Method and system for intelligent generation and optimization of whole process of vocational education based on multi-agent cooperation

Abstract

The application provides a multi-agent collaboration-based vocational education whole-process intelligent generation and optimization method and system, and particularly relates to the field of intelligent teaching technology based on artificial intelligence. The method acquires a real-time learning state of a learner and a teaching task target; based on the learning state, the target and a context memory network, a meta-coordination agent generates a coordination instruction, specifies a main interaction agent and an auxiliary interaction agent from a set of collaborative teaching agents and configures a strategy; the main interaction agent is controlled to interact with the learner and collect results; based on the results, a strategy effectiveness index is generated and the context memory network is updated; an intervention adjustment instruction is generated by the meta-coordination agent to dynamically adjust the agent role, timing or strategy, including switching the dominant right. The application solves the problem that static role configuration is difficult to dynamically contextualize interaction, realizes smooth cooperation like educational drama, and improves the immersion and long-term effect of personalized teaching.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching technology based on artificial intelligence, and more specifically, to a method and system for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration. Background Technology

[0002] In recent years, generative artificial intelligence and multi-agent system technologies have developed rapidly, injecting new momentum into the digital transformation of vocational education. By constructing intelligent agents that include functions such as instructional design, content generation, dialogue-based practice, and assessment feedback, and enabling them to work collaboratively, the aim is to achieve personalized intelligent teaching covering the entire process of "learning, practice, assessment, and diagnosis." However, in practical applications, it has been found that such systems face significant challenges in supporting the acquisition of complex vocational skills that require long-term, gradual development. Existing architectures typically assign different teaching functions to fixed agent roles and rely on pre-set, relatively static rules or processes for collaboration, making it difficult to adapt to the dynamically changing micro-contexts and learners' cognitive states during the teaching process. This results in collaboration between agents often exhibiting a mechanical, modular relay, lacking the coherence, flexibility, and deep interaction necessary for the overall teaching narrative, thus affecting the construction of immersive learning experiences and the achievement of long-term teaching goals.

[0003] To enhance the intelligence and adaptability of the teaching process, existing technical solutions primarily attempt optimization in two aspects: First, at the multi-agent collaboration level, a task allocation mechanism based on a workflow engine or central scheduler is adopted to trigger sequential responses from corresponding functional agents (such as explanation agents and exercise generation agents) based on learners' answers or pre-defined teaching paths. Second, at the personalized adaptation level, appropriate teaching content or learning paths are selected by constructing learner knowledge graphs or state models and setting decision rules based on states (such as knowledge point mastery). However, these methods are essentially still an "if-then" rule-driven or discrete state-triggered mode. Their core limitation lies in decoupling the complex, continuous teaching interaction process into a series of discrete, predefined decision points. The roles of agents and their interaction modes are usually fixed or subject to limited change in a single session or even the entire teaching cycle. This approach cannot support the delicate and dynamic reallocation of agent roles and adjustment of real-time interaction strategies based on the teaching "contextual flow" (such as the learner's real-time thought process, emotional fluctuations, and operational difficulties in a complex fault diagnosis task). As a result, the collaboration between agents appears stiff and lacks "humanity." It is difficult to simulate the flexible, appropriate, and supportive interactive atmosphere in real apprenticeship or collaborative learning, thus having inherent defects in supporting long-term, gradual skill internalization.

[0004] In summary, the urgent problem to be solved is how to address the technical challenges in multi-agent collaborative systems, where static, modular role configurations and discrete, rule-based decision-making mechanisms make it difficult to achieve nuanced, dynamic, and contextualized real-time role coordination and maintain consistency in teaching during the long-term, progressive acquisition of vocational skills that aligns with cognitive patterns. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration. This aims to at least solve the technical problem in multi-agent collaborative teaching where static role configuration and rule-based collaboration mechanisms make it difficult for agents to engage in dynamic and contextualized interactions to adapt to the long-term skill acquisition process. As a result, the collaboration between agents can be as smooth, coherent, and deeply adapted to the learner's real-time state as an "educational drama," significantly improving the immersion, adaptability, and long-term effectiveness of personalized teaching.

[0006] To achieve the above objectives, this invention provides a method and system for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration.

[0007] In a first aspect, the present invention provides a method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration, the method comprising:

[0008] To obtain learners' real-time learning status and the objectives of the current teaching task;

[0009] Based on the real-time learning state, the goal of the current teaching task, and the contextual memory network constructed for the learner, a coordination instruction is generated through a unary coordinating agent. The unary coordinating agent generates decision instructions for scheduling teaching agents based on the real-time learning state, the goal of the current teaching task, and the contextual memory network. The contextual memory network stores the learner's current learning progress, task-solving paths for acquired skills, past intervention strategies, and corresponding learning feedback. The coordination instruction dynamically specifies the current primary interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents, configures a primary interactive strategy for the primary interactive agent, and configures corresponding auxiliary interactive strategies for at least one of the auxiliary interactive agents. The set of collaborative teaching agents includes multiple teaching agents capable of independently performing specific teaching functions.

[0010] According to the coordination instructions, the current main interactive agent is controlled to conduct teaching interactions with the learner based on the main interaction strategy, and the interaction results generated by this teaching interaction are collected.

[0011] Based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy, evaluate and generate corresponding strategy effectiveness indicators;

[0012] Update the contextual memory network based on the interaction results;

[0013] Based on the updated contextual memory network and the policy effectiveness index, the meta-coordinating agent generates intervention adjustment instructions; wherein, the intervention adjustment instructions are used to indicate the adjustment of the role, intervention timing or interaction strategy of at least one of the current main interaction agent and at least one of the auxiliary interaction agents;

[0014] According to the intervention and adjustment instructions, in the subsequent teaching interaction process, dynamic adjustments are performed on the current main interactive agent and at least one of the auxiliary interactive agents; wherein, the dynamic adjustment includes switching the leadership of the teaching interaction from the current main interactive agent to a designated auxiliary interactive agent when the goal of the current teaching task is not completed.

[0015] Specifically, the generation of coordination instructions through a unary coordinating agent based on the real-time learning state, the objective of the current teaching task, and the contextual memory network constructed for the learner includes:

[0016] Based on the real-time learning status and the objectives of the current teaching task, assess the learner's cognitive load level and current skill mastery.

[0017] Based on the cognitive load level, the current skill mastery, and the learner's past successful intervention strategy sequence extracted from the contextual memory network, the current main interactive agent and at least one auxiliary interactive agent are determined from the set of collaborative teaching agents through the strategy matching module in the meta-coordinating agent, and a main interactive strategy corresponding to the current main interactive agent and an auxiliary interactive strategy corresponding to at least one of the auxiliary interactive agents are generated.

[0018] An instruction data packet containing the current main interactive agent identifier, the identifier of at least one auxiliary interactive agent, the main interactive strategy, and the auxiliary interactive strategy is encapsulated as the coordination instruction.

[0019] Specifically, the step of controlling the current main interactive agent to conduct teaching interactions with the learner based on the main interaction strategy according to the coordination instruction, and collecting the interaction results generated by this teaching interaction, includes:

[0020] Parse the coordination command to obtain the current main interactive agent identifier and the main interaction strategy;

[0021] The corresponding current main interactive agent is invoked according to the current main interactive agent identifier, and the main interactive strategy is loaded. The main interactive strategy includes a dialogue template, a question sequence, and expected skill operation steps.

[0022] The system controls the current main interactive agent, which has been loaded with the main interaction strategy, to conduct multiple rounds of dialogue or guidance operations with the learner, and records the learner's answers, operation sequences, response times, and emotional feedback signals during the interaction process. The answers, operation sequences, response times, and emotional feedback signals are then integrated into the interaction result.

[0023] Specifically, the step of evaluating and generating corresponding strategy effectiveness indicators based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy includes:

[0024] Key feedback metrics are extracted from the interaction results, including task completion accuracy, step completion time, number of points of confusion, and intensity of emotional feedback signals.

[0025] The key feedback indicators are compared with the expected indicators preset in the main interaction strategy and the pre-triggered conditions preset in the auxiliary interaction strategy to calculate the immediate effect score of the main interaction strategy and the pre-association score of the auxiliary interaction strategy.

[0026] Based on the immediate effect score and the preliminary association score, a weighted calculation is performed to generate a main strategy effectiveness index corresponding to the main interaction strategy and an auxiliary strategy effectiveness index corresponding to the auxiliary interaction strategy. The main strategy effectiveness index and the auxiliary strategy effectiveness index together constitute the strategy effectiveness index.

[0027] Specifically, updating the contextual memory network based on the interaction result includes:

[0028] The interaction results are processed in a structured manner to generate a structured interaction record containing a timestamp, the associated goal of the current teaching task, the learner's behavior sequence, and key feedback indicators of the interaction results;

[0029] The structured interaction records are associated with the past interaction records stored in the contextual memory network to generate or update learning path nodes that describe the evolution of the learner's skill mastery status under the goal of the current teaching task.

[0030] The learning path nodes, the structured interaction records, and the main interaction strategy and the auxiliary interaction strategy that triggered the current interaction result are stored in the contextual memory network as new historical intervention strategies and corresponding learning feedback.

[0031] Specifically, the step of generating intervention adjustment instructions through the meta-coordinating agent based on the updated contextual memory network and the policy effectiveness index includes:

[0032] The updated contextual memory network is used to query the learner’s successful intervention strategy sequence in similar historical tasks when the main strategy effectiveness index is lower than the first threshold, as a historical successful adjustment reference.

[0033] Based on the historical successful adjustment reference, the current strategy effectiveness index, and the remaining sub-goals of the current teaching task, the adjustment decision module in the meta-coordinating agent determines whether it is necessary to initiate a switch of leadership, adjust the interaction strategy, or adjust the intervention timing of the auxiliary interaction agent.

[0034] When it is determined that adjustment is needed, the intervention adjustment instruction is generated, which includes the target agent identifier, role change indication, policy replacement content, or intervention timing correction parameters.

[0035] Specifically, the step of dynamically adjusting the current main interactive agent and at least one of the auxiliary interactive agents during subsequent teaching interactions, based on the intervention adjustment instruction, includes:

[0036] The intervention and adjustment instructions are parsed to identify the agent identifier to be adjusted and the type of adjustment operation.

[0037] If the adjustment operation type is a role change and involves a switch of control, a dormant command is sent to the current main interactive agent, and one of the auxiliary interactive agents specified in the intervention adjustment command is activated. The control of the teaching interaction, the unfinished sub-tasks and the context are synchronized to the activated auxiliary interactive agent.

[0038] The auxiliary interactive agent, having gained control, operates based on a synchronized context and loads the new interaction strategy specified in the intervention adjustment instructions, continuing subsequent teaching interactions with the learner toward the goal of the current teaching task.

[0039] Secondly, the present invention provides an intelligent generation and optimization system for the entire vocational education process based on multi-agent collaboration. The generation and optimization system applies the generation and optimization method described in the first aspect, and the generation and optimization system includes:

[0040] The status acquisition module is used to acquire the learner's real-time learning status and the objectives of the current teaching task;

[0041] An instruction generation module, connected to the state acquisition module, includes a meta-coordinating agent. The instruction generation module generates coordination instructions based on the real-time learning state, the goal of the current teaching task, and a contextual memory network constructed for the learner. The meta-coordinating agent generates decision instructions for scheduling teaching agents based on the real-time learning state, the goal of the current teaching task, and the contextual memory network. The contextual memory network stores the learner's current learning progress, task-solving paths for acquired skills, past intervention strategies, and corresponding learning feedback. The coordination instructions dynamically specify the current main interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents, configuring a main interaction strategy for the current main interactive agent and corresponding auxiliary interaction strategies for at least one of the auxiliary interactive agents. The set of collaborative teaching agents includes multiple teaching agents capable of independently performing specific teaching functions.

