TWI787669B - System and method of automated machine learning based on model recipes - Google Patents

Abstract

A machine learning system at least includes an application program interface, an optimization module, and atrial execution module. The optimization module is composed of database, recipe pool and advisor pool. The machine learning method selects the best recipe from the recipe pool by the user, recommends a project setting based on the selected recipe, and then selects an advisor from the advisor pool based on the project setting. The selected advisor generates an optimized hyperparameter combination, and provides user an optimized machine learning model.

Description

System and method for automatic machine learning based on model prescription

The present invention relates to a system and method for machine learning, and in particular to a system and method for automatically optimizing machine learning.

Due to the vigorous development of the information industry, artificial intelligence has become a prominent science in the industry in recent years, especially in the industrial application of Industry 4.0, artificial intelligence plays an important role. Machine learning is an important branch of artificial intelligence in industrial applications. It is widely used in data mining, computer vision, natural language processing, biometric identification, search engines, medical diagnosis, detection of financial fraud, securities market analysis, DNA sequencing, speech and Handwriting recognition, strategic games and robotics and other fields. Generally speaking, the process of machine learning mainly includes seven steps: gathering data, preparing the data, choosing model, training machine, evaluating and analyzing, and parameter tuning. Hyperparameter tuning and prediction. Among them, parameter tuning has a great impact on the performance and effectiveness of machine learning models.

Common open source hyperparameter tuning tools include Hyperopt, NNI, AutoKeras, etc.; in addition, software companies such as Google, Microsoft, Amazon, and SigOpt also provide their own hyperparameter tuning services. Regardless of the hyperparameter tuning tool, the user still needs to determine a set of suitable models, corresponding parameter search ranges, and algorithms for the machine learning target in order to obtain an optimal combination of hyperparameters. In addition to the decision of the model/search scope/algorithm, the user also needs to build the model for subsequent training, testing, deployment (deploy) and maintenance. Each of the above-mentioned links requires users to have considerable professional knowledge and accumulated experience; therefore, it is a high-threshold and long-term process to introduce machine learning into actual field applications. The present invention greatly reduces the threshold and time for machine learning introduction by developing a machine learning system and method that accumulates numerical experience, systematizes, and automatically generates optimized (optimized) hyperparameter combinations.

One aspect of the present invention is to provide a machine learning system and method, which can store development experience numerically in a database, and provide a machine learning model/search space/search algorithm combination from the recipe pool.

Another aspect of the present invention is to provide a machine learning system and method, which can automatically generate optimized hyperparameter combinations.

Another aspect of the present invention is to provide a machine learning system and method, which has an executable library for performing simulation tests to optimize hyperparameter combinations, and can be imported into field devices.

According to an embodiment of the present invention, a machine learning system includes: an API, an optimization module, and a test execution module. API for inputting a project setting; optimized module connection API. The optimization module further includes: a database for storing multiple project information; a consultant pool for storing multiple consultants; and a prescription pool for storing multiple prescriptions. The test execution module connects the API and optimization modules. After the optimization module receives the project requirements, it selects a prescription from the prescription pool according to the project requirements; the prescription generates suggested project settings and sends them back to the user; the user adjusts and confirms the project settings and then submits them to the optimization module through the API; optimization The module selects a consultant from the consultant pool according to the project settings. The execution steps of the machine learning system include: A. Optimizing the module according to According to the selected consultant, a hyperparameter combination is generated based on the project information; B. The test execution module receives the hyperparameter combination, executes a training test process according to the hyperparameter combination, and generates a new project information, and stores and updates the data Library; C. Repeat steps A and B in a predetermined loop interval; D. Output an optimization project result.

The method of optimizing the module selection prescription includes the reinforcement learning method. The machine learning system further includes a user interface connected to the application programming interface through a network link, and project requirements are input through the user interface.

The test execution module further includes at least a plurality of executable functions, and the test execution module uses the executable functions to execute the training and testing process.

According to another embodiment of the present invention, a machine learning method is suitable for generating optimized hyperparameter combinations, wherein the machine learning method is executed in an optimization module and a test execution module. The optimization module includes: a database for storing multiple project information; a consultant pool for storing multiple consultants; and a prescription pool for storing multiple prescriptions. The machine learning method includes the following steps: A. receiving a project requirement; B. selecting a prescription from the prescription pool according to the project requirement, generating a project setting according to the prescription, and selecting a consultant from the consultant pool according to the project setting; C. optimizing the model The group generates a hyperparameter combination based on the project information based on the selected consultant; D. The test execution module receives the hyperparameter combination, executes a training test process according to the hyperparameter combination, and generates a new project information, and stores and updates it Database; E. Repeat steps C and D in a predetermined loop interval; F. Generate optimized hyperparameter combinations.

