JP2020201870A - Hyper-parameter management device, hyper-parameter management method and hyper-parameter management program product - Google Patents

Abstract

To solve the problem that there are more than tens of thousands of combinations of hyper-parameters, and when machine learning is applied to a problem to be solved, it is required that many combinations having higher precision are put in operation and a combination with the highest precision is specified.SOLUTION: A hyper-parameter management device according to the present invention includes: a reinforced learning part which generates a set of first hyper-parameters including a hyper-parameter for processing and a hyper-parameter for model; a processing part which processes object data based upon the hyper-parameter for processing to generate information for training and information for test; a model generation part which uses the information for training to generate a machine learning model based upon the hyper-parameter for model; an evaluation part which uses the information for test to verify a machine learning model to calculate evaluation points as to the first hyper-parameters; and a modification part which generates a set of second hyper-parameters based upon high-order hyper-parameters stored in the high-order hyper-parameter database.SELECTED DRAWING: Figure 2

Description

The present invention relates to hyperparameter management devices, hyperparameter management methods and hyperparameter management program products.

Machine learning technology is attracting attention as a technology that creates new social value for the analysis of data accumulated in various fields such as healthcare, finance, and industry. Machine learning has the types of algorithms such as support vector machine and deep learning, and the parameters required to determine the model in each algorithm. These parameters that are preset and control the behavior of the model are called "hyperparameters".

In general, there are tens of thousands or more combinations of hyperparameters. Therefore, when applying machine learning to the problem to be analyzed, it is necessary to try a number of hyperparameter combinations and identify the combination with the highest accuracy. In addition, since the optimum combination of hyperparameters differs depending on the problem to be analyzed, it is necessary to perform it every time the problem or data changes, which is a problem in utilizing machine learning technology. .. In this way, the process of comprehensively trying the combination of hyperparameters to obtain the optimum solution is called "grid search".

Several means have been studied to make it easier to identify the optimal hyperparameters. For example, WO2018223123A1 (Patent Document 1) states that "a method and device for optimizing a target system having measurable state values and adjustable parameters using a cloud service. The cloud tuning service is a cloud tuning service. Set and operated by a provider. Cloud tuning services include one or more machine learning or artificial intelligence methods using resources obtained from one or more cloud providers. State values and parameters of the target system are: It is identified by the owner of the target system and sent to the cloud tuning service for periodic analysis. Parameter adjustment instructions are generated by the cloud service and sent back to the target system on a regular basis. "

WO20182223123A1

The above-mentioned Patent Document 1 describes a means for identifying hyperparameters by grid search by utilizing the computing resources of a cloud service. However, since this merely shifts the processing load for specifying hyperparameters to the cloud, it does not save computing resources and requires enormous implementation costs and man-hours for operating the cloud service.

Therefore, an object of the present invention is to provide a machine learning model that can be applied to fields such as event prediction at low cost by specifying hyperparameters using a method of reinforcement learning.

In order to solve the above problems, one of the representative hyperparameter management devices of the present invention puts a first set of hyperparameters including a processing hyperparameter and a model hyperparameter into a candidate hyperparameter database. An enhanced learning unit that generates based on the stored candidate hyperparameter space, a processing unit that processes target data based on the processing hyperparameters and generates training information and test information, and the training information. A model creation unit that creates a machine learning model based on the hyperparameters for the model, and the machine learning model is verified using the test information, and the evaluation score for the first hyperparameter set is obtained. Generates a second set of hyperparameters based on the evaluation unit that calculates the data, the hyperparameter database that stores the hyperparameters that achieve the predetermined evaluation criteria, and the hyperparameters that are stored in the hyperparameter database. Includes changes to be made.

According to the present invention, by specifying hyperparameters using a method of reinforcement learning, it is possible to provide a machine learning model applicable to fields such as event prediction at low cost.
Issues, configurations and effects other than those described above will be clarified by the following description of the embodiments for carrying out the invention.

It is a block diagram of the computer system for carrying out the Example of this invention. It is a figure which shows the structure of the hyperparameter management system which concerns on Example 1 of this invention. It is a block diagram which shows the functional structure of the hyperparameter management system which concerns on Example 1 of this invention. It is a figure which shows an example of the candidate hyperparameter database which concerns on Example 1 of this invention. It is a figure which shows an example of the upper hyperparameter database which concerns on Example 1 of this invention. It is a flowchart which shows the processing by the processing part which concerns on Example 1 of this invention. It is a flowchart which shows the process by the model making part which concerns on Example 1 of this invention. It is a figure which shows the process by the evaluation part which concerns on Example 1 of this invention. It is a flowchart which shows the process by the change part which concerns on Example 1 of this invention. It is a flowchart which shows the process by the reinforcement learning part which concerns on Example 1 of this invention. It is a figure which shows the structure of the hyperparameter management system which concerns on Example 2 of this invention. It is a block diagram which shows the functional structure of the hyperparameter management system which concerns on Example 2 of this invention. It is a flowchart which shows the process by the prediction adjustment part which concerns on Example 2 of this invention. It is a figure which shows the structure of the hyperparameter management system which concerns on Example 3 of this invention. It is a block diagram which shows the functional structure of the hyperparameter management system which concerns on Example 3 of this invention. It is a figure which shows the structure of the hyperparameter management system which concerns on Example 4 of this invention. It is a block diagram which shows the functional structure of the hyperparameter management system which concerns on Example 4 of this invention. It is a flowchart which shows the process by the hyperparameter update part which concerns on Example 4 of this invention. It is a figure which shows the structure of the hyperparameter management system which concerns on Example 5 of this invention. It is a block diagram which shows the functional structure of the hyperparameter management system which concerns on Example 5 of this invention. It is a flowchart which shows the process by the hyperparameter synthesis part which concerns on Example 5 of this invention.

Hereinafter, conventional examples and embodiments of the present invention will be described with reference to the drawings. The present invention is not limited to this embodiment. Further, in the description of the drawings, the same parts are indicated by the same reference numerals.
(Overview)

As mentioned above, the present invention relates to hyperparameter optimization. The hyperparameters here are parameters that control the behavior of the machine learning algorithm. Hyperparameters include those that define various characteristics of a machine learning algorithm, such as learning rate, batch size, number of learning iterations, and so on.

When applying machine learning to a problem to be analyzed, it is desirable to identify the optimal hyperparameters of the model in order to build a machine learning model that can solve the problem most efficiently. By means such as grid search, which have been conventionally used to identify hyperparameters, when applying machine learning to the problem to be analyzed, hyperparameters in the hyperparameter space that define the hyperparameters that can be used. It was necessary to try a number of parameter combinations and identify the combination with the highest accuracy of the target machine learning model.

Moreover, when the hyperparameter space expands due to a design change of a machine learning model or the like, the number of hyperparameter combinations increases exponentially. In particular, when the hyperparameter space is wide and computing resources are limited, it is difficult to identify the optimum hyperparameters within a practical time by a conventional hyperparameter identification method such as grid search. Therefore, there is a need for a means for more efficiently identifying the optimum hyperparameters.