[0042] An interactive execution module, connected to the instruction generation module, is used to control the current main interactive agent to conduct teaching interaction with the learner based on the main interactive strategy according to the coordination instruction, and to collect the interaction results generated by this teaching interaction.

[0043] An evaluation module, connected to the interaction execution module, is used to evaluate and generate corresponding strategy effectiveness indicators based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy.

[0044] A memory update module is connected to the interaction execution module, and the memory update module is used to update the context memory network according to the interaction result;

[0045] An adjustment instruction generation module is connected to the evaluation module and the memory update module. The adjustment instruction generation module is used to generate intervention adjustment instructions through the meta-coordinating agent based on the updated context memory network and the policy effectiveness index. The intervention adjustment instructions are used to indicate the adjustment of the role, intervention timing or interaction strategy of at least one of the current main interaction agent and at least one of the auxiliary interaction agents.

[0046] A dynamic adjustment execution module is connected to the adjustment instruction generation module. The dynamic adjustment execution module is used to perform dynamic adjustments on the current main interactive agent and at least one of the auxiliary interactive agents in the subsequent teaching interaction process according to the intervention adjustment instruction. The dynamic adjustment includes switching the leadership of the teaching interaction from the current main interactive agent to a designated auxiliary interactive agent when the goal of the current teaching task has not been completed.

[0047] Specifically, the instruction generation module includes:

[0048] An assessment unit is used to assess the learner's cognitive load level and current skill mastery based on the real-time learning status and the objectives of the current teaching task.

[0049] A strategy matching unit, connected to the evaluation unit, is used to determine the current main interactive agent and at least one auxiliary interactive agent from the set of collaborative teaching agents based on the cognitive load level, the current skill mastery, and the learner's past successful intervention strategy sequences extracted from the contextual memory network, through the strategy matching module in the meta-coordinating agent, and to generate a main interaction strategy corresponding to the current main interactive agent and an auxiliary interaction strategy corresponding to at least one of the auxiliary interactive agents;

[0050] An encapsulation unit, connected to the policy matching unit, is used to encapsulate an instruction data packet containing the current main interactive agent identifier, the identifier of at least one auxiliary interactive agent, the main interactive policy, and the auxiliary interactive policy, as the coordination instruction.

[0051] Specifically, the interactive execution module includes:

[0052] The parsing unit is used to parse the coordination instruction and obtain the current main interactive agent identifier and the main interaction strategy;

[0053] A loading unit, connected to the parsing unit, is used to call the corresponding current main interactive agent according to the current main interactive agent identifier and load the main interactive strategy, which includes a dialogue template, a question sequence and expected skill operation steps.

[0054] An interaction control unit, connected to the loading unit, is used to control the current main interaction agent loaded with the main interaction strategy to conduct multi-round dialogues or guided operations with the learner, and to record the learner's answers, operation sequences, response times and emotional feedback signals during the interaction process, and to integrate the answers, operation sequences, response times and emotional feedback signals into the interaction result.

[0055] This application provides a method and system for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration. The method acquires the learner's real-time learning status and current teaching task objectives; based on the real-time learning status, current teaching task objectives, and contextual memory network, it generates coordination instructions through a meta-coordinating agent, dynamically designating the current main interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents and configuring interaction strategies; it controls the teaching interaction between the current main interactive agent and the learner and collects the interaction results; it evaluates the effectiveness index of the generated strategies based on the interaction results; it updates the contextual memory network according to the interaction results; and based on the updated contextual memory network and strategy effectiveness index, it generates intervention and adjustment instructions through a meta-coordinating agent, dynamically adjusting the agent's role, intervention timing, or interaction strategy, including switching the leadership from the current main interactive agent to the auxiliary interactive agent. This invention solves the problem of difficulty in dynamic contextualized interaction due to static role configuration in multi-agent collaborative teaching, achieving smooth and coherent agent collaboration like an educational drama, and improving the immersion, adaptability, and long-term effectiveness of personalized teaching. Attached Figure Description

[0056] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0057] Figure 1 A flowchart illustrating the intelligent generation and optimization method for the entire vocational education process based on multi-agent collaboration provided in this application;

[0058] Figure 2 A connection diagram for the intelligent generation and optimization system for the entire vocational education process based on multi-agent collaboration provided in this application.

[0059] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] The terms "first," "second," "third," "fourth," etc. (if present) in the specification and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein.

[0062] In this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0063] This application provides a method and system for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration. The method acquires the learner's real-time learning status and current teaching task objectives; based on the real-time learning status, current teaching task objectives, and contextual memory network, it generates coordination instructions through a meta-coordinating agent, dynamically designates the current main interactive agent and auxiliary interactive agents from a set of collaborative teaching agents, and configures interaction strategies; it controls the teaching interaction between the current main interactive agent and the learner, collects interaction results, and generates strategy effectiveness indicators; it updates the contextual memory network based on the interaction results, and generates intervention and adjustment instructions through a meta-coordinating agent to dynamically adjust agent roles, intervention timing, or interaction strategies, including switching the leadership from the current main interactive agent to the auxiliary interactive agent, achieving seamless agent collaboration like an educational drama.

[0064] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0065] Figure 1 A flowchart illustrating the intelligent generation and optimization method for the entire vocational education process based on multi-agent collaboration provided in this application is shown below. Figure 1 As shown in this embodiment, the method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration includes:

[0066] S101: Obtain the learner's real-time learning status and the goals of the current teaching task.

[0067] The specific steps of implementation S101 include:

[0068] S1011: Collect heterogeneous raw data from multiple sources from learners.

[0069] A front-facing camera deployed on the learner's interactive terminal continuously captures a sequence of red, green, and blue (RGB) images of the learner's face at a sampling rate of 30 frames per second. A microphone integrated into the interactive terminal captures a single-channel audio stream of the learner's speech during interaction at a sampling rate of 16 kHz. Through the human-computer interface of the interactive terminal, operation event logs are recorded regarding the learner's actions on buttons, menus, and 3D models in the graphical interface of the teaching software. These logs include the operation type, object identifier, timestamp, and coordinates. A smart bracelet worn by the learner collects photoplethysmography (PPG) and electrodermal activity (EDA) signals from the learner via Bluetooth 5.0 at a frequency of once per second. The image sequence, audio stream, operation event logs, PPG signals, and EDA signals are collectively tagged as the current session identifier and cached as raw data packets.

[0070] S1012: Process raw data and extract real-time learning state features.

[0071] For the image sequence, a convolutional neural network model from the OpenFace toolkit is used to detect and identify facial feature points of the learner frame by frame, calculate the motion amplitude of feature points between consecutive frames, and output a sequence of facial expression intensity values ​​including pleasure, focus, and confusion, as well as an eye-tracking fixation point coordinate sequence. For the audio stream, the librosa library is used to extract Mel-frequency cepstral coefficients, fundamental frequency, harmonic-to-noise ratio, and loudness features of each 250-millisecond audio frame, which are then input into a pre-trained support vector machine classifier. The support vector machine classifier outputs a sequence of voice emotion labels including positive, neutral, negative, and confused. For the operation event log, the success rate, average response time, and erroneous operation type distribution of effective operations within a preset time window are statistically analyzed. For the photoplethysmography (PPG) signal, after denoising using a moving average filter, the high-frequency power values ​​of heart rate and heart rate variability are calculated. For the electrodermal activity (EDA) signal, the number of times the signal amplitude rises per unit time is extracted as the EDA response frequency. The mean of the facial expression intensity value sequence, the mode of the voice emotion label sequence, the operation success rate, the average response time, the error operation type distribution vector, the heart rate, the high-frequency power value of the heart rate variability, and the skin conductance response frequency are concatenated into a multi-dimensional real number vector, which serves as the learner's real-time learning state vector S_t, where t represents the current timestamp.

[0072] S1013: Obtain structured teaching task data from the teaching management system.

[0073] The task scheduling service of the teaching management system is invoked via an application programming interface (API), passing in the learner's unique identifier. The system receives return data from the teaching management system, which is in JavaScript object representation format and includes the following fields: task_id (unique identifier for the current teaching task), task_title (task title), core_skills (a list of core skill points the task points to, each skill point including a skill identifier and target proficiency level), evaluation_criteria (a list of evaluation criteria for task completion, each criterion including an indicator name, measurement method, and pass threshold), and prerequisite_tasks (a list of prerequisite task identifiers). The returned data is parsed and encapsulated into a target object, Goal_current, for the current teaching task.

[0074] S1014: Construct or load the learner's contextual memory network.

[0075] Using the learner's unique identifier as the key, a graph database is queried. If the query result shows a subgraph with the learner's identifier as the root node, the subgraph structure and all attributes are loaded from the graph database to form the learner's contextual memory network G_learner. The contextual memory network G_learner is a directed attribute graph, whose node types include learner nodes, historical teaching task nodes, skill nodes, and historical intervention strategy nodes, and edge types include "mastery," "involvement," "application," and "leading to feedback." If the query result does not show a node, an empty graph containing only one learner node (with the learner's identifier as its attribute) is initialized as the contextual memory network G_learner.

[0076] S1015: Output the result obtained in step S101.

[0077] The real-time learning state vector S_t obtained in step S1012, the target object Goal_current of the current teaching task obtained in step S1013, and the contextual memory network G_learner obtained or loaded in step S1014 are used as the final output of step S101 and provided for use in the subsequent step S102.

[0078] Step S101 integrates various specific technologies such as computer vision, speech signal processing, human-computer interaction log analysis, and physiological signal sensing to objectively and continuously collect and quantify learners' external performance and internal state from multiple dimensions including vision, hearing, behavior, and physiology, forming a structured real-time learning state vector. Simultaneously, through standardized interface interaction with external teaching management systems, it obtains accurate and structured descriptions of teaching objectives. Furthermore, it establishes or associates a unique contextual memory network for each learner, storing their personal learning history and strategy feedback in a graph structure for long-term storage. These three parts (state, objective, and memory) together constitute the accurate, comprehensive, and personalized data foundation for all subsequent intelligent decisions and personalized interventions, ensuring the system's deep perception of learners' instantaneous states and long-term processes, which is a prerequisite for achieving dynamic adaptation and optimization throughout the entire process.

[0079] S102: Based on the real-time learning state, the goal of the current teaching task, and the contextual memory network constructed for the learner, a coordination instruction is generated through a unary coordinating agent.

[0080] The meta-coordinating agent is used to generate decision instructions for scheduling teaching agents based on the real-time learning state, the goal of the current teaching task, and the contextual memory network. The contextual memory network is used to store the learner's current learning progress, task-solving paths of acquired skills, past intervention strategies, and corresponding learning feedback. The coordination instructions are used to dynamically specify the current main interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents, and configure a main interactive strategy for the current main interactive agent and a corresponding auxiliary interactive strategy for at least one of the auxiliary interactive agents. The set of collaborative teaching agents includes multiple teaching agents capable of independently performing specific teaching functions.

[0081] Specifically, the step of generating coordination instructions through a unary coordinating agent based on the real-time learning state, the goal of the current teaching task, and the contextual memory network constructed for the learner includes: assessing the learner's cognitive load level and current skill mastery based on the real-time learning state and the goal of the current teaching task; determining the current main interaction agent and at least one auxiliary interaction agent from the set of collaborative teaching agents based on the cognitive load level, the current skill mastery, and the learner's past successful intervention strategy sequence extracted from the contextual memory network, and generating a main interaction strategy corresponding to the current main interaction agent and an auxiliary interaction strategy corresponding to at least one of the auxiliary interaction agents through the strategy matching module in the unary coordinating agent; and encapsulating an instruction data packet containing the identifier of the current main interaction agent, the identifier of the at least one auxiliary interaction agent, the main interaction strategy, and the auxiliary interaction strategy as the coordination instructions.

