KR102257082B1 - Apparatus and method for generating decision agent - Google Patents

Abstract

Disclosed are a device and method for generating a decision-making agent. When a creation of optimization and automation models related to a corporate decision-making is requested, the present invention can create and provide a model thereof. The device for generating the decision-making agent comprises: a training agent part; and a deploy agent part.

Description

Device and method for generating decision agents {APPARATUS AND METHOD FOR GENERATING DECISION AGENT}

The present invention relates to an apparatus and method for generating a decision agent, and more particularly, to an apparatus and method for generating a decision agent for generating a model when an optimization and automation model related to a company's decision-making is requested. will be.

Artificial intelligence (AI) systems are computer systems that implement human-level intelligence. Unlike conventional rule-based smart systems, AI is a system that learns, judges, and becomes smart.

As the recognition rate and user preferences are more accurately understood using artificial intelligence, existing rule-based smart systems are gradually being replaced by deep-based artificial intelligence systems.

Artificial intelligence technology consists of machine learning (deep learning) and element technology using machine learning.

Machine learning (ML) is an algorithmic technology that classifies/learns features of input data by itself.

Elemental technology is a technology that simulates human brain functions such as recognition and judgment using machine learning algorithms such as deep learning, and consists of technical fields such as language understanding, visual understanding, reasoning/prediction, knowledge expression, and motion control. .

On the other hand, reinforcement learning, a kind of machine learning, is a program that achieves an optimal goal using an agent, and it achieves optimization by performing reinforcement learning using rewards (compensation).

However, there is a problem that the concept of reinforcement learning is difficult and cases are insufficient, so that it cannot be sufficiently used in the decision-making process of a company through an agent.

In addition, the conventional agent for reinforcement learning is configured so that a user with specialized knowledge can perform a complex coding task using a simulator, etc., so it is difficult to use a user without specialized knowledge in reinforcement learning. have.

In addition, there is a problem in that an agent for reinforcement learning in the related art cannot check progress information on a subsequent process after performing a task for generating an optimization or automation model.

Korean Patent Application Publication No. 10-2019-0087635 (title of the invention: method and apparatus for automated decision making)

In order to solve this problem, an object of the present invention is to provide an apparatus and method for generating a decision agent that generates a model for optimization and automation model generation related to a company's decision-making request.

In order to achieve the above object, an embodiment of the present invention is an apparatus for generating a decision agent, which generates and trains an arbitrary reinforcement learning target model based on input data for a business domain. A training agent unit for training the reinforcement learning model by reflecting user setting data; And a deployment agent unit for deploying the reinforcement learning model generated by the training agent unit.

In addition, the apparatus for generating a decision agent according to the above embodiment includes an environment element of reinforcement learning input by drag & drop or coding from a user, optimization data and reinforcement learning to increase the learning speed and performance of reinforcement learning. And a user interface unit that outputs setting data consisting of at least one of algorithm selection data to the training agent unit.

In addition, the user interface unit according to the embodiment requests the training agent unit to search for learning information of the reinforcement learning model distributed through the deployment agent unit, and visually converts and displays evaluation data and monitoring data of the searched reinforcement learning model. It is done.

Further, the training agent unit according to the embodiment may include a data storage unit that stores input business domain data, and outputs environment information for generating a reinforcement learning model and performing reinforcement learning based on the stored data; A built-in model unit that generates a reinforcement learning model based on the stored data; A customization function unit configured to reflect a reward function input from a user in environmental information for performing the reinforcement learning; A training unit that generates an agent using a reinforcement learning model based on the environment information and a reward function, and learns the generated agent; And a learned agent storage unit storing the reinforcement learning model learned through the agent, wherein the training unit generates a plurality of agents to learn the reinforcement learning model in parallel, and visualizes the performance value of the reward for each agent's learning. It is characterized in that it is displayed.

In addition, the training agent unit according to the embodiment may include a catalyst learning unit configured to set optimization information for increasing the learning speed and performance of reinforcement learning and automatic compensation information to facilitate application of reinforcement learning; And an RL algorithm unit for selecting a reinforcement learning algorithm for learning an agent according to the business domain.

In addition, the training agent unit according to the embodiment may include a trained model unit that provides a model capable of quickly learning an agent using a model learned in advance; An explainable AI model unit that provides a model for a domain requiring explanation of decision making; And a generative AI model unit that provides a model for generating data in which the missing values are replaced by using the existing data distribution.