Therefore, the present invention provides a machine learning model that can be applied to fields such as event prediction at low cost by specifying hyperparameters using a reinforcement learning method.
Reinforcement learning is a type of machine learning that deals with the problem that an agent in a certain environment observes the current state and decides the action to be taken. The agent is then to be rewarded by the environment by choosing an action. Therefore, in reinforcement learning, it is necessary to learn a policy that gives the most reward through a series of actions.
As typical methods of reinforcement learning, for example, TD (Temporal Difference) learning and Q-learning are known. As will be described later, by combining the reinforcement learning model and the recurrent neural network, a system for more efficiently specifying the optimum hyperparameters becomes possible.

The present invention relates to identifying hyperparameters in machine learning models for predicting events such as loan defaults, fraudulent financial transactions, disease progression and the like. The hyperparameters here include, for example, a processing hyperparameter that defines how the target data should be processed, and a model hyperparameter that defines the design conditions of the machine learning model.

First, the reinforcement learning unit according to the present invention generates a first set of hyperparameters including a processing hyperparameter and a model hyperparameter based on the candidate hyperparameter space stored in the candidate hyperparameter database. Next, the processing unit according to the present invention processes the target data based on the generated processing hyperparameters, and generates training information and test information. Next, the model creation unit according to the present invention creates a machine learning model using the training information based on the generated model hyperparameters. Next, the evaluation unit according to the present invention verifies the machine learning model using the test information and calculates the evaluation score for the first hyperparameter. Next, the hyperparameters that achieve the predetermined evaluation criteria are stored in the upper hyperparameter database. Next, the modification part according to the present invention generates a second set of hyperparameters based on the hyperparameters stored in the upper hyperparameter database.

As a result, the optimum hyperparameters can be efficiently identified, and a machine learning model applicable to event prediction can be provided.
(Hardware configuration)

First, a computer system 300 for implementing the embodiments of the present disclosure will be described with reference to FIG. The mechanisms and devices of the various embodiments disclosed herein may be applied to any suitable computing system. The main components of the computer system 300 include one or more processors 302, memory 304, terminal interface 312, storage interface 314, I / O (input / output) device interface 316, and network interface 318. These components may be interconnected via memory bus 306, I / O bus 308, bus interface unit 309, and I / O bus interface unit 310.

The computer system 300 may include one or more general purpose programmable central processing units (CPUs) 302A and 302B collectively referred to as processors 302. In one embodiment, the computer system 300 may include a plurality of processors, and in another embodiment, the computer system 300 may be a single CPU system. Each processor 302 may execute an instruction stored in memory 304 and include an onboard cache.

In certain embodiments, memory 304 may include a random access semiconductor memory, storage device, or storage medium (either volatile or non-volatile) for storing data and programs. The memory 304 may store all or part of the programs, modules, and data structures that perform the functions described herein. For example, the memory 304 may store the hyperparameter management application 350. In certain embodiments, the hyperparameter management application 350 may include instructions or descriptions that perform functions described below on the processor 302.

In certain embodiments, the hyperparameter management application 350 replaces or in addition to a processor-based system, semiconductor devices, chips, logic gates, circuits, circuit cards, and / or other physical hardware. It may be implemented in hardware via a device. In certain embodiments, the hyperparameter management application 350 may include data other than instructions or descriptions. In certain embodiments, a camera, sensor, or other data input device (not shown) may be provided to communicate directly with the bus interface unit 309, processor 302, or other hardware of the computer system 300. ..

The computer system 300 may include a processor 302, a memory 304, a display system 324, and a bus interface unit 309 that communicates between the I / O bus interface unit 310. The I / O bus interface unit 310 may be connected to an I / O bus 308 for transferring data to and from various I / O units. The I / O bus interface unit 310, via the I / O bus 308, is a plurality of I / O interface units 312, 314, 316, also known as I / O processors (IOPs) or I / O adapters (IOAs). And 318 may be communicated.

The display system 324 may include a display controller, display memory, or both. The display controller can provide video, audio, or both data to the display device 326. The computer system 300 may also include devices such as one or more sensors configured to collect the data and provide the data to the processor 302.

For example, the computer system 300 includes a biometric sensor that collects heart rate data, stress level data, an environmental sensor that collects humidity data, temperature data, pressure data, etc., and a motion sensor that collects acceleration data, exercise data, etc. May include. Other types of sensors can also be used. The display system 324 may be connected to a stand-alone display screen, a display device 326 such as a television, tablet, or portable device.

The I / O interface unit has the ability to communicate with various storage or I / O devices. For example, the terminal interface unit 312 may be a user output device such as a video display device or a speaker TV, or a user input device such as a keyboard, mouse, keypad, touchpad, trackball, button, light pen, or other pointing device. It is possible to attach such a user I / O device 320. The user inputs input data and instructions to the user I / O device 320 and the computer system 300 by operating the user input device using the user interface, and receives the output data from the computer system 300. May be good. The user interface may be displayed on the display device via the user I / O device 320, reproduced by the speaker, or printed via the printer, for example.

The storage interface 314 is an array of disk drives or other storage devices configured to appear as a single disk drive, although one or more disk drives or a direct access storage device 322 (usually a magnetic disk drive storage device). It may be). In certain embodiments, the storage device 322 may be implemented as any secondary storage device. The contents of the memory 304 are stored in the storage device 322 and may be read out from the storage device 322 as needed. The I / O device interface 316 may provide an interface to other I / O devices such as printers and fax machines. The network interface 318 may provide a communication path so that the computer system 300 and other devices can communicate with each other. This communication path may be, for example, network 330.

In certain embodiments, the computer system 300 is a device that receives a request from another computer system (client) that does not have a direct user interface, such as a multi-user mainframe computer system, a single user system, or a server computer. There may be. In other embodiments, the computer system 300 may be a desktop computer, a portable computer, a laptop computer, a tablet computer, a pocket computer, a telephone, a smartphone, or any other suitable electronic device.

The configuration of the hyperparameter management system according to the first embodiment of the present invention will be described with reference to FIG.

FIG. 2 is a diagram showing a configuration of a hyperparameter management system 150 according to a first embodiment of the present invention. As shown in FIG. 2, the hyperparameter management system 150 includes a hyperparameter management server 100, client terminals 135A and 135B, and a communication network 225. The hyperparameter management server 100 is connected to the client terminals 135A and 135B via the communication network 225. The communication network 225 may be, for example, a LAN (Local Area Network), the Internet, or the like.

The hyperparameter management server 100 is a server device for carrying out various functions according to the embodiment of the present invention. As shown in FIG. 2, the hyperparameter management server 100 includes a processor 110, a memory 120, and a storage unit 130.

The processor 110 is a processor that reads various programs stored in the memory 120 as needed and executes processing in response to instructions from the programs. For example, the functions of the processing unit 121, the model creation unit 122, the changing unit 123, the reinforcement learning unit 124, and the evaluation unit 125, which will be described later, may be realized by the processing executed by the processor 110.

