A method and system for improving multi-agent collaboration using data flywheel-based reinforcement learning

Abstract

In the method for improving multi-agent collaboration using data flywheel-based reinforcement learning according to the present invention, the method comprises: a step in which a service provider server receives basic content information from an educator terminal; a step in which the service provider server analyzes natural language input based on the basic content information; a step in which the service provider server generates at least one hierarchical curriculum structure among a course, a chapter, and a lesson based on the result of the natural language input analysis and the basic content information; a step in which the service provider server evaluates the quality of the generated data to determine whether a standard is met; a step in which, if it is determined that the standard is met, the service provider server generates specific content for each lesson based on the hierarchical curriculum structure; a step in which the service provider server evaluates the quality of the generated data to determine whether the standard is met; a step in which, if it is determined that the standard is met, the service provider server generates evaluation items in conjunction with the specific content of each lesson generated; a step in which the service provider server evaluates the quality of the generated data to determine whether the standard is met; and a step in which, if it is determined that the standard is met, the service provider server integrates the generated hierarchical curriculum structure, the specific content of each lesson, and the evaluation items to store them as a single educational content and transmits it to the educator terminal. The method includes the steps of: an educator terminal requesting a service provider server to generate code corresponding to the educational content; receiving the generated code from the service provider server and distributing the code; a student terminal accessing the educational content based on the code and starting learning; a step in which, upon completion of learning, the student terminal transmits learning data to the service provider server; and a step in which the service provider server analyzes the received learning data to improve the educational content.

Description

Technology Field

[0001] The present invention relates to a method and system for improving multi-agent collaboration using data flywheel-based reinforcement learning. Specifically, it relates to a method and system for improving multi-agent collaboration using data flywheel-based reinforcement learning that utilizes multi-agent-based artificial intelligence to analyze basic content information provided by educators, automatically generates and manages hierarchical curriculum structures and evaluation elements, and continuously improves collaboration efficiency by reflecting learner interaction data. Background Technology

[0003] The modern educational environment is rapidly changing in conjunction with digital transformation. In particular, with the proliferation of online learning platforms and generative AI technologies, active attempts are being made to automatically generate educational content and provide it in a personalized manner for learners. However, existing content generation methods utilizing a single AI model have faced limitations, such as difficulties in quality control and frequent issues where generated results do not align with learning objectives or fail to maintain contextual consistency. Consequently, learners have been provided with results that are merely a list of information or have been exposed to content containing incorrect information.

[0004] Furthermore, existing educational AI systems have tended to focus heavily on generating content on an ad-hoc basis. In other words, feedback loops were not sufficiently incorporated during the content creation process, preventing learner responses or learning data from being systematically utilized for future content improvement. This acted as a factor hindering the personalization of education and limited the improvement of educational quality. Multi-agent based approaches are garnering attention as an alternative to address these issues. When multiple agents collaborate by dividing roles such as interpretation, structural design, content creation, evaluation, and quality control, errors or incompleteness that occur when a single AI handles the process can be reduced. However, existing multi-agent systems often remain at the level of simple role division, lacking systematic mechanisms to improve collaboration efficiency between agents over the long term or to continuously incorporate learner interaction data.

[0005] Meanwhile, the concept of the data flywheel for optimizing AI performance has been studied across various fields. A data flywheel refers to a cyclical structure that collects and analyzes user interaction data and incorporates it into model improvements, thereby enabling the gradual enhancement of system performance over time. However, the application of data flywheels in the education sector remains limited, and instances of its utilization to improve multi-agent collaboration structures are particularly rare.

[0006] As prior art, there is Korean Published Patent No. 10-2025-0122980 (Method and System for Task Processing through Collaboration between Agents), but it only provides a method for processing tasks that cannot be processed by a single agent alone through collaboration between different agents, and a system to which such method is applied. The problem to be solved

[0008] To address the aforementioned problems, the objective of the present invention is to provide a method and system for improving multi-agent collaboration using data flywheel-based reinforcement learning, which overcomes the limitations of failing to resolve quality degradation and hallucination issues associated with the generation of educational content based on a single artificial intelligence model; automatically generates and manages hierarchical curriculum structures and evaluation elements based on basic content information provided by educators; improves generated results through quality evaluation and iterative feedback; and continuously enhances collaboration efficiency among multiple agents by utilizing learner interaction data to calculate quality indicators and reflecting them in data flywheel-based reinforcement learning. means of solving the problem

[0010] In the method for improving multi-agent collaboration using data flywheel-based reinforcement learning according to the present invention, the method comprises: a step in which a service provider server receives basic content information from an educator terminal; a step in which the service provider server analyzes natural language input based on the basic content information; a step in which the service provider server generates at least one hierarchical curriculum structure among a course, a chapter, and a lesson based on the result of the natural language input analysis and the basic content information; a step in which the service provider server evaluates the quality of the generated data to determine whether a standard is met; a step in which, if it is determined that the standard is met, the service provider server generates specific content for each lesson based on the hierarchical curriculum structure; a step in which the service provider server evaluates the quality of the generated data to determine whether the standard is met; a step in which, if it is determined that the standard is met, the service provider server generates evaluation items in conjunction with the specific content of each lesson generated; a step in which the service provider server evaluates the quality of the generated data to determine whether the standard is met; and a step in which, if it is determined that the standard is met, the service provider server integrates the generated hierarchical curriculum structure, the specific content of each lesson, and the evaluation items to store them as a single educational content and transmits it to the educator terminal. The method includes the steps of: an educator terminal requesting a service provider server to generate code corresponding to the educational content; receiving the generated code from the service provider server and distributing the code; a student terminal accessing the educational content based on the code and starting learning; a step in which, upon completion of learning, the student terminal transmits learning data to the service provider server; and a step in which the service provider server analyzes the received learning data to improve the educational content.

[0011] In a multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning according to the present invention, the system comprises: an educator terminal that transmits basic content information to a service provider server, receives educational content from the service provider server, requests the service provider server to generate code corresponding to the educational content, receives the generated code, and distributes the code; a service provider server that analyzes natural language input based on the basic content information, generates at least one hierarchical curriculum structure among a course, a chapter, and a lesson based on the natural language input analysis result and the basic content information, evaluates the quality of the generated data to determine whether criteria are met, and if the criteria are determined to be met, generates specific content for each lesson based on the hierarchical curriculum structure, and if the criteria are determined to be met, generates evaluation items linked to the specific content of each lesson generated, and if the criteria are determined to be met, integrates the generated hierarchical curriculum structure, the specific content of each lesson, and evaluation items to store it as a single educational content, transmits it to the educator terminal, and analyzes the received learning data to improve the educational content; and a system that accesses the educational content based on the code to start learning, and when learning is completed, transmits the learning data to the service provider server. Includes student terminals. Effects of the invention

[0013] According to one embodiment of the present invention, the efficiency of educational course design can be increased by automatically generating a hierarchical curriculum structure and evaluation elements from basic content information.

[0014] In addition, hierarchical structures can be utilized as technical guardrails to reduce hallucinations and ensure contextual consistency.

[0015] In addition, content quality can be continuously managed through iterative verification and correction requests via the quality evaluation agent.

[0016] In addition, the efficiency of multi-agent collaboration can be improved by calculating quality indicators based on training data and applying reinforcement learning.

