Educational intelligent agent system and method based on multi-engine cooperation and dynamic knowledge evolution

Abstract

The application discloses an education intelligent agent system and method based on multi-engine cooperation and dynamic knowledge evolution, and the system comprises a data acquisition layer, which is responsible for collecting original data from various data sources and providing data support for the upper layer; a data layer is used for storing various data; a model layer is used for providing different levels of large model capabilities; an intelligent analysis layer uses an artificial intelligence layered large model as the technology, analyzes and processes the collected data, and generates high-level abstract data models and knowledge; an application service layer encapsulates the results of the intelligent analysis layer into reusable services; an application layer directly faces the application interface of the end user and provides specific functions; the application has the advantages that: a real personalized learning path planning is realized, through the deep cooperation of the six intelligent engines, the education mode innovation from "one face for thousands of people" to "one person thousands of faces" is realized, and the intelligent education experience of precise adaptation, real-time response and combination of learning and use is provided for each learner.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to an educational intelligent agent system and method based on multi-engine collaboration and dynamic knowledge evolution. Background Technology

[0002] In the field of digital higher and vocational education, a fragmented approach is prevalent, with mainstream models relying on a combination of large-scale online course platforms, static question bank systems, and isolated enterprise simulation software. These systems are isolated, forming "data silos" that fail to build a unified view of knowledge and skills. Their core flaws lie in the highly static and one-way transmission of content, lacking genuine personalization; interaction is primarily click-based, unable to support natural multi-turn dialogues that align with human cognitive habits; and, more importantly, a severe disconnect between theoretical teaching and practical application, failing to provide students with a closed-loop "learn-practice-assessment-application" experience. While early adaptive learning technologies attempted to achieve a degree of personalization through pre-defined rule paths, their flexibility and intelligence are limited, making them ill-suited for complex and open educational scenarios. Therefore, the market urgently needs an intelligent education architecture that deeply integrates knowledge systems, contextualized practice, and personalized guidance to overcome existing technological bottlenecks and achieve a paradigm shift from "tools" to "intelligent partners."

[0003] From a business motivation and value creation perspective, the reasons for launching educational intelligent agent products are very compelling: First: Adapt to technological trends: Seize the historic opportunities brought by generative AI and large language models, and deepen them from general chat tools into vertical, professional, and trustworthy experts in the field of education. This is an inevitable direction for technological development.

[0004] Second: Meeting market demand: For students: It addresses the pain points of "knowledge overload" and "choice paralysis" by providing a personalized learning partner who accompanies them throughout the process, deeply understands them, and offers full-chain support from "learning" to "practice" to "exams" to "employment".

[0005] For universities: help them improve teaching quality, implement individualized instruction, and more closely align teaching outcomes with the job market to enhance graduates' employment competitiveness.

[0006] For businesses: It provides a precise channel for identifying and developing the talent they need, reducing recruitment and training costs.

[0007] Third: Building core competitive advantages: By solving the aforementioned technical problems, we are not simply creating an "AI + Education" application, but rather building a deeply integrated educational ecosystem that is difficult to replicate. The knowledge graph, user data, and intelligent agent interaction logic within this system collectively form a powerful competitive moat.

[0008] The core requirements and functions of the educational intelligent agent include multimodal adaptive content generation and explanation, intelligent exercise generation and detailed explanation, personalized learning path generation, dynamic generation of user profiles, value verification and ability certification, and highly realistic job case simulation. The essence of this product is to build an AI educational intelligent agent that provides lifelong companionship and has the triple role of "domain knowledge expert", "personal tutor", and "career planning mentor".

[0009] Currently, products on the market mainly address the needs of higher education and vocational education through several isolated or combined solutions. These include: Massive Open Online Course (MOOC) platforms, which provide pre-recorded videos of courses from prestigious universities, fixed course outlines, PowerPoint presentations, and accompanying standardized assignments / quizzes; online question banks and exam preparation platforms, which gather massive amounts of questions, provide answers and explanations, and some offer simple error logs and chapter exercises; and traditional learning management systems, which serve as the "infrastructure" for teaching management, primarily used for issuing notices, uploading materials, submitting assignments, organizing discussions, and recording grades. Their core function is management rather than intelligent teaching.

[0010] The existing technology has the following problems: Data silos: Data between courses, exercises, simulations, and job requirements is completely disconnected. After completing a MOOC course, students need to jump to another platform to practice questions, and then to a third platform to do simulations, resulting in a fragmented experience and an inability to form a unified learning profile.

[0011] Lack of true personalization: None of the solutions can achieve dynamic, generative, and multimodal content creation and path planning based on a continuous, multi-dimensional user profile.

[0012] Weak interactivity: The interaction methods are mainly based on clicking, selecting, and submitting, and cannot support multi-turn dialogue in natural language, which is the most suitable way for human learning habits.

[0013] There is a disconnect between cognitive and skills development: theoretical learning and practical application are separate, failing to form a closed loop of "learning-practice-feedback-relearning".

[0014] Low intelligence: The system has no "memory", no "planning" ability, and cannot understand the context. It is essentially a collection of simple rules, rather than an intelligent agent with cognitive abilities.

[0015] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0016] The technical problem this invention aims to solve is to overcome the above-mentioned technical deficiencies and provide an educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution, comprising: The data acquisition layer is responsible for collecting raw data from various data sources to provide data support for the upper layer. After cleaning and preprocessing, the collected data is stored in the data layer for use by the intelligent analysis layer. The data layer is used to store various types of data, including student profile databases, knowledge graph databases, learning behavior databases, assessment result databases, question bank databases, etc. The model layer is used to provide large model capabilities at different levels, including general large models, industry large models, and proprietary large models; The intelligent analysis layer uses a layered artificial intelligence model as its technology to analyze and process the collected data, generating high-level abstract data models and knowledge to support the application service layer. The application service layer encapsulates the results of the intelligent analysis layer into reusable services for the application layer to call. The application layer is the application interface that directly faces the end user and provides specific functions.

[0017] Preferably, the data acquisition layer includes: The learning behavior collection module is used to collect students' behavioral data on the learning platform, such as click stream, page dwell time, and interaction records. The knowledge data acquisition module is used to collect teaching resource data, including recorded lessons, textbooks, courseware, test question banks, etc. The job data collection module is used to collect employment market data, such as job recruitment needs and skill requirements. The assessment data collection module is used to collect assessment data such as students' competition and exam results, project performance, and online and offline learning.

[0018] Preferably, the intelligent analysis layer includes: The user profile engine module extracts multi-dimensional features of students based on learning behavior, assessment data, etc., and builds and updates user profiles. The knowledge graph engine module automatically constructs knowledge graphs from knowledge data, including knowledge points, concepts, and the relationships between them, and supports dynamic updates. At the same time, it builds the association between knowledge points and job skills and designs an update mechanism to dynamically adjust the knowledge graph based on students' learning progress and mastery.

[0019] The learning assessment and diagnosis engine module dynamically monitors students' mastery of knowledge points, learning efficiency, and ability development through a real-time assessment model. It also uses in-depth diagnostic analysis to accurately identify knowledge gaps and learning obstacles. The system automatically generates individual ability reports and teaching suggestion reports, while establishing a complete diagnostic feedback loop to continuously track intervention effects and optimize the assessment model, forming a closed-loop system of "assessment-diagnosis-intervention-optimization" to provide data support for personalized teaching. The recommendation algorithm engine module, based on the input from other engine modules, performs learning path planning and resource matching, recommends personalized learning content and order to students, and supports intelligent recommendation services and personalized path planning services in the application service layer; The intelligent question bank engine module enables personalized and adaptive question generation and recommendation, provides a parameterized question template library, supports multi-dimensional assessment of the same knowledge point, features an adaptive difficulty adjustment algorithm that dynamically adjusts question difficulty based on students' real-time performance, a knowledge point coverage optimization mechanism to ensure effective coverage of target knowledge points, and a question quality evaluation and elimination mechanism to optimize the quality of the question bank based on answer data. The value verification engine module integrates past exam questions for postgraduate entrance exams and civil service exams with enterprise cases. It verifies learning outcomes through simulated exams and project practice, providing a multi-dimensional value verification system for educational effectiveness. At the same time, it provides mapping analysis from past exam questions to knowledge points, mapping analysis of technical requirements from real enterprise cases, and can adjust the learning path based on the results.