[0082] The specific steps of implementation S102 include:

[0083] S1021: Assess cognitive load levels and current skill mastery.

[0084] Obtain the real-time learning state vector S_t and the target object Goal_current of the current teaching task from step S101. Extract four components from the real-time learning state vector S_t: facial expression attention value f_att, average response time t_resp, heart rate hr, and heart rate variability high-frequency power hrv_hf. Using the calculation model of the NASA-TLX scale, normalize these four components and then sum them by weight to calculate the cognitive load level CL. The specific calculation formula is as follows:

[0085] CL = w1 * f_att_norm + w2 * t_resp_norm + w3 * hr_norm + w4 * (1 -hrv_hf_norm), where w1, w2, w3, and w4 are preset weight coefficients and their sum is 1, and *_norm represents the minimum-maximum normalized value of the corresponding component. Extract the list of all skill identifiers involved in this task, skill_list, from the core_skills field of the target object Goal_current of the current teaching task. Traverse skill_list, and for each skill identifier skill_id, query the contextual memory network G_learner for an edge originating from the learner node, of type "mastery", pointing to the skill node with skill identifier skill_id. If such an edge exists, obtain the attribute "mastery score" mastery_score of that edge; otherwise, set mastery_score to the initial value 0. Calculate the arithmetic mean of the mastery_scores corresponding to all skill_ids, and use this as the learner's current skill mastery MS_current.

[0086] S1022: Extract sequences of past successful intervention strategies from the episodic memory network.

[0087] In the contextual memory network G_learner, starting from the learner node, all historical intervention strategy nodes are searched by traversing backwards along the "cause feedback" edge. For each found historical intervention strategy node, the "feedback effect value" (feedback_score) attribute of its connected "cause feedback" edge is obtained. All historical intervention strategy nodes with a feedback_score greater than a preset success threshold (e.g., 0.7) are selected to form a set of successful intervention strategy nodes. Based on the timestamp attribute of these nodes, they are sorted in reverse chronological order, and the N most recent nodes (e.g., 5) are selected. The strategy content fields stored in each node (including historical main interaction agent identifier, historical auxiliary interaction agent identifier, historical main interaction strategy, and historical auxiliary interaction strategy) are extracted to form a list arranged by time, which serves as the past successful intervention strategy sequence Seq_success.

[0088] S1023: Determine the agent and configure the policy through the policy matching module.

[0089] The cognitive load level CL and current skill mastery MS_current calculated in step S1021, the previously successful intervention strategy sequence Seq_success extracted in step S1022, and the target object Goal_current of the current teaching task are jointly input into the pre-set strategy matching module in the meta-coordinating agent. The strategy matching module consists of a similarity matching model based on knowledge graph embedding and a set of predefined rule engines. First, the similarity matching model (such as the TransE model) encodes the current state (CL, MS_current) and task Goal_current into a query vector, and retrieves the k most similar tuples from a pre-built strategy knowledge graph containing a large number of <state, task, successful strategy> triples. The "successful strategy" part of the retrieved tuples is fused and deduplicated with the strategy in Seq_success to generate a set of candidate agent combinations and strategy configurations. Subsequently, the rule engine filters candidate configurations based on the value of CL: if CL is higher than the load threshold, strategies with a relaxed interaction pace and concise information presentation are prioritized; if MS_current is lower than the mastery threshold, agent combinations that include more explanations of basic concepts and step-by-step demonstrations are prioritized. Finally, the policy matching module outputs a definite current main interactive agent identifier ID_main, a list of auxiliary interactive agent identifiers ID_list_assist (containing at least one identifier), a specific main interactive policy object Policy_main corresponding to ID_main, and a specific list of auxiliary interactive policy objects Policy_list_assist corresponding to each identifier in ID_list_assist. The main interactive policy object Policy_main is a data structure that explicitly includes: a list of dialogue templates filled with specific knowledge points for guiding the dialogue; a series of specific questions and their standard answers for evaluating comprehension; and a decomposed, expected sequence of standard operation steps for the operation steps specified in evaluation_criteria in Goal_current. Each auxiliary interaction policy object, Policy_assist_x, is also a well-defined data structure containing: the preset trigger conditions for the activation of the auxiliary agent (e.g., when the learner answers incorrectly twice in a row when the main agent asks a question, or when the learner times out in a certain step); and the specific intervention action that the auxiliary agent should perform after activation (e.g., play a 3D animation demonstration targeting the current point of confusion, or provide a simpler guiding question).

[0090] S1024: Encapsulation coordination instructions.

[0091] Create a JavaScript object representation as the coordination instruction data package Cmd_coord. In Cmd_coord, create and assign the following fields: `main_agent_id` field, assigned the value of `ID_main` obtained in step S1023; `assist_agent_ids` field, assigned the value of `ID_list_assist` obtained in step S1023; `main_policy` field, assigned the value of the `Policy_main` object obtained in step S1023; `assist_policies` field, assigned a dictionary whose keys are the identifiers of each assistant agent in `ID_list_assist`, and whose values ​​are the corresponding `Policy_assist_x` objects in `Policy_list_assist`. Use the encapsulated Cmd_coord as the final output of step S102, i.e., the coordination instructions, for use in subsequent step S103.

[0092] Step S102 is a crucial step in the initial core decision-making of the meta-coordinating agent. It transforms the learner's real-time state into calculable cognitive load and skill mastery indicators through specific, quantifiable models (NASA-TLX derivative model, Sigmoid mastery model); it accurately extracts historically effective personalized intervention strategy patterns from the learner's contextual memory network via graph queries; and then, by integrating the current state, historical successes, and task objectives, it dynamically and accurately assigns appropriate "actors" (main / auxiliary agents) and "scripts" (specific interaction strategies) to the current teaching moment through a specific strategy matching module that combines knowledge graph-based similarity matching with rule-based logical filtering. This process completely abandons static, pre-set role allocation rules, achieving a high degree of contextualization and personalization in scheduling decisions. It provides a crucial initial arrangement for a smooth and adaptable "educational drama" to unfold subsequently, and is the core embodiment of solving the technical problem of dynamic, contextualized collaboration between agents.

[0093] S103: According to the coordination instruction, control the current main interactive agent to conduct teaching interaction with the learner based on the main interaction strategy, and collect the interaction results generated by this teaching interaction.

[0094] Specifically, the step of controlling the current main interactive agent to conduct teaching interactions with the learner based on the main interaction strategy according to the coordination instruction, and collecting the interaction results generated by this teaching interaction, includes: parsing the coordination instruction to obtain the identifier of the current main interactive agent and the main interaction strategy; calling the corresponding current main interactive agent according to the identifier of the current main interactive agent and loading the main interaction strategy, wherein the main interaction strategy includes a dialogue template, a question sequence and expected skill operation steps; controlling the current main interactive agent loaded with the main interaction strategy to conduct multiple rounds of dialogue or guidance operations with the learner, and recording the learner's answer content, operation sequence, response time and emotional feedback signal during the interaction process, and integrating the answer content, operation sequence, response time and emotional feedback signal into the interaction result.

[0095] The specific steps in step S103 during implementation include:

[0096] S1031: Parse the coordination instructions and invoke the main interactive agent.

[0097] The system receives the coordination instruction data packet Cmd_coord output from step S102. Using a JavaScript object representation parser, it parses Cmd_coord, extracting the value of the main_agent_id field as the current main interactive agent identifier ID_main; and extracting the value of the main_policy field as the main interactive policy object Policy_main. Based on ID_main, it queries the agent registry to obtain the corresponding agent deployment description information, which includes the agent type, server endpoint Uniform Resource Locator (URI), and API specification. Based on the deployment description information, it sends an activation request to the service container deploying the current main interactive agent via remote procedure call or message queue. The service container instantiates or wakes up the corresponding agent instance in memory and returns a usable session handle Handle_main. It extracts dialogue_templates (dialogue template list), question_sequence (question sequence), and expected_operation_steps (expected skill operation step sequence) from the Policy_main object and loads this data into the working memory of the current main interactive agent through the session handle Handle_main.

[0098] S1032: Execute instructional interactions based on the main interaction strategy.

[0099] The current main interactive agent, controlled by Policy_main, begins interacting with the learner via a graphical user interface, voice, or a combination of both. First, the agent selects a template from dialogue_templates that matches the current teaching task goal_current, such as "Welcome to [task title], we will learn [core skill points]. First, please observe this demonstration." It then replaces the placeholders in the template with the specific content from Goal_current, generating an opening text. This text is then converted to speech by a text-to-speech engine and played, while simultaneously displayed on the graphical interface. Next, the agent presents questions one by one in the order defined in question_sequence. For each question, the time interval between the learner's completion of the question presentation and the start of their response (detection of voice activity or clicking the "Submit" button) is recorded as the response time RT_i for that question. The learner's voice response is converted to text using a speech recognition engine (such as using a deep learning-based automatic speech recognition model, such as Wav2Vec 2.0 or the Conformer model), and this is recorded as the learner's response text AnswerText_i. Simultaneously, the learner's speech or text responses are analyzed using a sentiment analysis model (such as a pre-trained model wav2vec 2.0 fine-tuned on a speech sentiment dataset, or a text sentiment analysis model such as BERT), outputting a sentiment vector EmotionVec_i containing probability values ​​for dimensions such as positive, negative, neutral, and confused. During guided virtual operations, the current main interactive agent gradually highlights the tools or components to be operated on in the 3D virtual training environment according to the expected_operation_steps sequence. By monitoring the virtual environment's event system, each step of the learner's operation is recorded, including the object identifier ObjectID_j, the operation type ActionType_j (e.g., click, drag, assemble), and the operation timestamp Timestamp_j. This data is compared with the corresponding expected steps in expected_operation_steps to determine if the operation is correct, generating an operation correctness flag CorrectFlag_j.

[0100] S1033: Collect and record raw data of multimodal interaction.

[0101] Throughout the interaction in step S1032, data acquisition is performed in parallel. Using the same interface as in step S101, and with a synchronized timestamp base, the learner's facial image sequences, audio streams, human-computer interaction event logs, photoplethysmography (PPG) signals, and electrodermal activity signals are continuously acquired. However, in this step, each piece of acquired data is appended with a correlation identifier to indicate its associated teaching stage (e.g., "during answering question 1," "during performing step 3"). All raw data with correlation identifiers is temporarily stored as the raw data cache for this interaction.

[0102] S1034: Integrate and generate interactive results.

[0103] The teaching interaction ends when the preset number of interaction rounds is completed, the learner actively ends the interaction, or the termination condition defined in Policy_main is triggered. All recorded data is integrated to construct an interaction result object, Result_interaction. The Result_interaction object contains the following fields:

[0104] session_id: A unique identifier for this interactive session.

[0105] main_agent_id: ID_main, parsed from Cmd_coord.

[0106] interaction_type: The type of interaction, such as "question and answer", "virtual operation" or "hybrid".

[0107] qa_records: A list that records the details of each question-and-answer interaction. Each element in the list is an object containing question_text, answer_text (AnswerText_i), response_time (RT_i), and emotion_vector (EmotionVec_i).

[0108] operation_records: A list that records details of virtual operations. Each element in the list is an object containing step_index (step number), object_id (object ID), action_type (actionType_j), timestamp (timestamp_j), and correct_flag (correctness flag).

[0109] raw_data_refs: A set of references pointing to the locations of the multimodal raw data caches with associated identifiers temporarily stored in step S1033.

[0110] overall_metrics: Summary metrics calculated based on the above records, including: task completion accuracy (calculated as the number of correct_flag values ​​in operation_records divided by the total number of operation steps), average step completion time (avg_step_time), confusion_count (counting the number of times the probability of the confusion dimension in the emotion_vector exceeds the threshold, or the number of times the preset confusion keyword such as "I don't understand" is matched in the answer_text), and average emotion feedback signal intensity (avg_emotion_intensity, calculated as the norm average of all EmotionVec_i vectors).