[0017] In addition, it can provide educational content that is self-evolved and optimized over time. Brief explanation of the drawing

[0019] FIG. 1 is a flowchart illustrating a method for improving multi-agent collaboration using data flywheel-based reinforcement learning according to an embodiment of the present invention. FIG. 2 is a flowchart illustrating a method for improving educational content by analyzing learning data according to an embodiment of the present invention. Figure 3 is a configuration diagram of a multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning according to an embodiment of the present invention. FIGS. 4 to 7 are usage diagrams of a multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning according to an embodiment of the present invention. Specific details for implementing the invention

[0020] Specific structural or functional descriptions of embodiments according to the concept of the present invention disclosed herein are merely illustrative for the purpose of explaining embodiments according to the concept of the present invention, and embodiments according to the concept of the present invention may be implemented in various forms and are not limited to the embodiments described herein.

[0021] Since embodiments according to the concept of the present invention may be subject to various modifications and may take various forms, embodiments are illustrated in the drawings and described in detail in this specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes all modifications, equivalents, or substitutions that fall within the spirit and scope of the present invention.

[0022] The terms used herein are used merely to describe specific embodiments and are not intended to limit the invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as “comprising” or “having” are intended to indicate the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described herein, and should not be understood as precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0023] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings attached to this specification.

[0025] FIG. 1 is a flowchart illustrating a method for improving multi-agent collaboration using data flywheel-based reinforcement learning according to an embodiment of the present invention.

[0026] Referring to FIG. 1, a service providing server (200) receives basic content information from an educator terminal (100) (S101). The basic content information may be data including at least one of a course idea, educational purpose, target learner, and a brief description entered by the educator, and may be written in a free text format or in accordance with predefined input items such as a course name, learner level, and duration. The service providing server (200) analyzes natural language input based on the basic content information (S103). The service providing server (200) can identify not only the meaning at the word level but also contextual relationships by applying natural language processing techniques such as morphological analysis, syntactic analysis, and semantic role analysis to the sentences included in the basic content information. Even if the basic content information is a simply descriptive sentence, the service providing server (200) can extract at least one key element necessary for curriculum design from the sentence, such as the educational purpose, target learner, expected learning period, and difficulty level.

[0027] The service providing server (200) generates at least one hierarchical curriculum structure among a course, a chapter, and a lesson based on the natural language input analysis results and basic content information (S105). The service providing server (200) can automatically design a multi-stage structure leading from a higher-level course to lower-level units such as chapters and lessons by synthesizing at least one of the extracted learning objectives, the level of the learner, topic keywords, and the expected learning period. First, the service providing server (200) defines the overall learning objectives at the course level and can divide the subject areas necessary to achieve them into chapter units. Subsequently, each chapter can be subdivided into lesson units where specific learning activities are placed. The service providing server (200) evaluates the quality of the generated data to determine whether the criteria are met (S107). The service providing server (200) can verify at least one of whether the learning objectives and content are logically linked for each stage of the course, chapter, and lesson, whether the learning difficulty is suitable for the level of the target learner, and whether the learning flow is balanced. The service providing server (200) can comprehensively check the generated curriculum structure by referring to predefined quality evaluation criteria such as the specificity of the learning objectives, the diversity of learning activities, the gradual increase in difficulty, and whether evaluation elements are included.

[0028] The service providing server (200) generates specific content for each lesson based on the above-described hierarchical curriculum structure (S109). When generating specific content for each lesson, the service providing server (200) may generate it based on the detailed learning content generation request information received from the educator terminal (100) and the hierarchical curriculum structure. The service providing server (200) may automatically create at least one of the lesson unit learning materials, examples, and multimedia materials according to the detailed learning content generation request information. In one embodiment, if a specific lesson is on the subject of “drawing claims in a patent specification,” the service providing server (200) may generate learning materials that include at least one of a theoretical explanation, a representative case analysis, and step-by-step writing examples. The service providing server (200) evaluates the quality of the generated data to determine whether the criteria are met (S111). The service providing server (200) may receive at least one of the generated lesson unit learning materials, examples, practice assignments, and multimedia materials and evaluate its quality. The service providing server (200) verifies whether the generated content maintains logical consistency with the initial curriculum structure, whether the learning objectives and learning activities of each lesson are appropriately connected, and whether the level and difficulty of the learner match. The service providing server (200) can check the content by referring to predefined quality criteria such as factual accuracy, pedagogical validity, appropriateness of difficulty, clarity of expression, and whether learner participation elements are included.

[0029] The service providing server (200) generates evaluation items in conjunction with the specific content of each lesson generated above (S113). The service providing server (200) can design questions that directly correspond to the learning goals so as to verify whether the learning goals of each lesson have been achieved. The service providing server (200) can automatically generate questions in at least one format among multiple choice, short answer, descriptive, project assignment, and interactive quiz, and can generate questions to measure whether the learner has actually acquired the concepts and knowledge that need to be understood at the lesson level. Additionally, the evaluation item generation agent (234) can adjust the complexity of the questions by reflecting the learner's level and the set difficulty level. The service providing server (200) evaluates the quality of the generated data to determine whether the criteria are met (S115). The service providing server (200) can verify whether each evaluation item can appropriately measure the corresponding learning goal, whether the difficulty level of the questions matches the learner's level, and whether the expression of the questions is described clearly and consistently. The service providing server (200) can detect cases where the generated evaluation items are too simple or ambiguous and fail to reflect the learning objectives, and conversely, identify items with a difficulty level that is excessively high compared to the learner's level and classify them as items for modification. In addition, the service providing server (200) can ensure that the evaluation is not biased toward a specific type of problem by checking whether various types of items are evenly distributed for the same learning objective.

[0030] The service providing server (200) integrates the generated hierarchical curriculum structure, specific content of each lesson, and evaluation items to store them as a single educational content and transmits it to the educator terminal (100) (S117). When the evaluation items are finalized, the service providing server (200) can store the finalized evaluation items and the specific content of each lesson as detailed learning content. Additionally, the service providing server (200) can synthesize (integrate) the detailed learning content and the initial curriculum into a single educational content and store it.

[0031] An educator terminal (100) requests the service provider server (200) to generate a code corresponding to the educational content, receives the generated code from the service provider server (200), distributes the code, and a student terminal (300) accesses the educational content based on the code and begins learning (S119). The code can be used as an authentication means for a learner to access the educational content on the student terminal (300), and in one embodiment, it may be in the form of a random number-based unique identifier (UUID), a time / session-based token, or an alphanumeric string generated according to a predefined rule. The service provider server (200) performs a duplicate verification procedure to ensure that the generated code does not conflict with existing codes, and can automatically regenerate a new code if a conflict is detected. Additionally, the service provider server (200) can store mapping information between the generated code and the educational content in the form of metadata.

[0032] When learning is completed, the student terminal (300) transmits the learning data to the service provider server (200) (S121), and the service provider server (200) analyzes the received learning data to improve the educational content (S123). The student terminal (300) can be implemented in the form of at least one of various electronic devices such as a desktop computer, laptop, tablet PC, smartphone, or wearable device, and can communicate with the service provider server (200) in a wired or wireless network environment. The student terminal (300) can automatically access educational content corresponding to a code by entering a code through a connection interface provided by the service provider server (200) or by recognizing a visual medium such as a QR code. At this time, the student terminal (300) is allowed to access only permitted content after undergoing a code validity verification process, and if the code has expired or is not authenticated, it may receive an error message from the service provider server (200) and access may be restricted.