[0020] Preferably, the application service layer includes: The user profile service module provides interfaces for querying and updating user profiles; The knowledge graph service module provides services such as knowledge graph querying and visualization. The learning assessment and diagnostic service module provides learning assessment results and diagnostic reports. The intelligent recommendation service module provides students with the most suitable learning resources, learning paths, practice questions, courses, etc. The personalized learning path planning service module generates personalized learning paths for students based on recommendation algorithms. The intelligent question bank service module provides personalized question bank generation and management services; The value verification service module provides value verification for learning outcomes, such as matching with past exam questions for postgraduate entrance examinations and civil service examinations, and matching with enterprise cases. Preferably, the application layer includes: a student-side application module, a teacher-side application module, and an administration-side application module, wherein the student-side application module includes: Personalized learning space: Displays personalized learning content and paths; Intelligent question bank practice: Provides personalized question practice; Learning progress dashboard: Displays learning progress and grades; Competency assessment report: Shows the results of learning assessments and diagnostics; The teacher-side application module includes: Class monitoring panel: Monitors the overall learning situation of the class; Learning analysis tools: Analyze students' learning situation and identify problems; Intervention management platform: Provides intervention for students with learning difficulties; The management application module includes: System configuration: manage system parameters and permissions; Data Insights: View system data statistics and analysis; Report generation: Generate various reports.

[0021] Preferably, in the intelligent analysis layer, the input dependencies of the user profiling engine include: The learning assessment and diagnostic engine provides ability assessment results to update the ability dimensions in the profile; The intelligent question bank engine uses answer data to analyze learning behavior and changes in ability. The knowledge graph engine provides knowledge point mastery data and knowledge point associations to help build profiles; The output of the user profiling engine has the following effects: The learning assessment and diagnostic engine provides users' historical and behavioral data as an assessment reference. The recommendation algorithm engine makes personalized recommendations based on user profile features and customizes the learning path; The intelligent question bank engine adjusts the difficulty of questions based on the user's ability level. The value verification engine uses student ability and skill data from the user profile engine to match and verify data using a verification model. The input dependencies of the knowledge graph engine include: The user profiling engine provides information on students' knowledge acquisition, which is used for personalized knowledge graph generation. The learning assessment and diagnostic engine provides learning diagnostic results and optimizes the association weights of knowledge points. The intelligent question bank engine provides data linking questions to knowledge points, thus improving the knowledge network. The output of the knowledge graph engine has the following impacts: The user profiling engine provides a knowledge structure framework to support capability dimension analysis. The learning assessment and diagnosis engine provides knowledge point dependencies to aid in the diagnosis of learning disabilities. The recommendation algorithm engine provides a knowledge topology structure to support learning path planning; The intelligent question bank provides a knowledge point system and guides the scope of question generation; The value verification engine provides a knowledge point-skill mapping relationship to support the verification of employability. The input dependencies of the learning evaluation and diagnostic engine include: The user profiling engine provides students' basic characteristics and learning history data. The knowledge graph engine provides a knowledge architecture and supports the diagnostic analysis framework. The intelligent question bank engine provides answer performance data for real-time ability assessment. The recommendation algorithm engine provides learning path execution data to evaluate learning effectiveness; The output impact of the learning evaluation and diagnostic engine includes: The user profiling engine provides ability assessment results and updates student ability dimensions. The knowledge graph engine provides information on knowledge point mastery and adjusts the strength of graph relationships. The recommendation algorithm engine provides diagnostic results and optimizes learning path recommendations. The intelligent question bank engine identifies weaknesses and guides the generation of targeted questions. The value validation engine provides learning performance data to support value validation analysis. The input dependencies of the recommendation algorithm engine include: The user profiling engine provides data on student characteristics, preferences, and ability levels. Knowledge graph engines provide knowledge structure and dependency relationship data; The learning assessment and diagnostic engine provides the current learning status and diagnostic results. The value verification engine provides verification feedback to optimize recommendation strategies. The output of the recommendation algorithm engine has the following effects: The user profiling engine provides recommendation behavior data to enrich learning behavior features; The learning evaluation and diagnostic engine provides recommended path execution data for evaluating learning effectiveness. The intelligent question bank engine provides question recommendation strategies and sequence planning; The value validation engine provides learning path planning and supports value pre-validation. The input dependencies of the intelligent question bank engine include: The user profiling engine provides data on students' ability levels and learning styles. A knowledge graph engine provides a knowledge point system and its relationships; The learning assessment and diagnostic engine provides diagnosis of knowledge gaps and weaknesses. The recommendation algorithm engine provides question recommendation strategies and difficulty requirements; The output impact of the intelligent question bank engine includes: The user profiling engine provides data on answering behaviors and updates learning behavior characteristics; The learning assessment and diagnostic engine provides answer performance data for real-time ability assessment. The value verification engine provides practice effect data to support the verification of skill mastery. The input dependencies of the value verification engine include: The user profiling engine provides student ability and skills data; The learning assessment and diagnostic engine provides data on learning effectiveness and progress. The recommendation algorithm engine provides learning path planning and execution data; The intelligent question bank engine provides data on practice effectiveness and skill mastery. The output impact of the value verification engine includes: The user profiling engine provides verification results and updates student competency certification status. The learning assessment and diagnostic engine provides valuable feedback and optimizes assessment criteria. The recommendation algorithm engine provides validation suggestions and adjusts the learning path strategy.

[0022] Preferably, the system also includes real-time collaboration and dynamic dependency graph collaboration. Real-time collaboration is used for collaboration in small-scale business scenarios, ensuring that interdependent engines quickly obtain information from other engines and can initiate their processing within a short time. Dynamic dependency graph collaboration is used for large-scale business scenarios. The real-time collaboration scheme includes: A unified data bus is used to realize data exchange between engines. Each engine publishes its processing results to the bus, and other engines can subscribe to the data they need. Standardize data formats and define a unified data exchange format, such as using Protobuf or Avro for serialization, to ensure that each engine can parse the data; Event-driven architecture: When an engine generates new data or a state change, it emits an event, and other engines respond to these events and trigger the corresponding processing flow. The proposed dynamic dependency graph collaboration method introduces a dynamic dependency graph to model and manage complex dependencies between engines in real time. The dynamic dependency graph is a directed graph G = (V, E), where vertices V represent each engine and its output state data, edges E represent dependencies, and the weights of the edges can be dynamically adjusted according to the system's running state. Engine state management involves designing a state management service to record the current state and latest output of each engine. It maintains the latest state of each engine, and the engine updates the state management service after completing its processing so that other engines can obtain it as needed. The construction and maintenance of the dynamic dependency graph includes: Vertex definition: Each vertex is a triple (Engine_ID, Data_Type, Data_Version), representing (engine ID, data status, and version status), respectively. Edge definition: A dependency edge E(A->B) indicates that a certain execution of engine B depends on the latest or a specific version of the output of engine A. The edge weight W(A->B) represents the degree of criticality of this dependency; Dynamic updates: Initial graph, the dependency graph is initialized based on business logic; runtime evolution, the system adjusts dependencies based on real-time feedback; The dependency graph-based collaborative interaction process includes the following steps: S51. Event Trigger: When a student logs into the system and requests a new learning path, this event is captured by the recommendation algorithm engine (set as vertex R). S52. Dependency Resolution: The recommendation algorithm engine R queries the dynamic dependency graph G to resolve all the prerequisite vertices required to generate the learning path. Let the set of these prerequisite vertices be D={D1, D2, ..., Dk}, where each Di corresponds to an engine's output state; S53, Task Orchestration: R checks the state management service to obtain the current state of each dependent vertex Di; for each Di, there may be three states: data is ready and fresh, data is being computed, or data is missing or expired; S54. Parallel Data Acquisition and Conditional Waiting: For dependency vertices in the state that the data is ready and fresh, the data is directly acquired from the state management service; for dependency vertices in the state that the data is being computed, R registers a callback and waits for the data to be computed; for dependency vertices in the state that the data is missing or expired, R needs to trigger the corresponding engine to perform the computation. S55, Execution and Feedback Closed Loop: Once all the dependent data is ready, R executes its core recommendation algorithm to generate N candidate learning paths; R submits N candidate paths to the value verification engine (let's call them vertex V) for fast simulation verification; where V may already be in the dependency set D, but at this time V is required to verify N paths, which is a new computational task, so it can be regarded as a new vertex V', which depends on the output of R; V returns the value score V_score for each path; R selects the path with the highest V_score as the final recommendation and records the complete context of this decision. Adaptive optimization of dependency graph: The system background continuously monitors the effectiveness of path recommendations; by analyzing historical decision records, the system uses correlation analysis or reinforcement learning to adjust the edge weights in the dynamic dependency graph.