The method for selecting a prescription by the optimization module at least includes a reinforcement learning method. The test execution module includes: a plurality of executable functions, and the test execution module uses the executable functions to execute the training and testing process. The machine learning method further includes: performing machine learning on a field device by using an executable function and optimizing a combination of hyperparameters.

The above summary is intended to provide a simplified summary of the disclosure to provide readers with a basic understanding of the disclosure. This summary is not an extensive overview of the disclosure and it is not intended to identify key/critical elements of the embodiments of the invention or to delineate the scope of the invention. After referring to the following embodiments, those with ordinary knowledge in the technical field of the present invention can easily understand the basic spirit and other invention objectives of the present invention, as well as the technical means and implementation aspects adopted by the present invention.

100: Machine Learning Systems

102: Optimize the module

104: Database

106: Core

108: Prescription pool

110: Consultant pool

112:Application Programming Interface

114:Test Execution Module

116: User Portal Interface

118: user

120: field device

200: Machine Learning Methods

202, 203, 204, 206, 208, 210, 212, 214, 216, 218, 220: steps

FIG. 1 shows a schematic diagram of a machine learning system architecture according to an embodiment of the present invention.

FIG. 2 shows a flowchart of a machine learning method according to another embodiment of the present invention.

In the following description, an example structure and an example execution method of the above-mentioned machine learning system and method will be introduced. For the sake of easy understanding of the described embodiments, a number of technical details will be provided below. Of course, not all embodiments require these technical details. Meanwhile, some well-known structures or elements are only drawn schematically in the drawings to appropriately simplify the contents of the drawings.

In order to make the description of the disclosure more detailed and complete, the following provides illustrative descriptions of the implementation aspects and specific embodiments of the present invention; but this is not the only form of implementing or using the specific embodiments of the present invention. The description covers features of various embodiments as well as method steps and their sequences for constructing and operating those embodiments. However, other embodiments can also be used to achieve the same or equivalent functions and step sequences.

Referring to FIG. 1 , it shows a schematic diagram of a machine learning system architecture according to an embodiment of the present invention. According to the machine learning problem to be solved, set the learning model, search space, and search algorithm, and then conduct training, testing and deployment for on-site devices. In order to reduce the application threshold and improve application efficiency, the present invention proposes a machine learning system 100, including an optimization module 102 (Optimizer Module), application program interface 112 (Application Programming Interface, API), and test execution module 114 (Trial and Executing Module). The optimization module 102 is mainly controlled by the kernel 106 (Kernel). The core 106 mainly controls three components: a database 104 (Database), a prescription pool 108 (Recipe Pool) and a consultant pool 110 (Advisor Pool). For the convenience of explanation, we call the construction process of the machine learning model a project. The database 104 stores all project-related information, including but not limited to project number, creation time, owner, time limit, number of times limit, model type, model default parameters, consultant type, parameter search space, test parameter combination, quantitative model performance etc. The data stored in the database 104 provides the core 106 with access to applications for various learning or optimization operations.

A plurality of prescriptions are stored in the prescription pool 108 . The so-called prescription refers to the basic configuration of the project for each problem to be solved by machine learning, including machine learning model setting and parameter space setting. The user 118 inputs the problem to be solved by the machine learning through the user interface 116 , that is, project requirements, and some basic information. After completing the project requirement input, the core 106 will provide a more suitable prescription from the prescription pool 108 for the user to choose according to the project requirement input by the user 118 . The initial prescription data in the prescription pool 108 is the prescription record of the past project, that is, the record of the machine learning model and parameter space corresponding to the problem of the past project. Therefore, there are two methods for the core 106 to provide prescriptions: one is to select the pairing of existing project prescriptions, if the project demand input by the user 118 matches the existing project in the prescription pool 108, then directly output the prescription of the project, provided to the user 118. Another method is to use Reinforcement Learning to generate a better prescription, and store and update it in the prescription pool 108 as a reference for future projects. Machine learning models that can be processed include but are not limited to: supervised learning (linear regression, logistic regression, decision tree, support vector machine, K-neighbor algorithm), cluster analysis, pattern recognition (K-means algorithm, integrated learning AdaBoost, Bayesian classifiers), dimensionality reduction and metric learning (PCA, autoencoders, linear discriminant analysis), structure prediction, anomalies Detection, artificial neural network (feedforward neural network, radial basis function network, recurrent neural network), reinforcement learning type (Markaff chain, Q-learning, Monte Carlo, SARSA (State Action Reward State Action Learning), variational method).