[0034] FIG. 2 is a flowchart illustrating a method for improving educational content by analyzing learning data according to an embodiment of the present invention.

[0035] Referring to FIG. 2, a service providing server (200) analyzes learning data and calculates a quality indicator including at least one of a correct answer rate, an incorrect answer pattern, a dwell time, and a level of participation (S201). The service providing server (200) can obtain a reliable dataset by normalizing the quality indicator and removing outliers. In one embodiment, records of repeated responses to the same problem within a short period of time may be invalidated, and weights may be applied differently according to the learner's level to enable a more precise evaluation.

[0036] The service provider server (200) generates a reward signal based on the quality indicators (S203). The reward signal can be derived in detail through weight calculations that consider the importance of each indicator, rather than simply classifying it as positive if performance is high and negative if performance is low. In one embodiment, the accuracy rate is given high weight as it represents basic learning achievement, the time spent is a factor evaluating immersion, and the level of participation is a factor representing the voluntariness of learning, each with a differentiated weight set accordingly.

[0037] The service provider server (200) uses the reward signal to update the operation parameters of at least one agent using a reinforcement learning technique to improve the collaboration efficiency between agents (S205). Reinforcement learning is a technique that finds an optimal policy by repeatedly learning the actions of agents and the rewards resulting from those actions. The service provider server (200) can apply this to a multi-agent environment to improve collaboration efficiency. In one embodiment, if the core concept is under-extracted from the input text, thereby lowering the consistency of the entire curriculum, the service provider server (200) can adjust the parameter by reflecting a negative reward to increase the extraction accuracy. Conversely, if the service provider server (200) produces material that increases the learner's participation, it can update the method of generation to reinforce it through a positive reward. The service provider server (200) improves educational content based on the agent to which the updated operation parameters have been applied (S207). The service provider server (200) can proceed by reconstructing lesson-unit materials and evaluation items with the new parameters reflected, and verifying them repeatedly if necessary. In one embodiment, a lesson with difficulty adjustments applied is converted into a form more suitable for the learner's level, and the generated questions can reduce the learner dropout rate by readjusting the balance of difficulty.

[0039] Figure 3 is a configuration diagram of a multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning according to an embodiment of the present invention.

[0040] Referring to FIG. 3, a multi-agent collaboration improvement system (10) utilizing data flywheel-based reinforcement learning consists of an educator terminal (100), a service provider server (200), and a student terminal (300).

[0041] The educator terminal (100) can transmit basic content information to the service provider server (200). The basic content information may be data including at least one of a course idea, educational purpose, target learner, and a brief description entered by the educator, and may be written in a free text format or in accordance with predefined input items such as a course name, learner level, and period. Additionally, the educator terminal (100) can receive additional content information from the service provider server (200), review the received additional content information, and modify it as necessary, and can request the creation of an initial curriculum by transmitting the finally confirmed additional content information to the service provider server (200) regardless of whether it is modified. Additionally, the educator terminal (100) may receive a request from the service provider server (200) for at least one of the following: Learning Objectives, Focus Areas, Content Types, Custom Instructions, and AI Tutor Personality information, and transmit the information to the service provider server (200). Additionally, the educator terminal (100) may receive an initial curriculum from the service provider server (200), review it, and request the service provider server (200) to create detailed learning content. The educator terminal (100) may receive educational content from the service provider server (200).

[0042] The educator terminal (100) may request the service provider server (200) to generate a code and receive the generated code from the service provider server (200) to distribute the code. The educator terminal (100) may distribute the code received from the service provider server (200) in various ways. In one embodiment, the educator terminal (100) may link with an external educational service environment, such as a Learning Management System (LMS), an online lecture platform, or a dedicated application, so that the corresponding code is automatically mapped to a learner account. At this time, the student terminal (300) can immediately access educational content through the distributed code without a separate registration or manual input procedure. Additionally, the educator terminal (100) may use various delivery means, such as email, messenger, public chat room, web bulletin board, or QR code generator, to distribute the code. In one embodiment, when an educator sends a code for a specific course via email, the student terminal (300) can immediately access the educational content by clicking the link included in the received email, or in an offline lecture environment, the code can be converted into a QR code and accessed by the learner scanning it with their own terminal. Additionally, the educator terminal (100) may impose constraints on the code, such as an expiration date or a limit on the number of accesses, to prevent unauthorized or duplicate use of the code. Furthermore, the educator terminal (100) may restrict the use of the code to a specific group of learners only, thereby blocking access by outsiders who are not participants in the class. In one embodiment, the code may be linked with a list of students registered in advance by the educator, allowing access to the content only after undergoing an identity verification process upon entering the code. Additionally, the educator terminal (100) may monitor the usage status of the distributed code in real time. In one embodiment, it may track how many students have entered the code to access the content, the time and frequency of access, and whether there are users who repeatedly enter a specific code on multiple terminals.Through this, educators can objectively verify learning participation and, if necessary, reissue new codes or reclaim existing ones.

[0043] The educator terminal (100) may generally be a terminal device carried by a user who performs the role of designing and delivering educational content, such as at least one of an educator, instructor, professor, or training manager. The educator terminal (100) may be implemented as various types of electronic devices, such as a desktop computer, laptop, tablet PC, or smartphone, and may operate in a general-purpose environment without being limited to a specific operating system. In one embodiment, the creation of educational content may be requested via a laptop used in a classroom, or simple course ideas may be entered and results checked via a smartphone application while on the move. The educator terminal (100) may receive the generated educational content through communication with a service provider server (200), and this communication path may include at least one of an internet network, a mobile communication network, or a wireless network (Wi-Fi). The generated educational content may be delivered via a web-based user interface, a dedicated application screen, or a transmission means such as email. In one embodiment, when an educator accesses a service provider server (200) on a web browser and requests a curriculum, the server may display automatically generated content on a web interface or provide it in a downloadable file format (PDF, PPT, DOCX, etc.). Additionally, the educator terminal (100) may not merely receive content but may include a function to check the content delivered from the service provider server (200) in real time, modify parts of it, or reflect opinions. Through this function, the educator can check the quality of the generated content and provide immediate feedback if necessary.

[0044] The service provision server (200) is composed of a basic content information preprocessing unit (210), an environment setting unit (220), an agent unit (230), a data flywheel learning unit (240), a storage unit (250), a communication unit (260), and a control unit (270).

[0045] The basic content information preprocessing unit (210) can receive basic content information from the educator terminal (100). The basic content information may be data including at least one of a course idea, educational purpose, target learner, and a brief description entered by the educator, and may be written in a free text format or in accordance with predefined input items such as a course name, learner level, and period. The basic content information preprocessing unit (210) can perform preprocessing of the basic content information. The basic content information preprocessing unit (210) can ensure the accuracy of the data to be analyzed by removing or correcting unnecessary special characters, spaces, duplicate words, and incomplete expressions within the basic content information. In addition, errors that may occur during the automatic processing based on learning data can be reduced in advance through spell checking and typographical correction. Furthermore, the basic content information preprocessing unit (210) can perform various preprocessing functions to simplify the sentence structure of the basic content information and ensure data consistency. In one embodiment, various expressions or abbreviations having the same meaning can be converted into a standardized form so that they can be recognized without confusion in subsequent stages. Additionally, the basic content information preprocessing unit (210) can remove elements that are irrelevant to curriculum design or unnecessary in advance by comparing and contrasting key words and phrases included in the basic content information with a predefined set of education-related terms. In one embodiment, if the content simply includes the educator's personal notes or symbols that have no meaning in context, it can be automatically filtered out so that only meaningful data remains. At this time, the basic content information preprocessing unit (210) can apply keyword extraction techniques or pattern recognition to ensure that at least one of the learning purpose, target learners, and period is emphasized. Additionally, the basic content information preprocessing unit (210) can perform a preliminary operation to convert each element included in the basic content information into a format suitable for a structured database.In one embodiment, the basic content information preprocessing unit (210) can organize string data into a fixed length and encoding method, and standardize numeric data through unit conversion or range verification.