[0023] Preferably, the learning path recommendation process includes the following steps: S11. Student login system: The recommendation algorithm engine needs to generate personalized learning paths for students. S12. The recommendation algorithm engine requests the student's user profile from the user profile engine. S13. The recommendation algorithm engine requests the student's current learning assessment results from the learning assessment and diagnosis engine. S14. The recommendation algorithm engine requests the knowledge structure of relevant disciplines from the knowledge graph engine; S15. The recommendation algorithm engine integrates this information, uses algorithms to generate learning paths, and returns them to the application layer. S16. At the same time, the value verification engine can evaluate the value of the recommended learning path and feed the evaluation results back to the recommendation algorithm engine to optimize the recommendation. The dynamic question generation process includes the following steps: S21. When students begin practicing, the intelligent question bank engine needs to generate a set of practice questions. S22. The intelligent question bank engine requests students' ability levels and learning styles from the user profile engine. S23. The intelligent question bank engine requests the student's knowledge gaps and recent learning difficulties from the learning assessment and diagnosis engine. S24. The intelligent question bank engine requests question templates and difficulty information for relevant knowledge points from the knowledge graph engine; S25. The intelligent question bank engine generates questions based on the above information and returns them to the application layer; S26. After students answer the questions, the learning assessment and diagnosis engine will update the learning assessment results based on the answers and update the user profile in the user profile engine. The learning assessment and update process includes the following steps: S31. After a student completes a learning task, the learning assessment and diagnostic engine will receive new learning data. S32. The learning assessment and diagnosis engine obtains students' historical profiles from the user profile engine and analyzes them in conjunction with new data. S33. The learning assessment and diagnosis engine uses the knowledge structure provided by the knowledge graph engine to analyze students' mastery of the knowledge graph. S34. The learning assessment and diagnosis engine updates the assessment results and sends the new assessment results to the user profile engine to update the profile; S35. Simultaneously, the learning evaluation and diagnosis engine sends the diagnosed knowledge gaps to the recommendation algorithm engine and the intelligent question bank engine so that they can be given priority in subsequent recommendations and question generation; The value verification process includes the following steps: S41. The value verification engine regularly collects student ability data from the user profile engine, learning outcome data from the learning assessment and diagnosis engine, and knowledge value data from the knowledge graph engine; S42. The value verification engine evaluates the learning value of students based on a pre-set verification model; S43. The value verification engine feeds back the value verification results to the student and teacher ends, and may also feed them back to the recommendation algorithm engine to adjust the recommendation strategy; Preferably, the correlation analysis algorithm includes the following steps: S61. Record historical data: For each dependency edge E(A->B), record the processing triggered by engine B after each data update of engine A, and record the following data: the degree of change in the output data of engine A, and the output effect of engine B; S62. Calculate correlation: Periodically analyze data over a period of time, and for each dependency edge, calculate the correlation between the two variables: Variable X: The degree of change in the output data of engine A; Variable Y: The output effect of engine B; Calculate the Pearson correlation coefficient ρ between X and Y.

[0024] S63. Adjusting weights: If ρ > θ_high, then increase the weight: w(e) = w(e) * (1 + η), where η is the increase factor; If ρ < θ_low, then reduce the weight: w(e) = w(e) * (1 - η); The reinforcement learning method includes the following steps: S71. Initialization: For each dependent edge i, initialize the weight W_i to 1.0. Initialize the preference functions H_i(1) and H_i(0) of the two actions to 0.

[0025] S72. Repeat until the weights converge or the preset number of iterations is reached: For each dependent edge i, select the action using the softmax strategy: P(a=1) = exp(H_i(1)) / (exp(H_i(0)) + exp(H_i(1))); P(a=0) = 1 - P(a=1)b; Perform the action and update the weights: W_i = W_i + α * (a - 0.5); Where: a=1 represents an increase, a=0 represents a decrease, so (a-0.5) is 0.5 or -0.5. After multiplying by the step size α, the weight increases or decreases by 0.05. Observe the reward R and the effect of engine B's processing this time, normalize to [0,1].

[0026] Update the average reward R_avg (used as the baseline): R_avg = R_avg + η * (R - R_avg) Update the action preference function: For the chosen action a: H_i(a) = H_i(a) + β * (R - R_avg) * (1 - P(a)) For the unselected action (1-a): H_i(1-a) = H_i(1-a) - β * (R - R_avg) * P(1-a).

[0027] Preferably, an adaptive hierarchical triggering mechanism is also included, the specific algorithm of which is as follows: A significance threshold θ is dynamically set for each student-engine combination, so that the downstream engine is only triggered to update when the amount of data change Δ exceeds θ. This can optimize the triggering frequency based on individual student differences and system load, balancing real-time performance and system overhead. S81. Define a change Δ. For each student and each engine, when the engine generates a new output for that student, calculate the difference between the new output and the old output. For example, for a learning assessment engine, the output might be a knowledge mastery vector K. The change is measured using Euclidean distance or cosine similarity. Assuming Euclidean distance is used, then Δ = ||K_new - K_old||. S82. Record historical changes. For each student-engine combination, maintain a fixed-size historical window to record the most recent m changes Δ. For example, the most recent 100 changes can be stored, denoted as the sequence D = [Δ1, Δ2, ..., Δm]. S83. Calculate the mean and standard deviation using sequence D; S84. Dynamically set the threshold θ, set θ = μ + k * σ, where k is an adjustable parameter. Intuitively, this means that the threshold is the upper bound of the historical normal range of change. If the current change exceeds this upper bound, the change is considered significant.

[0028] S85. Dynamically adjust k: The parameter k can be dynamically adjusted according to the system load. For example, when the system load is high, increase k to raise the threshold and reduce the triggering frequency; when the system load is low, decrease k to lower the threshold and make the system more sensitive. S86. Update the history window: After each calculation of Δ, add it to the history window and remove the oldest record to keep the window size fixed; S87. Trigger decision: Compare the current change Δ_current with the threshold θ. If Δ_current > θ, then trigger the downstream engine; otherwise, do not trigger.

[0029] The advantages of this invention compared to existing technologies are: 1. This invention addresses the core pain points of traditional education systems through a multi-engine collaborative architecture: For the problem of "homogenized teaching," the system achieves truly personalized learning path planning through a user profiling engine and a recommendation algorithm engine, dynamically adjusting teaching content based on students' cognitive characteristics, learning styles, and real-time status; for the pain point of "delayed feedback," the learning assessment engine, combined with an event-driven architecture, enables learning status diagnosis and intervention; for the problem of "fragmented knowledge," the dynamic knowledge graph engine constructs a systematic knowledge association network, ensuring the integrity and coherence of learning content; and for the dilemma of "disconnect between learning and application," the value verification engine, through job skill mapping and real-world case practice, establishes a closed loop from knowledge learning to practical application. Compared to the single-dimensional recommendation and static content system of traditional online education platforms, this invention's system, through the deep collaboration of six intelligent engines, achieves a revolutionary educational model from "one-size-fits-all" to "personalized learning," providing each learner with a precisely tailored, real-time responsive, and integrated intelligent educational experience. Attached Figure Description

[0030] Figure 1 This is a diagram of the intelligent agent architecture of the present invention. Detailed Implementation

[0031] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0032] like Figure 1 As shown, the educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution includes: The first layer, the data acquisition layer, is responsible for collecting raw data from various data sources to provide data support for the upper layers.

