An intelligent quality inspection effect optimization and evaluation system based on multi-agent cooperation

Abstract

The application relates to the technical field of artificial intelligence, and particularly discloses an intelligent quality inspection effect optimization and evaluation system based on multi-agent cooperation, an education marketing world model is constructed to form a deducible marketing state space, a dynamic graph agent cooperation network corresponding to a current consultation task is generated, the type, quantity, connection relationship and execution sequence of participating agents are adaptively determined, multiple functional agents generate multiple groups of candidate marketing strategies through self-game mechanism iteration in the dynamic graph network, the world model performs digital twin deduction on the future multi-round consultation process of each strategy, and the transaction probability, complaint and violation risk, refund risk and customer value are predicted, a decision agent selects a target marketing strategy according to the multi-objective optimization result, and outputs real-time marketing guidance, quality inspection results and evidence chains; the application realizes adaptive correction of the world model and closed-loop evolution of the agent cooperation topology through the deviation feedback of real execution results and twin prediction results, and is suitable for online education marketing and customer service quality inspection scenes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent quality inspection systems have been widely applied to service quality management in industries such as online education, customer service, and e-commerce retail. In the field of education marketing, intelligent quality inspection systems need to conduct real-time or offline quality checks on communication dialogues between planners, tutors, and parents to identify any improper marketing, over-promises, violations of advertising laws and national policies (such as the "double reduction" policy). However, the education marketing environment is highly dynamic—regulatory policies are constantly updated, market demands are constantly changing, and user profiles are evolving in real time—which places extremely high demands on the adaptive capabilities of intelligent quality inspection systems.

[0003] Currently, intelligent quality inspection systems mainly adopt the following three types of technical solutions: (I) Quality Inspection Method Based on Rule Engine These methods match and detect customer service conversations using pre-defined keywords and regular expressions. For example, keywords like "promise," "gift," and "compensation" can be used to detect over-promising behavior, and keyword matching can identify inappropriate language. While simple to implement and efficient, these methods have weak semantic understanding capabilities, resulting in low coverage and a high false negative rate for complex, unstructured quality control items. More importantly, when business rules or regulations change, manual updates are required, leading to long response times, high maintenance costs, and continuous performance degradation as the environment changes.

[0004] (ii) Quality inspection methods based on a single model These methods use a single machine learning or deep learning model to classify and score dialogues. Compared to rule engines, they offer improved semantic understanding, but a single model struggles to cover multiple types of quality control items simultaneously (such as attitude detection, commitment detection, and advertising law violation detection), and the model update cycle is long, making it difficult to adapt to rapid changes in business rules.

[0005] (III) Quality Inspection Method Based on Large Language Model and Multi-Agent Collaboration In recent years, constructing multi-agent collaborative systems using large language models has become a research hotspot. The following are relevant prior art documents: Comparative document 1 (CN121523815A) – “Software Multi-Agent Collaborative Method and System Based on Large Language Model” discloses a software multi-agent collaborative method based on a large language model. This method uses a pre-trained large language model central hub for semantic understanding and intent recognition, generating a structured task element set. Based on the topology and resource requirement characteristics of the task dependency graph, a two-layer graph attention network is used to dynamically match specialized agents with specific domain capabilities, forming a distributed collaborative network. When execution conflicts are detected between specialized agents, the large language model central hub generates a conflict resolution strategy based on the global state context. However, this scheme still has the following shortcomings: Its dynamic matching is based solely on the static topology of the task dependency graph, lacking a real-time perception and response mechanism for environmental changes. It cannot adaptively adjust the overall topology of the agent collaboration network according to the dynamic evolution of the external environment (such as policy changes and market fluctuations). It also lacks a systematic evaluation and closed-loop feedback optimization mechanism for the multi-agent collaboration effect, only recording collaboration process data for system adaptive optimization, failing to form a complete closed loop of "evaluation → feedback → reconstruction". Its application scenarios are limited to task collaboration in the field of software development, and do not involve the dynamic regulatory environment and multi-dimensional compliance requirements unique to the field of education marketing quality inspection.

[0006] Prior art document 2 (US 2026 / 0057331 A1) – “MULTI-AGENT SIMULATION SYSTEM ANDMETHOD” – discloses a multi-agent simulation system. This system uses a first language model to simulate a sales agent generating sales scripts, a second language model to simulate a customer responding to the dialogue, and a third language model to simulate a reviewer scoring the dialogue. This multi-agent simulation is used to create a digital twin of the real dialogue between the sales agent and the customer. However, this approach has the following significant drawbacks: the roles and interaction topologies of each agent are preset and fixed (a linear chain structure of sales agent → customer agent → reviewer agent), lacking the ability to dynamically adjust the agent type, number, and collaborative relationships based on task complexity; its digital twin is only used for post-event scoring simulation of “already occurred” dialogues, lacking the ability to prospectively extrapolate and multi-path search for future multi-round consultation processes; and its reviewer agent only outputs a single score, failing to provide multi-dimensional quality control results and an interpretable chain of evidence. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide an intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration, which solves the problems of existing intelligent quality inspection systems lacking real-time perception and adaptive capabilities to dynamically changing educational marketing environments, resulting in rigid quality inspection rules, fixed decision-making collaboration structures, inability to proactively assess marketing strategy risks, and continuous degradation of system performance as the environment and business change.

[0008] This invention provides an intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration, comprising: S1: Construct a world model for education marketing, integrating user profiles, course information, historical transaction cases, marketing rules and regulatory policies to form a predictable marketing state space; S2: Generate a dynamic graph agent collaboration network corresponding to the current consulting task based on the world model, and adaptively determine the type, number, connection relationship and execution order of the participating agents; S3: Based on the marketing generation agent, marketing quality inspection agent, regulatory constraint agent and transaction prediction agent, multiple sets of candidate marketing strategies are generated iteratively in a dynamic graph network through a self-game mechanism, and the output of each agent is adjusted in real time. S4: Using a world model, conduct multiple rounds of future consultations for each candidate marketing strategy to predict the probability of closing a deal, complaint risk, violation risk, refund risk, and long-term customer value. S5: Based on the comprehensive multi-objective optimization results of the decision agent, select the target marketing strategy and output real-time marketing guidance, quality inspection results and evidence chain.

[0009] Compared with the prior art, the present invention has the following beneficial effects: This invention first constructs an educational marketing world model, establishing a dynamic evolutionary relationship between the environment layer, state layer, behavior layer, and result layer. This allows the system to perceive changes in the educational marketing environment in real time. Then, based on feedback from the deviation between the actual execution results and the digital twin prediction results, the state transition matrix is ​​adaptively corrected layer by layer, completing the self-evolutionary update of quality inspection rules and the state space. This solves the rigidity problem in existing technologies where quality inspection rules rely on manual pre-setting and cannot dynamically change with the environment. Secondly, this invention decouples tasks based on the state complexity of the current marketing task. An initial Agent collaboration graph is generated based on the dynamic matching degree between Agent capability vectors and sub-task requirements. Then, a topology evolution event is triggered based on the collaboration entropy value of real-time interaction data between Agents. Redundant edges are deleted and enhanced Agents are added based on the marginal contribution value of the connecting edges. Simultaneously, an atomic snapshot mechanism is used to achieve seamless hot migration of the topology structure. This invention enables multi-agent collaboration topology to adaptively adjust based on task complexity, historical collaboration success rate, and real-time decision contribution, overcoming the structural defects of existing technologies where agent collaboration relationships are pre-fixed and cannot be self-optimized with the task. Finally, this invention constructs a digital twin environment corresponding to the real marketing environment, automatically generating multiple counterfactual marketing strategies with the target marketing strategy as the initial path. It then constructs multi-layered marketing decision search trees for multiple rounds of recursive expansion and state search, proactively and comprehensively evaluating multi-dimensional indicators such as transaction probability and violation risk. Simultaneously, it introduces a Pareto multi-objective evaluation mechanism to perform dynamic pruning and retain the optimal candidate path. This allows the system to perform multi-path compliance risk and marketing benefit deduction and optimization before the actual execution of the marketing strategy, changing the passive mode of existing quality inspection systems that can only determine after the fact, and realizing a shift from post-event detection to pre-event deduction and prevention. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, and these are all within the protection scope of the present invention.

[0011] Figure 1 This is a flowchart of an intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. In the description of the present invention, it should be understood that the terms "center," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements, but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, the element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Where there is no conflict, embodiments of the present invention and the various features thereof can be combined with each other, all of which are within the scope of protection of the present invention.

[0013] Example 1 Please see Figure 1 This invention provides an intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration, comprising: S1: Construct a world model for education marketing, integrating user profiles, course information, historical transaction cases, marketing rules and regulatory policies to form a predictable marketing state space; Specifically, the system collects multi-source heterogeneous data from multiple data sources during the education marketing process. This data includes at least the consultation dialogue texts of parents or users, user profile information (such as child's grade, learning pain points, and spending power), course product information (course system, pricing schemes, promotional activities, etc.), historical transaction case data, marketing rule manuals, and education industry regulatory policy documents (such as the Advertising Law and the "double reduction" policy). After uniformly cleaning and formatting the above data, natural language processing technology is used to identify various marketing behavior events (including course consultation, price inquiry, trial class reservation, objection handling, and transaction confirmation). An event association network is established based on the chronological order, contextual reference relationships, and marketing goal correlations between events, generating an education marketing state diagram. Each node in this state diagram represents a marketing state, and each marketing state includes at least the current marketing stage, user demand tags, course matching degree, transaction intention level, and marketing risk level; the directed edges between nodes represent the evolution direction between states. Based on this, the frequency of transitions between states is statistically analyzed to establish a state transition matrix. Combined with historical transaction cases, the probability of a transaction, the probability of violation, the probability of a complaint, and the probability of user churn corresponding to each state are calculated, forming a marketing state prediction parameter set. Finally, a complete educational marketing world model is constructed based on the above marketing state diagram, state transition matrix, and state prediction parameters. Logically, this world model includes an environment layer (describing external constraints such as policies, regulations, and market competition), a state layer (describing the user's current marketing stage and demand characteristics), a behavior layer (defining the set of executable marketing behaviors in each state), and a result layer (stores historical result data after different behaviors are executed). These four layers form a dynamic evolutionary relationship, enabling the prediction of future state evolution results given the current state and marketing behavior. Simultaneously, the world model provides a unified operating interface for each functional agent, including a state reading interface, a behavior execution interface, and a result feedback interface. This allows each agent to operate collaboratively within the world model environment, using marketing state as input, marketing behavior as output, and marketing results as feedback.