[0111] Step S103 is a crucial execution stage that transforms the intelligent scheduling decision of step S102 into concrete, perceptible teaching behaviors and collects feedback. Through specific parsing, invocation, and loading processes, it instantiates the abstract "main interactive agent" and "main interactive strategy" in the coordination instructions into actually running services and executable teaching scripts. By integrating various specific technologies such as speech recognition, sentiment analysis, and virtual environment event monitoring, it achieves concrete support and refined process recording for the two core teaching interaction forms: "multi-turn dialogue" and "guided operation." By synchronously collecting multimodal raw data and associating it with the teaching context, it provides rich, context-sensitive underlying signals for subsequent in-depth analysis. Finally, by structurally integrating process records and summary indicators, it generates an "interaction result" object that comprehensively represents the interaction process and its effects. This step ensures the accurate implementation of teaching intentions and the complete acquisition of interaction process data. It serves as a data bridge connecting intelligent decision-making with subsequent evaluation and optimization stages, providing indispensable factual evidence for the system to achieve evidence-based dynamic adjustments.

[0112] S104: Based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy, evaluate and generate corresponding strategy effectiveness indicators.

[0113] Specifically, the step of evaluating and generating corresponding strategy effectiveness indicators based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy includes: extracting key feedback indicators from the interaction results, including task completion accuracy, step completion time, number of confusion points, and emotional feedback signal strength; comparing the key feedback indicators with the preset expected indicators in the main interaction strategy and the preset pending trigger conditions in the auxiliary interaction strategy, and calculating the immediate effect score of the main interaction strategy and the preliminary association score of the auxiliary interaction strategy; and weightedly calculating and generating the main strategy effectiveness indicator corresponding to the main interaction strategy and the auxiliary strategy effectiveness indicator corresponding to the auxiliary interaction strategy based on the immediate effect score and the preliminary association score, wherein the main strategy effectiveness indicator and the auxiliary strategy effectiveness indicator together constitute the strategy effectiveness indicator.

[0114] The specific steps in step S104 during implementation include:

[0115] S1041: Extract key feedback metrics from interaction results.

[0116] Receive the interaction result object Result_interaction from the output of step S103. Extract the pre-calculated summary metrics directly from the overall_metrics field of Result_interaction as four key feedback metrics: task completion accuracy A_act (i.e., accuracy), average step completion time T_act (i.e., avg_step_time), number of confusion points C_act (i.e., confusion_count), and average emotion feedback signal intensity E_act (i.e., avg_emotion_intensity).

[0117] S1042: Compare with the preset expected indicators of the main interaction strategy to calculate the immediate effect score.

[0118] Extract the preset expected metrics from the currently executing main interaction policy object, Policy_main. These expected metrics include: target task completion accuracy A_exp and target step completion time T_exp. Calculate the immediate effect score Score_imm of the main interaction policy using the following formula:

[0119] Score_imm = w_a * (A_act / A_exp) + w_t * (T_exp / T_act) - w_c * (C_act / C_norm) - w_e * (E_act / E_norm);

[0120] in:

[0121] A_act, T_act, C_act, and E_act are the key feedback indicators extracted in step S1041.

[0122] A_exp and T_exp are the target accuracy and target time preset in the main interaction policy Policy_main.

[0123] C_norm and E_norm are preset normalization benchmarks for the number of confusion points and the emotional intensity, used to make C_act and E_act dimensionless.

[0124] w_a, w_t, w_c, and w_e are preset weight coefficients, and w_a + w_t + w_c + w_e = 1. w_a and w_t are positive values, rewarding high accuracy and short execution time; w_c and w_e are positive values, used to penalize more confusion points and stronger negative sentiment. Theoretically, Score_imm is a real number, but in practice, it is constrained to a reasonable range (e.g., [-1, 1]) through parameter settings. A higher value indicates better immediate execution performance of the main policy.

[0125] S1043: Compare the pre-triggered conditions with the assisted interaction strategy and calculate the preliminary association score.

[0126] Iterate through each Policy_assist_x in the currently configured list of auxiliary interaction policy objects, Policy_list_assist. Extract the preset trigger conditions from each Policy_assist_x, such as "when the number of consecutive incorrect answers >= 2" or "when the single-step operation time > 60 seconds". Check the detailed records (qa_records and operation_records) of the interaction result Result_interaction generated in step S103 to determine whether the trigger conditions of Policy_assist_x were met (i.e., "triggered") during this interaction. For each Policy_assist_x, calculate its pre-association score Score_prep_x.

[0127] Score_prep_x = 1 / (1 + exp(-k * (N_trigger_x - θ)));

[0128] in:

[0129] N_trigger_x represents the number of times the trigger condition of Policy_assist_x is actually met in this interaction.

[0130] θ is a preset threshold (e.g., set to 1) that represents the baseline for "being triggered at least once".

[0131] k is a scaling parameter that controls the steepness of the sigmoid function.

[0132] exp() is the natural exponential function.

[0133] This sigmoid function maps the number of triggers to a score between 0 and 1. The larger N_trigger_x is, the closer Score_prep_x is to 1, indicating a stronger correlation (preparatoryness) between the assistive policy and the current learning impasse. If the conditions for Policy_assist_x to be triggered are not met even once in this interaction, then N_trigger_x = 0, and Score_prep_x will be a decimal close to 0.

[0134] S1044: Weighted calculation generates the effectiveness index of the main strategy and the effectiveness index of the auxiliary strategy.

[0135] The immediate effect score Score_imm of the main interaction strategy calculated in step S1042 is normalized to the [0, 1] interval by linear scaling to obtain the main strategy effectiveness index E_main. The specific formula is: E_main = (Score_imm - Score_min) / (Score_max - Score_min), where Score_max and Score_min are the theoretically possible maximum and minimum values ​​of Score_imm, calculated according to the weight and parameter settings.

[0136] For each auxiliary interaction strategy calculated in step S1043, the preliminary correlation score Score_prep_x is directly used as the auxiliary strategy effectiveness index E_assist_x of the corresponding auxiliary interaction strategy. Since Score_prep_x itself is in the interval [0,1], it represents the strength of its correlation.

[0137] Add E_main and all E_assist_x (assuming there are m auxiliary policies) to a list to form the policy effectiveness index E_all, i.e., E_all = [E_main, E_assist_1, ..., E_assist_m]. E_all is the final output of step S104.

[0138] Step S104 is the core of the system's quantitative self-assessment and reflection. It uses specific, calculable mathematical formulas (such as weighted scoring and the Sigmoid function) to finely compare and map the multi-dimensional, multi-granular results (overall indicators, process records) generated from a single teaching interaction with pre-set teaching intentions (the expected goals of the main strategy and the triggering conditions of auxiliary strategies). The calculated main strategy effectiveness index E_main objectively reflects the degree to which the current dominant teaching strategy achieves its preset goals; while the auxiliary strategy effectiveness index E_assist_x quantifies the correlation and potential necessity between each prepared intervention plan and the actual learning situation. This set of indicators transforms the complex quality of the teaching process into a set of comparable and transferable scalar values, providing crucial and objective data-driven basis for the meta-coordinating agent in the subsequent step (S106) to determine whether and how to adjust. It is a vital decision-making support link connecting the preceding (assessment of effectiveness) and the following (guidance of adjustment) stages in the dynamic optimization cycle of the entire process.

[0139] S105: Update the contextual memory network based on the interaction results.

[0140] Specifically, updating the contextual memory network based on the interaction results includes: performing structured processing on the interaction results to generate a structured interaction record containing a timestamp, the associated goal of the current teaching task, the learner's behavior sequence, and key feedback indicators of the interaction results; associating the structured interaction record with past interaction records stored in the contextual memory network to generate or update learning path nodes describing the evolution of the learner's skill mastery status under the goal of the current teaching task; and storing the learning path nodes, the structured interaction record, and the main interaction strategy and the auxiliary interaction strategy that triggered the current interaction result as new historical intervention strategies and corresponding learning feedback in the contextual memory network.

[0141] The specific steps in step S105 during implementation include:

[0142] S1051: Perform structured processing on the interaction results to generate structured interaction records.

[0143] Receive the interaction result object Result_interaction from step S103, the coordination instruction data packet Cmd_coord (containing main_policy and assistant_policies) from step S102, and the policy effectiveness index E_all from step S104. Create a new structured interaction record object Record_struct. Populate Record_struct with the following fields:

[0144] record_id: Generates a unique identifier based on timestamp and session identifier.

[0145] timestamp: Taken from the session start time of Result_interaction.

[0146] associated_goal: Associates the goal of the current teaching task, storing the task_id and task_title of Goal_current.

[0147] learner_behavior_sequence: Integrates the detailed sequence of qa_records and operation_records in Result_interaction to form a complete behavioral chain that is sorted by time and contains the "question-answer-operation" event.

[0148] key_feedback_metrics: Stores four key feedback metrics extracted from overall_metrics of Result_interaction: A_act, T_act, C_act, and E_act.

[0149] effectiveness_vector: Stores the main policy effectiveness index E_main and all auxiliary policy effectiveness indices E_assist_x extracted from E_all.

[0150] raw_data_reference: Points to the storage path of the multimodal raw data cache referenced by the raw_data_refs field in Result_interaction.

[0151] S1052: Associate structured interactive records with past records to generate or update learning path nodes.

[0152] Retrieve the current task identifier (task_id) from the associated_goal field. In the learner's contextual memory network G_learner, search for all nodes of type "historical teaching task node" whose task_id attribute is equal to the current task_id. If found, record it as the target task node Node_task; otherwise, create a new "historical teaching task node" as Node_task in G_learner and set its task_id, task_title, and other attributes.

[0153] In G_learner, create a new node of type "historical interaction record node" and store the entire contents of the Record_struct object generated in step S1051 as the attribute of this node. This node is denoted as Node_record_new.

[0154] Create a directed edge between Node_record_new and Node_task, with the edge type "belonging to", indicating that this record is about this task.

[0155] Next, association and path generation are performed: In G_learner, all other "historical interaction record nodes" connected to Node_task are found. A graph neural network (e.g., a graph attention network GAT implemented using the PyTorch Geometric library) is used to learn node embeddings for all "historical interaction record nodes" (including the new Node_record_new) under Node_task. Based on these node embeddings, the cosine similarity between Node_record_new and each historical record node is calculated. The K historical record nodes with the highest cosine similarity (e.g., K=3) are selected. For each selected historical record node Node_record_hist, a directed edge is created between Node_record_new and Node_record_hist, with the edge type being "follow-up" and the edge weight set to the calculated cosine similarity value. This process establishes an explicit association between the current interaction record and similar past records in G_learner.

[0156] Based on the key_feedback_metrics (especially A_act) and effectiveness_vector (E_main) in the Record_struct, and combined with the core_skills of the Goal_current task, the learner's mastery score for the relevant skills in this task is calculated or updated. In G_learner, a "skill node" corresponding to each skill identifier in core_skills is found or created. Between Node_record_new and each relevant "skill node," a directed edge of type "representation" is created or updated, and the skill mastery score calculated in this interaction is recorded in the edge's attribute. This process generates or updates learning path nodes describing the evolution of the learner's skill mastery status under the goal of the current teaching task. This "learning path" is a subgraph structure composed of Node_task, a series of "historical interaction record nodes" connected by "subsequent" edges (including Node_record_new), and the skill nodes and mastery scores "represented" by these record nodes.

[0157] S1053: Store nodes, records, and strategies as new historical intervention strategies and feedback.

[0158] Extract `main_policy` (the main interaction policy object `Policy_main`) and `assist_policies` (a dictionary corresponding to the list of auxiliary interaction policy objects) from `Cmd_coord`. In `G_learner`, create a new node of type "historical intervention policy node", denoted as `Node_strategy_new`. Store the complete contents of `Policy_main` and `assist_policies`, along with their corresponding agent identifiers (`ID_main` and `ID_list_assist`), as attributes of `Node_strategy_new`.