In addition, an embodiment of the present invention is a) an exploratory data analysis for the business domain to set an environment for the application of reinforcement learning by the training agent unit based on data on an arbitrary business domain and to proceed with learning (Exploratory Data Analysis, Performing EDA); b) designing a work space by a training agent as a user setting is input through a user interface unit; c) defining, by the training agent unit, environment information and a compensation function for generating a reinforcement learning agent input through a user interface unit, and learning an agent generated by using a reinforcement learning model based on the environment information and the compensation function; And d) deploying, by the deployment agent unit, the reinforcement learning model generated by the training agent unit; wherein, in step c), the training agent unit generates a plurality of agents to parallelize the reinforcement learning model. It is characterized by training with.

In addition, the step d) according to the embodiment further comprises a step of analyzing and outputting the agent configuration management information, the reinforcement learning model, and learning parameter information through the user interface unit by the training agent.

In addition, step b) according to the above embodiment is characterized in that the previous experiment results are searched in the designed workspace, and experiments are sorted, searched, and analyzed according to the search results.

In addition, step c) according to the above embodiment is characterized in that at least some of the environmental information and the reward function set through the user interface unit are automatically set based on the metadata of the business domain data set.

In addition, step c) according to the above embodiment is characterized in that the performance value of the reward is visualized and displayed for each learning of each agent through the user interface unit.

The present invention has the advantage of generating and providing an optimization and automation model related to a company's decision-making.

In addition, the present invention has the advantage that a user without knowledge of reinforcement learning can learn by easily setting and applying core elements of machine learning.

In addition, the present invention has the advantage of being able to easily create a reinforcement learning agent with only a user's domain knowledge and general machine learning knowledge through the reinforcement learning agent.

In addition, the present invention has the advantage of being able to create a high-level decision agent by constructing various reinforcement learning designs for decision-making problems with minimal effort.

1 is a block diagram schematically illustrating an apparatus for generating a decision agent according to an embodiment of the present invention.
FIG. 2 is a block diagram showing the configuration of a training agent unit of the apparatus for generating a decision agent according to the embodiment of FIG. 1;
3 is a flowchart illustrating a method of generating a decision agent according to an embodiment of the present invention.
4 is an exemplary diagram for explaining an environment setting process of a method for generating a decision agent according to the embodiment of FIG. 3.
5 is an exemplary view for explaining a learning result of the method for generating a decision agent according to the embodiment of FIG. 3.
FIG. 6 is another exemplary view for explaining a screen for searching a learning result for a completed experiment of a method for generating a decision agent according to the embodiment of FIG. 3.
7 is an exemplary diagram for explaining a component design process of the method for generating a decision agent according to the embodiment of FIG. 3.
8 is another exemplary view for explaining a component design process of the method for generating a decision agent according to the embodiment of FIG. 3.

Hereinafter, the present invention will be described in detail with reference to a preferred embodiment of the present invention and the accompanying drawings, but it will be described on the premise that the same reference numerals refer to the same elements.

Prior to describing specific details for the implementation of the present invention, it should be noted that configurations that are not directly related to the technical gist of the present invention have been omitted within the scope not disturbing the technical gist of the present invention.

In addition, terms or words used in the present specification and claims are meanings and concepts consistent with the technical idea of the invention based on the principle that the inventor can define the concept of an appropriate term to describe his invention in the best way. Should be interpreted as.

In the present specification, the expression that a certain part "includes" a certain component does not exclude other components, but means that other components may be further included.

In addition, terms such as "... unit", "... group", and "... module" mean units that process at least one function or operation, which can be classified into hardware, software, or a combination of the two.

In addition, the term “at least one” is defined as a term including the singular and plural, and even if the term “at least one” does not exist, each component may exist in the singular or plural, and may mean the singular or plural. Will say self-explanatory.

In addition, it will be said that each component is provided in a singular or plural number and may be changed according to embodiments.

Hereinafter, a preferred embodiment of an apparatus and method for generating a decision agent according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram schematically illustrating an apparatus for generating a decision agent according to an embodiment of the present invention, and FIG. 2 is a block diagram showing a configuration of a training agent unit of the apparatus for generating a decision agent according to the embodiment of FIG. 1.

1 and 2, the apparatus for generating a decision agent according to an embodiment of the present invention includes a training agent unit 100 and a deployment agent to generate and provide an optimization and automation model related to a company's decision making. It may be configured to include an agent unit 200 and a user interface unit 300.