An example of the enhanced learning process of the core 106 to the prescription pool 108 is as follows: 1. The core 106 summarizes the initial parameter space for the specific problem and the machine learning model in the prescription pool 108 according to the specific problem and the machine learning model. 2. Define all possible parameter spaces and possible adjustment actions (for example: increase/decrease/maintain) in each parameter space, and give each possible adjustment action the same probability. 3. Execute the project with the current parameter space (the first parameter space) setting, and record the quantitative performance of the model (the first result). 4. Randomly perform the above adjustment action A1 to obtain a new parameter space (second parameter space). 5. Execute the project with the second parameter space setting, and record the quantitative performance of the model (second result). 6. Compare the first result with the second result, if the first result is better, lower the probability of adjusting action A1, if the second result is better, increase the probability of adjusting action A1, if the two results are the same, keep it constant. 7. Repeat steps 3-6 to converge the parameter space to a better value. The core 106 obtains the optimal parameter space setting according to the reinforcement learning method, that is, a new prescription is created, stored and updated in the prescription pool 108, and the new prescription is provided to the user. The selected recipe provides the user with suitable project settings, including but not limited to model/search space/search algorithm combinations, and is input into the optimization module 102 through the API 112 .

A plurality of consultants are stored in the consultant pool 110 , and the so-called consultants include parameter search algorithms and resource allocation strategies that may be used for the project, so as to optimize the combination of hyperparameters. Parameter search algorithms include but are not limited to: Bayesian optimization, genetic algorithm, annealing method, tree-structured Parzen Estimator (TPE), continuous halving process, overfrequency method (Hyperband), Bayesian The optimized overclocking method (Bayesian Optimization Hyperband, BOHB), the median stop method (Median Stop), curve fitting method (Curve Fitting), etc. In an embodiment of the present invention, the above-mentioned parameter search algorithm in the consultant pool 110 is written as an executable function in C language or Python language, and provided to the core 106 to optimize the combination of hyperparameters.

In one embodiment of the present invention, the core 106 selects an appropriate consultant (ie hyperparameter search algorithm and resource allocation strategy) from the consultant pool 110 according to the project settings selected by the user 118 and input by the prescription suggestion. For example, when the core 106 selects the consultant of the genetic algorithm according to the project setting input by the user 118, the hyperparameter search process includes: 1. Select a set of fixed parameters from the parameter space in the prescription provided by the prescription pool 108 The number of starting hyperparameter combinations to add to the list to be tested. 2. Perform training and testing according to the combination of hyperparameters in the test list and the machine learning model provided by the prescription pool 108, and record the performance results (performance). 3. Select a certain proportion of excellent hyperparameter combinations based on performance results. 4. Remix and arrange the values of excellent hyperparameter combinations to generate new hyperparameter combinations. 5. Put the new hyperparameter combination through a certain degree of variation (for example: give a specific hyperparameter a slight up and down perturbation) and add it to the list to be tested. 6. Repeat steps 2~5 in a certain loop interval. Finally, select the combination of hyperparameters that yields the best results.

The test execution module 114 includes a plurality of executable functions, these executable functions include the library of the above-mentioned various machine learning models, and according to various project settings, the binary execution file, virtual machine image file or virtualization The container is packaged to form an executable function, which can be used for training and testing calculations on the user's computing device, and can also be used for execution on the field device 120 . The test execution module communicates with the optimization module 102 through the application programming interface 112, and transmits data in both directions. According to the hyperparameter combination and project settings provided by the core 106, with appropriate executable functions, the training and testing process can be carried out, and feedback The quantified test results are transmitted to store and update the database 104 .

In an embodiment of the present invention, the API 112 may be composed of a Representational State Transfer API, which uses the HTTP/HTTPS protocol to establish a network between the user 118 and the optimization module 102 Connections, including virtual networks, local area networks, or the Internet. Specifically, the user 118 sends requests and receives feedback messages through a limited number of communication modes defined by the API 112 . These communication modes include: adding, deleting, executing, and restoring projects, project setting suggestions returned by optimization modules, hyperparameter combination requests and suggested results, quantitative performance feedback of training tests, error messages and code reports, etc.