[0159] In G_learner, a directed edge is created between Node_strategy_new and Node_record_new, with the edge type "Causes Feedback". An attribute "Feedback Score" is set for this "Causes Feedback" edge. The feedback_score is calculated as follows: feedback_score = α * E_main + (1 - α) * (1 / m) * Σ E_assist_x, where E_main is the effectiveness index of the main strategy, E_assist_x is the effectiveness index of each auxiliary strategy, m is the number of auxiliary strategies, and α is a preset weight parameter (e.g., 0.7) used to balance the comprehensive evaluation of the correlation between the main strategy effect and the auxiliary strategies. This feedback_score quantifies the overall effect produced by the combination of intervention strategies used in this study.

[0160] Finally, all new nodes (Node_record_new, Node_strategy_new, and any newly created skill nodes) and new edges ("belongs to", "follows", "embodies", "leads to feedback") created in S1052 and S1053 are persistently saved to the graph database (such as Neo4j) storing the contextual memory network G_learner through graph database transaction operations. At this point, the process of storing learning path nodes, structured interaction records, and the main and auxiliary interaction strategies that triggered the current interaction results as new historical intervention strategies and corresponding learning feedback in the contextual memory network is complete. The updated G_learner will be used in the subsequent step S106.

[0161] Step S105 is the core step in realizing the system's long-term memory and experience accumulation. Through specific, structured graph database operations, it transforms one-off teaching interaction events into persistent, associative knowledge within the learner's contextual memory network. By creating "historical interaction record nodes" and linking them to detailed structured data, a holographic archive of each interaction process is achieved. By calculating similarity and establishing "subsequent" edges through graph neural networks, an evolutionary path between multiple attempts under the same task is automatically constructed, forming a visualized "learning trajectory." By creating "historical intervention strategy nodes" and connecting them to record nodes with "leading to feedback" edges, and quantifying and storing feedback effect values, a clear causal chain between "teaching strategies" and "learning outcomes" is established, enabling the system to learn from historical experience "which strategies are more effective in which situations." This step is not only a record of the past but also provides a structured, searchable experience base based on personal historical data for future intelligent decision-making (S106), serving as the memory cornerstone for the system's continuous personalized optimization and adaptive evolution.

[0162] S106: Based on the updated contextual memory network and the policy effectiveness index, the meta-coordinating agent generates intervention adjustment instructions.

[0163] The intervention adjustment instruction is used to indicate the adjustment of the role, intervention timing, or interaction strategy of at least one of the current main interactive agent and at least one of the auxiliary interactive agents.

[0164] Specifically, the step of generating intervention adjustment instructions through the meta-coordinating agent based on the updated contextual memory network and the strategy effectiveness index includes: querying the updated contextual memory network for a sequence of successful intervention strategies adopted by the learner in similar historical tasks when the main strategy effectiveness index was lower than a first threshold, as a historical successful adjustment reference; based on the historical successful adjustment reference, the current strategy effectiveness index, and the remaining sub-goals of the current teaching task, determining, through the adjustment decision module in the meta-coordinating agent, whether it is necessary to initiate a switch of leadership, adjust the interaction strategy, or adjust the intervention timing of the auxiliary interaction agent; when it is determined that adjustment is needed, generating the intervention adjustment instructions containing the target agent identifier, role change indication, strategy replacement content, or intervention timing correction parameters.

[0165] The specific steps in step S106 during implementation include:

[0166] S1061: Successfully adjusted references by querying history from the updated contextual memory network.

[0167] Receive the updated contextual memory network G_learner from step S105 and the policy effectiveness index E_all from step S104. Extract the main policy effectiveness index E_main from E_all. Determine whether E_main is lower than a preset first threshold θ_low (e.g., θ_low = 0.5).

[0168] If E_main >= θ_low, the current main strategy is considered to be effective and no major intervention is needed based on historical successful adjustments to the reference. An empty reference or one marked "no adjustment required" can be generated, and the process can jump to S1062.

[0169] If E_main < θ_low, then the query is initiated. In G_learner, the following query operation is executed (its logic is described using a Cypher-like query language):

[0170] 1. Locate the current learner node.

[0171] 2. Find "historical teaching task nodes" that are connected to the learner node, whose task_id is different from the current task Goal_current.task_id, but whose core_skills list overlaps with the current task, and use them as candidate similar task nodes.

[0172] 3. For each candidate similar task node, find all "historical interaction record nodes" that are connected to it through the "belongs to" edge.

[0173] 4. For each found "historical interaction record node", check the historical value of the main policy effectiveness index E_main_hist in its effectiveness_vector attribute.

[0174] 5. Filter out the "historical interaction record nodes" where E_main_hist < θ_low.

[0175] 6. For each selected node, find the "historical intervention strategy node" connected to it along the "caused feedback" edge.

[0176] 7. Further filter out nodes whose "feedback effect value" (feedback_score) is greater than the preset success threshold θ_success (e.g., 0.7) among the "feedback-leading" edges connected to these "historical intervention strategy nodes". These nodes represent intervention strategies that were subsequently adopted and ultimately achieved good feedback effects in similar predicaments where the "main strategy was ineffective".

[0177] 8. Sort these policy nodes in descending order of feedback_score, select the top R (e.g., R=3), extract the complete policy content stored in these nodes (including historical master / slave agent identifiers and historical master / slave interaction policy objects), and form a list as the historical successful adjustment reference sequence Seq_adj_ref.

[0178] S1062: Determine the necessity and type of adjustment by adjusting the decision-making module.

[0179] The Seq_adj_ref obtained in step S1061, the current policy effectiveness index E_all, the target object of the current teaching task Goal_current, and the list of remaining sub-goals_remaining parsed from Goal_current are jointly input into the pre-set adjustment decision module in the meta-coordinating agent. This module can be a rule-based expert system or a lightweight classification model (such as a gradient boosting decision tree). This embodiment uses a rule-based expert system as an example to describe its judgment logic:

[0180] Input: Seq_adj_ref, E_all(E_main, E_assist_1, ...), subgoals_remaining.

[0181] Example of a rule set:

[0182] 1. Dominance switch judgment: If E_main < θ_switch (θ_switch is a lower threshold than θ_low, such as 0.3), and Seq_adj_ref is not empty and at least one record in Seq_adj_ref contains a different main agent identifier than the current one, and there exists an E_assist_x > θ_high (such as 0.7), then it is judged that "a dominance switch needs to be initiated", and the new main agent identifier candidate_main_id used by the record with the highest feedback_score in Seq_adj_ref is recorded.

[0183] 2. Main interaction strategy adjustment judgment: If θ_low <= E_main < θ_medium (e.g., θ_medium=0.7) or (E_main < θ_low but the dominance switching condition is not met), and Seq_adj_ref is not empty, then it is judged that "the main interaction strategy needs to be adjusted", and the historical main interaction strategy content candidate_main_policy used by the record with the highest feedback_score in Seq_adj_ref is recorded.

[0184] 3. Judgment on the timing of intervention of the auxiliary interaction strategy: If the value of a certain E_assist_x is very low (e.g., <0.2), but the preset triggering condition in the Policy_assist_x corresponding to the auxiliary strategy has never been met in S103, while the number of learners’ confusion points C_act is very high, it is judged that “the timing of intervention of the auxiliary interaction agent needs to be adjusted”, and it is recommended to lower the triggering condition threshold (such as the response time threshold) of the auxiliary strategy.

[0185] 4. Judgment on adjustment of auxiliary interaction strategy content: If the value of a certain E_assist_x is very low, and its preset conditions have been met but the effect after intervention is not good (in combination with the detailed analysis of Result_interaction), and there are historical records of similar auxiliary roles using different strategy content in Seq_adj_ref, then it is judged that "the interaction strategy content of the auxiliary interaction agent needs to be adjusted".

[0186] Output: The adjustment decision module comprehensively applies the above rules and outputs an adjustment decision result, Decision_adj. Decision_adj is a data structure that can contain a combination of the following fields: need_adjustment (boolean), adjustment_type (enumeration, such as MAIN_AGENT_SWITCH, MAIN_POLICY_UPDATE, ASSIST_TIMING_MOD, ASSIST_POLICY_UPDATE), target_agent_id (target agent identifier), suggested_new_main_agent_id (suggested new main agent identifier), suggested_policy_content (suggested new policy content object), and timing_mod_params (intervention timing adjustment parameters).

[0187] S1063: Generate intervention and adjustment instructions.

[0188] Check the need_adjustment field in Decision_adj. If it is False, generate an intervention adjustment instruction Cmd_adj that is empty or indicates "maintain the current state".

[0189] If need_adjustment is True, then the specific instruction content is generated based on adjustment_type. A new JavaScript object representation object is created as Cmd_adj, whose fields are dynamically set according to the adjustment type:

[0190] If it is MAIN_AGENT_SWITCH, then Cmd_adj contains: instruction_type: “SWITCH_LEAD”, current_main_id: [current main agent identifier], new_main_id: [Decision_adj.suggested_new_main_agent_id], context_to_transfer:{subgoals_remaining, current dialogue history, current operation state}.

[0191] If it is MAIN_POLICY_UPDATE, then Cmd_adj contains: instruction_type: "UPDATE_POLICY", target_agent_id: [current main agent identifier], new_policy: [Decision_adj.suggested_policy_content].

[0192] If it is ASSIST_TIMING_MOD, then Cmd_adj contains: instruction_type: "MOD_TIMING", target_agent_id: [specific auxiliary agent identifier], new_trigger_condition: [modified trigger condition logical expression].

[0193] If it is ASSIST_POLICY_UPDATE, then Cmd_adj contains: instruction_type: “UPDATE_POLICY”, target_agent_id: [specific agent identifier], new_policy: [Decision_adj.suggested_policy_content].

[0194] Among them, Decision_adj is an adjustment decision result data structure generated by the adjustment decision module in S1062. Its fields include: need_adjustment (a boolean field indicating whether any adjustment is needed), adjustment_type (an enumerated field indicating the specific adjustment operation type, such as MAIN_AGENT_SWITCH, MAIN_POLICY_UPDATE, etc.), target_agent_id (a string indicating the unique identifier of the target agent to be adjusted), suggested_new_main_agent_id (a string; when the adjustment type is MAIN_AGENT_SWITCH, this field stores the identifier of the new main agent recommended from historical successful experience), suggested_policy_content (an object; when the adjustment type is policy update, this field stores the recommended new policy content object), and timing_mod_params (an object or parameter list; when the adjustment type is ASSIST_TIMING_MOD, this field stores the corrected intervention timing parameters, such as the new response time threshold). The intervention adjustment instruction Cmd_adj generated based on Decision_adj is a serializable instruction object. Its core fields include: instruction_type (a string, directly taken from Decision_adj.adjustment_type, used to indicate what operation the receiver should perform), current_main_id (a string, existing only during a change of leadership, representing the identifier of the main agent to be replaced), new_main_id (a string, existing only during a change of leadership, representing the identifier of the new agent that will take over leadership), and context_to_transfer (an object, existing only during a change of leadership, encapsulating the context information to be passed to the new main agent). The information includes `subgoals_remaining` (a list of unfinished sub-goals remaining in the current teaching task), `current dialogue history` (a list of the most recent dialogue records in this session), `current operation status` (a snapshot of the learner's operation progress and status in the virtual training environment), `target_agent_id` (a string representing the specific recipient agent identifier of this adjustment instruction), `new_policy` (an object; when the instruction type is policy update, this field contains the new policy content to be loaded), and `new_trigger_condition` (a logical expression string; when the instruction type is ASSIST_TIMING_MOD, this field defines the modified conditional logic used to trigger the intervention of the assistant agent). These fields together constitute a complete and executable adjustment instruction, ensuring the accurate communication and implementation of dynamic adjustments.