The training agent unit 100 generates and trains an arbitrary reinforcement learning target model based on input data for the business domain.

The business domain may include a state and a reward, and may be inputs to be responded to by the agent and knowledge provided to the agent. For example, in the case of fixed automation of automobile manufacturing, the process of the manufacturing process, materials, etc. This is the business information you need to know when modeling.

In addition, the training agent unit 100 learns the reinforcement learning model by reflecting user setting data input through the user interface unit 300 in learning the reinforcement learning target model.

In addition, the training agent unit 100 outputs the reinforcement learning model in which the training has been completed to the deployment agent unit 200, and the data storage unit 110, the built-in model unit 111, the customization function unit 112, and , Catalyst learning unit 113, RL algorithm unit 114, learned model unit 115, explainable AI model unit 116, generative AI model unit 117, and training unit ( 118) and a learned agent storage unit 119 may be included.

The data storage unit 110 stores input business domain data (eg, status, rewards, actions, real-time observations, placement observations, etc.), and reinforcement learning in the training agent unit 100 based on the stored data. Generates the model and outputs it as environmental information for performing reinforcement learning.

The built-in model unit 111 generates a reinforcement learning model based on data stored in the data storage unit 110.

In addition, the built-in model unit 111 prepares an environment for performing reinforcement learning and reinforcement learning models by linking not only data stored when generating various reinforcement learning models, but also a specific layer such as a catalyst layer. It can also be built-in.

The customization function unit 112 is set to reflect a reward function input from a user in environmental information for performing reinforcement learning.

That is, the customization function unit 112 is a user-customized reward, a wizard-based reward, and a simple learning method that utilizes the correct answer so that the reward function, purpose, etc. desired by the user different from the prior art can be reflected. And reinforcement learning baselines and auto rewards that can be used for verification purposes may be configured to be set.

The catalyst learning unit 113 sets optimization information for increasing the learning speed and performance of reinforcement learning and automatic compensation information so that application of reinforcement learning can be easily performed.

That is, the catalytic learning unit 113 can set a quick understanding of the simulated models by the agent, a good state configuration, an optimal architecture configuration, and an automatic compensation function system using the user interface 300. To be there.

In addition, the catalytic learning unit 113 performs pre-processing according to the type of state such as structured data, image data, and text data, and automatically avoids dimensional overfitting for a given state through an algorithm. To be able to.

In other words, through modules such as automatic state configuration, missing value substitution, continuous variable, categorical variable, dimension reduction, variable selection, outlier removal, image noise removal, data augmentation, resizing, tokenizer, filtering, cleansing, etc. Various pretreatments can be performed.

In addition, the catalyst learning unit 113 may search for an optimal neural network architecture through reinforcement learning, evolutionary, Bayesian optimization, gradient-based optimization, and the like.

In addition, the catalytic learning unit 113 determines the hyper-parameters that have a great influence on the agent's performance, such as Grid-Search, Bayesian Optimization, Gradient-based Optimization, and Population. based Optimization) may be used to search, and an optimal hyperparameter combination may be provided based on the search result.

In addition, the catalytic learning unit 113 may automatically set a reward required for reinforcement learning according to a preset reward pattern.

The RL algorithm unit (Reinforcement Learning Algorithm, 114) selects a reinforcement learning algorithm for learning an agent according to the business domain.

That is, the RL algorithm unit 114 may select model-free reinforcement learning, model-based algorithms, hierarchical RL algorithms, multi-agent algorithms, and the like.

Here, model-free reinforcement learning is a value-based algorithm composed of DQN (Deep Q Networks), DDQN (Double Deep Q Networks), Dueling DDQN (Dueling Double Deep Q Networks), etc., and A2C (Advantage Actor Critic), TRPO (Trust Region). It may be composed of AC-based algorithms such as Policy Optimization), Proximal Policy Optimization (PPO), Deep Deterministic Policy Gradient (DDPG), Soft Actor Critic (SAC), and policy-based algorithms such as Direct Policy Search (DPS).

In addition, model-based reinforcement learning, unlike model-free reinforcement learning, provides an algorithm that the model learns in the presence of information about the environment, and it is an algorithm that learns the agent using the transition model of the model-based algorithm. , DynA, PILCO (Probabilistic Inference for Learning Control), MCTS (Monte-Carlo Tree Search), World Models, etc.