[0033] Module division: Learning behavior data collection: Collect students' behavioral data on the learning platform, such as click stream, page dwell time, and interaction records (comments, likes, favorites, etc.).

[0034] Knowledge data collection: Collect teaching resource data, including recorded lessons, textbooks, courseware, test question banks, etc.

[0035] Job data collection: Collect employment market data, such as job recruitment needs and skill requirements.

[0036] Assessment data collection: Collect assessment data such as students' competition and exam results, project performance, and online and offline learning.

[0037] Data flow: After being cleaned and preprocessed, the collected data is stored in the data layer for use by the intelligent analysis layer.

[0038] Second, the data layer stores various types of data, including student profile databases, knowledge graph databases, learning behavior databases, assessment result databases, and question bank databases. It supports the layers below.

[0039] Third, the model layer provides large model capabilities at different levels, including general large models, industry large models, and dedicated large models.

[0040] Fourth, the intelligent analysis layer uses a layered artificial intelligence model to analyze and process the collected data, generating high-level abstract data models and knowledge to support the application service layer.

[0041] Module division: User profile engine: Based on learning behavior, assessment data, etc., extract multi-dimensional characteristics of students (such as learning ability, interest preferences, knowledge mastery, etc.) to build and update user profiles.

[0042] Knowledge Graph Engine: Automatically constructs knowledge graphs from knowledge data, including knowledge points, concepts, and the relationships between them, and supports dynamic updates. It also establishes connections between knowledge points and job skills; and designs an update mechanism to dynamically adjust the knowledge graph based on students' learning progress and mastery (e.g., marking key points and difficulties).

[0043] Learning Assessment and Diagnostic Engine: This engine dynamically monitors students' knowledge mastery, learning efficiency, and ability development through a real-time assessment model. Deep diagnostic analysis accurately identifies knowledge gaps and learning obstacles. The system automatically generates individual ability reports and teaching suggestion reports, while establishing a complete diagnostic feedback loop to continuously track intervention effects and optimize the assessment model, forming a closed-loop system of "assessment-diagnosis-intervention-optimization" to provide data support for personalized teaching.

[0044] Recommendation algorithm engine: Based on input from other engine modules, it performs learning path planning and resource matching to recommend personalized learning content and order to students. It supports intelligent recommendation services and personalized path planning services at the application service layer.

[0045] Intelligent Question Bank Engine: Enables personalized and adaptive question generation and recommendation, provides a parameterized question template library, supports multi-dimensional assessment of the same knowledge point; adaptive difficulty adjustment algorithm, dynamically adjusts question difficulty based on students' real-time performance; knowledge point coverage optimization mechanism, ensures that questions effectively cover the target knowledge points; question quality evaluation and elimination mechanism, optimizes the quality of the question bank based on answer data.

[0046] Value Validation Engine: Integrating past exam questions for postgraduate entrance exams and civil service exams with enterprise case studies, this system verifies learning outcomes through simulated exams and project practice, providing a multi-dimensional value validation system for educational effectiveness. It also offers mapping analysis from past exam questions to knowledge points, mapping analysis of technical requirements from real enterprise case studies, and allows for adjustments to learning paths based on the results.

[0047] Fifth, the application service layer encapsulates the results of the intelligent analysis layer into reusable services for the application layer to call.

[0048] Module division: User profiling service: Provides interfaces for querying and updating user profiles.

[0049] Knowledge graph services: Provides services such as knowledge graph querying and visualization.

[0050] Learning assessment and diagnostic services: Provide learning assessment results and diagnostic reports.

[0051] Intelligent recommendation service: Provides students with the most suitable learning resources, learning paths, practice questions, courses, etc.

[0052] Personalized learning path planning service: Based on recommendation algorithms, it generates personalized learning paths for students.

[0053] Intelligent question bank service: Provides personalized question bank generation and management services.

[0054] Value Validation Service: Provides value validation for learning outcomes, such as linking with past exam questions for postgraduate entrance exams and civil service exams, and matching with corporate cases.

[0055] VI. Application Layer: This layer consists of the application interface directly facing the end user, providing specific functions. It can be a traditional website interaction method or a language-based interactive method (such as the Deepseek interaction style).

[0056] Module division: Student application: Personalized learning space: Displays personalized learning content and paths.

[0057] Intelligent question bank practice: Provides personalized question practice.

[0058] Learning progress dashboard: Displays learning progress and grades.

[0059] Competency assessment report: Shows the results of learning assessments and diagnostics.

[0060] Teacher-side application: Class monitoring panel: Monitors the overall learning situation of the class.

[0061] Learning analysis tools: Analyze students' learning situation and identify problems.

[0062] Intervention Management Platform: To provide intervention for students with learning difficulties.

[0063] Management application: System configuration: Manage system parameters and permissions.

[0064] Data Insights: View system data statistics and analysis.

[0065] Report generation: Generate various reports.

[0066] Multi-engine interaction process In this solution, the various engines work together to ensure that they can collaborate efficiently and in an orderly manner to provide personalized support for students.

[0067] For a user profiling engine, its input dependencies and output impacts are as follows: Input dependencies: The learning assessment and diagnostic engine provides ability assessment results to update the ability dimensions in the profile. The intelligent question bank engine uses answer data to analyze changes in learning behavior and abilities. The knowledge graph engine provides knowledge point mastery data and knowledge point associations to help build profiles.

[0068] Output impact: The learning assessment and diagnostic engine provides users' historical and behavioral data as assessment references. The recommendation algorithm engine performs personalized recommendations based on user profile features and customizes the learning path. The intelligent question bank engine adjusts the difficulty of questions based on the user's ability level. The value verification engine uses student ability and skill data from the user profile engine to match and verify data using a verification model.

[0069] For knowledge graph engines, the input dependencies and output impacts are as follows: Input dependencies: The user profiling engine provides information on students' knowledge acquisition, which is used for personalized knowledge graph generation. The learning assessment and diagnostic engine provides learning diagnostic results and optimizes the association weights of knowledge points. The intelligent question bank engine provides data linking questions to knowledge points, thus improving the knowledge network.

[0070] Output impact: The user profiling engine provides a knowledge structure framework to support capability dimension analysis. The learning assessment and diagnostic engine provides knowledge point dependencies to aid in the diagnosis of learning disabilities. The recommendation algorithm engine provides a knowledge topology structure to support learning path planning. The intelligent question bank engine provides a knowledge point system to guide the scope of question generation. The value verification engine provides a knowledge point-skill mapping relationship to support the verification of employability.

[0071] For the learning evaluation and diagnostic engine, its input dependencies and output impacts are as follows: Input Dependency The user profiling engine provides basic student characteristics and learning history data. Knowledge graph engines provide knowledge architecture structure to support diagnostic analysis frameworks. The intelligent question bank engine provides answer performance data for real-time ability assessment. The recommendation algorithm engine provides learning path execution data to evaluate learning effectiveness. Output impact: The user profiling engine provides competency assessment results and updates student competency dimensions. The knowledge graph engine provides information on knowledge point mastery and adjusts the strength of graph relationships. The recommendation algorithm engine provides diagnostic results and optimizes learning path recommendations. The intelligent question bank engine identifies weaknesses and guides the generation of targeted questions. The value validation engine provides learning performance data to support value validation analysis. For recommendation algorithm engines, their input dependencies and output impacts are as follows: Input dependencies: The user profiling engine provides data on student characteristics, preferences, and ability levels. Knowledge graph engines provide knowledge structure and dependency relationship data. The learning assessment and diagnostic engine provides the current learning status and diagnostic results. The value verification engine provides verification feedback to optimize recommendation strategies. Output impact: The user profiling engine provides recommendation behavior data to enrich learning behavior features. The learning evaluation and diagnostic engine provides recommended path execution data for evaluating learning effectiveness. The intelligent question bank engine provides question recommendation strategies and sequence planning. The value validation engine provides learning path planning to support value pre-validation. For intelligent question bank engines, the input dependencies and output impacts are as follows: Input dependencies: The user profiling engine provides data on student ability levels and learning styles. Knowledge graph engine provides a knowledge point system and relationships. The learning assessment and diagnostic engine provides diagnostics for knowledge gaps and weaknesses. The recommendation algorithm engine provides question recommendation strategies and difficulty requirements. Output impact: The user profiling engine provides data on answering behaviors and updates learning behavior characteristics. The learning assessment and diagnostic engine provides answer performance data for real-time ability assessment. The value verification engine provides practice performance data to support the verification of skill mastery. For the value validation engine, its input dependencies and output impacts are as follows: Input Dependency User profiling engine provides student ability and skills data. The learning assessment and diagnostic engine provides data on learning effectiveness and progress. The recommendation algorithm engine provides learning path planning and execution data. The intelligent question bank engine provides data on practice effectiveness and skill mastery. Output impact: The user profiling engine provides verification results and updates student competency certification status. The learning assessment and diagnostic engine provides valuable feedback and optimizes assessment criteria. The recommendation algorithm engine provides validation suggestions and adjusts the learning path strategy. Example of interaction flow: Learning path recommendation process: a. When students log into the system, the recommendation algorithm engine needs to generate personalized learning paths for each student.