[0014] S2: Generate a dynamic graph agent collaboration network corresponding to the current consulting task based on the world model, and adaptively determine the type, number, connection relationship and execution order of the participating agents; Specifically, the system reads the current marketing state node from the world model, extracting information such as user demand tags, course matching degree, marketing risk level, sales intention level, and marketing stage contained within that node. It then performs state complexity analysis on the current marketing task to identify key factors influencing subsequent marketing decisions. Subsequently, based on the complexity analysis results, the system decouples the current marketing task, dividing the overall marketing objective into multiple interconnected marketing sub-tasks, including but not limited to user demand analysis, course recommendation, marketing compliance analysis, marketing script generation, sales prediction, and marketing risk assessment. The system reads the capability descriptions of each agent from the Agent capability resource pool, calculates the capability matching degree between each agent and sub-task based on their capability requirements, and maps sub-tasks to agents based on this matching degree. Building upon this, based on the data dependencies, temporal constraints, and capability collaboration relationships among the sub-tasks, a dynamic connection relationship is established between multiple agents, generating a dynamic graph agent collaboration network corresponding to the current marketing state. Nodes in this network represent functional agents, and edges between nodes represent information interaction, execution dependencies, and feedback relationships between agents. Subsequently, the system configures shared context, running priority, input / output interfaces, long-term memory access permissions, and collaborative communication strategies for each Agent, and initializes the shared state space among Agents. Furthermore, after the dynamic graph Agent collaborative network is initialized, the network undergoes topology self-adjustment based on historical collaboration success rates, historical decision contributions, and current marketing status changes among Agents. This involves dynamically adding, deleting, or reconnecting Agent relationships to form a self-evolving Agent collaborative network that adapts to changes in the current marketing status.

[0015] S3: Based on the marketing generation agent, marketing quality inspection agent, regulatory constraint agent and transaction prediction agent, multiple sets of candidate marketing strategies are generated iteratively in a dynamic graph network through a self-game mechanism, and the output of each agent is adjusted in real time. Specifically, the dynamic graph agent collaborative network includes at least a marketing generation agent, a marketing quality inspection agent, a regulatory constraint agent, and a sales prediction agent. First, the marketing generation agent generates multiple candidate marketing strategies based on the current marketing state in the world model. Each candidate strategy corresponds to a different course recommendation method, marketing script combination, discount strategy, and communication path. Then, the system controls each functional agent in the dynamic graph agent collaborative network to independently analyze each candidate marketing strategy. The marketing quality inspection agent outputs evaluation results from a service quality dimension, the regulatory constraint agent outputs evaluation results from a policy compliance dimension, and the sales prediction agent outputs evaluation results from a conversion probability dimension. Each agent generates corresponding strategy modification suggestions based on its evaluation results and feeds these suggestions back to the marketing generation agent. The marketing generation agent readjusts the candidate marketing strategies based on the received modification suggestions and resends them to the functional agents for re-evaluation. The process of strategy generation, evaluation, and correction described above is executed iteratively. In each iteration, the system calculates the current degree of collaborative convergence based on the consistency of the output results of each Agent, the magnitude of evaluation changes, and the changes in overall returns. When the degree of collaborative convergence reaches a preset convergence threshold, the iteration stops, and the target marketing strategy optimized through multiple rounds of self-game is determined. Simultaneously, the system records the information interaction content, strategy correction trajectory, and evaluation change trajectory between Agents during each round of negotiation, generating corresponding Agent collaborative decision-making trajectories and storing them in the long-term memory module for subsequent decision-making reference.

[0016] S4: Using a world model, conduct multiple rounds of future consultations for each candidate marketing strategy to predict the probability of closing a deal, complaint risk, violation risk, refund risk, and long-term customer value. Specifically, the system first replicates the current marketing state in a world model, constructing a digital twin marketing environment corresponding to the real marketing environment. This twin environment maintains consistent user profiles, course information, marketing stages, and constraints with the real environment. Using the target marketing strategy as the initial marketing path, and combining it with the world model, multiple counterfactual marketing strategies are automatically generated. Each counterfactual strategy employs different course recommendation methods, marketing script combinations, discount schemes, and user communication paths, forming multiple marketing scenarios to be simulated. Then, using each simulated marketing scenario as an initial node, the system performs multi-step state simulations of the subsequent marketing process based on the state transition matrix. A multi-layered marketing decision search tree is constructed based on the transition relationships between states, performing multiple rounds of path expansion and state search for each marketing path to obtain the future marketing state sequence corresponding to each path. The system calculates the comprehensive marketing revenue for each corresponding marketing path based on the future marketing state sequence. This comprehensive marketing revenue includes at least multiple dimensions such as course conversion probability, violation risk, complaint probability, refund probability, long-term customer value, brand influence value, and user satisfaction, and establishes a comprehensive revenue evaluation result for each path.

[0017] S5: Based on the comprehensive multi-objective optimization results of the decision agent, select the target marketing strategy and output real-time marketing guidance, quality inspection results and evidence chain; Specifically, the decision agent obtains the comprehensive benefit evaluation results of each marketing path. These results include at least the predicted conversion probability, violation risk level, complaint probability, refund probability, and long-term customer value for each path. The decision agent optimizes multiple objectives—maximizing conversion probability, minimizing violation risk, maximizing long-term customer value, and optimizing brand safety—using a multi-objective decision algorithm to comprehensively rank candidate paths and select the target marketing strategy with the optimal comprehensive benefit. After selection, the decision agent generates three types of output: The first is real-time marketing guidance, including the optimal recommended script, recommended courses, recommended discounts, and suggestions for the next communication strategy, directly pushed to the customer service or planner's interface for reference; the second is quality inspection results, including compliance judgments of the current marketing scripts and communication strategies, violation risk point markings, and quality inspection scores, used to guide real-time compliance adjustments to marketing behavior; the third is a chain of evidence, including the original text of the specific script that triggered the quality inspection judgment, contextual references, applicable legal clauses or marketing rules, explanations of the judgment basis, and suggested correction directions, for management review and customer service personnel to learn and improve.

[0018] Further, step S1 includes the following steps: S11: Collect multi-source heterogeneous data in the education marketing process. The multi-source heterogeneous data includes at least parent consultation dialogue data, student basic information data, course product information data, marketing rule data, education industry regulatory rule data, and historical transaction case data. The data is then standardized according to a unified data format to form the original marketing dataset. Specifically, the system connects to multiple data source interfaces, including a customer service dialogue record database, a user information management system, a course product management system, a marketing rules knowledge base, a policy and regulation database, and an order transaction database. It extracts complete records of parent consultation dialogues, student basic information, course product information, marketing rules manual content, education industry regulatory policy documents, and historical transaction cases from these data sources. Since the data originates from different systems, their data formats, encoding methods, and field definitions differ. Therefore, the system first performs format unification processing on the data from each data source: for structured data, its field names and data types are mapped to a unified data model; for semi-structured data, a parser extracts key terms and structured fields; and for unstructured data, its original text content is preserved and metadata tags are attached.

[0019] S12: Perform semantic parsing on the original marketing dataset to identify the corresponding marketing behavior events during the consultation process, and generate corresponding marketing event nodes according to the preset marketing event types. The marketing behavior events include at least course consultation events, price inquiry events, trial class reservation events, course comparison events, objection handling events, transaction confirmation events, and after-sales consultation events. Specifically, the system utilizes a natural language processing engine to perform deep semantic analysis on the dialogue text in the original marketing dataset. This involves sentence and word segmentation, followed by intent classification and slot filling for each sentence using a pre-trained semantic understanding model. Intent classification identifies the type of marketing behavior expressed in the current statement. The system pre-defines seven types of marketing behavior events: course inquiry, price inquiry, trial lesson reservation, course comparison, objection handling, transaction confirmation, and after-sales consultation. Slot filling extracts key parameters related to the marketing behavior from the statement, such as course name, price, reservation time, and comparison object. Each identified marketing behavior event is assigned a corresponding event node, along with a unique event identifier, a timestamp recording the event's occurrence, the associated session ID, and the extracted key parameter information. All event nodes are stored in a structured format in the system's graph database.

[0020] S13: Establish an event association network based on the time sequence, contextual reference relationship and marketing goal association relationship between each marketing event, calculate the state transition relationship between each marketing event, and generate an education marketing state diagram. Each marketing state node includes at least the current marketing stage, user demand tag, course matching degree, transaction intention level and marketing risk level. Specifically, based on the timestamps, consultation session identifiers, and marketing stages corresponding to each marketing event, the time sequence of all event nodes is reconstructed to generate a marketing event time series. Events within the same consultation session are arranged sequentially according to their actual occurrence order, while different consultation sessions are linked by establishing continuous time relationships based on the historical consultation trajectory of the same user, enabling the system to track all marketing events experienced by a user throughout the entire consultation lifecycle.

