0064] 2. Calculation of initial chromosome fitness: Calculate the fitness based on the information represented in the chromosome and the fitness function.

[0065] 3. Selection: The chromosome with the highest fitness is selected using the Roulette wheel selection method, which is a method in which the higher the fitness, the higher the probability of it being selected as the parent solution.

[0066] 4. Cross-over: Using the cross-over operation, the selected chromosomes are crossed together to generate offspring chromosomes with new structures.

[0067] 5.Mutation: Add and apply mutation operations to prevent local optimum problems.

[0068] The optimum value search unit 310A can define a fitness function for the target KPI, and can determine how much and how to optimize the target KPI through the fitness function. In the case of an optimization problem based on an evolutionary algorithm, basic functions such as calculations related to the problem domain and the sum of a parameter set can be used as the fitness function. In the first embodiment, when optimization of ticket sales is required, the total ticket sales can be set to the sum of the ticket sales of weekday events and the ticket sales of weekend events, and the optimization problem can be solved by adding a fitness function that maximizes the total ticket sales.

[0069] The already configured digital twin model 320 can receive at least one type of environmental factor value as input from the pre-processing unit 200, can receive at least one type of action value as input from the optimal value search unit 310A, and can predict and output at least one type of KPI leading indicator value.

[0070] The KPI calculation unit 330 can calculate at least one type of KPI value by calculating the action value output by the optimal value search unit 310A and the KPI leading indicator predicted by the digital twin model 320. In the first embodiment, the KPI calculation unit 330 can calculate the KPI value by multiplying the ticket price (action value) and the number of spectators (KPI leading indicator).

[0071] When the KPI calculation unit 330 calculates a large number of KPI values ​​for a particular type, the large number of KPI values ​​may be composed only of KPI values ​​that contribute to KPI optimization, or only of KPI values ​​that do not contribute to KPI optimization, or may be composed of KPI values ​​that contribute to KPI optimization and KPI values ​​that do not contribute to KPI optimization.

[0072] When the KPI calculation unit 330 calculates a large number of KPI values ​​for a specific type, and the large number of KPI values ​​include at least one, preferably at least two, KPI values ​​that contribute to the optimization of the KPI, the optimal value search unit 310A can update the action value output by the optimal value search unit 310A by performing an evolutionary algorithm (crossover, displacement, etc.) on the combination of action values. Here, the action value to be combined refers to the action value that calculates the KPI value that contributes to the optimization of the KPI. As a result, even if the input environmental factor value is the same, the optimal value search unit 310A can update and output the action value. In the first embodiment, the target KPI is the maximization of ticket sales. When the KPI calculation unit 330 calculates a large number of ticket sales, the optimal value search unit 310A can select at least one, preferably at least two, ticket sales in descending order of ticket sales, and can combine, crossover, or mutate the selected ticket sales with the calculated ticket price to update the ticket price.

[0073] The already-constructed digital twin model 320 may receive at least one type of environmental factor value input from the preprocessing unit 200, and may receive at least one type of updated action value input from the optimal value search unit 310A. Here, the environmental factor value input to the already-constructed digital twin model 320 may be the same as the environmental factor value input to the optimal value search unit 310A. The already-constructed digital twin model 320 may output at least one type of updated KPI leading indicator. The KPI calculation unit 330 may calculate the updated action value output by the optimal value search unit 310A and the updated KPI leading indicator output by the digital twin model 320 to calculate at least one type of updated KPI value. The update of the action value of the optimal value search unit 310A, the update of the KPI leading indicator of the already-constructed digital twin model 320, and the update of the KPI value of the KPI calculation unit 330 may be repeated until the optimal value search unit 310A outputs an action value that can achieve the target KPI.

[0074] A case where the action value output unit 310 is configured with the optimizer model 310B will be described. However, explanations overlapping with the explanation of the optimal value search unit 310A will be omitted, and differences from the optimal value search unit 310A will be described. The optimizer model 310B can learn to receive an input of an environmental factor value and output an action value, as shown in the following formula 3. The optimizer model 310B can be a machine learning model or a deep learning model.

[0075]

number

[0076] When the optimizer model 310B is configured as a machine learning model, the hyperparameters can be updated by executing an evolutionary algorithm instead of the conventional back propagation method. When the optimizer model 310B is configured as a deep learning model, the weights can be updated by executing an evolutionary algorithm instead of the conventional back propagation method.