The user portal interface 116 is used to 1. receive the project requirements from the user 118 and return the suggested project settings to the user 118, and 2. collect the project results and present them visually. As shown in FIG. 1 , the user 118 can directly connect to the application programming interface. After the user 118 inputs the project requirements through the user interface 116 , the optimization module 102 returns the suggested project settings. On the other hand, through the application programming interface 112 , the user 118 can submit project settings, execute the training and testing process through the test execution module 114 , and then return the training and testing results. The optimization module 102 does not need to access the data related to training and testing of the user 118, which can ensure the confidentiality of the data of the user 118. In addition, the user interface 116 can present the training and test results in a visualization interface, for example: combining the hyperparameters with the quantified results of the training and testing, using line graphs, distribution bar graphs, two-dimensional coordinate relationship graphs, parallel Coordinate diagrams, 3D point diagrams and surface diagrams are presented.

Please refer to FIG. 2 , which shows a flowchart of a machine learning method according to another embodiment of the present invention. Machine learning method of the present invention is carried out in above-mentioned machine learning system, please refer to Fig. 1 simultaneously, machine learning method 200 of the present invention comprises the following steps:

Step 202 is the beginning of the method. When the user 118 wants to construct a machine learning model, a project is opened. As mentioned above, the user creates project requirements. User 118 enters the The port interface 116 transmits the project requirements to the optimization module 102 . In step 203, the optimization module 102 will provide a suitable prescription for the user to select according to the project requirement input by the user. There are two methods for the core 106 to select prescriptions: one is to select the matching of existing project prescriptions. If the project requirements input by the user 118 match the existing project in the prescription pool 108, the prescription of the project will be directly output and provided to User 118. Another method is to use the reinforcement learning method to generate a better prescription, and store and update it in the prescription pool 108 as a reference for future projects. The reinforcement learning method is as mentioned above, and will not be repeated here. Step 204 , according to the prescription selected by the user, a suggested project setting is provided and input through the API 112 , and the core 106 of the optimization module 102 receives the project setting. Project settings include but not limited to: model/search space/search algorithm combination, project number, creation time, owner, time limit, number of times limit, project category, project problem description, etc. Then proceed to step 206 , select a consultant from the consultant pool 110 according to the project setting.

Step 208 , the optimization module 102 generates a hyperparameter combination based on the project information in the database 104 according to the selected consultant. As previously mentioned, the core 106 uses the selected consultant to search for a better combination of hyperparameters from the parameter space defined in the project settings. The detailed steps are as mentioned above to carry out the hyperparameter optimization process according to the hyperparameter search algorithm.

In step 210, the test execution module 114 receives the hyperparameter combination, and executes a training and testing process according to the hyperparameter combination. Step 212 , according to the training and testing results, a new project information is produced, that is, the quantitative testing results. Step 214 , store the test result and update the database 104 . For example, according to the project settings input by the user 118, the test execution module can select an appropriate executable function. As mentioned above, according to the hyperparameter combination and project settings provided by the core 106, the appropriate executable function can be matched. , the training and testing process can be carried out, and the quantified test results can be returned to store and update the database 104 . Since the project information in the database 104 has been updated and incorporated into the above training and test results, it is possible to generate A better combination of hyperparameters. Relevant implementation methods have been described in the above-mentioned genetic algorithm embodiment, but those skilled in the art should know that the present invention is not limited to the use of genetic algorithm, and other hyperparameter search algorithms can also be used. Even if the system resources are sufficient, It is also possible to use multiple consultants in the consultant pool to search for hyperparameters to achieve better optimization results.

Step 216 , repeatedly execute steps 208 , 210 , 212 , and 214 in predetermined loop intervals. The setting of the loop interval may be the number of loop executions, the loop execution time, or the upper limit of the resources consumed by the loop execution. In step 218, an optimized hyperparameter combination is produced. As mentioned above, when executing steps 208, 210, 212, 214, the selected consultant will filter the hyperparameter combination with better test results, and store and update it in the database 104, so the optimal hyperparameter combination generated according to the updated database 104, It will outperform the best combination of hyperparameters produced by the database 104 before the update. After multiple rounds of execution, an optimized hyperparameter combination will be produced. Step 220, the machine learning method of the present invention is completed so far, the user 118 can obtain the optimized hyperparameter combination of the project to be constructed, which can be matched with the executable function of the test execution module 114, and can be directly applied to the field device 120 to obtain Better machine learning models.

In summary, the optimization module of the present invention can construct an initial prescription pool through experienced experts, and can automatically generate better prescriptions for users through the optimization module, and the machine learning method of the present invention can Generate an optimized hyperparameter combination and match it with an executable function to obtain a complete and optimized machine learning model, which can be easily imported into the field device to perform machine learning.