[0021] For each event node, its semantic content vector is extracted and its similarity is calculated with the semantic vectors of all other event nodes in the same user's historical session. When two event nodes have a clear referencing or successive relationship in semantic content (e.g., a price inquiry event in the current session and a course recommendation event in a historical session mention the same course name or the same price), the system establishes a contextual reference edge between the two nodes. This reference edge is accompanied by a reference type label (e.g., direct reference, indirect reference, or topic continuation) and a reference strength weight. Through the above analysis, the system forms a cross-session event association structure covering the entire consultation cycle.

[0022] The system performs goal correlation analysis based on the marketing objective type corresponding to each marketing event. Each event node is labeled with its corresponding marketing objective dimension, which includes course recommendation objectives, sales promotion objectives, user trust building objectives, and marketing risk control objectives. The system calculates the degree of correlation between each pair of event nodes across each objective dimension, generating a correlation strength value between each event node and each marketing objective. Based on this, a marketing objective correlation network is established, enabling the system to identify which event sequences collectively serve the same marketing objective.

[0023] Based on the three types of relationships mentioned above, a state aggregation operation is performed. Using a clustering algorithm based on semantic similarity and goal consistency, multiple marketing behavior events with the same marketing semantics, user intent, and marketing goals are aggregated into a single marketing state node. The purpose of state aggregation is to abstract fine-grained event sequences into macro-level marketing stages, with each state node representing a stable marketing psychological or behavioral stage that a user is in over a continuous period of time. The system assigns multi-dimensional attributes to each aggregated state node, including the current marketing stage identifier (e.g., awareness stage, consideration stage, decision-making stage, after-sales stage), a set of user demand tags, a course matching score, a transaction intention level, and a marketing risk level.

[0024] S14: Based on the marketing state diagram, perform statistical analysis on the transition probabilities between each state node, establish a marketing state transition matrix, and combine historical transaction cases to calculate the transaction probability, violation probability, complaint probability and user churn probability corresponding to each state, forming marketing state prediction parameters. Specifically, directed transition edges and their corresponding transition frequencies are extracted from the state graph. The state transition frequency is divided by the total outgoing edge frequency of that state to calculate the conditional transition probability from each state to all its possible subsequent states. The conditional transition probabilities between all states are arranged in rows (each row corresponds to a current state, and each column corresponds to a subsequent state) to construct a complete marketing state transition matrix. The sum of the elements in each row of this matrix is ​​1, representing the probability distribution of transitioning from that state to all possible subsequent states.

[0025] While constructing the transition matrix, the system calculates marketing prediction parameters for each state node by combining historical transaction case data. For each state node in the state graph, the system retrieves all complete marketing session records that passed through that state from historical data, and calculates the percentage of sessions that ultimately resulted in a sale, the percentage of customer complaints, the percentage of sessions marked as having violations by the quality inspection system, and the percentage of refunds after a sale. These four percentages are respectively used as the probability of sale, the probability of complaint, the probability of violation, and the probability of user churn for that state. A prediction parameter vector is generated for each state node, containing the probability of sale, the probability of violation, the probability of complaint, and the probability of user churn for that state. The prediction parameter vectors of all state nodes constitute a complete marketing state prediction parameter set.

[0026] S15: Based on the marketing state diagram, marketing state transition matrix and marketing state prediction parameters, construct an educational marketing world model, and use each marketing state in the world model as the operating environment of the dynamic graph agent network; Specifically, the education marketing state diagram serves as the state space skeleton of the world model, the marketing state transition matrix serves as the dynamic rule base for state evolution, and marketing state prediction parameters serve as the evaluation basis for state value. These three elements together constitute the core logical layer of the world model. Based on this, a multi-layered structure of the world model is further constructed to achieve a comprehensive digital mapping of the real marketing environment. An environment layer is constructed to collect and store external constraints on the current operation of education marketing tasks, including but not limited to currently effective course pricing schemes, existing promotional policies, applicable regulatory requirements such as advertising laws and laws protecting minors, market competition information, and complete user profiles. The parameters of the environment layer directly affect the effective scope of each state node and the legality of transition conditions in the state layer. For example, when the "double reduction" policy is implemented, the environment layer automatically updates the regulatory constraint parameters, and state nodes in the state layer involving weekend subject-based training recommendations and their associated transition edges are marked as unavailable.

[0027] A marketing state layer is constructed, and each state node in the state diagram is standardized and encoded, so that each state node is represented by a feature vector of fixed dimensions. The dimensions of the feature vector include at least the user demand feature sub-dimension, the marketing goal sub-dimension, the user psychological state sub-dimension, and the transaction stage sub-dimension. The standardized state feature vectors facilitate rapid state matching and similarity calculation for each functional agent in the subsequent decision-making process. At the same time, the system establishes an environmental constraint relationship mapping between the state layer and the environment layer to determine the effective range of each marketing state under different combinations of environmental parameters.

[0028] A marketing behavior layer is constructed. Based on the marketing objectives corresponding to each marketing state, a set of executable marketing behaviors is configured for each state node. These marketing behaviors include at least course recommendations, trial lesson reservations, promotional pushes, case studies, faculty introductions, customized learning plans, and objection handling. The system defines the mode of action for each behavior, including the expected duration of the behavior, the expected direction of its impact on the user's psychological state, and the expected adjustment range of the state transition probability. The system establishes the state interaction relationship between the behavior layer and the state layer, enabling the system to deduce the distribution of the next state based on the current state and the executed behavior through the state transition matrix.

[0029] A marketing results layer is constructed, and a results mapping table is established based on historical marketing results data. This table associates the historical results of each marketing activity executed under different states (including course sales results, user satisfaction ratings, refund results, complaint results, course renewal results, and marketing violation results) with corresponding (state, behavior) key-value pairs. The system establishes a mapping relationship between the behavior layer and the results layer, enabling the system to retrieve the average distribution of results after executing the behavior historically, given the current state and the behavior to be executed.

[0030] A dynamic evolutionary relationship is established between the aforementioned environment layer, state layer, behavior layer, and result layer: the environment layer constrains the effective scope of the state layer, the state layer, in conjunction with the behavior layer, drives state transitions, the transition results are mapped to the result layer, and the statistical data of the result layer inversely corrects the prediction parameters of the state layer. These four layers form a complete dynamic closed loop, collectively constituting a complete educational marketing world model.

[0031] Specifically, three types of standard interfaces are opened to each functional Agent: a marketing status reading interface, which allows the Agent to query the complete attribute information of the current marketing status; a marketing behavior execution interface, which allows the Agent to submit the marketing behavior to be executed, and the world model calculates the state transition result based on the current state and behavior parameters and returns it; and a marketing result feedback interface, which allows the Agent to feed back the execution results observed in the real marketing environment to the world model, which is used to update the state prediction parameters and result mapping table.

[0032] Furthermore, the step of establishing an event association network based on the temporal sequence, contextual reference relationships, and marketing objective correlations among various marketing events, calculating the state transition relationships among these events, and generating an educational marketing state diagram includes the following steps: S131: Reconstruct the time sequence of all marketing events according to their occurrence time, consultation session, and marketing stage to generate a marketing event time sequence. Marketing events within the same consultation session are arranged in the actual order of occurrence, and continuous time relationships are established between different consultation sessions based on the user's historical consultation trajectory. Specifically, the system reads all marketing event nodes, each recording a timestamp, a unique identifier for its associated consultation session, and a marketing stage label. Using users as the dimension, all events belonging to the same user are globally sorted according to the order of their timestamps. During the sorting process, events within the same consultation session are strictly arranged in chronological order of their actual occurrence. For events belonging to different consultation sessions, the system arranges them according to the chronological order of the sessions, and simultaneously establishes continuous temporal relationships between events across sessions based on the user's continuous behavioral patterns across different sessions.

[0033] S132: Based on the semantic content, historical reference information and contextual semantic similarity of each marketing behavior event, perform reference association analysis on marketing behavior events with cross-session reference relationships, establish contextual reference edges between marketing behavior events, and form a marketing event association structure across consultation cycles; Specifically, based on the time sequence, the system identifies the referencing relationships between different sessions and extracts the semantic content feature vector for each marketing event node. This feature vector is obtained by jointly encoding the event's text description, key parameters, and event type label. Subsequently, the system calculates the cosine similarity of the semantic feature vectors of each pair of event nodes belonging to different consultation sessions and performs a comprehensive score based on the time interval between events and historical referencing patterns. When the comprehensive score between two events exceeds a preset referencing threshold, the system determines that a cross-session referencing relationship exists between them, i.e., a contextual referencing edge is established between the two event nodes. For example, if a user inquires about the "XX English summer course" schedule in the first session, the system generates a course consultation event; three days later, in the second session, the user mentions "that summer course I mentioned last time" again, and the system generates another course consultation event and identifies a clear referencing relationship between this event and the event in the first session. A contextual referencing edge is then established between the two, and the referencing type is labeled as "direct continuation."