In addition, model-based reinforcement learning uses both real data and data from a simulation environment when updating policies, and it is possible to train a transfer model using real data or use a mathematical model such as a linear quadratic regulator (LQR).

The hierarchical RL algorithm is a business domain that is too complex to solve a problem with a single agent, divides the agent into multiple layers, and the agents of each layer learn with their own reinforcement learning algorithm, and help the master agent learn. It provides an algorithm with a structure that can be used.

The multi-agent algorithm provides an algorithm to learn through competition or collaboration among agents when a plurality of agents exist in one environment.

In addition, the RL algorithm unit 114 is an algorithm used for learning an unlabeled data set by learning an agent like supervised learning or by inversely finding a reward function with a labeled data set, Long Short Term Memory (LSTM). , MAML (Model-Agnostic Meta Learning), MQL (Meta Q Learning), and other meta-RL algorithms, batch RL algorithms that learn using offline data in business domains where real-time interaction with the environment is difficult, and algorithms using A2GAN are provided. May be.

The trained model unit 115 is a component that provides a model capable of rapidly learning an agent by using a model learned in advance from the learned agent 500, and allows reuse of an existing trained model.

Explainable AI Models (116) explain whether these models work well when reinforcement learning models are operated, or provide models for domains that require explanations for decision making.

In other words, the explainable AI model unit 116 provides a model for a domain that requires explanation of decision making because the neural network algorithm including reinforcement learning lacks explanatory power for the learning result.

The generative AI model unit (Generative Models, 117) provides a model that generates data with missing values by using the existing data distribution for data with missing values.

In other words, it provides a model that generates data with missing values by using the existing data distribution, and the data can be augmented to solve the data shortage problem. It can also be provided as a model with this.

The training unit 118 generates an agent using a reinforcement learning model based on environmental information and a reward function, and learns the generated agent.

In addition, the training unit 118 generates a plurality of agents to learn the reinforcement learning model in parallel.

That is, since it takes a lot of time to generate one reinforcement learning model, several reinforcement learning models are generated and then the reinforcement learning model is trained in parallel.

In addition, the training unit 118 stores the learned reinforcement learning model in the learned agent storage unit 219, and when additional learning is required, the reinforcement learning model stored in the learned agent storage unit 119 is retrained.

In addition, the training unit 118 visualizes the performance value of the reward for each learning of each agent and displays it through the user interface unit 300, so that the user can more easily check the learning process of the reinforcement learning model.

The learned agent storage unit 119 stores the reinforcement learning model learned through the agent in the training unit 118.

The deployment agent unit 200 deploys the reinforcement learning model generated by the training agent unit 100 to the business domain unit 400.

That is, the deployment agent unit 100 controls an arbitrary business domain, such as finance, energy, logistics, material, marketing, robotics, and system from the business domain unit 400 accessed through an API (Application Programming Interface) service. In response to requests for reinforcement learning models, such as variable prices, optimized and automated reinforcement learning models can be distributed.

The user interface unit 300 includes at least one of an environmental element of reinforcement learning input by drag & drop or coding from a user, optimization data for increasing the learning speed and performance of reinforcement learning, and data for selecting a reinforcement learning algorithm. The configured setting data is output to the training agent unit 100.

That is, the user interface unit 300 allows the user to selectively configure environment information and rewards whenever necessary to generate a new reinforcement learning model.

In addition, the user interface unit 300 can sample and use information from what it wants to perform, and when simulating, it is possible to use it by using a set value, etc., so that an environment called a simulator is not configured. Instead, the data itself is set as an environment, and the data can be retrieved and set through sampling.

The user interface unit 300 is free from coding using a conventional simulator, and performs a task through a click using the user interface unit 300, so that the user can intuitively recognize the task performance and the result.

In addition, the user interface unit 300 is a customized reward that is defined and used from use, a wizard reward that uses a variable existing in data or a key performance indicator (KPI) of each company as a weight control method, The user may request the training agent unit 100 to set any one of auto rewards that can be used for simple learning and reinforcement learning baseline verification purposes.

In addition, the user interface unit 300 requests the training agent unit 100 to search for learning information of the reinforcement learning model distributed through the deployment agent unit 200, and evaluates data including analysis of the results of the retrieved reinforcement learning model, and The monitoring data is visually converted and displayed.

Next, a method of generating a decision agent according to an embodiment of the present invention will be described.

3 is a flowchart illustrating a method of generating a decision agent according to an embodiment of the present invention.