[0046] Subsequently, the basic content information preprocessing unit (210) can automatically determine whether the received basic content information is in a predefined input item format or a free-form format. If it is determined to be in a predefined input item format, the basic content information preprocessing unit (210) can extract the items as they are and convert them into structured data that can be directly utilized in the initial curriculum creation process. If it is determined to be in a free-form format, the basic content information preprocessing unit (210) can analyze it using machine learning-based natural language processing techniques. The basic content information preprocessing unit (210) can extract key keywords from the basic content information written in a free-form format and analyze the context to automatically extract at least one additional content information among the educational purpose, expected learner level, expected learning period, course name, target learner, and difficulty level. The basic content information preprocessing unit (210) can interpret the meaning of an input sentence based on a pre-trained Large Language Model (LLM) and separate verb phrases (learn, understand, apply, etc.) indicating educational purposes and noun phrases (elementary student, graduate student, new engineer, etc.) referring to the learner. Additionally, the basic content information preprocessing unit (210) can recognize temporal expressions such as a 3-month course and twice a week, convert the learning period into numerical data, and map words indicating difficulty to a pre-defined grade system. The basic content information preprocessing unit (210) can identify the target represented by each keyword by utilizing context similarity analysis and a dependency phrase tree, and derive core information by considering semantic association rather than based on simple keyword frequency. In addition, the basic content information preprocessing unit (210) is connected to a domain-specific ontology and can convert items that refer to the same learning topic, such as 'machine learning', 'deep learning', and 'neural network', into standardized additional content information by integrating them.Additionally, the basic content information preprocessing unit (210) may suggest a correction value through inference based on similar literature when an incomplete or ambiguous expression is input, or select a value with high reliability among multiple candidates. The basic content information preprocessing unit (210) may convert the additional content information into a structured form and transmit it to the educator terminal (100). The basic content information preprocessing unit (210) may receive the additional content information, whether modified or not, from the educator terminal (100), reflect it in the basic content information, and finally store the confirmed additional content information as new basic content information. The basic content information preprocessing unit (210) may transmit the basic content information to the agent unit (230).

[0047] The environment setting unit (220) can support the educator in directly setting various environmental conditions of the educational content through the educator terminal (100). The environment setting unit (220) can ensure that the generated educational content aligns with the purpose and situation by integrally managing at least one element among the language, cultural context, learning objectives, learning areas, tutor personality, and content type of the educational content. The environment setting unit (220) can request information on Learning Objectives and Focus Areas from the educator terminal (100). The above Learning Objectives are items that define the competencies that a learner must achieve after completing the educational course, and the above Focus Areas may be major topics to be covered during the learning process, but are not necessarily limited to these. The environment setting unit (220) can receive information on Learning Objectives and Focus Areas from the educator terminal (100) and store them by matching them with the above Basic Content Information.

[0048] Additionally, the configuration unit (220) may request information on Content Types and Custom Instructions. The above Content Type refers to a detailed form of learning content including at least one of a practice assignment or a quiz project, and various options may be delivered to the educator terminal (100) so that the educator can select or combine various types as needed. The above Custom Instructions may be items that allow the educator to directly input specific instructions to control content creation more precisely, but are not necessarily limited to this. The configuration unit (220) may receive the Content Type and Custom Instruction information from the educator terminal (100) and store them by matching them with the above Basic Content Information.

[0049] Additionally, the environment setting unit (220) may request information on the personality of the AI ​​tutor. The personality of the AI ​​tutor may be the tone, tone, and intimacy to be used when the AI ​​tutor and the learner interact, but is not necessarily limited thereto. In one embodiment, if the personality of the AI ​​tutor is 'Formal & Academic', it may provide academic and formal explanations, and if it is 'Casual & Friendly', it may provide conversational and friendly feedback. The environment setting unit (220) may receive the personality information of the AI ​​tutor from the educator terminal (100) and store it by matching it with the basic content information.

[0050] The agent unit (230) consists of an analysis agent (231), a structural design agent (232), a content creation agent (233), an evaluation item creation agent (234), and a quality evaluation agent (235).

[0051] The interpretation agent (231) can analyze natural language input based on the basic content information. The interpretation agent (231) can identify not only the meaning at the word level but also contextual relationships by applying natural language processing techniques, such as morphological analysis, syntactic analysis, and semantic role analysis, to sentences included in the basic content information. Even if the basic content information is simply a descriptive sentence, the interpretation agent (231) can extract at least one key element necessary for curriculum design from the sentence, such as educational purpose, target learner, expected learning period, and difficulty level. Additionally, the interpretation agent (231) can perform the function of classifying which category a keyword included in the basic content information belongs to. In one embodiment, if the input "Python programming basics" is included, it can identify that "Python" corresponds to the learning topic, "basics" to the difficulty level, and "programming" to the technical field, and convert this into data that can be utilized by the subsequent structure design agent (232). Additionally, the interpretation agent (231) can correct unclear or ambiguous expressions through mapping between basic content information and external knowledge bases, such as educational ontologies and learning classification systems. In one embodiment, if the terms “deep learning” and “neural network” are used interchangeably, they can be integrated into the same topic and processed consistently in subsequent steps. Additionally, the interpretation agent (231) can perform the role of verifying the semantic consistency of each item when the basic content information is provided in a predefined input item format. In one embodiment, if the difficulty level is set to “advanced” but the target learner is entered as “middle school student,” this can be detected as a potential discrepancy and a warning or a correction suggestion can be generated.