[0034] S133: Based on the marketing objective type, course recommendation objective, sales promotion objective, user trust building objective, and marketing risk control objective corresponding to each marketing activity event, calculate the degree of objective correlation between the marketing activity event and each marketing objective, and establish a marketing objective correlation network; Specifically, five marketing goal dimensions are preset: course recommendation goal, sales promotion goal, user trust building goal, marketing risk control goal, and user needs mining goal. For each marketing event node, the system calculates the association strength value between the event and the above five goal categories based on the event's type tag, semantic content, and the actual effect of executing the event in historical marketing processes. The calculation of the association strength value comprehensively considers three factors: the semantic similarity between the event's semantics and the goal definition, the actual contribution of the event to goal achievement in historical data, and the rule weights preset by experts, and finally normalizes it to an association strength score between 0 and 1. After completing the goal association calculation for all event nodes, event nodes with the same goal association pattern are grouped into the same goal cluster, and the node connection relationships within each goal cluster are established to form a marketing goal association network.

[0035] S134: Based on the marketing event time sequence, context reference edges and marketing goal association network, aggregate the states of multiple marketing behavior events with the same marketing semantics, the same user intent and the same marketing goal to generate corresponding marketing state nodes. Each marketing state node includes at least user demand characteristics, course focus direction, transaction intention level, marketing risk level and user psychological state. Specifically, by integrating the three types of relationships established in the preceding steps—temporal proximity provided by marketing event time series, cross-session semantic relationships provided by contextual reference edges, and goal consistency relationships provided by the marketing goal association network—a comprehensive correlation metric is constructed between events. The system employs a graph clustering-based algorithm to group event nodes with a comprehensive correlation exceeding a preset clustering threshold into the same event cluster. Multiple marketing behavior events within each event cluster are temporally continuous or adjacent, semantically consistent, and serve the same marketing objective.

[0036] S135: Statistically analyze the continuous transitions between marketing state nodes in the historical marketing process, calculate the frequency of state transitions, duration of state transitions, and stability of state transitions between marketing state nodes, and establish marketing state transition relationships based on the state transition results. Specifically, the system iterates through all historical marketing session records and arranges each marketing state node in chronological order according to the user's actual consultation path, forming multiple state sequence samples. The system counts the number of times a transition occurs from state node A to state node B across all samples, denoted as the transition frequency. Simultaneously, the system calculates the average duration of each state node before being transitioned to the next state, as well as the stability coefficient for transitions from that state to subsequent states. The stability coefficient is determined by calculating the proportion of the transition frequency from that state to a certain subsequent state to the total outgoing edge transition frequency of that state, and combining this proportion with the variance over different time intervals; a higher proportion and a smaller variance result in a higher stability coefficient.

[0037] S136: Establish an education marketing state diagram based on marketing state nodes and marketing state transition relationships. Each marketing state node represents a dynamic marketing state in the marketing process, and the connecting edges between each state node represent the evolution direction between marketing states. Each connecting edge is associated with the corresponding state transition probability, marketing revenue change trend, and marketing risk change trend, so as to form an education marketing state diagram that can be used for digital twin simulation. Specifically, a basic state graph structure is established using marketing state nodes as the vertex set and state transition relationships as the directed edge set, associating multi-dimensional evolutionary information with each directed edge. The statistically obtained transition frequency is normalized into a transition probability value, serving as the first weight attribute of the edge. Changes in average order value, user satisfaction, and user lifetime value observed in subsequent marketing processes after state transitions along this path in historical data are comprehensively calculated as the marketing revenue trend value, serving as the second weight attribute of the edge. The average proportion and severity of quality inspection violations triggered in cases of state transitions along this path are used as the marketing risk trend value, serving as the third weight attribute of the edge.

[0038] After completing the above associations, a complete education marketing state diagram is output. In this state diagram, nodes represent identifiable dynamic marketing states during the marketing process. Node attributes include user demand characteristics, course focus, transaction intention level, marketing risk level, and user psychological state. Directed edges represent the evolution direction between marketing states. Edge attributes include state transition probability, marketing revenue change trend, and marketing risk change trend.

[0039] Further, step S15 includes the following steps: S151: Read each marketing state node and marketing state transition relationship in the marketing state diagram, and establish a marketing state space based on the state transition relationship between marketing state nodes. Each marketing state node represents the current marketing state of the user in the educational marketing process, and the connection relationship between each marketing state node represents the evolution direction between marketing states. Specifically, the system reads all marketing state nodes and their directed transition relationships from the state graph database, mapping each state node to a discrete state in the state space. The system traverses all nodes in the state graph, extracting the unique identifier and complete attribute information of each node to form a set of vertices in the state space; at the same time, it reads the directed edges between each node and their associated transition probabilities to form a set of transition relationships in the state space.

[0040] S152: Collect course information, enrollment policies, marketing rules, education supervision requirements, market competition information and user profile information corresponding to the current education marketing task, and construct a marketing environment layer based on the environmental information to form a set of environmental states describing the operating environment of the current education marketing task. Specifically, the system retrieves currently valid course product information from the course management system; current enrollment deadlines, enrollment quotas, and special enrollment policies from the enrollment policy database; currently effective marketing script guidelines, pricing rules, and promotional activity rules from the marketing rules knowledge base; currently applicable education industry regulatory policy clauses from the regulatory rules database, including restrictive clauses of the Advertising Law regarding education and training advertising, relevant clauses of the Law on the Protection of Minors, and specific requirements for the scheduling and fee management of subject-based training in the "double reduction" policy; competitor price benchmarking data and industry average conversion rate reference data from the market intelligence module; and current user demographic characteristics, historical purchase records, preference tags, and spending power levels from the user profiling system.

[0041] S153: Based on the marketing state nodes in the marketing state diagram, standardize the coding of marketing states with the same user demand characteristics, marketing goals, user psychological state and transaction stage, construct a marketing state layer, and establish the environmental constraint relationship between the marketing state layer and the marketing environment layer to determine the effective range of each marketing state under different marketing environments. Specifically, each attribute is discretized into a fixed-dimensional feature vector: user demand characteristics are mapped to demand type codes; course focus is mapped to course category codes; transaction intention levels are mapped to high / medium / low three-level code values; marketing risk levels are mapped to corresponding risk scores; and user psychological states are mapped to emotional polarity and intensity values. Standardized coding enables unified similarity calculation and state matching between different state nodes.

[0042] After completing the state coding, an environmental constraint relationship is established between the state layer and the environment layer constructed in step S152. The environmental constraint relationship defines the effective range of each marketing state under different combinations of environmental parameters. For example, when the regulatory policy parameter in the environment layer indicates "The 'double reduction' policy regarding the ban on weekend subject training has taken effect," the marketing state node involving weekend subject course recommendations in the state layer is marked as unavailable and removed from the available state space; when the promotional activity parameter in the environment layer indicates "There is currently a discount promotion," the state nodes involving price inquiries and discount applications in the state layer are given a higher effective priority.

[0043] S154: Based on the marketing objectives corresponding to each marketing state, configure a set of corresponding marketing behaviors for each marketing state. Each marketing behavior includes at least course recommendation, trial class reservation, discount push, case display, teacher introduction, learning plan and objection handling. Establish the state interaction relationship between marketing behaviors and marketing states to form a marketing behavior layer. Specifically, a predefined standard marketing behavior library is established. These standard marketing behaviors include at least course recommendations, trial lesson bookings, promotional offers, case studies, faculty introductions, customized learning plans, and objection handling. For each state node, the system selects behaviors from the standard marketing behavior library that match the user's needs and purchase intention level, forming a candidate behavior set for that state. Simultaneously, for each state, a state-interaction relationship is established between behaviors. This relationship defines the expected direction and magnitude of change in the user's state after executing the behavior in that state, including the expected change in the user's psychological state, the expected increase in purchase intention level, and the expected change in the probability of transition to subsequent states. Through this configuration and mapping, the system forms a complete behavior layer.

[0044] S155: Establish a marketing results layer based on historical marketing results, link the course transaction results, user satisfaction, refund results, complaint results, course renewal results and marketing violation results corresponding to each marketing behavior to the corresponding marketing behavior, and establish a result mapping relationship between the marketing behavior layer and the marketing results layer; Specifically, the system extracts all historical marketing session records from the historical marketing database. For each record, the system extracts the specific marketing behavior performed, the marketing status at the time of the behavior, and the actual marketing results observed after the behavior is performed. The actual marketing results include at least course sales results, user satisfaction ratings, refund results, complaint results, course renewal results, and marketing violation results.

[0045] S156: Based on the marketing state transition matrix and marketing state prediction parameters, perform state deduction on the marketing state change process after the marketing behavior takes effect, generate prediction results of the future marketing state under the corresponding marketing behavior, and establish dynamic evolution relationship between the marketing environment layer, marketing state layer, marketing behavior layer and marketing result layer to form an educational marketing world model. Specifically, the marketing state transition matrix serves as the core calculation engine for state evolution, and marketing state prediction parameters are used as the basis for state value assessment. The system's deduction and calculation process is as follows: Given the current state S and the marketing behavior A to be executed, the system queries the state interaction relationship corresponding to (S, A) in the behavior layer to obtain the parameters of the effect of behavior A on state S, including the adjustment coefficient of the state transition probability and the expected change in the user's psychological state; the behavior interaction parameters are superimposed on the transition probability distribution corresponding to state S in the state transition matrix to calculate the corrected transition probability distribution after the action of behavior A; then, the system samples and predicts the next state based on the corrected transition probability distribution to obtain the prediction result of the future state; the system queries the result mapping table corresponding to (S, A) in the result layer to obtain the expected result distribution after the execution of the behavior, and associates the predicted future state with the expected result distribution.