[0195] Optionally, Cmd_adj may also include a decision_basis field, which briefly records the main basis for triggering this adjustment (such as E_main being too low). The generated Cmd_adj is used as the final output of step S106, i.e., the intervention adjustment instruction, and provided to subsequent step S107.

[0196] Step S106 is a crucial decision-making step in achieving "intelligent optimization throughout the entire process," enabling the system to dynamically adjust based on real-time feedback during instruction. This step first utilizes an updated contextual memory network to proactively search learners' personal historical experiences for situations similar to the current "teaching dilemma" (poor performance of the main strategy) and previously validated solutions, providing valuable personalized experience references for current decision-making. Subsequently, through a structured adjustment decision module, considering the effectiveness of the current strategy, remaining teaching objectives, and historical successes, it intelligently determines the most urgent and potentially effective adjustment type (such as changing the lead teacher, fine-tuning teaching methods, or altering the timing of teaching assistant intervention). Ultimately, the generated intervention instruction is a specific and actionable plan, rather than a vague suggestion. This process profoundly reflects the "director" role of the meta-coordinating agent, which is no longer limited to initial choreography but can decisively call in "substitute actors" or modify the "script" based on the real-time effects of the "performance" (teaching interaction), thereby ensuring that the teaching "drama" can overcome obstacles and continue smoothly and effectively. This fundamentally solves the pain point that static role configuration cannot adapt to dynamic teaching processes and is the core mechanism for achieving long-term, adaptive teaching.

[0197] S107: According to the intervention adjustment instruction, in the subsequent teaching interaction process, perform dynamic adjustment of the current main interactive agent and at least one of the auxiliary interactive agents.

[0198] The dynamic adjustment includes switching the control of the teaching interaction from the current main interactive agent to a designated auxiliary interactive agent when the goal of the current teaching task has not been achieved.

[0199] Specifically, according to the intervention adjustment instruction, in the subsequent teaching interaction process, the dynamic adjustment of the current main interactive agent and at least one of the auxiliary interactive agents includes: parsing the intervention adjustment instruction, identifying the agent identifier to be adjusted and the adjustment operation type; if the adjustment operation type is a role change and involves a switch of dominance, then sending a dormant instruction to the current main interactive agent, simultaneously activating one of the auxiliary interactive agents specified in the intervention adjustment instruction, and synchronizing the dominance of the teaching interaction, unfinished sub-tasks, and context to the activated auxiliary interactive agent; controlling the auxiliary interactive agent that has obtained dominance to continue the subsequent teaching interaction with the learner with the goal of the current teaching task based on the synchronized context and loading the new interaction strategy specified in the intervention adjustment instruction.

[0200] The specific steps in step S107 during implementation include:

[0201] S1071: Parse intervention and adjustment instructions and identify adjustment operations.

[0202] The system receives the intervention adjustment instruction object Cmd_adj generated in step S106. By parsing the instruction_type field of Cmd_adj, the specific adjustment operation type is identified. Simultaneously, fields such as target_agent_id (target agent identifier), new_main_id (new master agent identifier, if it exists), context_to_transfer (context to be synchronized), new_policy (new interaction policy object), or new_trigger_condition (new trigger condition) are extracted from Cmd_adj. Which fields are extracted depends on the instruction_type. For example, if the instruction_type is "SWITCH_LEAD", then current_main_id, new_main_id, and context_to_transfer are extracted.

[0203] S1072: Dynamic adjustment of the shift in leadership.

[0204] If the instruction_type identified in step S1071 is "SWITCH_LEAD" (i.e., a role change involving a switch of leadership), then the following specific operation sequence is executed:

[0205] a) Suspend the current main interactive agent: Based on Cmd_adj.current_main_id, send a structured slump command message to the service container corresponding to the agent. This message contains the command type "SUSPEND" and the current session identifier. Upon receiving the message, the service container of the current main interactive agent suspends its dialogue thread, saves the current short-term session state (such as the last few lines of dialogue history) to persistent storage, and releases computing resources, but keeps the service instance alive to receive possible wake-up commands. After the slump command is successfully sent, remove the agent from the system's active interactive session registry.

[0206] b) Activate the specified auxiliary interactive agent: Query the agent registry based on `Cmd_adj.new_main_id` to obtain its service endpoint information. Send a structured activation command message to the service container of the auxiliary interactive agent. The message contains the command type "ACTIVATE_AS_LEAD", the session identifier, and the synchronization context extracted from `Cmd_adj.context_to_transfer`. The synchronization context specifically includes: `subgoals_remaining` (a list of incomplete sub-goals in `Goal_current`), `recent_dialogue_history` (text records of the last N rounds of dialogue loaded from the state saved by the previous main interactive agent), and `current_operation_state` (the learner's current operation state in the virtual training environment, such as completed steps, current tool position, etc.).

[0207] c) Synchronizing Dominance and Context: The activated auxiliary interactive agent service container instantiates or restores an agent session in memory, loads the received `recent_dialogue_history` and `current_operation_state` into its working memory, and explicitly sets its role as the "dominant interactor". The system updates the dominant agent identifier of the current session in the active interactive session registry to `Cmd_adj.new_main_id`, completing the switch of teaching interaction dominance.

[0208] d) Loading the new interaction strategy and continuing the interaction: The auxiliary interaction agent, which has now gained control (and is now the new master interaction agent), loads a new interaction strategy specifically configured for the remaining teaching task goal `subgoals_remaining`, based on the synchronized context, from its local policy library or from the `new_policy` field attached to `Cmd_adj`. The new interaction strategy object also contains a specific dialogue template, question sequence, and operation step guidance. The agent then naturally continues from the previously interrupted dialogue point or operation step, for example, saying, "We just reached step three; now let's understand this operation in a different way. Please see this breakdown demonstration...", and guides the learner to continue the subsequent teaching interaction for the current teaching task's goal `Goal_current`.

[0209] S1073: Perform dynamic adjustments for non-dominant power switching.

[0210] If the instruction_type identified in step S1071 is not "SWITCH_LEAD", then adjustments will be performed according to the specific type:

[0211] a) Update the main interaction policy: If instruction_type is "UPDATE_POLICY" and target_agent_id is the current main agent identifier, then send a policy update instruction to the agent's service container. The instruction contains a new Policy_main object (from Cmd_adj.new_policy). The current main interaction agent dynamically replaces the policy in its working memory, and subsequent interactions will immediately be based on the new policy.

[0212] b) Adjust the intervention timing of the auxiliary agent: If instruction_type is "MOD_TIMING", then find the corresponding auxiliary interaction agent configuration based on target_agent_id, modify the trigger condition judgment logic in its listening module, and replace the old logic described by Cmd_adj.new_trigger_condition with the new logic. For example, change "response time > 60 seconds" to "response time > 45 seconds".

[0213] c) Update the auxiliary interaction policy: If instruction_type is “UPDATE_POLICY” and target_agent_id is the identifier of a certain auxiliary agent, then update the policy library of that auxiliary agent and replace its original policy content with Cmd_adj.new_policy.

[0214] Step S107 is the final execution stage that transforms the optimization decisions (intervention and adjustment instructions) of the meta-coordinating agent into actual system behavior, achieving "dynamic adjustment." It precisely controls the agent's state transitions (dormancy, activation), role changes (transfer of leadership), and policy updates through a concrete and reliable inter-service communication mechanism (sending structured instruction messages). In leadership switching scenarios, it ensures the continuity of the teaching process, achieving "seamless handover" between different agents through context synchronization, making learners almost unaware of the change in background "actors," maintaining a consistent experience. For non-switching adjustments, it can fine-tune the agent's behavioral parameters or policy content in real time and with precision. This step directly addresses the problem of rigid interaction caused by static role configuration, giving the teaching system the ability to dynamically reorganize and flexibly adapt during operation. This allows multi-agent collaboration to function like a well-coordinated theatrical performance, smoothly switching "roles" and "scripts" based on real-time feedback from the "audience" (learners), thereby ensuring the continuous and efficient achievement of teaching objectives and greatly improving the system's adaptability and teaching effectiveness.

[0215] This embodiment provides a method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration. The method acquires learners' real-time learning status and teaching task objectives; based on the learning status, objectives, and contextual memory network, it generates coordination instructions through a meta-coordinating agent, designates the main interactive agent and auxiliary interactive agents from the set of collaborative teaching agents, and configures strategies; it controls the interaction between the main interactive agent and the learner and collects results; it generates strategy effectiveness indicators based on the results and updates the contextual memory network; and it generates intervention and adjustment instructions through a meta-coordinating agent to dynamically adjust agent roles, timing, or strategies, including switching leadership. This invention solves the problem of static role configuration making dynamic contextualized interaction difficult, achieving smooth collaboration like an educational drama, and enhancing the immersive experience and long-term effectiveness of personalized teaching.

[0216] Figure 2 The connection diagram of the intelligent generation and optimization system for the entire vocational education process based on multi-agent collaboration provided in this application is as follows: Figure 2 As shown, this embodiment provides a vocational education full-process intelligent generation and optimization system based on multi-agent collaboration. This system applies... Figure 1 The embodiment describes a method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration. The generation and optimization system includes:

[0217] The status acquisition module is used to acquire the learner's real-time learning status and the objectives of the current teaching task;

[0218] An instruction generation module, connected to the state acquisition module, includes a meta-coordinating agent. The instruction generation module generates coordination instructions based on the real-time learning state, the goal of the current teaching task, and a contextual memory network constructed for the learner. The meta-coordinating agent generates decision instructions for scheduling teaching agents based on the real-time learning state, the goal of the current teaching task, and the contextual memory network. The contextual memory network stores the learner's current learning progress, task-solving paths for acquired skills, past intervention strategies, and corresponding learning feedback. The coordination instructions dynamically specify the current main interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents, configuring a main interaction strategy for the current main interactive agent and corresponding auxiliary interaction strategies for at least one of the auxiliary interactive agents. The set of collaborative teaching agents includes multiple teaching agents capable of independently performing specific teaching functions.

[0219] An interactive execution module, connected to the instruction generation module, is used to control the current main interactive agent to conduct teaching interaction with the learner based on the main interactive strategy according to the coordination instruction, and to collect the interaction results generated by this teaching interaction.

[0220] An evaluation module, connected to the interaction execution module, is used to evaluate and generate corresponding strategy effectiveness indicators based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy.

[0221] A memory update module is connected to the interaction execution module, and the memory update module is used to update the context memory network according to the interaction result;

[0222] An adjustment instruction generation module is connected to the evaluation module and the memory update module. The adjustment instruction generation module is used to generate intervention adjustment instructions through the meta-coordinating agent based on the updated context memory network and the policy effectiveness index. The intervention adjustment instructions are used to indicate the adjustment of the role, intervention timing or interaction strategy of at least one of the current main interaction agent and at least one of the auxiliary interaction agents.

[0223] A dynamic adjustment execution module is connected to the adjustment instruction generation module. The dynamic adjustment execution module is used to perform dynamic adjustments on the current main interactive agent and at least one of the auxiliary interactive agents in the subsequent teaching interaction process according to the intervention adjustment instruction. The dynamic adjustment includes switching the leadership of the teaching interaction from the current main interactive agent to a designated auxiliary interactive agent when the goal of the current teaching task has not been completed.

[0224] Specifically, the instruction generation module includes:

[0225] An assessment unit is used to assess the learner's cognitive load level and current skill mastery based on the real-time learning status and the objectives of the current teaching task.