Memory 120 is a random access semiconductor memory, storage device, or storage medium (either volatile or non-volatile) for storing data and programs. The memory 120 may store all or part of the programs, modules, and data structures that perform the functions described herein. For example, the memory 304 may include an instruction or description that executes the functions of the processing unit 121, the model creation unit 122, the change unit 123, the reinforcement learning unit 124, and the evaluation unit 125, which will be described later, on the processor 110.

The reinforcement learning unit 124 is a functional unit that generates a first hyperparameter set including a processing hyperparameter and a model hyperparameter based on a predetermined hyperparameter space. Reinforcement learning unit 124 may use a recurrent neural network to generate hyperparameters. This recurrent neural network is trained by the policy gradient method to generate better hyperparameters with the evaluation score described later as a reward.

The processing unit 121 is a functional unit that preprocesses the target data using the processing hyperparameters generated by the reinforcement learning unit 124. The preprocessing here includes converting the target data into a format that can be interpreted by the subsequent functional unit, supplementing the missing value in the target data, and deleting the redundant target data. The preprocessing executed by the processing unit 121 is defined by the processing hyperparameters. The processing unit 121 generates training information for training the machine learning model and test information for testing the trained machine learning model by processing the target data based on the processing hyperparameters. To do.

The model creation unit 122 is a functional unit that creates a machine learning model based on the model hyperparameters generated by the reinforcement learning unit 124 and trains the machine learning model using the training information from the processing unit 121. .. The machine learning model here may be a machine learning model for predicting events such as loan defaults, fraudulent financial transactions, and disease progression.

The evaluation unit 125 verifies the machine learning model created by the model creation unit 122 using the test information, and calculates the evaluation score for the first hyperparameter set. This evaluation score is a scale that quantitatively indicates the performance of each hyperparameter, and may be represented by a number of 0 to 1, for example. As a general rule, hyperparameters with higher evaluation scores mean that they exhibited better accuracy, performance, efficiency, etc. when verifying machine learning models. Hyperparameters that achieve a predetermined evaluation criterion are stored in the upper hyperparameter database 132 described later.
For convenience of explanation, hyperparameters that achieve a predetermined evaluation criterion are also referred to as "upper-level hyperparameters".

The change unit 123 is a functional unit that generates a second hyperparameter set based on the upper hyperparameters. By generating hyperparameters for machine learning models based on higher-level hyperparameters that have already been determined to have high performance, rather than the entire hyperparameter space, it is highly possible that hyperparameters with high performance can be obtained. Therefore, it is possible to efficiently identify excellent hyperparameters as compared with conventional hyperparameter generation means such as grid search.
Specifically, the change unit 123 may generate a second hyperparameter set by exchanging the values of the upper hyperparameters or synthesizing the upper hyperparameters, for example.

The storage unit 130 is a storage device for storing various data used by the above-mentioned functional unit. The storage unit 130 may be any storage medium such as a flash memory or a hard disk drive. Further, as shown in FIG. 2, the storage unit 130 may include the candidate hyperparameter database 131 and the upper hyperparameter database 132.

The candidate hyperparameter database 131 is a database that defines a predetermined hyperparameter space. The hyperparameter space here is a logical space that defines the range of hyperparameters that can be used for a certain field or problem and includes all the hyperparameters that can be used. This hyperparameter space may be predefined by the user. Further, the hyperparameter space included in the candidate hyperparameter database 131 may define a range of hyperparameters that can be used for both the processing hyperparameters and the model hyperparameters.

The upper hyperparameter database 132 is a database for storing upper hyperparameters. As described above, the upper hyperparameter database 132 may be a hyperparameter of the first set of hyperparameters that achieves a predetermined evaluation criterion when the machine learning model is verified using the test information. Good. This evaluation standard is a standard that defines the threshold of hyperparameters of "good" and "bad", and has a predetermined evaluation score (for example, 0.85 or more), a predetermined percentage (top 10%), and a predetermined number (top). It may be expressed as 20) or the like.

The client terminals 135A and 135B are terminals that transmit target data to be analyzed by the machine learning model to the hyperparameter management server 100 via the communication network 225. After the analysis (for example, event prediction) by the hyperparameter management device 205 is completed, the information indicating the analysis result is returned to the client terminals 135A and 135B. These client terminals 135A and 135B may be any device such as a desktop personal computer, a laptop computer, a tablet, or a smartphone.

Each functional unit included in the hyperparameter management server 100 may be a software module constituting the hyperparameter management application 350 shown in FIG. 1, or may be an independent dedicated hardware device. Further, the above-mentioned functional unit may be implemented in the same computing environment, or may be implemented in a distributed computing environment.

Next, the functional configuration of the hyperparameter management system according to the first embodiment of the present invention will be described with reference to FIG.

FIG. 3 is a block diagram showing a functional configuration 250 of the hyperparameter management system according to the first embodiment of the present invention. As shown in FIG. 3, this functional configuration 250 includes the above-mentioned processing unit 121, model creation unit 122, change unit 123, reinforcement learning unit 124, evaluation unit 125, candidate hyperparameter database 131, and upper hyperparameter database 132. Indicates the transmission and reception of data between.

First, the reinforcement learning unit 124 generates a first hyperparameter set including a processing hyperparameter and a model hyperparameter based on the candidate hyperparameter space stored in the candidate hyperparameter database 131. Specifically, the reinforcement learning unit 124 uses the so-called Epsilon-Greedy method to determine whether the first hyperparameter set is generated by exploration or exploitation.
In the case of exploitation, the reinforcement learning unit 124 uses the calculation of forward propagation to generate a sequence of hyperparameters as a first set of hyperparameters. In the case of search, the reinforcement learning unit 124 generates hyperparameters uniformly extracted from the candidate hyperparameter space of the candidate hyperparameter database 131 as a first set of hyperparameters. After that, the reinforcement learning unit 124 transmits the generated first hyperparameter set to the processing unit 121 and the model creation unit 122.
Also, at a later stage, the reinforcement learning unit 124 is trained by backpropagation using the policy gradient method based on the evaluation score calculated for the first set of hyperparameters. As a result, the reinforcement learning unit 124 learns so that better hyperparameters can be generated.

Next, the processing unit 121 processes arbitrary target data based on the processing hyperparameters included in the first hyperparameter set. The target data here may be, for example, data to be analyzed (data related to financial transactions of the analysis target person, etc.) received from the client terminals (135A, 135B) shown in FIG. The specific steps of the processing performed by the processing unit 121 on the target data are defined by the processing hyperparameters.
For example, when the processing hyperparameter has a value of "avage_filling", the processing unit 121 may fill the missing value in the target data with the average value of the values of similar data, and the processing hyperparameter When has a value of "median_filling", the processing unit 121 may fill the missing value in the target data with the median value of the similar data.

After the processing for the target data is completed, the processing unit 121 generates training information and test information. Specifically, the processing unit 121 may divide the received target data into training information for training the machine learning model and test information for testing the trained machine learning model. .. By dividing the target data into training information and test information, it is possible to train and test the machine learning model separately, and it is learned for the training information, but it is unknown test information. On the other hand, it is possible to avoid over-fitting that is not fit. For example, when the target data is data indicating a financial transaction conducted between 2000 and 2010, the processing unit 121 uses the data from 2000 to 2005 as training information, and from 2005 to 2010. Data may be used as test information.