[0046] S157: Establish an operational interface between the education marketing world model and the dynamic graph agent collaborative network. Open marketing status reading interface, marketing behavior execution interface and marketing result feedback interface to each functional agent in the dynamic graph agent collaborative network, so that each functional agent can take marketing status as input, marketing behavior as decision output and marketing result as feedback information, and complete collaborative operation in the unified education marketing world model. Specifically, three types of standard interfaces are provided. The first type is the marketing status reading interface: each functional agent calls this interface, passing in the user identifier or session identifier as query parameters. The system returns complete attribute information of the current user's marketing status node, including all structured data such as status identifier, user demand characteristics, course focus, transaction intention level, marketing risk level, and user psychological state, enabling the agent to obtain complete environmental awareness information needed for decision-making. The second type is the marketing behavior execution interface: each functional agent calls this interface, passing in the current status identifier and the identifier of the marketing behavior to be executed as execution parameters. Based on the current status and behavior parameters, the system performs state extrapolation calculations in the world model and returns the predicted future state and expected return assessment, allowing the agent to try strategies and predict effects without affecting the real marketing environment. The third type is the marketing result feedback interface: After each functional agent observes the results of the behavior execution in the real marketing environment, it calls this interface to feed back the real execution result data to the world model. Based on the feedback data, the transition probabilities in the state transition matrix, the result distribution in the result mapping table, and the state interaction parameters are incrementally updated, so that the world model can continuously evolve and become more accurate with the accumulation of marketing practices.

[0047] Further, step S2 includes the following steps: S21: Read the state node corresponding to the current marketing state in the education marketing world model, and perform state complexity analysis on the current marketing task based on the user demand tags, course matching degree, marketing risk level, transaction intention level and marketing stage information contained in the state node, and identify multiple decision factors that affect the current marketing decision. Specifically, after normalizing the attribute values ​​for each dimension, the overall complexity score of that state node is calculated. The complexity score calculation comprehensively considers the number of attribute dimensions, the dispersion of values ​​for each dimension, and the mutual influence relationships between dimensions. When the overall complexity score exceeds a preset threshold, the current state is marked as a high-complexity state. Simultaneously, based on the values ​​of each attribute dimension, key factors influencing the current marketing decision are identified: if multiple conflicting needs exist in the user's demand tags, "demand priority ranking" is identified as a key decision factor; if the course matching degree is at a medium level and the transaction intention level is low, "course matching optimization" is identified as a key decision factor; if the marketing risk level is medium or high, "compliance risk control" is identified as a key decision factor; if the transaction intention level is high but the user is still hesitant, "transaction promotion strategy" is identified as a key decision factor.

[0048] S22: Based on the state complexity analysis results, the current marketing task is decoupled and the current marketing goal is divided into multiple interrelated marketing sub-tasks. The marketing sub-tasks include at least the user needs analysis sub-task, the course recommendation sub-task, the marketing compliance analysis sub-task, the marketing script generation sub-task, the sales prediction sub-task, and the marketing risk assessment sub-task. Specifically, taking the current overall marketing goal as the top-level objective, it is decomposed layer by layer into multiple interconnected but clearly defined marketing sub-tasks. Decoupling and matching are performed based on a pre-defined marketing sub-task type library, which includes at least user needs analysis, course recommendation, marketing compliance analysis, marketing script generation, sales prediction, and marketing risk assessment sub-tasks. Based on the complexity analysis of the current state, the set of sub-tasks to be activated is determined: if the state complexity is low, only the basic course recommendation and script generation sub-tasks are activated; if the state complexity is high and involves multiple key decision factors, all or most of the sub-tasks are activated. Each sub-task logically has a clear input-output relationship; for example, the output of the user needs analysis sub-task is the input of the course recommendation and script generation sub-tasks, and the outputs of the marketing compliance analysis and marketing risk assessment sub-tasks together serve as constraints for strategy selection.

[0049] S23: Read the capability description information corresponding to each Agent in the Agent capability resource pool, calculate the capability matching degree between each Agent and the marketing sub-task according to the capability requirements of each marketing sub-task, and complete the mapping relationship between each marketing sub-task and the corresponding Agent based on the capability matching degree. Specifically, the system maintains an Agent capability resource pool, in which multiple functional Agents are registered. Each functional Agent has corresponding capability description information, which includes at least the Agent's functional type label, applicable input data types, output result format specifications, historical accuracy, average inference latency, and the Agent's capability embedding vector. The system obtains the capability requirement vector for each marketing subtask, which describes the capability dimensions required to execute the subtask and the requirement intensity value in each dimension. The system calculates the cosine similarity between each Agent's capability embedding vector and each subtask's capability requirement vector, and uses the cosine similarity score as the preliminary capability matching degree between the Agent and the subtask.

[0050] S24: Based on the data dependency relationship, time sequence constraint relationship and capability synergy relationship between each marketing sub-task, establish a dynamic connection relationship between multiple Agents and generate a dynamic graph Agent synergy network corresponding to the current marketing state. In the dynamic graph Agent synergy network, the nodes represent the corresponding functional Agents, and the edges between each node represent the information interaction relationship, execution dependency relationship and feedback relationship between the corresponding Agents. Specifically, the system analyzes the data flow relationships between subtasks: If the output data of subtask A is necessary input data for the execution of subtask B, the system determines that there is a data dependency relationship between A and B, and the corresponding Agent A should provide data output to Agent B. The system then analyzes the temporal constraints between subtasks: If the execution of subtask B must wait for the completion of subtask A before it can begin, the system determines that there is a sequential temporal constraint between A and B, and the corresponding Agent A should execute before Agent B. Finally, the system analyzes the capability synergy relationships between subtasks: If the output results of subtask A and subtask B need to be jointly evaluated at the same decision node (e.g., compliance analysis results and risk assessment results jointly determine the feasibility of a strategy), the system determines that there is a capability synergy relationship between A and B, and the corresponding Agent should establish a bidirectional or star-shaped collaborative connection to facilitate cross-validation and joint judgment of results.

[0051] In this embodiment, based on the comprehensive identification results of the three relationships mentioned above, a dynamic connection relationship is established between Agents. For data dependency relationships, directed edges with arrows are established between the corresponding Agents, with the arrow pointing from the data provider to the data receiver, and the data transmission format and update frequency are marked on the edges. For time-series constraint relationships, this is achieved by configuring the execution order of Agents and the synchronization waiting mechanism, which are marked with time sequence numbers in the graph. For capability collaboration relationships, the system establishes undirected or bidirectional collaborative edges between the corresponding Agents, with the collaboration type and negotiation protocol marked on the edges.

[0052] S25: Based on the generated dynamic graph Agent collaborative network, configure the shared context, running priority, input / output interface, long-term memory access permission and collaborative communication strategy for each Agent, and complete the initialization of the shared state space between Agents so that Agents can complete collaborative operation under the unified education marketing world model; Specifically, a shared context is configured for each Agent. The shared context includes basic information of the current session, a complete attribute snapshot of the current state node, and a set of constraint parameters of the environment layer. All Agents can obtain consistent global information by accessing the shared context, avoiding the problem of inconsistent information caused by each Agent maintaining its own independent state.

[0053] Each Agent is configured with a running priority, determined based on its position and role in the network topology: Agents at the starting point of data flow have higher startup priority and need to generate basic output for downstream Agents; Agents at the decision endpoint have higher resource priority to ensure sufficient computing power support for their inference process. The system configures input / output interfaces for each Agent, specifying the format requirements for input data and the structural specifications for output data, ensuring correct data transmission and parsing between upstream and downstream Agents.

[0054] Each Agent is configured with long-term memory access permissions. This long-term memory module stores successful and failed cases, collaborative decision-making processes, and quality inspection records from the historical marketing process. The system grants different access scopes based on the functional needs of each Agent: for example, the script generation Agent can access the historical successful script library, the compliance review Agent can access the historical violation records and policy clause library, and the sales prediction Agent can access the status sequences of historical sales cases. Finally, the system configures collaborative communication strategies for each Agent, including communication protocols, message format standards, timeout retry mechanisms, and exception handling strategies.

[0055] The shared state space among the agents is initialized by allocating data structures for the state space in the shared memory region. Snapshot information of the current state node, environment layer constraint parameters, and initial input data for each agent are written into the shared state space, and the read / write pointers of each agent are set to the corresponding areas of the shared state space. Through these configuration and initialization operations, each agent can complete collaborative decision-making and execution based on its own runtime configuration within a unified shared state space.

[0056] S26: After the dynamic graph agent collaborative network is initialized, the dynamic graph agent collaborative network is topologically adjusted according to the historical collaboration success rate, historical decision contribution and current marketing status changes of each agent. The connection relationship between corresponding agents is dynamically added, deleted or reconnected to form a self-evolving agent collaborative network that adapts to the current marketing status changes. Specifically, the system monitors the collaborative interaction process between agents in real time and retrieves the historical collaboration success rate and historical decision contribution of each agent from the long-term memory module. Based on this historical data, the system calculates the overall performance score of each connection edge. When the overall performance score of two connection edges is consistently below a preset threshold, the system determines the connection relationship to be inefficient and marks it as a connection to be adjusted.

[0057] Meanwhile, the system continuously monitors changes in the current marketing status node. When the set of key decision factors identified in step S21 changes significantly, the system determines that the current task requirement structure has changed. In response to the change in the task requirement structure, the system re-executes the capability matching and connection relationship establishment process from steps S22 to S24 to identify whether a new Agent needs to be introduced or existing connection relationships need to be adjusted.

[0058] At the execution level of topology self-adjustment, the system supports three types of adjustment operations: First, the addition operation, which means that when a new subtask requirement arises, the system selects a matching new Agent from the Agent capability resource pool to establish a connection relationship and connect it to the existing network; second, the deletion operation, which means that when the overall performance score of certain connections is consistently below a threshold and the corresponding subtask is no longer active in the current state, the system deletes the connection relationship and switches the corresponding Agent to a standby state; and third, the reconnection operation, which means that when the data flow or collaborative relationship changes, the system disconnects the original connection edge and establishes a new connection edge, so that the structure of the collaborative network can reflect the latest task requirement relationship.