[0052] The above-described structure design agent (232) can generate at least one hierarchical curriculum structure among courses, chapters, and lessons based on the results of the natural language input analysis and basic content information. The structure design agent (232) can automatically design a multi-stage structure leading from a higher-level course to lower-level units such as chapters and lessons by synthesizing at least one of the learning objectives, learner level, subject keywords, and expected learning period extracted by the interpretation agent (231). First, the structure design agent (232) can define the overall learning objectives at the course level and divide the subject areas necessary to achieve them into chapter units. Subsequently, each chapter can be subdivided into lesson units where specific learning activities are placed. The structure design agent (232) can reasonably distribute the volume and order of each chapter and lesson by considering at least one of the educator's input values ​​among the total learning period, weekly learning time, and learning difficulty. Additionally, the structural design agent (232) can ensure that the generated curriculum structure maintains educational consistency by referring to predefined educational design guidelines or domain-specific standard curriculum templates. In one embodiment, the structural design agent (232) can design the structure to automatically reflect the sequence of basic concepts, practical examples, and project execution if it is an “introduction to programming course.” Furthermore, the structural design agent (232) can support customized curriculum design for different learner types. Even for the same topic, the curriculum can be structured to focus on theoretical explanations and basic practice for “beginners,” and on application projects or problem-solving tasks for “advanced learners.” The structural design agent (232) can adjust the difficulty and volume of each layer element to reflect the learner's level and goals.In addition, the structure design agent (232) may not merely define a hierarchical curriculum structure, but may also set constraints that the content creation agent (233) must adhere to when creating specific content for each lesson in a subsequent stage. These constraints may include, but are not limited to, at least one of the following: rules ensuring that each lesson does not deviate from the learning objectives and scope defined in the upper course and chapter, the use of terminology and expression methods suitable for the difficulty level and learner level, and the adjustment of volume based on the estimated learning time. The structure design agent (232) may transmit these constraints to the content creation agent (233). By constraining the output of the content creation agent (233) to always maintain consistency with the upper hierarchical structure, the structure design agent (232) can ensure the logical consistency and educational reliability of the entire educational content. Furthermore, the structure design agent (232) may not only set these constraints statically but may also dynamically update them as needed based on the evaluation results of the quality evaluation agent (235). In one embodiment, if the initially set constraints are judged to be too difficult or ambiguous for the learner, the constraints may be relaxed or supplemented by reflecting the feedback of the quality evaluation agent (235). The structure design agent (232) can function as a key control module that goes beyond simply defining the framework of the curriculum and provides strong guidelines and constraints throughout the lesson content creation process. Once the creation of the hierarchical curriculum structure is complete, the structure design agent (232) can transmit the generated hierarchical curriculum structure data to the quality evaluation agent (235).

[0053] If the evaluation results received from the quality evaluation agent (235) meet the criteria, the structural design agent (232) can finalize the corresponding curriculum structure data as an initial curriculum and transmit it to the educator terminal (100). At this time, the finalized initial curriculum includes a hierarchical structure in the units of Course, Chapter, and Lesson, and can be used as basic data for creating detailed learning content in subsequent stages. If the evaluation results received from the quality evaluation agent (235) do not meet the criteria, the structural design agent (232) can redesign the curriculum structure by reflecting specific modification recommendations provided by the quality evaluation agent (235). The redesigned curriculum structure is transmitted back to the quality evaluation agent (235) for re-verification, and through this iterative feedback process, an initial curriculum that finally meets the criteria can be finalized.

[0054] The content creation agent (233) can create specific content for each lesson based on the hierarchical curriculum structure, which is the confirmed initial curriculum. At this time, when creating specific content for each lesson, the content creation agent (233) can create it based on the detailed learning content creation request information received from the educator terminal (100) and the hierarchical curriculum structure. The content creation agent (233) can automatically create at least one of learning materials, examples, and multimedia materials per lesson unit according to the detailed learning content creation request information. In one embodiment, if a specific lesson is on the topic of “drawing claims in a patent specification,” the content creation agent (233) can create learning materials that include at least one of theoretical explanations, representative case analyses, and step-by-step writing examples. Additionally, the content creation agent (233) can create educational materials in various formats according to the content type requested by the educator, such as text-based materials, images or diagrams, and video scripts. The content creation agent (233) can appropriately adjust the depth of the content, difficulty level, and method of expression by reflecting predefined pedagogical design principles and the learner level. Additionally, the content creation agent (233) can automatically map learning activities corresponding to the learning objectives of each lesson so that the created content is consistently connected to the initial curriculum structure. In particular, the content creation agent (233) can create content by applying the constraints defined by the structure design agent (232). The said constraints may include at least one of rules to ensure that the content does not deviate from the learning objectives and scope set in the higher-level course and chapter, vocabulary and expression methods according to the difficulty level, and volume limits according to the estimated learning time. Through this, the content creation agent (233) can ensure that the specific content of each lesson maintains contextual consistency with the hierarchical curriculum structure.In one embodiment, if a specific chapter is set to an 'introductory level,' the corresponding lesson content may be generated by limiting it to materials suitable for the basic level without including advanced concepts. The content generation agent (233) may generate content and evaluation elements in a structure that are organically combined so that learners can achieve learning goals for each lesson unit. Additionally, the content generation agent (233) may incorporate the latest information and cases by linking with external knowledge bases or specialized materials. In one embodiment, in rapidly changing fields such as law, science, and engineering, the timeliness and reliability of the content may be enhanced by supplementing educational materials by referring to the latest papers, precedents, and technical standards. Once the specific content generation for each lesson is completed, the content generation agent (233) may transmit the specific content data for each lesson to the quality evaluation agent (235).

[0055] If the evaluation result received from the quality evaluation agent (235) meets the criteria, the content creation agent (233) can save the specific content of each lesson. If the evaluation result received from the quality evaluation agent (235) does not meet the criteria, the content creation agent (233) can regenerate the specific content of each lesson by reflecting the specific correction recommendations provided by the quality evaluation agent (235). The regenerated specific content of each lesson is sent back to the quality evaluation agent (235) for re-verification, and through this iterative feedback process, the specific content of each lesson that finally meets the criteria can be determined.

[0056] The evaluation item generation agent (234) can generate evaluation items based on the specific content of each confirmed lesson. The evaluation item generation agent (234) can design questions that directly correspond to the learning objectives of each lesson to verify whether the learning objectives of each lesson have been achieved. The evaluation item generation agent (234) can automatically generate questions in at least one format among multiple-choice, short-answer, descriptive, project assignment, and interactive quiz, and can generate questions to measure whether the learner has actually acquired the concepts and knowledge that must be understood at the lesson level. In addition, the evaluation item generation agent (234) can adjust the complexity of the questions by reflecting the learner's level and the set difficulty level. In one embodiment, multiple-choice or short-answer questions to verify basic concepts are given to beginner learners, while actual case analysis or application problem solving may be included for advanced learners. To this end, the evaluation item generation agent (234) can automatically recognize key concepts and cases within the content and map questions according to difficulty levels to enable evaluation at various levels. In addition, the evaluation item generation agent (234) can combine various types of questions that can evaluate concept understanding, problem-solving ability, application ability, and critical thinking in order to measure the learner's understanding of the same lesson from multiple perspectives. In one embodiment, in the “Writing Patent Specification Claims” lesson, multiple-choice questions asking for definitions of key terms, descriptive tasks that require writing actual claims, and analytical questions that present incorrect claims and require finding errors can be generated together. In addition, the evaluation item generation agent (234) can verify whether the generated questions correspond exactly to the lesson learning objectives and automatically adjust questions that are unnecessarily too difficult or too difficult in order to ensure the quality of the questions.The evaluation item generation agent (234) can distribute items in a balanced manner according to the stages of knowledge, comprehension, application, analysis, synthesis, and evaluation by referring to predefined pedagogical evaluation guidelines or a taxonomy of educational objectives. Once the evaluation item generation agent (234) completes the generation of evaluation items, it can transmit the generated evaluation item data to the quality evaluation agent (235).

[0057] The evaluation item generation agent (234) can save the evaluation item if the evaluation result received from the quality evaluation agent (235) meets the criteria. If the evaluation result received from the quality evaluation agent (235) does not meet the criteria, the evaluation item generation agent (234) can regenerate the evaluation item by reflecting specific correction recommendations provided by the quality evaluation agent (235). The regenerated evaluation item is transmitted back to the quality evaluation agent (235) for re-verification, and through this iterative feedback process, the evaluation item that finally meets the criteria can be determined.