Next, the model creation unit 122 creates a machine learning model based on the model hyperparameters included in the first hyperparameter set. The type of the machine learning model created here may be preset by the user or may be specified by the model hyperparameters. This machine learning model may be an event prediction machine learning model for predicting events such as loan defaults, fraudulent financial transactions, and disease progression. In addition, the characteristics and settings of the machine learning model are defined by the hyperparameters for the model.
For example, when the model hyperparameters have the values of "number_of_layers = 3", "number_of_nodes = 3", "learning_rate = 0.001", and "activation_function = linear_activation", the model creation unit 122 sets these conditions. Build a machine learning model according to.
After creating the machine learning model based on the model hyperparameters, the model creation unit 122 trains the machine learning model using the training information received from the processing unit 121. This training method may be appropriately selected according to the type of machine learning model.

Next, the evaluation unit 125 verifies the machine learning model created by the model creation unit 122 using the test information received from the processing unit 121, and calculates the evaluation score for the first hyperparameter. As described above, this evaluation score is a scale that quantitatively indicates the performance of the first hyperparameter, and may be represented by a number of 0 to 1, for example. As a general rule, hyperparameters with higher evaluation scores mean that they exhibited better accuracy, performance, efficiency, etc. in the verified machine learning model. Hyperparameters that achieve a predetermined evaluation criterion are stored in the upper hyperparameter database 132 described later.

Next, the change unit 123 generates a second hyperparameter set based on the upper hyperparameters stored in the upper hyperparameter database 132. For example, the change unit 123 may generate a second hyperparameter set by exchanging the values of a plurality of upper hyperparameters, or may generate a second hyperparameter set by synthesizing a plurality of upper hyperparameters. May be generated.
The second set of hyperparameters generated here may be used for training the reinforcement learning unit 124, may be set in the machine learning model, or may be stored in the upper hyperparameter database 132. ..
Here, as a method of generating the second hyperparameter set, exchanging the values of the upper hyperparameters and synthesizing the upper hyperparameters have been described as examples, but the present invention is limited to this. However, other means of generating new hyperparameters using higher-level hyperparameters are also possible.

Next, an example of the candidate hyperparameter database according to the first embodiment of the present invention will be described with reference to FIG.

FIG. 4 is a diagram showing an example of the candidate hyperparameter database 131 according to the first embodiment of the present invention. As shown in FIG. 4, the candidate hyperparameter database 131 includes parameter name 351, first option 352, second option 353, and other option 354.
Although the contents of the candidate hyperparameter database 131 are omitted for convenience of explanation, the candidate hyperparameter database 131 may store a large number of hyperparameter information. These hyperparameters constitute the hyperparameter space described above.

The parameter name 351 specifies the type of hyperparameter. For example, "missing_value" refers to processing hyperparameters indicating a method for filling in missing values, and "number_of_layers" refers to model hyperparameter data that specifies the number of layers in a machine learning model.

The first option 352, the second option 353, and the other option 354 specify the validity or range of possible values of each hyperparameter. For example, when the first option 352 of "missing_value" is "True", the processing for filling the missing value is performed by the processing unit, and when it is "False", it is not performed. Further, when the second option 353 of "number_of_layers" is "3", the model creation unit constructs a machine learning model having three layers.

As described above, the reinforcement learning unit may generate hyperparameters randomly or uniformly selected from the candidate hyperparameter database 131 as a first set of hyperparameters.

Next, the upper hyperparameter database according to the first embodiment of the present invention will be described with reference to FIG.

FIG. 5 is a diagram showing an example of the upper hyperparameter database 132 according to the first embodiment of the present invention. As shown in FIG. 5, the upper hyperparameter database 132 includes a set ID 401, a parameter name 402, a value 403, and an evaluation score 404.
As described above, the upper hyperparameter database 132 is a database that stores upper hyperparameters that achieve a predetermined evaluation criterion.

The set ID 401 is an identifier of the set to which the parameter belongs. The set here consists of one processing hyperparameter and one model hyperparameter. For example, hyperparameters to which the same set ID 401 is assigned belong to the same set and are paired.

The parameter name 402 specifies the type of hyperparameter. Since the parameter name 402 is substantially the same as the parameter name 351 described with reference to FIG. 4, the description thereof will be omitted.

The value 403 specifies the value of the hyperparameter. The value 403 here indicates a value when it is actually used in the processing unit or the model creation unit, unlike the first option and the second option that indicate possible values.

The evaluation score 404 is a scale that quantitatively indicates the performance of the hyperparameter, and may be represented by a number of 0 to 1, for example. As a general rule, hyperparameters with higher evaluation scores mean that they have demonstrated better accuracy, performance, efficiency, etc. in machine learning models. The evaluation score 404 is a value determined by the evaluation unit 125 described above when evaluating a machine learning model trained and verified based on a specific hyperparameter set.

Next, with reference to FIG. 6, the processing by the processing unit according to the first embodiment of the present invention will be described.

FIG. 6 is a flowchart showing processing 500 by the processing unit according to the first embodiment of the present invention. This process 500 indicates each step of the process performed on the target data by the process unit based on the process hyperparameters.

First, in step 501, the processing unit (for example, the processing unit 121 shown in FIG. 2) receives from the reinforcement learning unit a first set of hyperparameters including hyperparameters for processing and hyperparameters for models.

Next, in step 502, the processing unit processes the target data based on the processing hyperparameters included in the first hyperparameter set. For example, when the processing hyperparameter has a value of "average_filling", the processing unit may fill in the missing values in the target data with the average value of the values of similar data.

Next, in step 503, the processing unit outputs the processed data to the model creation unit. This processed data includes training information and test information processed based on the hyperparameters for processing.

Next, with reference to FIG. 7, the processing by the model creation unit according to the first embodiment of the present invention will be described.

FIG. 7 is a flowchart showing the process 600 by the model creation unit according to the first embodiment of the present invention. This process 600 shows a process of creating a machine learning model based on the model hyperparameters.

First, in step 601 the model creation unit includes a set of first hyperparameters including a hyperparameter for processing from the reinforcement learning unit and a hyperparameter for the model, and training information and test information from the processing unit. To receive.

Next, in step 602, the model creation unit creates a machine learning model based on the model hyperparameters included in the first hyperparameter set. For example, the model creation unit creates a machine learning model according to the model type (for example, "Support Vector Machine" or "Deep Natural Network") defined by the hyperparameters for the model, the number of layers, the number of nodes in the hidden layer, and the like. You may.

Next, in step 603, after the machine learning model is created based on the model hyperparameters, the model creation unit 122 trains the machine learning model using the training information received from the processing unit 121. Methods of this training include, for example, linear regression, logistic regression, naive Bayes classification and k-nearest neighbors. This training improves the accuracy of machine learning models in accurately predicting specific events such as loan defaults and disease progression.