[0226] A strategy matching unit, connected to the evaluation unit, is used to determine the current main interactive agent and at least one auxiliary interactive agent from the set of collaborative teaching agents based on the cognitive load level, the current skill mastery, and the learner's past successful intervention strategy sequences extracted from the contextual memory network, through the strategy matching module in the meta-coordinating agent, and to generate a main interaction strategy corresponding to the current main interactive agent and an auxiliary interaction strategy corresponding to at least one of the auxiliary interactive agents;

[0227] An encapsulation unit, connected to the policy matching unit, is used to encapsulate an instruction data packet containing the current main interactive agent identifier, the identifier of at least one auxiliary interactive agent, the main interactive policy, and the auxiliary interactive policy, as the coordination instruction.

[0228] Specifically, the interactive execution module includes:

[0229] The parsing unit is used to parse the coordination instruction and obtain the current main interactive agent identifier and the main interaction strategy;

[0230] A loading unit, connected to the parsing unit, is used to call the corresponding current main interactive agent according to the current main interactive agent identifier and load the main interactive strategy, which includes a dialogue template, a question sequence and expected skill operation steps.

[0231] An interaction control unit, connected to the loading unit, is used to control the current main interaction agent loaded with the main interaction strategy to conduct multi-round dialogues or guided operations with the learner, and to record the learner's answers, operation sequences, response times and emotional feedback signals during the interaction process, and to integrate the answers, operation sequences, response times and emotional feedback signals into the interaction result.

[0232] Example: Intelligent generation and optimization system for the entire vocational education process based on multi-agent collaboration.

[0233] This embodiment details a specific implementation scheme for a vocational education end-to-end intelligent generation and optimization system based on multi-agent collaboration, to ensure that those skilled in the art can fully understand and implement the invention. This system serves as the physical and software carrier for the aforementioned method embodiments, achieving the end-to-end intelligent generation and optimization through specific hardware deployment, software module construction, and network connection.

[0234] 1. System Overall Architecture and Physical Deployment

[0235] The generation and optimization system adopts a distributed client-server architecture. The learner interactive terminal, acting as the client, is deployed at the learning site. Its hardware includes: a front-facing camera with at least 2 megapixels and supporting a 30 frames per second acquisition rate; an omnidirectional microphone supporting a 16kHz sampling rate; a touchscreen or keyboard and mouse supporting graphical interface interaction; a spatial positioning device supporting 3D model manipulation (such as a VR controller or Leap Motion gesture recognizer); and a smart bracelet supporting Bluetooth 5.0 protocol and capable of acquiring photoplethysmography (PPG) and electrodermal activity (EDA) signals. The learner interactive terminal runs learner interaction software, which integrates computer vision libraries (such as OpenCV), audio processing libraries (such as librosa), and device drivers for raw data acquisition and preliminary processing.

[0236] The core of the system, the intelligent agent collaboration platform, is deployed in a backend cloud computing environment or a high-performance server cluster. Logically, this platform is divided into the seven modules mentioned above, and physically hosted by multiple servers. These servers are connected via a high-speed local area network (such as 10 Gigabit Ethernet) and managed using load balancing and service mesh technologies. Data exchange between all modules is accomplished through predefined application programming interfaces and message queues (such as RabbitMQ or Kafka) to ensure loose coupling and high reliability.

[0237] 2. Specific structure, connection and operation of each module

[0238] 2.1 Status Acquisition Module

[0239] The status acquisition module includes a data acquisition submodule deployed on the learner's interactive terminal and a data processing submodule deployed on the server.

[0240] The data acquisition submodule is a software process that continuously acquires raw data streams in parallel from the front-facing camera, microphone, interactive devices, and smart bracelet by calling the operating system's application programming interfaces (APIs) for camera video streams, audio input, human-computer interaction event hooks, and Bluetooth serial port configuration files. Its function is to capture the learner's multimodal behaviors and physiological signals.

[0241] The data processing submodule is a microservice running on a server. The data acquisition submodule pushes raw, timestamped data packets to the data processing submodule in real time via the WebSocket protocol. Internally, the data processing submodule runs multiple processing threads in parallel: one thread calls a convolutional neural network model from the OpenFace toolkit to process image sequences, outputting facial expression intensity and eye-tracking coordinates; another thread calls the librosa library and a pre-trained support vector machine model to process audio, outputting speech emotion tags; another thread parses operation event logs, calculating success rate and response time; and a third thread filters and extracts features from physiological signals. All features are concatenated into the real-time learning state vector S_t. Simultaneously, this submodule initiates an HTTP request to an external teaching management system through a RESTful application programming interface (API) client, using the learner's unique identifier as a parameter to obtain and parse the returned JavaScript object representation data, constructing the target object Goal_current for the current teaching task. Furthermore, it queries a graph database server (such as a server deploying Neo4j) using a graph query language (such as Cypher) through another API client to obtain or initialize the contextual memory network G_learner. The data processing submodule encapsulates S_t, Goal_current, and G_learner into a data structure and sends it to the instruction generation module via an internal message bus. Its function is to transform raw sensor data into high-level semantic states and provide complete contextual information for decision-making.

[0242] 2.2 Instruction Generation Module

[0243] The instruction generation module is a core decision-making service deployed on a standalone server or container. It includes a meta-coordinating agent and evaluation unit, a policy matching unit, and an encapsulation unit.

[0244] The meta-coordinating agent is a web service built on the Python Flask or FastAPI framework. Internally, this service loads and runs a pre-trained policy matching model. In this embodiment, the model employs the knowledge graph embedding model TransE, which is trained offline on a policy knowledge graph containing a large number of <learner state, task, successful policy> triples, capable of mapping input states and tasks to neighboring points in the policy space. The meta-coordinating agent receives data from the state acquisition module, and its role is to coordinate the work of internal units and execute the final decision.

[0245] Evaluation Unit: A functional function serving the meta-coordinating agent. This function receives S_t and Goal_current, extracts specific components from S_t, applies the computational logic of the NASA-TLX scale (a set of pre-defined weighted summation formulas) to calculate the cognitive load level CL; and calculates the current skill mastery MS_current by querying the "mastery" edge attributes related to the Goal_current skill in G_learner.

[0246] The policy matching unit is the core function serving the meta-coordinating agent. This function first performs a graph traversal query from the G_learner to extract the sequence of past successful intervention policies, Seq_success. Then, CL, MS_current, Seq_success, and Goal_current are input into the loaded TransE model for encoding and similarity matching, resulting in a set of candidate policies. Subsequently, a built-in rule engine (based on the Drools framework or hard-coded if-else logic) filters and sorts the candidate policies based on the values ​​of CL and MS_current, ultimately outputting the determined current primary interacting agent identifier ID_main, the list of auxiliary interacting agent identifiers ID_list_assist, the primary interacting policy object Policy_main, and the list of auxiliary interacting policy objects Policy_list_assist. Policy_main is a concrete data structure, such as a Python dictionary containing key-value pairs {"dialogue_templates": ["template 1 content", ...], "question_sequence": ["question 1", ...], "expected_steps": ["step 1 description", ...]}.

[0247] The encapsulation unit is a simple function within the meta-coordinating agent service that encapsulates the aforementioned identifiers and policy objects into a JavaScript object representation object, `Cmd_coord`. The instruction generation module ultimately publishes `Cmd_coord` to a designated topic via a message queue for consumption by the interactive execution module. Its role is to generate executable, detailed scheduling instructions based on specific algorithms and rules.

[0248] 2.3 Interactive Execution Module

[0249] The interactive execution module mainly includes various teaching agent instances running in the agent container cluster, as well as a central interactive scheduling and coordination submodule.

[0250] The collaborative teaching agent set consists of multiple independent Docker containers or Kubernetes Pods, each running a specific teaching agent service (e.g., "theoretical explanation agent," "practical demonstration agent," "troubleshooting agent"). Each agent service exposes a unified gRPC or HTTP application programming interface. These containers are deployed on a server cluster and managed by a container orchestration platform.

[0251] The Interaction Scheduling and Coordination Submodule is a resident process comprising a parsing unit, a loading unit, and an interaction control unit. The parsing unit retrieves and parses the Cmd_coord from the message queue. The loading unit, based on ID_main, uses a service discovery mechanism (such as Consul) to find the network address of the corresponding agent container, and then sends an "activation" request and Policy_main to the container via a remote procedure call. The target agent container returns a ready signal after loading the policy. The interaction control unit maintains the WebSocket connection to the learner's interaction terminal. It controls the flow: pushing the dialogue templates and questions in Policy_main to the terminal interface for display or playing them via text-to-speech conversion; receiving learner speech returned from the terminal and converting it to text using an integrated automatic speech recognition service (such as a service based on the Wav2Vec 2.0 model); recording response times; and, in the virtual training scenario, listening for 3D operation events from the terminal and comparing them with the expected steps in Policy_main. Simultaneously, this unit continuously receives multimodal raw data streams (corresponding to S1033) synchronously sent from the terminal's data acquisition submodule. Finally, the interaction control unit integrates all recorded data into a Result_interaction object and sends it to the evaluation module and memory update module via a message queue. Its function is to accurately execute the teaching script and comprehensively record the interaction process.

[0252] 2.4 Evaluation Module

[0253] The evaluation module is a standalone microservice that receives `Result_interaction`, `Policy_main`, and `Policy_list_assist`. Its internal logic implements the steps described in S104: extracting key feedback metrics from `Result_interaction.overall_metrics`; extracting expected metrics and trigger conditions from the policy object; calculating `Score_imm` using a specific weighted calculation formula; calculating each `Score_prep_x` by iterating through the detailed records of `Result_interaction` and applying the Sigmoid function; and finally performing normalized weighted calculations to generate the policy effectiveness metric `E_all`. This module outputs `E_all` through a message queue. Its function is to quantitatively evaluate teaching effectiveness using a mathematical model.

[0254] 2.5 Memory Update Module

[0255] The memory update module is a service specifically designed for interacting with graph databases. It receives `Result_interaction`, `Cmd_coord`, and `E_all`. Its internal logic implements the steps described in S105: creating a `Record_struct` object; executing complex Cypher queries using Neo4j's official driver; creating new nodes and edges in `G_learner`, including "historical interaction record nodes" and "historical intervention strategy nodes," and establishing relational edges such as "belongs to," "follow-up," "manifests," and "leads to feedback," while simultaneously calculating and storing `feedback_score`. Its function is to persist the experience of each interaction in the form of an attribute graph, forming a queryable contextual memory.

[0256] 2.6 Adjust the instruction generation module

[0257] The adjustment instruction generation module is another service of the meta-coordinating agent or a different endpoint of the same service. It receives the updated G_learner (notified by the memory update module or queried directly) and E_all. Its internal logic implements the steps described in S106: First, it compares E_main with the threshold θ_low; if it is lower than the threshold, it sends a Cypher query to the graph database to retrieve the historical successful adjustment reference sequence Seq_adj_ref; subsequently, its built-in adjustment decision module (an inference engine based on the CLIPS expert system shell or a hard-coded rule tree) infers based on Seq_adj_ref, E_all, and the remaining sub-objectives to generate Decision_adj; finally, it encapsulates the decision into a specific Cmd_adj instruction. This module publishes Cmd_adj through a message queue. Its function is to perform secondary optimization decisions based on memory and real-time evaluation.

[0258] 2.7 Dynamically Adjust the Execution Module

[0259] The dynamic adjustment execution module is a process that works closely with or is integrated with the interactive scheduling and coordination submodule. It listens for Cmd_adj messages from the adjustment instruction generation module. Its internal logic implements the steps described in S107: parsing Cmd_adj to identify the operation type. If it is "SWITCH_LEAD", it sends a "sleep" instruction to the agent container corresponding to current_main_id via a remote procedure call; simultaneously, it activates the agent container corresponding to new_main_id through service discovery and a remote procedure call, and synchronizes context_to_transfer there; it updates the dominant agent registration information in the interactive scheduling and coordination submodule; and it notifies the new dominant agent to load new_policy and continue the interaction. For other adjustment types, it sends the corresponding policy update or parameter modification instruction to the target agent container. Its function is to reliably execute system reconstruction instructions, achieving dynamic and seamless adjustment of the teaching process.