[0077] The optimizer model 310B can update the hyperparameters or weights by executing an evolutionary algorithm (crossover, variation, etc.) on the combination of hyperparameters or the combination of weights. As a result, even if the input environmental factor values ​​are the same, the optimizer model 310B can update and output the action values. Here, the combination of hyperparameters or the combination of weights refers to the combination of hyperparameters or the combination of weights that outputs the action value that calculates the KPI value that contributes to the optimization of the KPI.

[0078] Unlike the learning step, the application unit 400 can perform the same pre-processing on the environmental factor values ​​as that performed by the pre-processing unit 200, and when the environmental factor values ​​are input to the optimized optimizer model 310B, the optimizer model 310B can output optimal action values ​​that can achieve the target KPI using the final generated algorithm.

[0079] The re-learning unit 500 may be included in the system 10 because, even though the predicted value of the digital twin shows high reliability, an error may occur between the predicted value and the actual measured value due to various causes. The error may be caused by a lack of learning data, a change in trend, a change in awareness, etc. The re-learning unit 500 repeats a cycle of [learning the digital twin model]-[application of the optimal value search unit] or a cycle of [learning the digital twin model]-[learning the optimizer model]-[application] using error (=|predicted KPI-actual KPI|) data, so that the system 10 can track changes in the operating environment.

[0080] Hereinafter, a cloud-based large-scale job processing system 70 for providing the KPI optimization system 10 will be described. FIG. 4 shows a schematic configuration of a cloud-based large-scale job processing system according to an embodiment of the present invention. Referring to FIG. 4, the cloud-based large-scale job processing system 70 can include an API 710, a Supervisor 720, a Job Manager 730, a Queue 740, a Container Orchestration 750, a number of Job Containers 760-1, 760-2, . . . , 760-n (n is a natural number equal to or greater than 2, hereinafter, parentheses are omitted), an instance 770, and a database 780.

[0081] The API 710 functions as an interface to enable communication between the user terminal 600 and the large-scale job processing system 70. The user terminal 600 can transmit large-scale job processing request information to the API 710, and the API 710 can transmit the large-scale job processing request information of the user terminal 600 to the monitor 720. Here, the large-scale job processing request information can include data or information to be processed, etc.

[0082] The supervisor 720 can assign a large-scale job to the work manager 730. The supervisor 720 can instruct the work manager 730 to determine the necessity of dividing the job. The large-scale job processing system 70 can include multiple work managers, and the jobs handled by each work manager can be different from each other. The supervisor 720 can assign the large-scale job to the work manager 730 who is in charge of the same type of job as the large-scale job processing requested by the user terminal 600.

[0083] The work manager 730 can analyze the data or information to be processed and determine whether or not work division is necessary. If the work manager 730 determines that work division is necessary, the work manager 730 can determine a criterion for dividing the large-scale work Jobtotal into a number of small-scale works Job1, Job2, ..., Jobn (n is a natural number of 2 or more, hereinafter, parentheses are omitted). The work manager 730 can store a number of small-scale works Job1, Job2, ..., Jobn in a queue 740. The queue 740 is one of the material structures in which data is stored, and has a first-in, first-out structure in which the first input is processed first, and can store (record) a number of small-scale works Job1, Job2, ..., Jobn.

[0084] The container orchestration 750 can pool a large number of small tasks Job1, Job2, ..., Jobn stored in the queue 740, generate a large number of work containers 760-1, 760-2, ..., 760-n) (n is a natural number of 2 or more, hereinafter parentheses are omitted) that execute a large number of small tasks, and can assign each small task to each work container 760-1, 760-2, ..., 760-n. It is preferable that the number of small tasks and the number of work containers are the same.

[0085] The container orchestration 750 can allocate resources of instances 770 to each of the work containers 760-1, 760-2, ..., 760-n. The instances 770 can include fixed instances (Fixed Instances) 771 and spot instances (Spot Instances) 772. The fixed instances 771 refer to fixed resources (CPU, GPU, memory, etc.) that are allocated according to a contract concluded between a user and a large-scale work processing service provider, and the spot instances 772 refer to spare resources (CPU, GPU, memory, etc.) that can be dynamically allocated by the container orchestration 750 when the work is executed.

[0086] The work container can execute the assigned small-scale work. The work container can execute the work by using the resources of the fixed instance 771 preferentially, and when the work container needs additional resources in consideration of the operating rate, progress rate, data size, etc., but determines that the additional resources are insufficient for the fixed instance 771, the container orchestration 750 can assign the resources of the spot instance 772 to the work container that needs the additional resources, and the work container can continue the work using the assigned resources of the spot instance 772.

[0087] The work containers 760-1, 760-2, ..., 760-n can store work execution results in the database 780. The monitor 720 can compile the work execution results stored in the database 780 and transmit them to the API 710, which can transmit the compiled work execution results to the user terminal 600.