[0059] When performing the aforementioned topology adjustments, the system employs a smooth transition strategy. This means maintaining the stable operation of the existing network during the adjustment period, and gradually transferring decision-making power to the optimized network structure only after new connections are established and new agents have synchronized their states. This avoids decision interruptions or data loss due to abrupt topology changes. Through this self-adjusting mechanism, the dynamic graph agent collaborative network can adaptively evolve its topology based on changes in the current marketing status, always maintaining an optimal match between the network structure and current task requirements.

[0060] Furthermore, step 24 includes the following steps: S241: Obtain the data input items, data output items, target output results and capability requirements information corresponding to each marketing subtask, and calculate the data dependency degree between each marketing subtask based on the data input and output relationship between each marketing subtask, and establish a marketing subtask data dependency network. Specifically, the system acquires information on data inputs, data outputs, target outputs, and capability requirements for each marketing subtask. For each marketing subtask, the system defines the list of data inputs required for its execution and the list of data outputs generated upon completion. The system iterates through all marketing subtask pairs. For each pair (task i, task j), it calculates the intersection between the data inputs of task j and the data outputs of task i. The ratio of this intersection to the total number of inputs in task j is used as the data dependency value from task i to task j. When the data dependency value exceeds a preset dependency threshold, the system establishes a directed data dependency edge between task i and task j, pointing from the data provider to the data consumer, with the dependency strength weight marked on the edge.

[0061] S242: Based on the execution stage and execution order of each marketing subtask, calculate the pre-constraint relationship, post-constraint relationship and parallel execution relationship between each marketing subtask, and establish a marketing subtask timing constraint network. Specifically, the execution stage information and the predetermined execution order within each marketing subtask are obtained. For any two marketing subtasks A and B, the type of temporal constraint between them is determined based on their execution stage and predetermined execution order: if A must be completed before B starts, then a pre-constraint relationship exists between A and B; if B must be completed before A starts, then a pre-constraint relationship exists between B and A, and a post-constraint relationship exists between A and B; if A and B can be executed independently without waiting for each other's results, then a parallel execution relationship exists between them, and the corresponding functional agents can concurrently execute their respective reasoning processes; intra-stage synchronization constraints are established between multiple subtasks within the same execution stage, meaning that the next execution stage is only allowed after all subtasks within the stage are completed.

[0062] S243: Read the capability description information corresponding to each functional agent, calculate the degree of capability synergy between each functional agent based on the capability complementarity, capability coverage and capability conflict between each marketing sub-task, and establish a functional agent capability synergy network. Specifically, the system reads complete capability description information for each functional agent from the Agent capability resource pool. This capability description information includes the agent's functional type, preferred input data types, output formats, applicable marketing sub-task type tags, and capability embedding vectors. For any two functional agents, the system calculates the degree of capability complementarity between them based on their capability embedding vectors: the two capability embedding vectors are compared dimension by dimension. If one vector has a lower value in a certain dimension while the other vector has a higher value in that dimension, then the two are considered complementary in that dimension. The degree of complementarity is measured by the absolute value of the difference between the two vectors; the larger the difference, the higher the degree of complementarity. The system calculates a weighted sum of the complementarity degrees in each dimension as the capability complementarity score between the two agents.

[0063] S244: Based on the marketing sub-task data dependency network, the marketing sub-task temporal constraint network, and the functional agent capability collaboration network, establish a multi-layer agent collaboration constraint model, and calculate the connection priority between each functional agent according to the multi-layer agent collaboration constraint model. Specifically, directed edges in the data dependency network are mapped to the intensity of data-driven demand for connections between agents; directed edges in the temporal constraint network are mapped to the time urgency demand for connections; and undirected edges in the capability synergy network are mapped to the intensity of capability complementarity demand for connections. For any pair of functional agents, the system comprehensively evaluates connection demand across three dimensions: First, if there is a data dependency between the subtasks corresponding to X and Y, a higher connection demand score is assigned based on the degree of data dependency; second, if there is a temporal constraint between the subtasks corresponding to X and Y, a medium connection demand score is assigned based on the strictness of the time constraint; and third, if there is a high degree of capability synergy between X and Y, a supplementary connection demand score is assigned.

[0064] S245: Based on the connection priority between each functional Agent, establish information transmission edge, collaborative execution edge and feedback correction edge between each functional Agent. Among them, the information transmission edge is used for marketing status sharing, the collaborative execution edge is used for collaborative processing of marketing tasks, and the feedback correction edge is used for marketing strategy correction and collaborative optimization. Specifically, the following steps are taken: First, establish information transmission edges: These edges are used to transmit shared marketing status information among agents, including global status data such as current user profile, current marketing stage, and environmental constraint parameters. Information transmission edges primarily connect multiple agents that need to perceive the same status information; they are typically undirected or unidirectional broadcast edges with high data transmission frequency and large data volume. Second, establish collaborative execution edges: These edges connect agents executing the same marketing sub-task or having upstream / downstream collaborative relationships to support collaborative processing of marketing tasks. For example, connecting a script generation agent and a compliance review agent allows for real-time compliance feedback during script generation. Collaborative execution edges are typically bidirectional, supporting synchronous communication in a request-response pattern. Third, establish feedback correction edges: These edges feed back the execution results or evaluation results of downstream agents to upstream agents to support the correction and collaborative optimization of marketing strategies. For example, feeding back violation markings from the compliance review agent to the script generation agent drives script regeneration; feeding back prediction results from the sales prediction agent to the course recommendation agent drives adjustments to the recommendation strategy.

[0065] S246: Based on the established information transmission edge, collaborative execution edge and feedback correction edge, topological connections are made to each functional agent to generate a dynamic graph agent collaborative network corresponding to the current marketing state. Each functional agent is a node in the dynamic graph agent collaborative network, and the connection relationship between each node is dynamically adjusted as the marketing state changes. Specifically, using all functional agents selected in step S23 as the network node set, and the information transmission edges, collaborative execution edges, and feedback correction edges established in step S245 as the network edge set, a complete dynamic graph agent collaborative network is constructed. During the construction process, the system configures the running status attribute for each node and the communication protocol attribute for each edge. After construction, the system performs structural verification on the generated dynamic graph agent collaborative network: checking for isolated nodes in the network; if so, determining that they cannot participate in collaborative decision-making and triggering an exception handling process; checking for loop dependencies in the network; if so, removing the loop by adjusting the direction of the edges or adding intermediate nodes; checking whether key nodes in the network have sufficient connection redundancy; if the number of connections of key nodes is lower than the minimum redundancy threshold, adding backup connections. After successful verification, a complete dynamic graph agent collaborative network corresponding to the current marketing state is output, and the topology of this network is dynamically adjusted as the marketing state changes.

[0066] S247: Map the dynamic graph agent collaborative network to the corresponding marketing state node in the education marketing world model, so that each functional agent can read shared state information, execute marketing behaviors and receive marketing feedback according to the current marketing state, thereby completing collaborative operation in the unified education marketing world model. Specifically, each functional agent in the dynamic graph agent collaborative network is registered as a subscriber to the current marketing state node in the world model, enabling each agent to obtain real-time updates of the current state through the world model's state reading interface. Simultaneously, the system registers each agent's execution capabilities as callable marketing behaviors within the world model. When a marketing behavior needs to be executed during the world model's deduction process, it can be generated and executed by calling the corresponding agent interface. Furthermore, the system registers each agent's feedback receiving channel as the world model's result feedback channel, allowing agents to receive result data and update their own state after real marketing results are generated through the feedback interface. Through these mapping relationships, each functional agent in the dynamic graph agent collaborative network uses the current marketing state of the world model as input, the world model's behavior execution interface as output, and the world model's result feedback interface as learning signals. Within a unified educational marketing world model environment, it completes a full closed-loop operation from state perception and collaborative decision-making to behavior execution and result feedback, ensuring real-time synchronization and consistency between the multi-agent collaborative process and the real marketing environment.

[0067] The specific implementation method of "performing topology self-adjustment of the dynamic graph agent cooperative network" in S26 is as follows: S261: Map Agent capabilities and subtask requirements to high-dimensional embedding vectors respectively, calculate the dynamic capability matching degree between each Agent and the corresponding marketing subtask based on cosine similarity, and generate an initial Agent collaboration graph, where the node weights correspond to the Agent's current confidence score. Specifically, a unified high-dimensional semantic space is pre-constructed, with dimensions including functional type dimension, data type processing capability dimension, inference depth dimension, and output granularity dimension. For each functional agent in the Agent capability resource pool, the system generates a corresponding capability embedding vector in the aforementioned semantic space based on its capability description information. Each dimension of this vector represents the capability strength of the agent in that dimension. For each marketing subtask partitioned in step S22, a corresponding requirement embedding vector is generated based on its capability requirement information. Each dimension of this vector represents the requirement strength for executing the subtask in that dimension. The cosine similarity between each agent's capability embedding vector and each subtask's requirement embedding vector is calculated, and the cosine similarity score is used as the dynamic capability matching degree between the agent and the subtask. For each subtask, the system selects the agent with the highest matching degree to establish a mapping relationship with it, and uses the matching degree score as the confidence score of the agent in that subtask. All agents with established mapping relationships are used as nodes, and the confidence scores of each node are used as node weights to generate an initial agent collaboration graph. In this initial collaboration graph, the edges between nodes are currently empty. Subsequent steps will gradually fill in the connecting edges based on the collaboration relationships between Agents.