[0058] When the evaluation item generation agent (234) determines the evaluation item, it can save the determined evaluation item and the specific content of each lesson as detailed learning content. Additionally, the evaluation item generation agent (234) can synthesize (integrate) the detailed learning content and the initial curriculum into a single educational content and save it. The evaluation item generation agent (234) can deliver the generated educational content to the educator terminal (100).

[0059] The above quality evaluation agent (235) performs the role of verifying the quality of the data and evaluation elements generated by each agent (231, 232, 233, 234) of the agent unit (230). If the generated result falls short of the standard, it does not simply return whether it is unsatisfactory, but rather generates feedback data including the item of discrepancy and the direction of correction, and transmits it to the relevant agent. Subsequently, by repeatedly performing re-evaluation on the corrected data, high-quality educational content that finally meets the standard is produced.

[0060] The quality evaluation agent (235) can evaluate the quality of the hierarchical curriculum structure data generated by the structure design agent (232) to determine whether the criteria are met. The quality evaluation agent (235) can verify at least one of the following for each stage of Course, Chapter, and Lesson: whether the learning objectives and content are logically linked, whether the learning difficulty is appropriate for the level of the target learner, and whether the learning flow is balanced. The quality evaluation agent (235) can comprehensively inspect the generated curriculum structure by referring to predefined quality evaluation criteria, such as the specificity of learning objectives, the diversity of learning activities, the gradual increase in difficulty, and whether evaluation elements are included. In one embodiment, by checking whether learning objectives are clearly set for every chapter and whether practice or evaluation elements are included in the lesson unit, if items that do not meet the criteria are found, it can automatically generate a warning or provide a recommendation for correction. Additionally, the quality evaluation agent (235) can evaluate whether the designed structure maintains consistency with generally accepted educational models by comparing and contrasting the generated structure data with existing high-quality curriculum examples in the database. In this process, the quality evaluation agent (235) can detect cases where the curriculum is excessively biased or excessively focused on specific topics by utilizing semantic similarity analysis, statistical distribution verification, etc. Additionally, the quality evaluation agent (235) can verify the consistency between the educator's input values, such as the total learning period, weekly learning hours, and learner level reflected by the structural design agent (232), and the generated curriculum results. In one embodiment, if more than 10 chapters are excessively placed in the curriculum despite the total learning period being set to 4 weeks, this can be determined as a discrepancy and the need for correction can be suggested.If the evaluation results meet predefined quality standards, the quality evaluation agent (235) may determine that the curriculum structure is suitable and transmit this to the structure design agent (232). If the evaluation results do not meet the standards, the quality evaluation agent (235) may notify the structure design agent (232) of the specific items that were not met and the reasons for such failure, thereby enabling corrections to be made. The quality evaluation agent (235) may not stop at merely notifying of "non-compliance" but may also provide specific recommendations for correction. In one embodiment, if a learning objective is missing in a specific chapter, a suggestion to "add a learning objective to that chapter" may be provided, and if the difficulty level is excessively high compared to the learner's level, a recommendation for correction to "reduce the amount of lessons or adjust the difficulty level" may be provided.

[0061] Additionally, the quality evaluation agent (235) can evaluate the quality of the detailed learning content data generated by the content creation agent (233) to determine whether the criteria are met. The quality evaluation agent (235) can receive at least one of the lesson unit learning materials, examples, practice tasks, and multimedia materials generated by the content creation agent (233) and evaluate their quality. The quality evaluation agent (235) verifies whether the generated content maintains logical consistency with the initial curriculum structure, whether the learning objectives and learning activities of each lesson are appropriately connected, and whether the difficulty level matches the learner's level. The quality evaluation agent (235) can inspect the content by referring to predefined quality criteria such as factual accuracy, pedagogical validity, appropriateness of difficulty, clarity of expression, and whether learner participation elements are included. In one embodiment, for a lesson titled “Drafting Claims in a Patent Specification,” the quality evaluation agent can evaluate whether the generated theoretical explanation aligns with legal definitions, whether the case analysis is based on actual patent literature, and whether the practice problems are written at a level capable of measuring the learning objectives. Additionally, the quality evaluation agent (235) can verify whether the generated content, such as text, images, and video scripts, is written in multiple formats and whether each format plays a complementary role while aiming for the same learning goal. This prevents the learner from failing to gain sufficient understanding due to an overemphasis on a specific format. If the quality evaluation agent (235) determines that the evaluation result meets the criteria, it can determine that the detailed learning content is suitable and transmit this to the content creation agent (233). If the evaluation result does not meet the criteria, the quality evaluation agent (235) can transmit the detailed items that were not met and the reasons therefor to the content creation agent (233) so that corrections can be made.The quality evaluation agent (235) may not merely notify of “non-compliance” but may also provide specific recommendations for correction. In one embodiment, if the learning objective of a specific lesson is defined as “understanding the drafting of patent specification claims” but the generated content does not include practice examples, the quality evaluation agent (235) may provide a recommendation for correction to “add practice examples to the lesson so that the learner can directly experience the process of drafting claims.”

[0062] Additionally, the quality evaluation agent (235) can evaluate the quality of the evaluation item data generated by the evaluation item generation agent (234) to determine whether the criteria are met. The quality evaluation agent (235) can verify whether each evaluation item can appropriately measure the corresponding learning objective, whether the difficulty level of the item matches the learner's level, and whether the expression of the item is described clearly and consistently. The quality evaluation agent (235) can detect cases where the generated evaluation item is too simple or ambiguous and fails to reflect the learning objective, and conversely, identify items with an excessively high difficulty level compared to the learner's level and classify them as items for modification. Additionally, the quality evaluation agent (235) can ensure that the evaluation is not biased toward a specific type of problem by verifying whether various types of items are evenly distributed for the same learning objective. Furthermore, the quality evaluation agent (235) can check from various angles whether the generated items are pedagogically valid and fair by referring to predefined evaluation criteria such as validity, reliability, difficulty distribution, and guidelines for eliminating bias. In one embodiment, if the learning objective of a specific lesson is “understanding the principles of drafting claims,” it may be checked whether simple rote memorization questions are excessively included, and if necessary, recommendations may be made to reinforce them with analytical or application-type questions. The quality evaluation agent (235) structurally determines whether the criteria are met, and if the evaluation result is determined to meet the criteria, it may determine that the evaluation item is appropriate and transmit this to the evaluation item creation agent (234). If the evaluation result does not meet the criteria, the quality evaluation agent (235) may transmit the detailed items that were not met and the reasons therefor to the evaluation item creation agent (234) so ​​that corrections can be made. The quality evaluation agent (235) may not stop at merely notifying of “non-compliance” but may also provide specific recommendations for correction.In one embodiment, even though the learning objective of a specific lesson is defined as “understanding the principles of drafting patent specification claims,” if the generated evaluation items consist only of simple definition memorization type questions, the quality evaluation agent (235) may provide a modification recommendation to the evaluation item generation agent (234) to “add analytical type questions that apply the structure of claims to actual cases or identify errors.”