1 to 3, in a method for generating a decision agent according to an embodiment of the present invention, the training agent unit 100 uploads and stores data for an arbitrary business domain, and sets an environment for applying reinforcement learning. To proceed with the learning process, exploratory data analysis (EDA) for the business domain is performed (S100).

That is, in step S100, data is stored in the data storage unit 110 in a drag-and-drop method, statistics and graphs summarized for each variable are provided, or if there is an incorrect metadata specification, correction or visualization for each variable is performed. do.

After performing step S100, as user settings are input through the user interface unit 300, the training agent unit 100 designs a work space (S200).

As shown in FIG. 4, on the environment setting screen 600, the user may set his own work space through the work space input unit 610, and set the data set to be used in the current experiment through the data search unit 620.

In addition, the results of the current experiment may be displayed by sorting information such as performance values through Trial No (630) and Reward (640).

In addition, in step S200, the user may create and analyze an experiment.

As shown in FIG. 5, analysis of the experiment may be performed through the learning result screen 700 and the learning result search screen 710 for the completed experiment as shown in FIG. 6.

Subsequently, the training agent unit 100 defines environment information and a compensation function for generating the reinforcement learning agent input through the user interface unit 300, and is generated using a reinforcement learning model based on the environment information and the compensation function. One agent is learned (S300).

In addition, at least a part of the environment information and the compensation function set through the user interface unit 300 in step S300 may be automatically set based on the metadata of the business domain data set.

That is, the part to be set by the user is automatically filled based on the metadata of the data set, and the status or statistical value of the corresponding data in the current experiment can be checked.

In addition, in step S300, elements for creating a reinforcement learning agent by the user, for example, environmental information such as state, action, and reward, are defined through the user interface unit 300 through click by click. can do.

In addition, the reward is a customized reward setting, such as the user-defined reward screen 800, as shown in FIG. 7, so that the reward function, purpose, etc. desired by the user different from the prior art can be reflected, and a wizard-based reward screen as shown in FIG. It can also be configured to set the reward in a way that uses the correct answer and a wizard-based reward setting such as (810).

That is, if the user does not code the entire reward function and only defines a specific reward function, it is internally combined and operated, so that even a user without specialized knowledge can easily define a reward through a click using the user interface unit 300. have.

In addition, in step S300, the training agent unit 100 generates a plurality of agents to learn the reinforcement learning model in parallel.

That is, the training agent unit 100 generates various reinforcement learning models and agents by using a parallel processing function to learn at the same time, and visualizes the performance value of the reward through the user interface unit 300 whenever each agent learns. By making it displayed, the user can intuitively check and analyze.

After performing step S300, the deployment agent unit 200 distributes the reinforcement learning model generated by the training agent unit 100 to the business domain unit 400 (S400).

In addition, the training agent unit 100 provides configuration management information of the agent through the user interface unit 300 so that information of a specific agent can be reused.

In addition, step S400 analyzes and outputs the reinforcement learning model and learning parameter information through an interactive visualization function using the user interface unit 300 in the corresponding experiment, so that the user can later apply it according to his or her domain.

Therefore, optimization and automation models related to corporate decision-making can be created and provided, and users without knowledge of reinforcement learning can easily set and apply key elements of machine learning to learn.

In addition, reinforcement learning agents can be easily created with only the user's domain knowledge and general machine learning knowledge through the reinforcement learning agent, and a high-level decision agent is created by constructing various reinforcement learning designs for decision-making problems with minimal effort. You can do it.

As described above, although it has been described with reference to a preferred embodiment of the present invention, those skilled in the art will variously modify and change the present invention within the scope not departing from the spirit and scope of the present invention described in the following claims. You will understand that you can do it.

In addition, the reference numbers in the claims of the present invention are provided for clarity and convenience of description, and are not limited thereto. In the process of describing the embodiments, the thickness of the lines shown in the drawings, the size of components, etc. May be exaggerated for clarity and convenience of description.

In addition, the above-described terms are terms defined in consideration of functions in the present invention and may vary according to the intentions or customs of users and operators, so interpretation of these terms should be made based on the contents throughout the present specification. .

In addition, even if not explicitly shown or described, a person having ordinary knowledge in the technical field to which the present invention pertains can make various modifications including the technical idea according to the present invention from the description of the present invention. It is obvious, and this still belongs to the scope of the present invention.