[0088] A specific embodiment of the cloud-based large-scale task processing system 70 will be described below with reference to Figures 5 to 7. The allocation and use of the instance 770 has been described above, so a detailed description thereof will be omitted.

[0089] Figure 5 shows the configuration of a large-scale work processing system for analyzing big data. First, we will explain the big data shown in Figure 5. Big data consists of 9999 rows and three columns when the column items (A, B, C) are excluded. It is assumed that there is a correlation between the data in column B and the data in column C.

[0090] The user terminal 600 transmits the big data analysis request information to the API 710, and the API 710 transfers the big data analysis request information to the monitor 720. Here, the big data analysis request information includes the above-mentioned big data.

[0091] The supervisor 720 assigns the big data analysis work to the work manager 730 in charge of data analysis. The work manager 730 determines that the big data analysis work Jobtotal needs to be divided because of the large amount of data, and decides to analyze the big data by column and to analyze the correlation between the data in column B and the data in column C separately from the analysis of the columns. The work manager 730 divides the big data analysis work Jobtotal into a column A analysis work Job1, a column B analysis work Job2, a column C analysis work (not shown), and a correlation analysis work Job4 between the data in column B and the data in column C, and stores the divided works Job1, Job2, ..., Job4 in a queue 740.

[0092] Container orchestration 750 pools the tasks Job1, Job2, ..., Job4 stored in the queue 740, and generates task containers 760-1, 760-2, ..., 760-4 that execute each task. For example, the container orchestration 750 assigns column A analysis task Job1 to task container 760-1, and column B analysis task Job2 to task container 760-2. Then, it assigns column C analysis task to task container 760-3, and correlation analysis task Job4 to task container 760-4.

[0093] The working containers 760-1, 760-2, ..., 760-4 store the results of the analysis runs in a database 780. The monitor 720 collates the results of the analysis runs stored in the database 780 and communicates them to the API 710, which transmits the collated results of the analysis runs to the user terminal 600.

[0094] Fig. 6 shows the configuration of a large-scale task processing system for learning algorithms and selecting models with excellent performance. In the following, "model" means an algorithm that has completed learning. Referring to Fig. 6, a user terminal 600 transmits learning request information for algorithms A, B, ..., Z and selection request information for the model with the best performance to an API 710, and the API 710 transmits the learning request information for the algorithms and the selection request information for the best model to a monitor 720.

[0095] The supervisor 720 assigns learning tasks of algorithms A, B, ..., Z to an operation manager 730 in charge of learning digital twins. The operation manager 730 determines that the learning task of algorithms A, B, ..., Z, Jobtotal needs to be divided because there are many algorithms to be learned, and decides to perform the learning task for each algorithm. The operation manager 730 divides the learning task of algorithms A, B, ..., Z, Jobtotal, into an A-algorithm learning task Job1, a B-algorithm learning task Job2, ..., a Z-algorithm analysis task Job26, and stores the divided tasks Job1, Job2, ..., Job26 in a queue 740.

[0096] The container orchestration 750 pools the tasks Job1, Job2, ..., Job26 stored in the queue 740, and generates task containers 760-1, 760-2, ..., 760-26 that execute each task. For example, the container orchestration 750 assigns the A-algorithm learning task Job1 to the task container 760-1, the B-algorithm learning task Job2 to the task container 760-2, and the Z-algorithm learning task Job26 to the task container 760-26.

[0097] Working containers 760-1, 760-2, ..., 760-26 store the results of the learning runs and the models in database 780. Monitor 720 can communicate the results of the learning runs and the models stored in database 780, which may refer to the best model or all models, to API 710, which can transmit the results of the learning runs and the models to a user terminal.

[0098] 7 shows the configuration of a large-scale task processing system for learning chromosomes. Referring to FIG. 7, a user terminal 600 transmits learning request information for 100 chromosomes to an API 710, and the API 710 transmits the learning request information for 100 chromosomes to a monitor 720.

[0099] The supervisor 720 assigns a learning task of 100 chromosomes to the task manager 730 in charge of executing the evolutionary algorithm. The task manager 730 determines that the learning task Jobtotal of 100 chromosomes needs to be divided because there are a large number of chromosomes to be learned, and decides to execute the learning task by dividing the 100 chromosomes into 10 learning tasks Job1, 11th to 20th chromosome learning tasks Job2, ..., and 91st to 100th chromosome learning tasks Job10. The task manager 730 stores the divided tasks Job1, Job2, ..., and Job10 in the queue 740. Furthermore, the task manager 730 can request the container orchestration 750 to allocate a total of 10 CPUs to one learning task.