[0260] 3. Data storage

[0261] Contextual memory network: Physically stored in a separate graph database server (running Neo4j), which is interconnected with the agent collaboration platform via a high-speed network.

[0262] Strategy knowledge base, model files, and multimedia resources: stored in a shared distributed file system (such as HDFS) or object storage service (such as AWS S3) for loading by the corresponding modules.

[0263] Real-time data caching: Use an in-memory database (such as Redis) to cache temporary data such as session state and real-time messages.

[0264] This embodiment discloses a vocational education full-process intelligent generation and optimization system based on multi-agent collaboration. Through the specific hardware configuration, software module division, internal unit implementation details, clear network communication protocols and data interfaces between modules, and the specific application of key algorithm models (such as TransE, NASA-TLX calculation, Sigmoid function, and rule engine), it fully reproduces all the functions of the generation and optimization system described in the invention. Each module performs its own function and works collaboratively, forming an intelligent teaching system entity capable of personalized, dynamic, and long-term optimization, thus solving the static collaborative mechanism problem proposed in the background art.

[0265] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only.

[0266] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration, characterized in that, The method includes: To obtain learners' real-time learning status and the objectives of the current teaching task; Based on the real-time learning state, the goal of the current teaching task, and the contextual memory network constructed for the learner, a coordination instruction is generated through a unary coordinating agent. The unary coordinating agent generates decision instructions for scheduling teaching agents based on the real-time learning state, the goal of the current teaching task, and the contextual memory network. The contextual memory network stores the learner's current learning progress, task-solving paths for acquired skills, past intervention strategies, and corresponding learning feedback. The coordination instruction dynamically specifies the current primary interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents, configures a primary interactive strategy for the primary interactive agent, and configures corresponding auxiliary interactive strategies for at least one of the auxiliary interactive agents. The set of collaborative teaching agents includes multiple teaching agents capable of independently performing specific teaching functions. According to the coordination instructions, the current main interactive agent is controlled to conduct teaching interactions with the learner based on the main interaction strategy, and the interaction results generated by this teaching interaction are collected. Based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy, evaluate and generate corresponding strategy effectiveness indicators; Update the contextual memory network based on the interaction results; Based on the updated contextual memory network and the policy effectiveness index, the meta-coordinating agent generates intervention adjustment instructions; wherein, the intervention adjustment instructions are used to indicate the adjustment of the role, intervention timing or interaction strategy of at least one of the current main interaction agent and at least one of the auxiliary interaction agents; According to the intervention and adjustment instructions, in the subsequent teaching interaction process, dynamic adjustments are performed on the current main interactive agent and at least one of the auxiliary interactive agents; wherein, the dynamic adjustment includes switching the leadership of the teaching interaction from the current main interactive agent to a designated auxiliary interactive agent when the goal of the current teaching task is not completed.

2. The method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration as described in claim 1, characterized in that, The method of generating coordination instructions through a unary coordinating agent based on the real-time learning state, the objective of the current teaching task, and the contextual memory network constructed for the learner includes: Based on the real-time learning status and the objectives of the current teaching task, assess the learner's cognitive load level and current skill mastery. Based on the cognitive load level, the current skill mastery, and the learner's past successful intervention strategy sequence extracted from the contextual memory network, the current main interactive agent and at least one auxiliary interactive agent are determined from the set of collaborative teaching agents through the strategy matching module in the meta-coordinating agent, and a main interactive strategy corresponding to the current main interactive agent and an auxiliary interactive strategy corresponding to at least one of the auxiliary interactive agents are generated. An instruction data packet containing the current main interactive agent identifier, the identifier of at least one auxiliary interactive agent, the main interactive strategy, and the auxiliary interactive strategy is encapsulated as the coordination instruction.

3. The method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration as described in claim 1, characterized in that, The step of controlling the current main interactive agent to conduct teaching interactions with the learner based on the main interaction strategy according to the coordination instruction, and collecting the interaction results generated by this teaching interaction, includes: Parse the coordination command to obtain the current main interactive agent identifier and the main interaction strategy; The corresponding current main interactive agent is invoked according to the current main interactive agent identifier, and the main interactive strategy is loaded. The main interactive strategy includes a dialogue template, a question sequence, and expected skill operation steps. The system controls the current main interactive agent, which has been loaded with the main interaction strategy, to conduct multiple rounds of dialogue or guidance operations with the learner, and records the learner's answers, operation sequences, response times, and emotional feedback signals during the interaction process. The answers, operation sequences, response times, and emotional feedback signals are then integrated into the interaction result.

4. The method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration as described in claim 1, characterized in that, The step of evaluating and generating corresponding strategy effectiveness metrics based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy includes: Key feedback metrics are extracted from the interaction results, including task completion accuracy, step completion time, number of points of confusion, and intensity of emotional feedback signals. The key feedback indicators are compared with the expected indicators preset in the main interaction strategy and the pre-triggered conditions preset in the auxiliary interaction strategy to calculate the immediate effect score of the main interaction strategy and the pre-association score of the auxiliary interaction strategy. Based on the immediate effect score and the preliminary association score, a weighted calculation is performed to generate a main strategy effectiveness index corresponding to the main interaction strategy and an auxiliary strategy effectiveness index corresponding to the auxiliary interaction strategy. The main strategy effectiveness index and the auxiliary strategy effectiveness index together constitute the strategy effectiveness index.

5. The method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration as described in claim 1, characterized in that, Updating the contextual memory network based on the interaction result includes: The interaction results are processed in a structured manner to generate a structured interaction record containing a timestamp, the associated goal of the current teaching task, the learner's behavior sequence, and key feedback indicators of the interaction results; The structured interaction records are associated with the past interaction records stored in the contextual memory network to generate or update learning path nodes that describe the evolution of the learner's skill mastery status under the goal of the current teaching task. The learning path nodes, the structured interaction records, and the main interaction strategy and the auxiliary interaction strategy that triggered the current interaction result are stored in the contextual memory network as new historical intervention strategies and corresponding learning feedback.

6. The method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration as described in claim 4, characterized in that, The step of generating intervention adjustment instructions through the meta-coordinating agent based on the updated contextual memory network and the policy effectiveness index includes: The updated contextual memory network is used to query the learner’s successful intervention strategy sequence in similar historical tasks when the main strategy effectiveness index is lower than the first threshold, as a historical successful adjustment reference. Based on the historical successful adjustment reference, the current strategy effectiveness index, and the remaining sub-goals of the current teaching task, the adjustment decision module in the meta-coordinating agent determines whether it is necessary to initiate a switch of leadership, adjust the interaction strategy, or adjust the intervention timing of the auxiliary interaction agent. When it is determined that adjustment is needed, the intervention adjustment instruction is generated, which includes the target agent identifier, role change indication, policy replacement content, or intervention timing correction parameters.

7. The method for intelligent generation and optimization of the entire vocational education process based on multi-agent collaboration as described in claim 1, characterized in that, The step of dynamically adjusting the current main interactive agent and at least one of the auxiliary interactive agents in subsequent teaching interactions according to the intervention adjustment instruction includes: The intervention and adjustment instructions are parsed to identify the agent identifier to be adjusted and the type of adjustment operation. If the adjustment operation type is a role change and involves a switch of control, a dormant command is sent to the current main interactive agent, and one of the auxiliary interactive agents specified in the intervention adjustment command is activated. The control of the teaching interaction, the unfinished sub-tasks and the context are synchronized to the activated auxiliary interactive agent. The auxiliary interactive agent, having gained control, operates based on a synchronized context and loads the new interaction strategy specified in the intervention adjustment instructions, continuing subsequent teaching interactions with the learner toward the goal of the current teaching task.

8. A vocational education whole-process intelligent generation and optimization system based on multi-agent collaboration, characterized in that, The generation and optimization system applies the generation and optimization method according to any one of claims 1-7, and the generation and optimization system comprises: The status acquisition module is used to acquire the learner's real-time learning status and the objectives of the current teaching task; An instruction generation module, connected to the state acquisition module, includes a meta-coordinating agent. The instruction generation module generates coordination instructions based on the real-time learning state, the goal of the current teaching task, and a contextual memory network constructed for the learner. The meta-coordinating agent generates decision instructions for scheduling teaching agents based on the real-time learning state, the goal of the current teaching task, and the contextual memory network. The contextual memory network stores the learner's current learning progress, task-solving paths for acquired skills, past intervention strategies, and corresponding learning feedback. The coordination instructions dynamically specify the current main interactive agent and at least one auxiliary interactive agent from a set of collaborative teaching agents, configuring a main interaction strategy for the current main interactive agent and corresponding auxiliary interaction strategies for at least one of the auxiliary interactive agents. The set of collaborative teaching agents includes multiple teaching agents capable of independently performing specific teaching functions. An interactive execution module, connected to the instruction generation module, is used to control the current main interactive agent to conduct teaching interaction with the learner based on the main interactive strategy according to the coordination instruction, and to collect the interaction results generated by this teaching interaction. An evaluation module, connected to the interaction execution module, is used to evaluate and generate corresponding strategy effectiveness indicators based on the interaction results, the main interaction strategy, and the auxiliary interaction strategy. A memory update module is connected to the interaction execution module, and the memory update module is used to update the context memory network according to the interaction result; An adjustment instruction generation module is connected to the evaluation module and the memory update module. The adjustment instruction generation module is used to generate intervention adjustment instructions through the meta-coordinating agent based on the updated context memory network and the policy effectiveness index. The intervention adjustment instructions are used to indicate the adjustment of the role, intervention timing or interaction strategy of at least one of the current main interaction agent and at least one of the auxiliary interaction agents. A dynamic adjustment execution module is connected to the adjustment instruction generation module. The dynamic adjustment execution module is used to perform dynamic adjustments on the current main interactive agent and at least one of the auxiliary interactive agents in the subsequent teaching interaction process according to the intervention adjustment instruction. The dynamic adjustment includes switching the leadership of the teaching interaction from the current main interactive agent to a designated auxiliary interactive agent when the goal of the current teaching task has not been completed.

9. The system according to claim 8, characterized in that, The instruction generation module includes: An assessment unit is used to assess the learner's cognitive load level and current skill mastery based on the real-time learning status and the objectives of the current teaching task. A strategy matching unit, connected to the evaluation unit, is used to determine the current main interactive agent and at least one auxiliary interactive agent from the set of collaborative teaching agents based on the cognitive load level, the current skill mastery, and the learner's past successful intervention strategy sequences extracted from the contextual memory network, through the strategy matching module in the meta-coordinating agent, and to generate a main interaction strategy corresponding to the current main interactive agent and an auxiliary interaction strategy corresponding to at least one of the auxiliary interactive agents; An encapsulation unit, connected to the policy matching unit, is used to encapsulate an instruction data packet containing the current main interactive agent identifier, the identifier of at least one auxiliary interactive agent, the main interactive policy, and the auxiliary interactive policy, as the coordination instruction.

10. The system according to claim 8, characterized in that, The interactive execution module includes: The parsing unit is used to parse the coordination instruction and obtain the current main interactive agent identifier and the main interaction strategy; A loading unit, connected to the parsing unit, is used to call the corresponding current main interactive agent according to the current main interactive agent identifier and load the main interactive strategy, which includes a dialogue template, a question sequence and expected skill operation steps. An interaction control unit, connected to the loading unit, is used to control the current main interaction agent loaded with the main interaction strategy to conduct multi-round dialogues or guided operations with the learner, and to record the learner's answers, operation sequences, response times and emotional feedback signals during the interaction process, and to integrate the answers, operation sequences, response times and emotional feedback signals into the interaction result.

Images