In addition, the above embodiments described with reference to the accompanying drawings are described for the purpose of describing the present invention, and the scope of the present invention is not limited to these embodiments.

10: agent generating device 100: training agent unit
110: data storage unit 111: built-in model unit
112: customization function unit 113: catalyst learning unit
114: RL algorithm unit 115: learned model unit
116: Explainable AI model unit 117: Generative AI model unit
118: training unit 119: learned agent storage unit
200: deployment agent unit 300: user interface unit
400: business domain unit 500: learned agent
600: environment setting screen 610: workspace input unit
620: Data search unit 700: Learning result screen
710: Learning result search screen 800: User-defined reward screen
810: Wizard-based reward screen

Claims (11)

A training agent unit 100 for generating and training a reinforcement learning target model based on input data for the business domain, and training the reinforcement learning model by reflecting user setting data in learning of the reinforcement learning target model; And
Including; a deployment agent unit 200 for deploying (Deploy) the reinforcement learning model generated by the training agent unit 100; and
The training agent unit 100 may include a data storage unit 110 for storing inputted business domain data and outputting environment information for generating a reinforcement learning model and performing reinforcement learning based on the stored data;
A built-in model unit 111 that generates a reinforcement learning model based on the stored data;
A customization function unit 112 configured to reflect a reward function input from a user in environmental information for performing the reinforcement learning;
A catalyst learning unit 113 that sets optimization information for increasing the learning speed and performance of the reinforcement learning and automatic compensation information so that application of the reinforcement learning can be easily performed; And
An RL algorithm unit 114 for selecting a reinforcement learning algorithm for learning an agent according to the business domain;
A training unit 118 that generates an agent using the reinforcement learning model based on the environmental information and the reward function, and learns the generated agent; And
Including; a learned agent storage unit 119 for storing the reinforcement learning model learned through the agent,
The training unit 118 generates a plurality of agents, trains the reinforcement learning model in parallel, and visualizes and displays a performance value of a reward for each learning of each agent. The method of claim 1,
Setting data consisting of at least one of reinforcement learning environment elements, optimization data for increasing the learning speed and performance of reinforcement learning, and reinforcement learning algorithm selection data input by drag & drop or coding from the user The apparatus for generating a decision agent, further comprising: a user interface unit 300 outputting to the unit 100. The method of claim 2,
The user interface unit 300 requests the training agent unit 100 to search for learning information of the reinforcement learning model distributed through the deployment agent unit 200, and visually retrieves the evaluation data and monitoring data of the retrieved reinforcement learning model. An apparatus for generating a decision agent, characterized in that to convert and display. delete delete The method of claim 1,
The training agent unit 100 includes a trained model unit 115 that provides a model capable of quickly learning an agent using a model learned in advance;
An explainable AI model unit 116 that provides a model for a domain requiring explanation of decision making; And
And a generative AI model unit 117 that provides a model for generating data in which the missing values are replaced by using the existing data distribution. A method for generating a decision agent by a decision agent generating apparatus including a training agent unit 100, a deployment agent unit 200, and a user interface unit 300,
a) The training agent unit 100 performs an exploratory data analysis (EDA) on the business domain to set the environment for applying reinforcement learning and to proceed with learning based on the data on an arbitrary business domain. step;
b) as the user setting is input through the user interface unit 300, the training agent unit 100 designing a work space;
c) The training agent unit 100 defines environment information and a compensation function for generating a reinforcement learning agent input through the user interface unit 300, and uses a reinforcement learning model based on the environment information and the compensation function. Learning the created agent; And
d) deploying, by the deployment agent unit 200, the reinforcement learning model generated by the training agent unit 100; including,
In step c), the training agent unit 100 generates a plurality of agents to train the reinforcement learning model in parallel. The method of claim 7,
The step d) further comprises the step of analyzing and outputting, by the training agent unit 100, the agent configuration management information, the reinforcement learning model, and the learning parameter information through the user interface unit 300. How to create a decision agent. The method according to claim 7 or 8,
Step b) is to search for previous experiment results in the designed workspace,
A method of creating a decision agent, characterized in that it sorts, searches, and analyzes experiments according to search results. The method according to claim 7 or 8,
In step c), at least some of the environmental information and the reward function set through the user interface unit 300 are automatically set based on the metadata of the business domain data set. The method according to claim 7 or 8,
The step c) is a method for creating a decision agent, characterized in that the user interface unit 300 visualizes and displays the performance value of the reward for each learning of each agent.

Images