Next, the process by the evaluation unit according to the first embodiment of the present invention will be described with reference to FIG.

FIG. 8 is a diagram showing a process 700 by the evaluation unit according to the first embodiment of the present invention. This process 700 shows a process of verifying the machine learning model created by the model creation unit and evaluating the first set of hyperparameters.

First, in step 703, the evaluation unit verifies the machine learning model created by the model creation unit using the test information received from the processing unit. For example, if the machine learning model is a model that predicts loan defaults, the evaluation department analyzes test information that shows financial information such as financial transactions, annual income, and credit score of the analysis target person, and the analysis target person analyzes it. Calculate the default probability. After that, the accuracy of the machine learning model can be verified by comparing the probability calculated by the machine learning model with the result indicating whether or not the analysis target actually defaults.

Next, in step 704, the evaluation unit calculates the evaluation score for the first hyperparameter based on the verification result. The format of this evaluation score may be selected depending on the type of machine learning model. For example, when the machine learning model classifies and predicts, the accuracy and F1 score may be used as this evaluation score, and when the machine learning model is a regression model, the R2 score is used as this evaluation score. You may use it.
As an example, this evaluation score may be represented by a number from 0 to 1, for example. As a general rule, hyperparameters with higher evaluation scores mean that they exhibited better accuracy, performance, efficiency, etc. in the verified machine learning model.

Next, in step 705, the evaluation unit may store hyperparameters that achieve a predetermined evaluation criterion in the above-mentioned upper hyperparameter database. This evaluation standard is a standard that defines the threshold of hyperparameters of "good" and "bad", and has a predetermined evaluation score (for example, 0.85 or more), a predetermined percentage (top 10%), and a predetermined number (top). It may be expressed as 20) or the like.

Next, with reference to FIG. 9, the processing by the modified portion according to the first embodiment of the present invention will be described.

FIG. 9 is a flowchart showing the process 800 by the modified portion according to the first embodiment of the present invention. This process 800 indicates a process of generating a second hyperparameter based on the upper hyperparameter.

First, in step 801, the change unit reads out a set of arbitrary several K hyperparameters among the upper hyperparameters stored in the upper hyperparameter database.
Here, K is a number of 2 or more. This is because if there is only one set of hyperparameters, values cannot be exchanged and synthesis described later cannot be performed.

Next, in step 802, the change unit generates a second set of hyperparameters by exchanging the values of the set of K hyperparameters.
For example, if the upper hyperparameter database has {"missing_value": "median", "number_of_layers": 2} and {"missing_value": "average", "number_of_layers": 3} , The change unit 123 exchanges the values of the respective hyperparameters to generate {"missing_value": "average", "number_of_layers": 2} and {"missing_value": "median", "number_of_layers": 3}.
For convenience of explanation, the case where there are two hyperparameters in the upper hyperparameter database has been described, but the present invention is not limited to this, and it goes without saying that it can be applied even when there are three or more hyperparameters. .. In this case, the change part may generate all permutations of hyperparameter data stored in the upper hyperparameter database.
Further, here, an example of generating a second hyperparameter set by exchanging the values of a plurality of hyperparameters has been described, but the present invention is not limited to this, and as will be described later, a plurality of hyperparameters are described. It is also possible to generate a second set of hyperparameters by synthesizing hyperparameters.

Next, in step 803, the change unit randomly selects an arbitrary hyperparameter from the second set of hyperparameters and outputs it. The hyperparameters selected here may be used for training the reinforcement learning unit, may be set in the machine learning model, or may be stored in the upper hyperparameter database.

Next, with reference to FIG. 10, the processing by the reinforcement learning unit according to the first embodiment of the present invention will be described.

FIG. 10 is a flowchart showing the process 900 by the reinforcement learning unit according to the first embodiment of the present invention. This process 900 shows a process of generating hyperparameters and training the reinforcement learning unit based on the evaluation points of the hyperparameters.

First, in step 901, the reinforcement learning unit randomly generates 0 or more and less than 1 in order to determine whether the first hyperparameter set is generated by exploration or exploitation. Generate a value.

Next, in step 902, the reinforcement learning unit compares the randomly generated value with the first epsilon parameter ε1. If the randomly generated value is greater than or equal to the first epsilon parameter ε1, the process proceeds to step 903, and if the randomly generated value is less than the first epsilon parameter ε1, the process proceeds to step 904. move on.
The first epsilon parameter ε1 is a value set in advance by the user.

In step 903, the reinforcement learning unit uses forward propagation calculation of a recurrent neural network as a means of so-called "exploitation" to generate a sequence of hyperparameters as a first set of hyperparameters.

In step 904, the reinforcement learning unit randomly determines whether to use the above-mentioned change unit or the hyperparameters uniformly extracted from the hyperparameter space as the search means. Generate a value.

Next, in step 905, the reinforcement learning unit compares the randomly generated value with the second epsilon parameter ε2. If the randomly generated value is greater than or equal to the second epsilon parameter ε2, the process proceeds to step 907, and if the randomly generated value is less than the second epsilon parameter ε2, the process proceeds to step 906. move on.
The second epsilon parameter ε2 is a value set in advance by the user.

In step 906, the reinforcement learning unit calls the modification unit to generate a second hyperparameter based on the first hyperparameter set. This process corresponds to process 800 described above.

In step 907, the reinforcement learning unit generates hyperparameters uniformly extracted from the candidate hyperparameter space of the candidate hyperparameter database as a first set of hyperparameters.

Next, in step 908, the reinforcement learning unit transmits the set of hyperparameters generated in step 903, step 906, or step 907 to the processing unit and the model creation unit. After that, the processing unit and the model creation unit perform the above-mentioned processing (processing 500 and processing 600), respectively.

Next, in step 909, after the processing by the processing unit, the model creation unit, and the evaluation unit is completed, the reinforcement learning unit receives the evaluation score calculated for the hyperparameter set from the evaluation unit.

Next, in step 910, the reinforcement learning unit updates the weighting of the recurrent network based on the received evaluation score by using the policy gradient method. For example, when the evaluation score of the hyperparameter is low (the predetermined evaluation standard is not achieved), the reinforcement learning unit may reduce the weighting of the recurrent network. Further, when the evaluation score of the hyperparameter is high (achieving a predetermined evaluation standard), the reinforcement learning unit may increase the weighting of the recurrent network. This trains the recurrent network of the reinforcement learning department to generate better hyperparameters.

Next, the configuration of the hyperparameter management system according to the second embodiment of the present invention will be described with reference to FIG.

FIG. 11 is a diagram showing the configuration of the hyperparameter management system 1150 according to the second embodiment of the present invention. The hyperparameter management system 1150 shown in FIG. 11 is different from the hyperparameter management system 150 in that it includes a prediction adjustment unit 126 in addition to the components of the hyperparameter management system 150 described with reference to FIG. Other than this point, the configuration of the hyperparameter management system 1150 is substantially the same as the configuration of the hyperparameter management system 150, and thus the description thereof will be omitted.