[0072] b. The recommendation algorithm engine requests the student's user profile (including interests, historical learning behavior, etc.) from the user profile engine.

[0073] c. The recommendation algorithm engine requests the student's current learning assessment results (such as knowledge mastery, ability level, etc.) from the learning assessment and diagnosis engine.

[0074] d. The recommendation algorithm engine requests the knowledge structure of relevant subjects (the relationship between knowledge points, the progression of difficulty, etc.) from the knowledge graph engine.

[0075] e. The recommendation algorithm engine integrates this information, uses algorithms to generate learning paths, and returns them to the application layer.

[0076] f. At the same time, the value verification engine can evaluate the value of the recommended learning path (such as predicting learning effectiveness, matching degree with career goals, etc.) and feed the evaluation results back to the recommendation algorithm engine to optimize the recommendation.

[0077] Dynamic question generation process: a. When students begin practicing, the intelligent question bank engine needs to generate a set of practice questions.

[0078] b. The intelligent question bank engine requests students' ability levels and learning styles from the user profile engine.

[0079] c. The intelligent question bank engine requests information from the learning assessment and diagnosis engine regarding students' knowledge gaps and recent learning difficulties.

[0080] d. The intelligent question bank engine requests question templates and difficulty information for relevant knowledge points from the knowledge graph engine.

[0081] e. The intelligent question bank engine generates questions based on the above information and returns them to the application layer.

[0082] f. After students answer the questions, the learning assessment and diagnostic engine will update the learning assessment results based on the answers, and update the user profile in the user profile engine.

[0083] Learning assessment update process: a. After a student completes a learning task (such as watching a video or completing an exercise), the learning assessment and diagnostic engine will receive new learning data.

[0084] b. The learning assessment and diagnostic engine obtains students' historical profiles from the user profiling engine and analyzes them in conjunction with new data.

[0085] c. The learning assessment and diagnostic engine utilizes the knowledge structure provided by the knowledge graph engine to analyze students' mastery of the knowledge graph.

[0086] d. The learning assessment and diagnostic engine updates the assessment results and sends the new assessment results to the user profiling engine to update the profile.

[0087] e. At the same time, the learning assessment and diagnosis engine sends the diagnosed knowledge gaps to the recommendation algorithm engine and the intelligent question bank engine so that they can be given priority in subsequent recommendations and question generation.

[0088] Value validation process: a. The value verification engine regularly collects student ability data from the user profile engine, learning outcome data from the learning assessment and diagnosis engine, and knowledge value data (such as the correlation with job skills) from the knowledge graph engine.

[0089] b. The value verification engine evaluates students' learning value based on preset verification models (such as matching postgraduate entrance examination questions with job skills).

[0090] c. The value verification engine feeds back the value verification results to the student (e.g., generating a value report) and the teacher, and may also feed them back to the recommendation algorithm engine to adjust the recommendation strategy.

[0091] Multi-engine collaborative mechanism Through the above interaction process, it can be seen that there are many interactions between the various engines in this solution design, rather than isolated entities. In order to enable the various engines to collaborate better, two collaboration methods are provided. One is real-time collaboration, which is more suitable for collaboration under small business volume. It can ensure that interdependent engines can quickly obtain information from other engines and start their own processing in a short period of time. However, in the case of large-scale business, if only real-time collaboration is provided, it is easy for multiple engines to start a large amount of work at the same time, which will put a huge burden on resources. In severe cases, it will lead to resource exhaustion and system paralysis. Therefore, a dynamic dependency graph collaboration solution is also provided. Each engine chooses the timing of processing based on its own dependencies and tolerable latency.

[0092] Real-time collaboration solution 1. A unified data bus (such as a message queue or data bus service) is used to achieve data exchange between engines. Each engine publishes its processing results to the bus, and other engines can subscribe to the data they need.

[0093] Technology selection: Apache Kafka or RabbitMQ as message middleware to achieve asynchronous data transmission and decoupling.

[0094] 2. Standardize data formats and define a unified data exchange format, such as using Protobuf or Avro for serialization, to ensure that each engine can parse the data.

[0095] 3. Event-driven architecture: When an engine generates new data or a state change, it emits an event, and other engines respond to these events and trigger the corresponding processing flow.

[0096] For example, when the learning assessment engine completes an assessment, it will publish an "assessment completed" event. The recommendation algorithm engine and the intelligent question bank engine subscribe to this event, thereby triggering new recommendations and question generation.

[0097] Dynamic Dependency Graph Collaboration Scheme This system introduces a dynamic dependency graph to model and manage complex dependencies between engines in real time. This graph is a directed graph G = (V, E), where vertices V represent engines and their output state data, edges E represent dependencies, and edge weights can be dynamically adjusted according to the system's running state. Engine state management involves designing a state management service to record the current state and latest output of each engine. It maintains the latest state of each engine (e.g., engine version status, latest output data version, last processing timestamp, dependencies, etc.). After completing processing, each engine updates the state management service so that other engines can access it as needed.

[0098] Construction and maintenance of dynamic dependency graphs Vertex definition: Each vertex is not a simple engine, but a triple (Engine_ID, Data_Type, Data_Version), representing (engine ID, data status, and version status) respectively.

[0099] Edge definition: A dependency edge E(A->B) indicates that a certain execution of engine B depends on the latest or a specific version of the output of engine A. The edge weight W(A->B) can represent the criticality of the dependency (e.g., strong dependency = 1, weak dependency = 0.5) or the data freshness requirement.

[0100] Dynamically updated: 1) Initial graph: Initialize the dependency graph according to the business logic.

[0101] 2) Runtime Evolution: The system adjusts dependencies based on real-time feedback. For example, when the system detects that a student has a low completion rate for the "recommended path," it will automatically reduce the dependency weight of the recommendation engine on the profiling engine and increase the dependency weight of the "knowledge gaps" in the "diagnostic engine."

[0102] Dependency Graph-Based Collaborative Interaction Process (Algorithmic Description) Taking the "learning path recommendation process" as an example, this explains how dynamic dependency graphs drive engine interactions.

[0103] 1) Event Trigger: A student logs into the system and requests a new learning path. This event is captured by the recommendation algorithm engine (let's call it vertex R).

[0104] 2) Dependency Resolution: The recommendation algorithm engine R queries the dynamic dependency graph G to resolve all the prerequisite dependency vertices required to generate the learning path. Let the set of these prerequisite dependency vertices be D={D1, D2, ..., Dk}, where each Di corresponds to the output state of an engine (e.g., the output of the user profile engine, the output of the knowledge graph engine, etc.).