[0068] S262: Continuously collect real-time interaction data between agents within a preset time window and calculate the cooperation entropy value; when the cooperation entropy value exceeds the preset chaos threshold, determine that the current topology is in an inefficient state and trigger a topology evolution event. Specifically, a preset time window is used as the time window to continuously collect real-time interaction data between agents within that window. This real-time interaction data includes at least: the average message transmission delay time between agents, the frequency of message loss or retransmission, the degree of inconsistency between agent outputs, and the proportion of idle time for each agent waiting for responses from other agents. Based on the collected interaction data, the collaboration entropy value of the current collaborative topology is calculated. The calculation of the collaboration entropy value comprehensively considers interaction delay entropy, conclusion divergence entropy, and resource waiting entropy, and the weighted sum of these three factors yields the comprehensive collaboration entropy value. A higher collaboration entropy value indicates a more chaotic current collaborative state, less smooth information transmission, and lower collaboration efficiency between agents. The calculated collaboration entropy value is compared with a preset chaos threshold: when the collaboration entropy value continuously exceeds the chaos threshold for a preset duration, the system determines that the current collaborative topology is in an inefficient state, automatically triggers a topology evolution event, and initiates the subsequent topology adjustment process.

[0069] S263: In response to the topology evolution event, calculate the marginal contribution value of each connection edge in the current collaboration graph, that is, the change in the overall decision-making benefit of the system after removing the corresponding connection edge from the graph; mark the connection edge with a marginal contribution value lower than the first preset threshold as a redundant edge and delete it; at the same time, according to the capability vector space model, select the enhanced agent with a matching degree exceeding the second preset threshold from the candidate agents that have never participated in the collaboration, establish a new connection edge between it and the associated agent in the existing collaboration graph, and assign the weight of the new connection edge according to the historical collaboration success rate and capability complementarity of the agent; Specifically, for edge E_AB connecting Agent A and Agent B, this edge is temporarily removed from the current collaboration graph. A complete collaborative decision-making process is run before and after the removal, and the change in the overall system decision benefit is recorded in both processes. The overall system decision benefit is a comprehensive benefit index, including a weighted sum of sub-indicators such as decision accuracy, decision delay time, and consistency of results among agents. The decrease in the overall system decision benefit after removing edge E_AB is the marginal contribution value of that edge; a larger decrease indicates a greater contribution of the edge to the system's decision quality. The system compares the marginal contribution value of each edge with a preset first threshold. Edges below the first threshold are marked as redundant and deleted from the collaboration graph. This deletion operation is only formally executed after verification in the simulation environment that there is no negative impact.

[0070] Simultaneously, the system performs filtering and addition of enhanced agents. From the set of candidate agents that have never participated in the current collaboration, the system calculates the matching degree between the capability embedding vector of each candidate agent and the requirement vector of each current marketing sub-task. Candidate agents with a matching degree exceeding a second preset threshold are marked as enhanced agents. For the selected enhanced agents, the system analyzes the degree of capability complementarity between them and other agents in the existing collaboration graph, i.e., comparing the differences in their capability embedding vectors across various dimensions; a larger difference indicates higher complementarity. From the existing collaboration graph, the system selects several associated agents with the highest capability complementarity to the enhanced agent and establishes new connection edges between the enhanced agent and each associated agent. The initial weight of the new connection edge is assigned based on a comprehensive evaluation of the enhanced agent's historical collaboration success rate and capability complementarity with associated agents, with the weight ranging from 0 to 1, serving as the priority basis for subsequent use of the connection edge.

[0071] S264: During topology mutation, atomic snapshots are performed on the shared state space, long-term memory call pointers, and intermediate decision results of each Agent. After mutation, state backfilling and consistency verification are performed on the shared context of each Agent based on the snapshots, so that the dynamic graph Agent collaborative network can complete the seamless hot migration and self-evolution of the topology without interrupting the current marketing service process. Specifically, before executing step S263 (deleting redundant edges and adding enhanced agents and their connections), an atomic snapshot saving operation is performed. The atomic snapshot includes: all data in the shared state space of all agents, long-term memory call pointers for each agent, and intermediate decision results of each agent during execution. The snapshot saving operation is atomic, meaning all snapshot data is saved as a whole at the same point in time, ensuring that the system state at the snapshot time is completely consistent.

[0072] After the topology mutation operation is completed, the system performs a state restoration operation based on the saved atomic snapshot. The restoration process includes: filling the shared state space data from the snapshot into the shared memory region of the new topology, ensuring that the state information accessible to each agent is completely consistent with that before the mutation; restoring the long-term memory call pointers from the snapshot to the corresponding long-term memory access locations for each agent, allowing agents to continue reading historical information from the point of interruption and avoiding information gaps caused by topology adjustments; and restoring the intermediate decision results from the snapshot to the context cache of the corresponding agent, enabling agents currently performing inference to continue inference from the state at the time of the snapshot after the mutation, without having to start from scratch. After completing the state filling, the system performs a consistency check on the restored state of each agent, verifying whether the reference relationship between the context data and the shared state space of each agent is correct, and whether the versions of the intermediate results and input data of each agent match. After the consistency check passes, the system smoothly switches the decision traffic from the original topology to the new topology, without interrupting the currently ongoing marketing service process, achieving seamless hot migration and self-evolution of the dynamic graph agent collaborative network. Further, step S3 includes the following steps: S31: Based on the functional responsibilities of each Agent in the dynamic graph Agent collaborative network, the marketing strategy generation Agent generates multiple candidate marketing strategies by combining the current marketing status in the education marketing world model. Each candidate marketing strategy corresponds to a different course recommendation method, marketing script combination, discount strategy and communication path. Specifically, based on the dynamic graph agent collaborative network, the marketing strategy generation agent responsible for strategy generation is identified. This agent reads the complete attribute information of the current marketing state node in the education marketing world model, including user demand tags, course matching degree, transaction intention level, marketing risk level, and current marketing stage. The marketing strategy generation agent uses the current state as input and, combined with its built-in strategy generation model and historical success case library, generates multiple candidate marketing strategies. Each candidate strategy has differentiated configurations in key decision variables: in course recommendation methods, different strategies employ different approaches such as precise recommendation based on user historical preferences, popular recommendation based on course popularity, and collaborative filtering recommendation based on similar users in user profiles; in marketing rhetoric combinations, different strategies employ different styles such as rational persuasive rhetoric, emotional resonance rhetoric, and urgency-creating rhetoric; in discount strategies, different strategies employ different combinations of discounts such as direct discounts, gift value-added services, stacked discounts, and extended learning periods; and in communication paths, different strategies plan different conversation rhythms, different key decision points, and different objection handling contingency plans. The Marketing Strategy Generation Agent generates a preset number of candidate marketing strategies. Each strategy describes a complete configuration scheme for the above four dimensions in a structured data format, forming a set of candidate strategies for subsequent evaluation steps.

[0073] S32: Each functional agent in the control dynamic graph agent collaborative network independently analyzes each candidate marketing strategy. Each functional agent outputs evaluation results for the candidate marketing strategy based on its corresponding capabilities. The evaluation results include at least the conversion probability evaluation, marketing compliance evaluation, user demand matching evaluation, course suitability evaluation, brand risk evaluation, and marketing service quality evaluation. Specifically, the control dynamic graph agent collaborative network reads the complete configuration schemes of each candidate marketing strategy. Each agent independently analyzes each candidate strategy based on its own capabilities and outputs multi-dimensional quantitative evaluation results: the sales prediction agent predicts the probability of a sale after adopting the strategy based on the candidate strategy's course recommendation method, discount level, and communication path, combined with current user profiles and historical sales data; the marketing compliance agent reviews the candidate strategy's wording, pricing methods, and discount expressions against the marketing rule base and regulatory policy base, outputting a compliance score and a list of violation risk points; the user needs analysis agent outputs a user needs matching score based on whether the candidate strategy's course recommendation direction covers the user's core needs tags; the course recommendation agent outputs a course suitability score based on the degree of fit between the courses recommended by the candidate strategy and the user's grade, learning level, and learning goals; the brand risk assessment agent outputs a brand risk score based on the potential impact of the candidate strategy's wording style, promised content, and promotional scale on the brand image; and the marketing service quality agent outputs a marketing service quality score based on the candidate strategy's communication path planning, response rhythm design, and user care level.

[0074] S33: Establish a negotiation and interaction mechanism among the functional agents, generate corresponding strategy correction suggestions based on the evaluation results output by each functional agent, and feed the strategy correction suggestions back to the marketing strategy generation agent to drive the marketing strategy generation agent to readjust the corresponding candidate marketing strategies. Specifically, the evaluation results output by each Agent are aggregated and fed back to each Agent, allowing each Agent to not only know its own evaluation results but also the evaluation conclusions of other Agents on the same candidate strategy. For example, the marketing compliance Agent can see the transaction prediction Agent's evaluation of the transaction probability of a certain strategy and make a comprehensive judgment based on its own compliance evaluation of that strategy. Based on the evaluation results obtained from other Agents, each functional Agent generates strategy modification suggestions for the marketing strategy generation Agent according to its own reasoning logic. The specific content of the modification suggestions includes: the direction, magnitude, and priority of adjustments for specific dimensions within a candidate strategy.