Some machine learning models lack versatility, such as the polarization of the output probability distribution. For example, in the probability distribution of a machine learning model that predicts the probability that a loan will be defaulted, the probability that a loan will be defaulted may be divided into two extreme options such as almost 100% or almost 0%. Since it is difficult to draw meaningful conclusions from such polarized results, a machine learning model that shows a smoother probability distribution between "100%" and "0%" is desirable.

Therefore, the hyperparameter management system 1150 according to the second embodiment of the present invention includes a prediction adjustment unit 126 that adjusts the evaluation points of the hyperparameters according to a desired probability distribution. As a result, the machine learning model created by the model creation unit 122 is trained to generate a prediction result having a more meaningful probability distribution.
The function of the prediction adjustment unit 126 will be described later.

Next, with reference to FIG. 12, the functional configuration of the hyperparameter management system according to the second embodiment of the present invention will be described.

FIG. 12 is a block diagram showing a functional configuration 1250 of the hyperparameter management system according to the second embodiment of the present invention. The functional configuration 1250 shown in FIG. 12 differs from the functional configuration 250 in that it includes a predictive adjustment unit 126 in addition to the components of the functional configuration 250 described with reference to FIG. Other than this point, the configuration of the functional configuration 1250 is substantially the same as that of the functional configuration 250, and thus the description thereof will be omitted.

The prediction adjustment unit 126 receives the prediction result generated by the machine learning model (not shown) of the model creation unit 122 and the evaluation score of the hyperparameter. The prediction result here may be, for example, a probability distribution regarding events such as loan defaults, fraudulent financial transactions, and disease progression, which are generated based on the target data.

Next, the prediction adjustment unit 126 compares the prediction result generated by the machine learning model with the desired prediction distribution input in advance, and adjusts the evaluation score of the hyperparameter based on the result of this comparison. This desired prediction distribution may be, for example, reference data showing a probability distribution desired by the user. Among the hyperparameters whose evaluation points have been adjusted, the hyperparameters that achieve a predetermined evaluation criterion are stored in the upper hyperparameter database 132.
The details of hyperparameter adjustment will be described later.

Next, with reference to FIG. 13, the processing by the prediction adjustment unit according to the second embodiment of the present invention will be described.

FIG. 13 is a flowchart showing the process 1300 by the prediction adjustment unit according to the second embodiment of the present invention. This process 1300 shows each step of the process of adjusting the evaluation score of the hyperparameter based on the comparison between the prediction result of the machine learning model and the desired prediction distribution.

First, in step 1301, the prediction adjustment unit receives the prediction result generated by the machine learning model (not shown) of the model creation unit 122. This prediction result may be, for example, a collection of predictions generated based on a plurality of test data into one prediction distribution.

Next, in step 1302, the prediction adjustment unit compares the prediction result of the machine learning model with the desired prediction distribution input in advance, and calculates the distance indicating the difference between the prediction result and the desired prediction distribution. This distance may be calculated, for example, by means of calculating the Kullback-Leibler distance.

Next, in step 1303, the prediction adjustment unit adjusts the evaluation score of the hyperparameter based on the distance between the prediction result and the desired prediction distribution. Specifically, when the prediction result achieves a predetermined similarity criterion with respect to the prediction result distribution, the prediction adjustment unit increases the evaluation score of the hyperparameter. On the other hand, if the prediction result does not achieve a predetermined similarity criterion with respect to the prediction result distribution, the prediction adjustment unit may make the evaluation score of the hyperparameter phenomenon. The similarity standard here is, for example, a standard for designating a value of a predetermined distance.

Next, in step 1304, the prediction adjustment unit outputs hyperparameters adjusted for evaluation points. For example, as described above, the prediction adjustment unit may store the hyperparameters whose evaluation points have been adjusted to achieve the predetermined evaluation criteria in the upper hyperparameter database.

As described above, the hyperparameters that promote the generation of prediction results similar to the desired prediction distribution have a high evaluation score, and the hyperparameters that promote the generation of prediction results that do not resemble the desired prediction distribution have a low evaluation score. As a result, when the model creation unit creates a machine learning model, it is possible to realize a machine learning model that can generate a prediction result having a prediction distribution desired by the user by using a hyperparameter with a high evaluation score.

Next, the configuration of the hyperparameter management system according to the third embodiment of the present invention will be described with reference to FIG.

FIG. 14 is a diagram showing the configuration of the hyperparameter management system 1400 according to the third embodiment of the present invention. In Examples 1 and 2 described above, the configuration for generating both the higher-level processing hyperparameters and the model hyperparameters has been described, but the present invention is not limited thereto, and the higher-level processing hyperparameters and models are used. It is also possible to configure hyperparameters to be generated individually. Therefore, the hyperparameter management system 1400 according to the third embodiment of the present invention describes a configuration in which a hyperparameter for higher processing and a hyperparameter for a model are generated in parallel.
Here, the expression "generate higher-level processing hyperparameters and model hyperparameters individually" is a machine learning pipeline in which higher-level processing hyperparameters and higher-level model hyperparameters are independent. It means to generate with. This makes it easier to obtain hyperparameters with a higher evaluation score than generating higher-level processing hyperparameters and model hyperparameters in the same machine learning pipeline.

As shown in FIG. 14, in order to generate the hyperparameters for processing and the hyperparameters for models in parallel, the hyperparameter management server 100 of the hyperparameter management system 1400 has two processors 110, 210 and two memories 110. , 220, and two storage units 130, 230 may be included. The processors 110 and 210 may be physically independent processors or may be different cores of the same processor. Similarly, the memories 120, 220 and the storage units 130, 230 may be physically independent or may be different partitions of the same storage medium.

As described above, in the second embodiment, the process of generating the hyperparameters for processing and the process of generating the hyperparameters for the model are individually performed. Therefore, in the process of generating the hyperparameters for processing, the reinforcement learning unit 124 does not generate the hyperparameters for the model. However, in the process of generating the hyperparameters for processing, the hyperparameters for the model are required to create the machine learning model. Similarly, in the process of generating the model hyperparameters, the process hyperparameters are required in order to perform the preprocessing of the target data.
Therefore, the hyperparameter management system 1400 according to the present embodiment has a fixed model hyperparameter database 135 in which the model hyperparameters used in the process of generating the processing hyperparameters are stored in advance, and the model hyperparameters. It includes a fixed processing hyperparameter database 235 in which processing hyperparameters used in the processing for generating the above are stored in advance.
As a result, when other types of hyperparameters are required in the process of generating the processing hyperparameters or the model hyperparameters, the hyperparameters are read from the corresponding database and used as appropriate. Since the values of these fixed hyperparameters are fixed, they are not changed by the processing of the evaluation unit 125 and the change unit 123.

As shown in FIG. 14, the storage unit 130 includes a candidate processing hyperparameter database 133 that generates processing hyperparameters, and a higher-level processing hyperparameter database 134 that stores processing hyperparameters that achieve a predetermined evaluation criterion. , Includes a fixed model hyperparameter database 135 that stores fixed model hyperparameters used when creating machine learning models.
Further, the storage unit 230 processes the target data, the hyperparameter database 233 for the candidate model that generates the hyperparameters for the model, the hyperparameter database 234 for the upper model that stores the hyperparameters for the model that achieves the predetermined evaluation criteria, and the target data. It includes a fixed processing hyperparameter database 235 that stores the processing hyperparameters used in the process.