[0105] 3) Task Orchestration: R checks the state management service to obtain the current state of each dependent vertex Di. For each Di, there are three possible states: a) Data is ready and fresh (i.e., the data version meets the requirements and is not expired); b) Data is being computed (triggered by other requests); c) Data is missing or expired. 4) Parallel Data Acquisition and Conditional Waiting: For dependent vertices in state a), data is directly acquired from the state management service. For dependent vertices in state b), R registers a callback and waits for the data computation to complete. For dependent vertices in state c), R needs to trigger the corresponding engine for computation. This may result in recursively triggering dependencies of that engine, forming subtasks.

[0106] For efficiency, R processes these dependencies concurrently and enters a conditional wait state until all dependencies have data in place or time out.

[0107] 5) Execution and feedback closed loop: a) Once all the dependent data is ready, R executes its core recommendation algorithm to generate N candidate learning paths.

[0108] b) R submits N candidate paths to the value verification engine (let's call it vertex V) for fast simulation verification. Note that V may already be in the dependency set D, but now V is required to verify N paths, which is a new computational task. Therefore, it can be regarded as a new vertex V' that depends on the output of R.

[0109] c) V returns the value score V_score for each path.

[0110] d) R selects the path with the highest V_score as the final recommendation and records the complete context of this decision (including the version of the dependent data used, candidate paths, value scores, etc.).

[0111] 6) Adaptive optimization of dependency graph: The system backend continuously monitors the effectiveness of path recommendations (e.g., students' subsequent completion rate, grade improvement, etc.). By analyzing historical decision records, the system uses correlation analysis or reinforcement learning to adjust the edge weights in the dynamic dependency graph.

[0112] Correlation analysis algorithm: 1) Record historical data: For each dependent edge E(A->B), record the processing triggered by engine B after each data update of engine A, and record the following data: the degree of change of the output data of engine A (e.g., the update range of user profile, which can be measured by vector distance), and the output effect of engine B (e.g., the completion rate of recommended paths, the accuracy rate of questions in the question bank, etc.).

[0113] 2) Calculate correlation: Analyze data over a period of time (e.g., every 24 hours). For each dependency edge, calculate the correlation between the two variables: Variable X: The degree of change in the output data of Engine A (the amount of change at each trigger). Variable Y: The output performance of engine B (e.g., completion rate, accuracy rate). Calculate the Pearson correlation coefficient ρ between X and Y.

[0114] 3) Adjusting weights: If ρ > θ_high (e.g., 0.5), then increase the weight: w(e) = w(e) * (1 + η), where η is the increase factor (e.g., 0.1).

[0115] If ρ < θ_low (e.g., 0.1), then reduce the weight: w(e) = w(e) * (1 - η).

[0116] It should be noted that upper and lower limits should be set for weight adjustments to avoid them being too large or too small.

[0117] Reinforcement learning methods (using gradient methods): 1) Initialization: For each dependent edge i, initialize the weight W_i to 1.0. Initialize the preference functions H_i(1) and H_i(0) for the two actions (increase and decrease) to 0.

[0118] 2) Loop (at each time step, e.g., after each time engine B is triggered) until the weights converge or the preset number of iterations is reached: A. For each dependent edge i, select the action using the softmax strategy: P(a=1) = exp(H_i(1)) / (exp(H_i(0)) + exp(H_i(1))) P(a=0) = 1 - P(a=1)b B. Perform the action and update the weights: W_i = W_i + α * (a - 0.5) Note: a=1 means increase, a=0 means decrease, so (a-0.5) is 0.5 or -0.5. After multiplying by the step size α (e.g., 0.1), the weight increases or decreases by 0.05.

[0119] C. Observe the reward R (the effect of engine B's processing this time, normalized to [0,1]).

[0120] Update the average reward R_avg (used as the baseline): R_avg = R_avg + η * (R - R_avg) D. Update the preference function for actions: For the chosen action a: H_i(a) = H_i(a) + β * (R - R_avg) * (1 - P(a)) For the unselected action (1-a): H_i(1-a) = H_i(1-a) - β * (R - R_avg) * P(1-a) Adaptive hierarchical triggering mechanism Frequent real-time data synchronization between multiple engines incurs significant computational and communication overhead. Traditional fixed threshold triggering mechanisms have two major drawbacks: first, insufficient sensitivity, as fixed thresholds cannot adapt to the dynamic changes in the learning characteristics of different students; and second, resource waste, as frequent triggering of downstream engine updates leads to excessive system load.

[0121] This algorithm achieves precise triggering through personalized and adaptive threshold settings. It maintains high sensitivity to students with rapidly changing learning states, reduces unnecessary triggering for students with stable learning states, and dynamically adjusts the triggering strategy based on system load to ensure system stability. The specific algorithm is as follows: A significance threshold θ is dynamically set for each student-engine combination, ensuring that the downstream engine is only triggered to update when the data change Δ exceeds θ. This optimizes the triggering frequency based on individual student differences and system load, balancing real-time performance and system overhead.

[0122] Step 1: Define the change Δ For each student and each engine, when the engine generates a new output for that student, we need to calculate the difference between the new output and the old output. For example, for a learning evaluation engine, the output might be a knowledge mastery vector K. We can measure the change using Euclidean distance or cosine similarity. Assuming we use Euclidean distance, then Δ = ||K_new - K_old||.

[0123] Step 2: Record historical changes For each student-engine combination, maintain a fixed-size history window to record the most recent m changes Δ. For example, the most recent 100 changes can be stored, denoted as the sequence D = [Δ1, Δ2, ..., Δm].

[0124] Step 3: Calculate the mean and standard deviation The mean μ and standard deviation σ are calculated using sequence D.

[0125] Step 4: Dynamically set the threshold θ Let θ = μ + k * σ, where k is an adjustable parameter. Intuitively, this means the threshold is the upper bound of the historical normal range of change. If the current change exceeds this upper bound, the change is considered significant.

[0126] Step 5: Dynamically adjust k The parameter k can be dynamically adjusted based on the system load. For example, when the system load (such as CPU utilization or request queue length) is high, k is increased to raise the threshold and reduce the triggering frequency; when the system load is low, k is decreased to lower the threshold and make the system more sensitive.

[0127] Step 6: Update the history window After each Δ calculation, add it to the history window and remove the oldest record to keep the window size fixed.

[0128] Step 7: Trigger the decision The current change Δ_current is compared with the threshold θ. If Δ_current > θ, the downstream engine is triggered; otherwise, it is not triggered.

[0129] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. An educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution, characterized in that... ,include: The data acquisition layer is responsible for collecting raw data from various data sources to provide data support for the upper layer. After cleaning and preprocessing, the collected data is stored in the data layer for use by the intelligent analysis layer. The data layer is used to store various types of data, including student profile databases, knowledge graph databases, learning behavior databases, assessment result databases, question bank databases, etc. The model layer is used to provide large model capabilities at different levels, including general large models, industry large models, and proprietary large models; The intelligent analysis layer uses a layered artificial intelligence model as its technology to analyze and process the collected data, generating high-level abstract data models and knowledge to support the application service layer. The application service layer encapsulates the results of the intelligent analysis layer into reusable services for the application layer to call. The application layer is the application interface that directly faces the end user and provides specific functions.

2. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 1, characterized in that: The data acquisition layer includes: The learning behavior collection module is used to collect students' behavioral data on the learning platform, such as click stream, page dwell time, and interaction records. The knowledge data acquisition module is used to collect teaching resource data, including recorded lessons, textbooks, courseware, test question banks, etc. The job data collection module is used to collect employment market data, such as job recruitment needs and skill requirements. The assessment data collection module is used to collect assessment data such as students' competition and exam results, project performance, and online and offline learning.

3. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 1, characterized in that: The intelligent analysis layer includes: The user profile engine module extracts multi-dimensional features of students based on learning behavior, assessment data, etc., and builds and updates user profiles. The knowledge graph engine module automatically constructs knowledge graphs from knowledge data, including knowledge points, concepts, and the relationships between them, and supports dynamic updates; at the same time, it builds the association between knowledge points and job skills; and it designs an update mechanism to dynamically adjust the knowledge graphs according to students' learning progress and mastery. The learning assessment and diagnosis engine module dynamically monitors students' mastery of knowledge points, learning efficiency, and ability development through a real-time assessment model. It also uses in-depth diagnostic analysis to accurately identify knowledge gaps and learning obstacles. The system automatically generates individual ability reports and teaching suggestion reports, while establishing a complete diagnostic feedback loop to continuously track intervention effects and optimize the assessment model, forming a closed-loop system of "assessment-diagnosis-intervention-optimization" to provide data support for personalized teaching. The recommendation algorithm engine module, based on the input from other engine modules, performs learning path planning and resource matching, recommends personalized learning content and order to students, and supports intelligent recommendation services and personalized path planning services in the application service layer; The intelligent question bank engine module enables personalized and adaptive question generation and recommendation, provides a parameterized question template library, supports multi-dimensional assessment of the same knowledge point, features an adaptive difficulty adjustment algorithm that dynamically adjusts question difficulty based on students' real-time performance, a knowledge point coverage optimization mechanism to ensure effective coverage of target knowledge points, and a question quality evaluation and elimination mechanism to optimize the quality of the question bank based on answer data. The value verification engine module integrates past exam questions for postgraduate entrance exams and civil service exams with enterprise cases. It verifies learning outcomes through simulated exams and project practice, providing a multi-dimensional value verification system for educational effectiveness. At the same time, it provides mapping analysis from past exam questions to knowledge points, mapping analysis of technical requirements from real enterprise cases, and can adjust the learning path based on the results.

4. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 1, characterized in that: The application service layer includes: The user profile service module provides interfaces for querying and updating user profiles; The knowledge graph service module provides services such as knowledge graph querying and visualization. The learning assessment and diagnostic service module provides learning assessment results and diagnostic reports. The intelligent recommendation service module provides students with the most suitable learning resources, learning paths, practice questions, courses, etc. The personalized learning path planning service module generates personalized learning paths for students based on recommendation algorithms. The intelligent question bank service module provides personalized question bank generation and management services; The value verification service module provides value verification for learning outcomes, such as matching with past exam questions for postgraduate entrance examinations and civil service examinations, and matching with enterprise cases.

5. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 1, characterized in that: The application layer includes: a student-side application module, a teacher-side application module, and a management-side application module. The student-side application module includes: Personalized learning space: Displays personalized learning content and paths; Intelligent question bank practice: Provides personalized question practice; Learning progress dashboard: Displays learning progress and grades; Competency assessment report: Shows the results of learning assessments and diagnostics; The teacher-side application module includes: Class monitoring panel: Monitors the overall learning situation of the class; Learning analysis tools: Analyze students' learning situation and identify problems; Intervention management platform: Provides intervention for students with learning difficulties; The management application module includes: System configuration: manage system parameters and permissions; Data Insights: View system data statistics and analysis; Report generation: Generate various reports.

6. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 3, characterized in that: In the intelligent analysis layer, the input dependencies of the user profiling engine include: The learning assessment and diagnostic engine provides ability assessment results to update the ability dimensions in the profile; The intelligent question bank engine uses answer data to analyze learning behavior and changes in ability. The knowledge graph engine provides knowledge point mastery data and knowledge point associations to help build profiles; The output of the user profiling engine has the following effects: The learning assessment and diagnostic engine provides users' historical and behavioral data as an assessment reference. The recommendation algorithm engine makes personalized recommendations based on user profile features and customizes the learning path; The intelligent question bank engine adjusts the difficulty of questions based on the user's ability level. The value verification engine uses student ability and skill data from the user profile engine to match and verify data using a verification model. The input dependencies of the knowledge graph engine include: The user profiling engine provides information on students' knowledge acquisition, which is used for personalized knowledge graph generation. The learning assessment and diagnostic engine provides learning diagnostic results and optimizes the association weights of knowledge points. The intelligent question bank engine provides data linking questions to knowledge points, thus improving the knowledge network. The output of the knowledge graph engine has the following impacts: The user profiling engine provides a knowledge structure framework to support capability dimension analysis. The learning assessment and diagnosis engine provides knowledge point dependencies to aid in the diagnosis of learning disabilities. The recommendation algorithm engine provides a knowledge topology structure to support learning path planning; The intelligent question bank provides a knowledge point system and guides the scope of question generation; The value verification engine provides a knowledge point-skill mapping relationship to support the verification of employability. The input dependencies of the learning evaluation and diagnostic engine include: The user profiling engine provides students' basic characteristics and learning history data. The knowledge graph engine provides a knowledge architecture and supports the diagnostic analysis framework. The intelligent question bank engine provides answer performance data for real-time ability assessment. The recommendation algorithm engine provides learning path execution data to evaluate learning effectiveness; The output impact of the learning evaluation and diagnostic engine includes: The user profiling engine provides ability assessment results and updates student ability dimensions. The knowledge graph engine provides information on knowledge point mastery and adjusts the strength of graph relationships. The recommendation algorithm engine provides diagnostic results and optimizes learning path recommendations. The intelligent question bank engine identifies weaknesses and guides the generation of targeted questions. The value validation engine provides learning performance data to support value validation analysis. The input dependencies of the recommendation algorithm engine include: The user profiling engine provides data on student characteristics, preferences, and ability levels. Knowledge graph engines provide knowledge structure and dependency relationship data; The learning assessment and diagnostic engine provides the current learning status and diagnostic results. The value verification engine provides verification feedback to optimize recommendation strategies. The output of the recommendation algorithm engine has the following impacts: The user profiling engine provides recommendation behavior data to enrich learning behavior features; The learning evaluation and diagnostic engine provides recommended path execution data for evaluating learning effectiveness. The intelligent question bank engine provides question recommendation strategies and sequence planning; The value validation engine provides learning path planning and supports value pre-validation. The input dependencies of the intelligent question bank engine include: The user profiling engine provides data on students' ability levels and learning styles. A knowledge graph engine provides a knowledge point system and its relationships; The learning assessment and diagnostic engine provides diagnosis of knowledge gaps and weaknesses. The recommendation algorithm engine provides question recommendation strategies and difficulty requirements; The output impact of the intelligent question bank engine includes: The user profiling engine provides data on answering behaviors and updates learning behavior characteristics. The learning assessment and diagnostic engine provides answer performance data for real-time ability assessment. The value verification engine provides practice effect data to support the verification of skill mastery. The input dependencies of the value verification engine include: The user profiling engine provides student ability and skills data; The learning assessment and diagnostic engine provides data on learning effectiveness and progress. The recommendation algorithm engine provides learning path planning and execution data; The intelligent question bank engine provides data on practice effectiveness and skill mastery. The output impact of the value verification engine includes: The user profiling engine provides verification results and updates student competency certification status. The learning assessment and diagnostic engine provides valuable feedback and optimizes assessment criteria. The recommendation algorithm engine provides validation suggestions and adjusts the learning path strategy.

7. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 1, characterized in that: It also includes real-time collaboration and dynamic dependency graph collaboration. Real-time collaboration is used for collaboration in small business volumes to ensure that interdependent engines can quickly obtain information from other engines and start their own processing in a short time. Dynamic dependency graph collaboration is used for large-scale business scenarios. The real-time collaboration scheme includes: A unified data bus is used to realize data exchange between engines. Each engine publishes its processing results to the bus, and other engines can subscribe to the data they need. Standardize data formats and define a unified data exchange format, such as using Protobuf or Avro for serialization, to ensure that each engine can parse the data; Event-driven architecture: When an engine generates new data or a state change, it emits an event, and other engines respond to these events and trigger the corresponding processing flow. The proposed dynamic dependency graph collaboration method introduces a dynamic dependency graph to model and manage complex dependencies between engines in real time. The dynamic dependency graph is a directed graph G = (V, E), where vertices V represent each engine and its output state data, edges E represent dependencies, and the weights of the edges can be dynamically adjusted according to the system's running state. Engine state management involves designing a state management service to record the current state and latest output of each engine. It maintains the latest state of each engine, and after an engine completes its processing, it updates the state management service so that other engines can obtain it as needed. The construction and maintenance of the dynamic dependency graph includes: Vertex definition: Each vertex is a triple (Engine_ID, Data_Type, Data_Version), representing (engine ID, data status, and version status), respectively. Edge definition: A dependency edge E(A->B) indicates that a certain execution of engine B depends on the latest or a specific version of the output of engine A. The edge weight W(A->B) represents the degree of criticality of this dependency; Dynamic updates: Initial graph, the dependency graph is initialized based on business logic; runtime evolution, the system adjusts dependencies based on real-time feedback; The dependency graph-based collaborative interaction process includes the following steps: S51. Event Trigger: When a student logs into the system and requests a new learning path, this event is captured by the recommendation algorithm engine (set as vertex R). S52. Dependency Resolution: The recommendation algorithm engine R queries the dynamic dependency graph G to resolve all the prerequisite vertices required to generate the learning path. Let the set of these prerequisite vertices be D={D1, D2, ..., Dk}, where each Di corresponds to an engine's output state; S53, Task Orchestration: R checks the state management service to obtain the current state of each dependent vertex Di; for each Di, there may be three states: data is ready and fresh, data is being computed, or data is missing or expired; S54. Parallel Data Acquisition and Conditional Waiting: For dependency vertices in the state that the data is ready and fresh, the data is directly acquired from the state management service; for dependency vertices in the state that the data is being computed, R registers a callback and waits for the data to be computed; for dependency vertices in the state that the data is missing or expired, R needs to trigger the corresponding engine to perform the computation. S55, Execution and Feedback Closed Loop: Once all the dependent data is ready, R executes its core recommendation algorithm to generate N candidate learning paths; R submits N candidate paths to the value verification engine (let's call them vertex V) for fast simulation verification; where V may already be in the dependency set D, but at this time V is required to verify N paths, which is a new computational task, so it can be regarded as a new vertex V', which depends on the output of R; V returns the value score V_score for each path; R selects the path with the highest V_score as the final recommendation and records the complete context of this decision. Adaptive optimization of dependency graph: The system continuously monitors the effectiveness of path recommendations in the background; by analyzing historical decision records, the system uses correlation analysis or reinforcement learning to adjust the edge weights in the dynamic dependency graph.

8. The method of using the educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 3, characterized in that, The learning path recommendation process includes the following steps: S11. Student login system: The recommendation algorithm engine needs to generate personalized learning paths for students. S12. The recommendation algorithm engine requests the student's user profile from the user profile engine. S13. The recommendation algorithm engine requests the student's current learning assessment results from the learning assessment and diagnosis engine. S14. The recommendation algorithm engine requests the knowledge structure of relevant disciplines from the knowledge graph engine; S15. The recommendation algorithm engine integrates this information, uses algorithms to generate learning paths, and returns them to the application layer. S16. At the same time, the value verification engine can evaluate the value of the recommended learning path and feed the evaluation results back to the recommendation algorithm engine to optimize the recommendation. The dynamic question generation process includes the following steps: S21. When students begin practicing, the intelligent question bank engine needs to generate a set of practice questions. S22. The intelligent question bank engine requests students' ability levels and learning styles from the user profile engine. S23. The intelligent question bank engine requests the student's knowledge gaps and recent learning difficulties from the learning assessment and diagnosis engine. S24. The intelligent question bank engine requests question templates and difficulty information for relevant knowledge points from the knowledge graph engine; S25. The intelligent question bank engine generates questions based on the above information and returns them to the application layer; S26. After students answer the questions, the learning assessment and diagnosis engine will update the learning assessment results based on the answers and update the user profile in the user profile engine. The learning assessment and update process includes the following steps: S31. After a student completes a learning task, the learning assessment and diagnostic engine will receive new learning data. S32. The learning assessment and diagnosis engine obtains students' historical profiles from the user profile engine and analyzes them in conjunction with new data. S33. The learning assessment and diagnosis engine uses the knowledge structure provided by the knowledge graph engine to analyze students' mastery of the knowledge graph. S34. The learning assessment and diagnosis engine updates the assessment results and sends the new assessment results to the user profile engine to update the profile; S35. Simultaneously, the learning evaluation and diagnosis engine sends the diagnosed knowledge gaps to the recommendation algorithm engine and the intelligent question bank engine so that they can be given priority in subsequent recommendations and question generation; The value verification process includes the following steps: S41. The value verification engine regularly collects student ability data from the user profile engine, learning outcome data from the learning assessment and diagnosis engine, and knowledge value data from the knowledge graph engine; S42. The value verification engine evaluates the learning value of students based on a pre-set verification model; S43. The value verification engine feeds back the value verification results to the student and teacher ends, and may also feed them back to the recommendation algorithm engine to adjust the recommendation strategy.

9. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 7, characterized in that, The correlation analysis algorithm includes the following steps: S61. Record historical data: For each dependency edge E(A->B), record the processing triggered by engine B after each data update of engine A, and record the following data: the degree of change in the output data of engine A, and the output effect of engine B; S62. Calculate correlation: Periodically analyze data over a period of time, and for each dependency edge, calculate the correlation between the two variables: Variable X: The degree of change in the output data of engine A; Variable Y: The output effect of engine B; Calculate the Pearson correlation coefficient ρ between X and Y; S63. Adjusting weights: If ρ > θ_high, then increase the weight: w(e) = w(e) * (1 + η), where η is the increase factor; If ρ < θ_low, then reduce the weight: w(e) = w(e) * (1 - η); The reinforcement learning method includes the following steps: S71. Initialization: For each dependent edge i, initialize the weight W_i to 1.

0. Initialize the preference functions H_i(1) and H_i(0) for the two actions to 0; S72. Repeat until the weights converge or the preset number of iterations is reached: For each dependent edge i, select the action using the softmax strategy: P(a=1) = exp(H_i(1)) / (exp(H_i(0)) + exp(H_i(1))); P(a=0) = 1 - P(a=1)b; Perform the action and update the weights: W_i = W_i + α * (a - 0.5); Where: a=1 represents an increase, a=0 represents a decrease, so (a-0.5) is 0.5 or -0.

5. After multiplying by the step size α, the weight increases or decreases by 0.

05. Observe the reward R and the effect of engine B's processing in this instance, normalizing to [0,1]. Update the average reward R_avg (used as the baseline): R_avg = R_avg + η * (R - R_avg) Update the action preference function: For the chosen action a: H_i(a) = H_i(a) + β * (R - R_avg) * (1 - P(a)) For the unselected action (1-a): H_i(1-a) = H_i(1-a) - β * (R - R_avg) * P(1-a).

10. The educational intelligent agent system based on multi-engine collaboration and dynamic knowledge evolution according to claim 7, characterized in that, It also includes an adaptive hierarchical triggering mechanism, the specific algorithm of which is as follows: A significance threshold θ is dynamically set for each student-engine combination, so that the downstream engine is only triggered to update when the amount of data change Δ exceeds θ. This can optimize the triggering frequency based on individual student differences and system load, balancing real-time performance and system overhead. S81. Define a change Δ. For each student and each engine, when the engine generates a new output for that student, calculate the difference between the new output and the old output. For example, for a learning assessment engine, the output might be a knowledge mastery vector K. The change is measured using Euclidean distance or cosine similarity. Assuming Euclidean distance is used, then Δ = ||K_new - K_old||. S82. Record historical changes. For each student-engine combination, maintain a fixed-size historical window to record the most recent m changes Δ. For example, the most recent 100 changes can be stored, denoted as the sequence D = [Δ1, Δ2, ..., Δm]. S83. Calculate the mean and standard deviation using sequence D; S84. Dynamically set the threshold θ, set θ = μ + k * σ, where k is an adjustable parameter. Intuitively, this means that the threshold is the upper bound of the historical normal range of change. If the current change exceeds this upper bound, the change is considered significant. S85. Dynamically adjust k: The parameter k can be dynamically adjusted according to the system load. For example, when the system load is high, increase k to raise the threshold and reduce the triggering frequency; when the system load is low, decrease k to lower the threshold and make the system more sensitive. S86. Update the history window: After each calculation of Δ, add it to the history window and remove the oldest record to keep the window size fixed; S87. Trigger decision: Compare the current change Δ_current with the threshold θ. If Δ_current > θ, then trigger the downstream engine; otherwise, do not trigger.

Images