[0063] The data flywheel learning unit (240) can receive learning data from the student terminal (300). The data flywheel learning unit (240) can analyze the learning data to calculate a quality indicator including at least one of the correct answer rate, incorrect answer pattern, dwell time, and participation rate. The data flywheel learning unit (240) can obtain a reliable dataset by normalizing the quality indicator and removing outliers. In one embodiment, records of repeated responses to the same problem within a short period of time can be invalidated, and weights can be applied differently according to the learner's level to enable a more precise evaluation. Through this, the data flywheel learning unit (240) can calculate a quality indicator that reflects the entire learning process rather than a simple result value. The data flywheel learning unit (240) can generate a reward signal based on the calculated quality indicator. The reward signal can be derived in detail through weight calculation that considers the importance of each indicator, rather than simply classifying it as positive if performance is high and negative if performance is low. In one embodiment, the accuracy rate is given high weight as it represents basic learning achievement, the dwell time is a factor for evaluating immersion, and the participation level is a factor representing the voluntariness of learning, and differential weights can be set for each. The data flywheel learning unit (240) can distribute different types of signals to each agent by linking the reward signals generated through this multi-layered evaluation with the performance contribution of each agent.

[0064] The data flywheel learning unit (240) can update the operation parameters of at least one agent using a reinforcement learning technique by utilizing the generated reward signal. Reinforcement learning is a technique that finds an optimal policy by repeatedly learning the agent's behavior and the rewards resulting from it. The data flywheel learning unit (240) can apply this to a multi-agent environment to improve collaboration efficiency. In one embodiment, if the interpretation agent (231) of the agent unit (230) under-extracts key concepts from the input text, thereby lowering the consistency of the entire curriculum, the data flywheel learning unit (240) can adjust the parameters by reflecting a negative reward to increase extraction accuracy. Conversely, if the content creation agent (233) produces materials that increase the learner's participation, the data flywheel learning unit (240) can update the creation method to reinforce it through a positive reward. The data flywheel learning unit (240) can improve educational content based on the agent to which the updated operation parameters have been applied. The data flywheel learning unit (240) can proceed by reconstructing lesson-unit materials and evaluation items with new parameters reflected, and, if necessary, repeatedly verifying them through the quality evaluation agent (235) of the agent unit (230). In one embodiment, lessons with difficulty adjustments applied are converted into a form more suitable for the learner's level, and the questions generated by the evaluation item generation agent (234) can readjust the balance of difficulty to reduce the learner dropout rate. In this way, the data flywheel learning unit (240) can gradually improve the overall learning performance of the system by reflecting the results of parameter updates in actual content improvement. By repeatedly performing a series of processes such as learning data analysis, quality indicator calculation, reward signal generation, agent parameter update, and educational content improvement, the data flywheel learning unit (240) can form a cyclic structure that self-evolves the service quality over time.The above-mentioned cyclic structure enables educational content to continuously evolve through interaction with learners rather than remaining in a static state, and can provide the effect of optimizing the efficiency of multi-agent collaboration in the long term.

[0065] The storage unit (250) can safely store educational content and each content of the educational content generated in each configuration of the service provider server (200), and can generate and manage a unique code that matches each content. When the educational content is finalized, the storage unit (250) can generate a unique code corresponding to the content. The code can be used as an authentication means for a learner to access the educational content on a student terminal (300), and in one embodiment, it may be in the form of a random number-based unique identifier (UUID), a time / session-based token, or an alphanumeric string generated according to a predefined rule. The storage unit (250) can perform a duplicate verification procedure to ensure that the generated code does not conflict with existing codes, and can automatically regenerate a new code if a conflict is detected. In addition, the storage unit (250) can store mapping information between the generated code and the educational content in the form of metadata. In one embodiment, at least one of the content title, course / chapter / lesson hierarchy information, creation date and time, validity period, and access permission conditions can be recorded by mapping it together with the code. Through this, the service provider server (200) can quickly determine which educational content to connect to when the student terminal (300) enters a specific code. Additionally, the storage unit (250) can manage the validity and usage conditions of the code. In one embodiment, the code may be set to have an expiration period so that it automatically expires after a certain period of time, or the number of possible connections or the number of simultaneous connections may be limited to prevent unauthorized use. Additionally, the storage unit (250) can link the access rights of the code to a specific user group, thereby controlling the code so that only registered students can use it. Additionally, the storage unit (250) can record and analyze the usage logs of the code. Through this, the educator can check which code was used, when, and by whom, and can detect at least one of the connection frequency, whether duplicate use occurred, or abnormal connection attempts.In one embodiment, if the same code is used simultaneously in multiple regions, the storage unit (250) may detect this as an abnormal sign and immediately discard the code or require an administrator approval procedure.

[0066] The communication unit (260) enables all components of the educator terminal (100), student terminal (300), and service provider server (200) to communicate through the network. The network means may include at least one CDMA-based (or HSDPA-based) mobile communication network, and / or an IEEE 802.16x-based high-speed wireless internet, and / or an IEEE 802.11x-based wireless LAN communication network, but is not necessarily limited thereto. "Network" refers to a connection structure capable of exchanging information between each node, such as terminals and servers, and includes a Local Area Network (LAN), a Wide Area Network (WAN), the World Wide Web (WWW), a wired and wireless data communication network, a telephone network, a wired and wireless television communication network, etc. Examples of wireless data communication networks include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Bluetooth communication, infrared communication, ultrasonic communication, Visible Light Communication (VLC), and LiFi.

[0067] The control unit (270) can control all configurations of the service provider server (200).

[0068] The student terminal (300) may refer to a terminal device that receives codes distributed from the educator terminal (100) and code information generated by the service provider server (200), and performs the role of accessing educational content and starting learning based on the corresponding codes. The student terminal (300) may be implemented in the form of at least one of various electronic devices such as a desktop computer, laptop, tablet PC, smartphone, or wearable device, and may communicate with the service provider server (200) in a wired or wireless network environment. The student terminal (300) may automatically access educational content corresponding to the code by inputting the code through a connection interface provided by the service provider server (200) or by recognizing a visual medium such as a QR code. At this time, the student terminal (300) may access only allowed content after undergoing a code validity verification process, and if the code has expired or is not authenticated, it may receive an error message from the service provider server (200) and restrict access. In one embodiment, the student terminal (300) registers a code through a code input window of a dedicated application or an authentication screen of a web browser, and upon successful authentication, the corresponding educational content can be loaded immediately onto the learning screen. The student terminal (300) proceeds with learning based on the connected educational content. The learning process can be provided in various content formats, such as text, images, multimedia materials, and interactive quizzes per lesson unit, and the student terminal (300) performs the function of displaying this and collecting user input. In one embodiment, the student terminal (300) supports interactive functions such as playing video lectures, solving problems, submitting practice assignments, and participating in discussions, and the learner can actively utilize the educational content through this. Additionally, the student terminal (300) presents evaluation questions linked to learning objectives within the content, temporarily stores the results of the learner's responses on the local terminal, and then transmits them to the service provider server (200). When learning is completed, the student terminal (300) transmits the learning data to the service provider server (200).The learning data may include not only basic performance indicators such as the correct answer rate, incorrect answer patterns, dwell time, and participation, but also at least one quantitative or qualitative data such as the learner's input records, response time per question, whether a hint was requested, and the status of assignment submission. The student terminal (300) can upload the learning data in real-time or in batch form via a network, and can ensure the integrity and security of the data by applying encrypted communication (e.g., TLS / SSL) during the transmission process. Additionally, the student terminal (300) can display the transmission status of the data provided by the learner and provide a feedback screen that allows verification of whether the learning has been successfully completed. In one embodiment, the student terminal (300) can output a message stating, “Learning completion and data submission have been successfully processed,” or provide the learner with visualized learning progress rates, score summaries, and recommended learning paths. Through this, the learner can immediately check their learning performance, and the service provider server (200) can activate a data flywheel structure that improves educational content based on the transmitted data.