[0100] The container orchestration 750 pools the operations Job1, Job2, ..., Job10 stored in the queue 740, and generates operation containers 760-1, 760-2, ..., 760-10 that execute each operation. For example, the container orchestration 750 assigns the 1st to 10th chromosome learning operations Job1 to the operation container 760-1, the 11th to 20th chromosome learning operations Job2 to the operation container 760-2, and the 91st to 100th chromosome learning operations Job10 to the operation container 760-10.

[0101] The working containers 760-1, 760-2, ..., 760-10 store the results of the learning execution in a database 780. The monitor 720 collates the results of the learning execution stored in the database 780 and transmits them to the API 710, which transmits the collated results of the learning execution to the user terminal 600.

[0102] Although the present invention has been described in detail above through representative embodiments, those skilled in the art will understand that various modifications can be made to the above-described embodiments without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by not only the claims below, but also all modifications or variations derived from the equivalent concepts of the claims. [Explanation of symbols]

[0103] 10 KPI Optimization System 100 Analysis Department 200 Pretreatment section 300 Learning Department 310 Action value output section 310A Optimum value search section 310B Optimizer Model 320 Digital Twin Model 330 KPI Calculation Department 400 Application section 500 Relearning Section 600 User Terminals 70 Large-scale work processing system 710 API 720 Observer 730 Work Manager 740 Queue 750 Container Orchestration 760 Working Container 770 instances 780 Database

Claims

1. An analysis unit that sets a target KPI and selects a number of second feature vectors by analyzing a number of first feature vectors related to the target KPI, the target KPI being composed of a KPI to be optimized and an optimization mode, the optimization mode being any one of maximization, minimization, or convergence to a specific value; a pre-processing unit that classifies the second feature vectors into an environmental factor value, an action value, and a KPI leading indicator and pre-processes the second feature vectors, the second feature vectors including at least one each of a first environmental factor value, a first action value, and a first KPI leading indicator; an action value output unit that receives the first environmental factor value and outputs at least one type of second action value; A digital twin that has already been trained to predict the first KPI leading indicator in response to an input of the first environmental factor value and the second action value, or that is a model defined by a mathematical-physical formula, the digital twin outputting a KPI leading indicator that matches a feature vector that has been classified as a KPI leading indicator and preprocessed by the preprocessing unit; a KPI calculation unit that calculates at least one first KPI value by calculating the second action value and the predicted first KPI leading indicator; The action value output unit optimizes an action value based on an evolutionary algorithm to optimize the target KPI, and executes the evolutionary algorithm based on the first KPI value to output at least one updated action value; The digital twin receives the first environmental factor value and the updated at least one action value and outputs an updated at least one KPI leading indicator; The first KPI value includes at least one KPI value that contributes to optimization of the target KPI; The action value output unit updates the at least one action value by executing an evolutionary algorithm on a combination of action values ​​calculated to calculate a KPI value that contributes to optimizing the target KPI; A KPI optimization system that prescribes optimal action values ​​that optimize the target KPI.

2. The KPI optimization system of claim 1 , wherein the digital twin is a model trained to receive inputs of the first environmental factor value and the first action value and predict the first KPI leading indicator.

3. The action value output unit is an optimizer model, The KPI optimization system of claim 1, wherein the optimizer model updates at least one action value by running an evolutionary algorithm on a combination of hyperparameters or weights that output an action value calculated to calculate a KPI value that contributes to optimizing the target KPI, thereby updating the hyperparameters or weights of the optimizer model.

4. The KPI optimization system of claim 1 , wherein the multiple second feature vectors have a variance inflation factor value for diagnosing multicollinearity of 10 or less.

5. A cloud-based large-scale work processing system for providing the KPI optimization system according to claim 1, An API for receiving large-scale task processing request information from a user terminal; a monitoring unit that receives the large-scale job processing request information from the API and assigns the large-scale job to a job management unit in charge of the large-scale job; A cloud-based large-scale task processing system comprising: a task management unit that determines the need for division of the large-scale task, determines criteria for division to divide the large-scale task into multiple tasks, and stores the multiple tasks in a queue.

6. The cloud-based large-scale work processing system according to claim 5, further comprising a container orchestration that generates a number of work containers that perform the plurality of works by pooling the plurality of works, and preferentially allocates fixed instances among fixed instances and contract instances to the plurality of work containers.

7. 7. The cloud-based large-scale work processing system of claim 6, wherein the multiple work containers store results of work execution in a database, the monitoring unit transmits the results of the work execution stored in the database to the API, and the API transmits the results of the work execution to the user terminal.

Images