Other than the points described above, the configuration of the hyperparameter management system 1400 is practically the same as that of the hyperparameter management system 150 described above, and thus the description thereof will be omitted.

Next, with reference to FIG. 15, the functional configuration of the hyperparameter management system according to the third embodiment of the present invention will be described.

FIG. 15 is a block diagram showing a functional configuration 1450 of the hyperparameter management system according to the third embodiment of the present invention. As described above, in the third embodiment of the present invention, the processing hyperparameters and the model hyperparameters are generated by separate processing. Therefore, as shown in FIG. 15, the functions of the reinforcement learning unit, the processing unit, the model creating unit, the changing unit, and the evaluation unit described above are executed in parallel for each of the processing hyperparameters and the model hyperparameters. Since the specific processing procedure and the transmission / reception of data between the functional units are substantially the same as those described with reference to FIG. 3, the description thereof will be omitted here.

Next, the configuration of the hyperparameter management system according to the fourth embodiment of the present invention will be described with reference to FIG.

FIG. 16 is a diagram showing the configuration of the hyperparameter management system 1600 according to the fourth embodiment of the present invention. The hyperparameter management system 1600 shown in FIG. 16 has a configuration in which a processing hyperparameter and a model hyperparameter are processed in parallel, similarly to the hyperparameter management system 1400 described above. However, the hyperparameter management system 1600 is different from the hyperparameter management system 1400 in that it includes a hyperparameter update unit 321 that updates hyperparameters stored in a fixed hyperparameter database. Other than this point, the configuration of the hyperparameter management system 1600 is substantially the same as the configuration of the hyperparameter management system 1400, and thus the description thereof will be omitted.

When the processing hyperparameter and the model hyperparameter are optimized independently, one type of hyperparameter is fixed and the other type of hyperparameter is optimized. Therefore, in order to realize a hyperparameter set in which both the processing hyperparameters and the model hyperparameters are optimized, fixed hyperparameters are based on the hyperparameters optimized in another machine learning pipeline. It is desirable to update.

As shown in FIG. 16, the hyperparameter management system 1600 includes a third processor 313, a memory 323, and a hyperparameter update unit 321. This hyperparameter update unit 321 migrates the hyperparameters for the upper model stored in the hyperparameter database 234 for the upper model to the hyperparameter database 135 for the fixed model at a predetermined frequency, and hyperparameter database for upper processing. It is a functional part that migrates the hyperparameters for higher processing stored in 134 to the fixed hyperparameter database 235 for processing.
This replaces the fixed hyperparameters with the higher-level hyperparameters, so that the latest and highest-rated hyperparameters are used in each machine learning pipeline.

Next, with reference to FIG. 17, the functional configuration of the hyperparameter management system according to the fourth embodiment of the present invention will be described.

FIG. 17 is a diagram showing a functional configuration 1750 of the hyperparameter management system according to the fourth embodiment of the present invention. As shown in FIG. 17, the hyperparameter update unit 321 migrates the hyperparameters for the upper model stored in the hyperparameter database 234 for the upper model to the hyperparameter database 135 for the fixed model at a predetermined frequency. The hyperparameters for higher processing stored in the hyperparameter database 134 for higher processing are migrated to the fixed hyperparameter database 235 for processing.
Other than this point, the functional configuration 1750 of the hyperparameter management system shown in FIG. 17 is substantially the same as the functional configuration 250 of the hyperparameter management system shown in FIG. 3, and therefore the description thereof will be omitted here. ..

Next, the process by the hyperparameter update unit according to the fourth embodiment of the present invention will be described with reference to FIG.

FIG. 18 is a flowchart showing processing 1800 by the hyperparameter update unit according to the fourth embodiment of the present invention. This process 1800 indicates a process of updating the hyperparameters stored in the fixed hyperparameter database with higher-level hyperparameters.

First, in step 1601, the hyperparameter update unit determines whether or not a predetermined hyperparameter update condition is satisfied. This hyperparameter update condition is a condition that determines the time and situation for updating the hyperparameter. This hyperparameter update condition is an arbitrary condition such as the elapse of a predetermined time (1 hour, 1 day, 1 week) since the last update, the number of trained machine learning models, the number of upper hyperparameters, and the like. May be good. If the hyperparameter update condition is satisfied, this process proceeds to step 1602, and if the hyperparameter update condition is not satisfied, this process ends.

Next, in step 1602, the hyperparameter update unit shifts the hyperparameters for higher processing stored in the hyperparameter database for higher processing to the fixed hyperparameter database for processing. At this time, the hyperparameters stored in the fixed processing hyperparameter database may be overwritten by the higher-level processing hyperparameters, or may be stored as spare hyperparameters.

Next, in step 1603, the hyperparameter update unit shifts the hyperparameters for the upper model stored in the hyperparameter database for the upper model to the hyperparameter database for the fixed model. Further, as described above, in this case, the hyperparameters stored in the fixed model hyperparameter database may be overwritten by the hyperparameters for the upper model, or may be stored as spare hyperparameters. ..

Next, the configuration of the hyperparameter management system according to the fifth embodiment of the present invention will be described with reference to FIG.

FIG. 19 is a diagram showing a configuration of a hyperparameter management system 1900 according to a fifth embodiment of the present invention. The hyperparameter management system 1900 shown in FIG. 19 has a configuration in which a processing hyperparameter and a model hyperparameter are generated in parallel, similarly to the hyperparameter management system 1400 described above. However, the hyperparameter management system 1900 is different from the hyperparameter management system 1400 in that it includes a hyperparameter synthesis unit 325 that synthesizes hyperparameters for higher-level processing and hyperparameters for higher-level models that are independently generated. Other than this point, the configuration of the hyperparameter management system 1900 is substantially the same as the configuration of the hyperparameter management system 1400, and thus the description thereof will be omitted.

When the processing hyperparameters and the model hyperparameters are optimized independently, the higher-level processing hyperparameters and the higher-level model hyperparameters are stored in separate databases. Therefore, in order to realize a hyperparameter set in which both the processing hyperparameters and the model hyperparameters are optimized, it is desirable to synthesize the upper processing hyperparameters and the upper model hyperparameters.

As shown in FIG. 19, the hyperparameter management system 1900 includes a third processor 313, a memory 323, and a hyperparameter synthesis unit 325. The hyperparameter synthesis unit 325 extracts hyperparameters for higher processing from the hyperparameter database 134 for higher processing, extracts hyperparameters for higher model from hyperparameter database 234 for higher model, and extracts the hyperparameters for higher processing. And by synthesizing the hyperparameters for the upper model, a second hyperparameter set is generated. The second hyperparameter set synthesized here is stored in the higher-level synthetic hyperparameter database 335 of the storage unit 333 shown in FIG.