[0070] FIGS. 4 to 7 are usage diagrams of a multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning according to an embodiment of the present invention.

[0071] Referring to FIG. 4, a screen is shown in which a student checks their learning status through the learning dashboard of the student terminal (300), and summary information such as active courses, completed courses, and average progress rate is provided. It shows that the learner can enter the step of entering a code by pressing the "Join a Course" (401) button to participate in a new course.

[0072] Referring to FIG. 5, a screen is illustrated in which a student enters a code to participate in a private course. When the provided code (501) is entered, the student is registered in the course. The screen displays the course registration procedure, security and privacy guidelines, and benefit guidelines together, allowing the learner to participate safely.

[0073] Referring to FIG. 6, the screen for starting learning in the course registered by the student displays the course structure and a course start button (601) on the left, and provides a list of detailed learning for the selected part (603) in the center. The learner can start learning by pressing the “Start the course” button and sequentially check and perform learning materials and evaluation questions for each lesson.

[0074] Referring to FIG. 7, evaluation questions are provided to allow learners to check their understanding after completing educational content, and multiple-choice questions are presented in the quiz area (701). Learners can select and submit the correct answer from the presented options, thereby allowing them to self-check their understanding of the core concepts learned in the lessons of the educational content. This configuration can contribute to increasing learner participation and enhancing learning effectiveness, going beyond the simple provision of learning materials.

[0076] The invention has been described with reference to embodiments illustrated in the drawings, but this is merely illustrative, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible therefrom. Accordingly, the true technical scope of protection of the invention should be determined by the technical spirit of the appended claims. Explanation of the symbols

[0078] 10; Multi-Agent Collaboration Improvement System Using Data Flywheel-Based Reinforcement Learning 100; Educator terminal 200; Service provider server 210; Basic Content Information Preprocessing Unit 220; Environment Configuration Unit 230; Agent Department 231; Interpretation Agent 232; Structural Design Agent 233; Content Creation Agent 234; Evaluation Item Generation Agent 235; Quality Evaluation Agent 240; Data flywheel learning unit 250; Storage unit 260; Communication unit 270; Control unit

Claims

Claim 1 A method for improving multi-agent collaboration using data flywheel-based reinforcement learning comprises: a step in which a service provider server receives basic content information from an educator terminal; a step in which the service provider server analyzes natural language input based on the basic content information; a step in which the service provider server generates at least one hierarchical curriculum structure among a course, a chapter, and a lesson based on the result of the natural language input analysis and the basic content information; a step in which the service provider server evaluates the quality of the generated data and determines whether a standard is met; a step in which, if it is determined that the standard is met, the service provider server generates specific content for each lesson based on the hierarchical curriculum structure; a step in which the service provider server evaluates the quality of the generated data and determines whether the standard is met; a step in which, if it is determined that the standard is met, the service provider server generates evaluation items linked to the specific content of each lesson generated; a step in which the service provider server evaluates the quality of the generated data and determines whether the standard is met; a step in which, if it is determined that the standard is met, the service provider server integrates the generated hierarchical curriculum structure, the specific content of each lesson, and the evaluation items to store them as a single educational content and transmits it to the educator terminal; and the educator terminal The method comprises the steps of: requesting a service provider server to generate a code corresponding to the educational content, receiving the generated code from the service provider server and distributing the code, and having a student terminal access the educational content based on the code to begin learning; when learning is completed, having the student terminal transmit learning data to the service provider server; and the step of the service provider server analyzing the received learning data to improve the educational content, wherein the step of the service provider server analyzing the received learning data to improve the educational content includes: the service provider server analyzing the learning data to calculate a quality indicator including at least one of a correct answer rate, an incorrect answer pattern, dwell time, and participation; and the service provider server generating a reward signal based on the quality indicator.A method for improving multi-agent collaboration using data flywheel-based reinforcement learning, characterized by further including: a step in which a service providing server improves the collaboration efficiency between agents by updating the operation parameters of at least one agent using a reinforcement learning technique based on the reward signal; and a step in which the service providing server improves educational content based on an agent to which the updated operation parameters have been applied, wherein the service providing server repeatedly performs the processes of analyzing learning data, calculating quality indicators, generating a reward signal, updating agent parameters, and improving educational content to form a cyclic structure, wherein the service providing server includes an agent unit, wherein the agent unit is composed of an interpretation agent, a structural design agent, a content generation agent, an evaluation item generation agent, and a quality evaluation agent, wherein the quality evaluation agent verifies the quality of data and evaluation elements generated by each agent of the agent unit, and if the generation result falls below a standard, generates feedback data including the corresponding discrepancy item and correction direction and transmits it to the relevant agent, and repeatedly performs re-evaluation on the modified data. Claim 2 delete Claim 3 A method for improving multi-agent collaboration using data flywheel-based reinforcement learning, characterized in that, in the step of the service providing server evaluating the quality of generated data to determine whether a standard is met, if it is determined that the standard is not met, feedback data is generated to correct the inconsistent data and transmitted to a relevant agent, and a quality evaluation is performed again on the corrected data. Claim 4 An educator terminal that transmits basic content information to a service provider server, receives educational content from the service provider server, requests the service provider server to generate code corresponding to the educational content, receives the generated code, and distributes the code; a service provider server that analyzes natural language input based on the basic content information, generates at least one hierarchical curriculum structure among a course, chapter, and lesson based on the natural language input analysis result and the basic content information, evaluates the quality of the generated data to determine whether criteria are met, and if it is determined that criteria are met, generates specific content for each lesson based on the hierarchical curriculum structure, and if it is determined that criteria are met, generates evaluation items in conjunction with the specific content of each lesson generated, and if it is determined that criteria are met, integrates the generated hierarchical curriculum structure, the specific content of each lesson, and evaluation items to store it as a single educational content, transmits it to the educator terminal, and analyzes the received learning data to improve the educational content;The system includes a student terminal that accesses the educational content based on code to start learning and, upon completion of learning, transmits the learning data to a service provider server; the service provider server analyzes the learning data to calculate a quality indicator including at least one of a correct answer rate, an incorrect answer pattern, dwell time, and participation level; generates a reward signal based on the quality indicator; uses the reward signal to update the operation parameters of at least one agent using a reinforcement learning technique to improve collaboration efficiency between agents; and improves the educational content based on the agent with the updated operation parameters. The service provider server repeatedly performs the processes of analyzing learning data, calculating quality indicators, generating reward signals, updating agent parameters, and improving educational content to form a cyclic structure. The service provider server includes an agent unit, the agent unit is composed of an interpretation agent, a structural design agent, a content creation agent, an evaluation item creation agent, and a quality evaluation agent. The quality evaluation agent verifies the quality of the data and evaluation elements generated by each agent of the agent unit, and if the generated result falls below the standard, generates feedback data including the corresponding discrepancy item and correction direction and transmits it to the relevant agent, and repeatedly performs re-evaluation on the modified data. A multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning. Claim 5 delete Claim 6 A multi-agent collaboration improvement system utilizing data flywheel-based reinforcement learning, characterized in that, in the step of the service providing server evaluating the quality of generated data to determine whether the criteria are met, if it is determined that the criteria are not met, it generates feedback data to correct the inconsistent data and transmits it to a relevant agent, and performs a quality evaluation again on the corrected data.

Images