Although the present invention has been disclosed above in terms of implementation, it is not intended to limit the present invention. Anyone skilled in this art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be defined by the appended patent application scope.

100: Machine Learning Systems

102: Optimize the module

104: Database

106: Core

108: Prescription pool

110: Consultant pool

112:Application Programming Interface

114:Test Execution Module

116: User Portal Interface

118: user

120: field device

Claims (8)

A machine learning system at least includes: an application programming interface for inputting a project setting; an optimization module connected to the application programming interface, and the optimization module at least includes: a database for storing a plurality of project information ; a consultant pool for storing a plurality of consultants; and a prescription pool for storing a plurality of prescriptions; and a test execution module connected to the application programming interface and the optimization module, wherein the optimization module receives a project After the requirement, perform the following steps: (e) the optimization module summarizes the initial parameter space in the prescription pool according to the project requirement; (f) defines a first parameter space and the adjustment action of the first parameter space, and Give each adjustment action the same probability; (g) execute the project with the first parameter space setting, and record the model quantitative performance to obtain a first result; (h) randomly execute the above steps (e)~(g) to obtain a The second parameter space; (i) execute the project with the setting of the second parameter space, and record the quantitative performance of the model to obtain a second result; (j) compare the first result with the second result, if the first result is relatively If the second result is better, the probability of each adjustment action is lowered, and if the second result is better, the probability of each adjustment action is increased, and if the two results are the same, it remains unchanged; (k) Repeat steps (g)~ (j), converging the first parameter space to a better value; (l) the optimization module will converge to the first parameter space of a better value as a consultant, and store the consultant in the consultant pool ; The machine learning system includes performing the following steps: (a) the optimization module generates a hyperparameter combination based on the project information according to the selected consultant; (b) the test execution module receives the hyperparameter combination, and according to the The combination of hyperparameters executes a training and testing process, and generates a new project information, and stores and updates the database; (c) repeatedly execute steps (a) and (b) in a predetermined loop interval; and ( d) Outputting an optimization project result. The machine learning system as claimed in claim 1, wherein the method for selecting one of the prescriptions by the optimization module includes at least a reinforcement learning method. The machine learning system as described in claim 1 further includes at least: a user interface connected to the application program interface through a network link, and the project requirement is input through the user interface. The machine learning system as described in Claim 1, wherein the test execution module further includes: a plurality of executable functions, and the test execution module executes the training by at least one of the executable functions test process. A machine learning method suitable for generating an optimized hyperparameter combination, wherein the machine learning method is executed in an optimization module and a test execution module, and the optimization module further includes: a database for storing a plurality of project information; a consultant pool for storing a plurality of consultants; and a prescription pool for storing a plurality of prescriptions; wherein the machine learning method at least includes: (a) receiving a project requirement; (b) According to the project requirements, perform the following steps (g)~(n), and then perform step (c); (g) summarize the initial parameter space in the prescription pool; (h) define a first parameter space , and the adjustment actions in the first parameter space, and give each adjustment action the same probability; (i) execute the project with the first parameter space settings, and record the model quantification performance to obtain a first result; (j) execute randomly The above steps (g)~(i) obtain a second parameter space; (k) execute the project with the setting of the second parameter space, and record the quantitative performance of the model to obtain a second result; (l) compare the first result With the second result, if the first result is better, the probability of each adjustment action is lowered, if the second result is better, the probability of each adjustment action is increased, and if the two results are the same, the probability of each adjustment action is maintained. (m) repeat steps (i)~(l) to converge the first parameter space to a better value; (n) use the first parameter space converged to a better value as a consultant, and storing the consultant in the consultant pool; (c) the optimization module generates a hyperparameter combination based on the project information according to the selected consultant; (d) the test execution module receives the hyperparameter combination, and according to the The combination of hyperparameters executes a training and testing process, and generates a new project information, and stores and updates the database; (e) repeatedly executes steps (c) and (d) in a predetermined loop interval; and ( f) Output the optimized hyperparameter combination. The machine learning method as claimed in claim 5, wherein the method for selecting one of the prescriptions by the optimization module includes at least a reinforcement learning method. The machine learning method as described in Claim 5, wherein the test execution module further includes: a plurality of executable functions, and the test execution module uses at least one of the executable functions to execute the training test process. The machine learning method as described in Claim 7 further includes: using the executable function and the optimized hyperparameter combination to execute machine learning on a field device.

Images