Next, with reference to FIG. 20, the functional configuration of the hyperparameter management system according to the fifth embodiment of the present invention will be described.

FIG. 20 is a diagram showing a functional configuration 2050 of the hyperparameter management system according to the fifth embodiment of the present invention. As described above, in the fifth embodiment of the present invention, a second hyperparameter set is generated by synthesizing the independently generated hyperparameters for higher processing and hyperparameters for higher model.

As shown in FIG. 20, the hyperparameter synthesis unit 325 extracts hyperparameters for higher processing from the hyperparameter database 134 for higher processing, extracts hyperparameters for higher model from hyperparameter database 234 for higher model, and extracts the hyperparameters. A second hyperparameter set is generated by synthesizing the hyperparameters for higher processing and the hyperparameters for higher model.

After that, the second hyperparameter set is transmitted to the processing unit 121 and the model creation unit 122, and is used for processing the target data and creating the machine learning model. After that, the evaluation unit 125 calculates the evaluation points for the second hyperparameter set, and stores the hyperparameters that achieve the predetermined evaluation criteria in the upper synthetic hyperparameter database 335.
Other than this point, the functional configuration 2050 of the hyperparameter management system shown in FIG. 20 is substantially the same as the functional configuration 250 of the hyperparameter management system shown in FIG. 3, and therefore the description thereof will be omitted here. ..

Next, with reference to FIG. 21, the process by the hyperparameter synthesizer according to the fifth embodiment of the present invention will be described.

FIG. 21 is a flowchart showing the process 2100 by the hyperparameter synthesis unit according to the fifth embodiment of the present invention. This process 2100 shows a process of synthesizing independently generated hyperparameters for higher processing and hyperparameters for higher model.

First, in step 1901, the compositing unit determines whether or not there are a necessary number of processing hyperparameters and model hyperparameters for compositing. As a general rule, at least one hyperparameter for higher-level processing and at least one hyperparameter for higher-level model are required for compositing.
Therefore, the synthesis unit refers to the hyperparameter database for upper processing and the hyperparameter database for upper models described above, and confirms whether a sufficient number of hyperparameters are stored. If there are a sufficient number of hyperparameters, this process proceeds to step 1902, and if there are a sufficient number of hyperparameters, this process ends.

Next, in step 1902, the synthesis unit extracts the hyperparameters for higher processing from the hyperparameter database for higher processing, extracts the hyperparameters for higher model from the hyperparameter database for higher model, and extracts the hyperparameters for higher processing. Synthesize parameters and hyperparameters for higher model.
As an example, the hyperparameter database for higher processing has {"missing_value": "media"} hyperparameters for higher processing, and the hyperparameter database for upper model has {"number_of_layers": 2} hyperparameters for higher model. If there is, the synthesizer synthesizes these hyperparameters and generates the hyperparameters of {"missing_value": "media", "number_of_layers": 2} as a second set of hyperparameters.
For convenience of explanation, the case where there is one hyperparameter for higher processing and one hyperparameter for higher model has been described, but the present invention is not limited to this, and when there are a plurality of hyperparameters of each type. Needless to say that is also applicable. In this case, the synthesis unit may generate all the permutations of the hyperparameters for higher processing and the hyperparameters for higher models.

Next, in step 1903, the synthesis unit transmits the generated second hyperparameter set to the processing unit and the model creation unit. After that, as described above, the processing unit processes the target data using the processing hyperparameters included in the second hyperparameter set, and the model creation unit uses the model included in the second hyperparameter set. Generate a machine learning model using hyperparameters.

The functions of the embodiments according to the present invention described above may be realized as a program product. The program product here is a medium that can be read by a computer on which a computer program is recorded. Examples of the storage medium for recording the program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, optical disks, magneto-optical disks, CD-Rs, magnetic tapes, non-volatile memory cards, ROMs, and the like.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist of the present invention.

100 Hyperparameter management server 110 Processor 120 Memory 121 Processing unit 122 Model creation unit 123 Change unit 124 Reinforcement learning unit 125 Evaluation unit 130 Storage unit 131 Candidate hyperparameter database 132 Upper hyperparameter database 135A, 135B Client terminal 225 Communication network

Claims (8)

It is a hyperparameter management device
A reinforcement learning unit that generates a first set of hyperparameters including processing hyperparameters and model hyperparameters based on the candidate hyperparameter space stored in the candidate hyperparameter database.
A processing unit that processes target data based on the processing hyperparameters and generates training information and test information.
A model creation unit that creates a machine learning model based on the hyperparameters for the model using the training information.
An evaluation unit that verifies the machine learning model using the test information and calculates an evaluation score for the first hyperparameter set.
A high-level hyperparameter database that stores hyperparameters that meet certain evaluation criteria,
A change part that generates a second hyperparameter set based on the upper hyperparameters stored in the upper hyperparameter database, and
A hyperparameter management device characterized by. The changed part
Among the hyperparameters stored in the upper hyperparameter database
By exchanging the value of the first hyperparameter with the value of the second hyperparameter, the set of the second hyperparameter is generated.
The hyperparameter management device according to claim 1. The changed part
Among the hyperparameters stored in the upper hyperparameter database
The second hyperparameter set is generated by synthesizing the processing hyperparameters and the model hyperparameters.
The hyperparameter management device according to claim 1. The model creation unit
Create a second machine learning model based on the second set of hyperparameters.
The second machine learning model generates a prediction result based on the target data.
The hyperparameter management device according to claim 1. The hyperparameter management device
The prediction result is compared with the desired prediction result distribution, and
When the prediction result achieves the similarity criterion with respect to the desired prediction result distribution, the evaluation score of the second hyperparameter set is increased.
A prediction adjustment unit that reduces the evaluation score of the second hyperparameter set when the prediction result does not achieve the similarity criterion with respect to the desired prediction result distribution.
The hyperparameter management device according to claim 4, further comprising. The reinforcement learning department
Based on the second set of hyperparameters, trained using the policy gradient method,
The hyperparameter management device according to claim 1. It is a hyperparameter management method
A process of generating a first set of hyperparameters including processing hyperparameters and model hyperparameters based on the candidate hyperparameter space, and
A process of processing target data based on the processing hyperparameters and generating training information and test information, and
The process of creating a machine learning model based on the hyperparameters for the model using the training information, and
The process of verifying the machine learning model using the test information and calculating the evaluation score for the first hyperparameter set, and
Achieve certain evaluation criteria The process of generating a second set of hyperparameters based on the upper hyperparameter database, and
Hyperparameter management methods including. It is a hyperparameter management program product
The hyperparameter management program product
A process of generating a first set of hyperparameters including processing hyperparameters and model hyperparameters based on the candidate hyperparameter space, and
A process of processing target data based on the processing hyperparameters and generating training information and test information, and
The process of creating a machine learning model based on the hyperparameters for the model using the training information, and
The process of verifying the machine learning model using the test information and calculating the evaluation score for the first hyperparameter set, and
Achieve certain evaluation criteria The process of generating a second set of hyperparameters based on the upper hyperparameter database, and
A program product that records computer programs for executing.

Images