[0075] S34: The marketing strategy generation agent regenerates new candidate marketing strategies based on the strategy modification suggestions returned by each functional agent, and sends them again to each functional agent in the dynamic graph agent collaboration network for re-evaluation. The strategy generation, strategy evaluation and strategy modification process is repeated until the preset collaboration convergence condition is met. Specifically, the marketing strategy generation agent reads strategy correction suggestions returned by each functional agent from the message queue and summarizes and analyzes the correction suggestions for each candidate strategy. For strategy dimensions that are simultaneously identified as problematic by multiple agents, the marketing strategy generation agent marks them as high-priority correction items; for potential problems identified by only a single agent, they are marked as low-priority correction items. Based on the above correction suggestions, the marketing strategy generation agent regenerates the current candidate strategy set: for high-priority correction items, it directly adjusts the configuration parameters of the corresponding strategy dimension; for low-priority correction items, it generates multiple alternative configuration schemes for differential comparison in the next round of evaluation.

[0076] S35: Calculate the current degree of collaborative convergence based on the consistency of the output results of each functional agent, the magnitude of evaluation changes, and the changes in overall benefits during each round of negotiation. When the degree of collaborative convergence reaches the preset convergence threshold, stop the negotiation interaction between each functional agent and determine the target marketing strategy. Specifically, after each round of strategy evaluation, the degree of collaborative convergence for the current round is calculated. The convergence degree is calculated by considering the following three dimensions: The first dimension is the consistency between the outputs of each functional agent, i.e., the standard deviation between the multi-dimensional evaluation scores given by each agent for the same candidate strategy. The smaller the standard deviation, the more consistent the agents' understanding of the strategy, and the higher the convergence. The second dimension is the magnitude of evaluation change, i.e., the absolute value of the change in scores for the same candidate strategy between adjacent rounds. The smaller the change, the more stable the agents' opinions are, and the less significant the fluctuations. The third dimension is the change in overall return, i.e., the magnitude of change in the overall return score of each strategy in the candidate strategy set between adjacent rounds. The smaller the magnitude of change, the more stable the strategy solution has become the optimal solution. The values ​​of the above three dimensions are weighted and summed according to preset weights to obtain the comprehensive score of the degree of collaborative convergence for the current round. The overall score is compared with a preset convergence threshold: when the overall score reaches or exceeds the convergence threshold, the collaborative process is determined to have converged, the negotiation and interaction between the functional agents is stopped, and the candidate strategy with the highest overall benefit score is selected from the current candidate strategy set as the target marketing strategy; when the overall score does not reach the convergence threshold, the system continues to execute the next round of iterations until convergence.

[0077] S36: Record the information interaction content, strategy correction trajectory, evaluation change trajectory and final negotiation result between each functional agent during each round of negotiation, and generate the corresponding agent collaborative decision trajectory, and store the agent collaborative decision trajectory in the long-term memory module. Specifically, during each round of negotiation iterations, the following information is continuously recorded: all information interactions between functional agents, the correction trajectory of each candidate strategy in each round, the evaluation change trajectory of each strategy by each functional agent in each round, and the final negotiation result. The complete data recorded above is structured and organized in chronological order to generate the corresponding agent collaborative decision-making trajectory. This collaborative decision-making trajectory is encapsulated in a standardized data format, including header information. The encapsulated collaborative decision-making trajectory is stored in a long-term memory module, enabling the trajectory data to be retrieved and referenced in similar marketing situations in the future, to assist subsequent decision-making, optimize agent parameter configuration, and serve as source data for historical success cases.

[0078] Further, step S4 includes the following steps: S41: Based on the target marketing strategy, replicate the current marketing state in the education marketing world model and construct a digital twin marketing environment corresponding to the real marketing environment, wherein the digital twin marketing environment maintains the same user profile, course information, marketing stage, marketing state and marketing constraints as the real marketing environment. Specifically, a complete snapshot of the current real marketing state is obtained by calling the marketing state reading interface of the world model. This complete snapshot includes all attribute fields of the current user profile, the currently discussed course product information, the current marketing stage identifier, all attribute information of the current state node, and all currently effective marketing constraints. This snapshot data is then completely copied to an independent simulation partition of the world model. This simulation partition maintains complete data consistency with the real marketing environment, forming an initial digital twin marketing environment. This twin environment runs independently in subsequent simulations, its state evolution unaffected by subsequent actual operations in the real marketing environment, and it also does not negatively impact the operational state of the real marketing environment, thus ensuring the isolation and security between the simulation process and real business operations.

[0079] S42: Using the target marketing strategy as the initial marketing path, and combining it with the education marketing world model to automatically generate multiple counterfactual marketing strategies. Each counterfactual marketing strategy adopts different course recommendation methods, marketing script combinations, preferential schemes, trial class arrangements and user communication paths to form multiple marketing scenarios to be deduced. Specifically, using the target marketing strategy as the baseline initial path, the system extracts the configuration parameters of this strategy across key decision dimensions, including course recommendation methods, marketing script combinations, discount schemes, trial lesson arrangements, and user communication paths. The system employs a counterfactual generation method to systematically decouple and perturb the configuration parameters of each dimension: In the course recommendation method dimension, the system generates a counterfactual version that switches the recommended target from the current course to an alternative course of the same level; in the marketing script combination dimension, the system generates a counterfactual version that switches persuasive scripts to resonant scripts, or short scripts to long scripts; in the discount scheme dimension, the system generates a counterfactual version that switches discounts to gift offers, or full-reduction offers to extended-period offers; in the trial lesson arrangement dimension, the system generates a counterfactual version that switches immediate trial lessons to scheduled trial lessons, or live trial lessons to pre-recorded trial lessons; and in the user communication path dimension, the system generates a counterfactual version that switches high-frequency, fast-paced communication to low-frequency, in-depth communication, or a direct sales path to a gradual trust-building path.

[0080] S43: Using each marketing scenario to be simulated as the initial node, perform multi-step state simulation of the subsequent marketing process based on the marketing state transition matrix, and construct a marketing decision search tree based on the transition relationship between each marketing state. Perform multiple rounds of path expansion and state search for each marketing path to obtain the future marketing state sequence corresponding to each marketing path. Specifically, the initial marketing state corresponding to each marketing scenario to be simulated is taken as the root node of the marketing decision search tree. Starting from this root node, the subsequent marketing process is iteratively simulated in multiple steps based on the marketing state transition matrix in the education marketing world model. In each step of the simulation, the system queries the state transition matrix to obtain the conditional transition probability distribution from the current state to each candidate subsequent state based on the current state node and the currently executed marketing behavior. The system then performs compliance correction on the probability distribution by combining the constraint parameters of the current environment layer, filtering out state transition paths that are unreachable in the current environment. The subsequent marketing states are sampled and expanded according to the corrected probability distribution, generating corresponding child nodes in the search tree and establishing directed transition edges between parent and child nodes. The edges record the marketing behavior identifier, transition probability, and expected change in revenue corresponding to the transition.

[0081] S44: Calculate the comprehensive marketing revenue of the corresponding marketing path according to each future marketing state sequence. The comprehensive marketing revenue includes at least the course transaction probability, violation risk, complaint probability, refund probability, long-term customer value, brand influence value and user satisfaction, and establish the comprehensive revenue evaluation result of the corresponding marketing path. Specifically, for each marketing path's corresponding future marketing state sequence, the system reads the prediction parameters corresponding to each state node, including the predicted values ​​for the probability of conversion, violation, complaint, and user churn at that state. The system accumulates these prediction parameters along the time dimension: the conversion probability is comprehensively evaluated using the cumulative maximum of the conversion probabilities across all states on the path; violation risk is evaluated using the weighted sum of the violation probabilities across all states on the path; the refund probability is calculated using the predicted refund probability at the end of the path; long-term customer value is jointly evaluated using the cumulative sum of user satisfaction scores along the path and the course renewal probability; and brand influence value is calculated using the cumulative brand friendliness scores of each marketing activity along the path. The system also calculates a comprehensive user satisfaction score based on the changing trends of user psychological states corresponding to each state node on the path.

[0082] S45: Compare the comprehensive revenue evaluation results corresponding to each marketing path, determine the target marketing path according to the preset revenue optimization rules, and output the marketing strategy execution order, marketing risk level and marketing revenue prediction results corresponding to the target marketing path. Specifically, the comprehensive revenue evaluation results of all marketing paths generated in step S44 are sorted from highest to lowest according to their comprehensive scores. Path selection is performed according to preset revenue optimization rules, which are multi-objective optimization rules. The core logic is: prioritize the path with the highest comprehensive revenue score; when the comprehensive scores of multiple paths are similar (the difference is less than a preset tolerance threshold), further compare the violation risk scores of each path and prioritize the path with lower violation risk; when both the comprehensive score and violation risk are similar, prioritize the path with higher long-term customer value. Through the layer-by-layer screening of the above multi-level optimization rules, the system determines a comprehensively optimal target marketing path from all candidate paths. Based on the marketing behavior identifiers corresponding to each transition edge on the target marketing path, the system outputs the marketing strategy execution sequence corresponding to the path in chronological order (i.e., the entire sequence of behaviors from the current state to the final marketing result); the system takes the maximum value of the violation risk level of each state node on the path as the marketing risk level output for the target marketing path; and generates a marketing revenue prediction result output based on the predicted transaction probability and long-term customer value of the path's final state.

[0083] S46: Compare the consistency between the actual marketing execution results and the predicted results in the digital twin marketing environment, calculate the prediction deviation between each marketing state, and adaptively correct the state transition relationship, state prediction parameters and marketing state transition matrix in the education marketing world model based on the prediction deviation in order to update the education marketing world model. Specifically, after executing the marketing strategy corresponding to the target marketing path in a real marketing environment, the system collects real marketing execution result data through the marketing result feedback interface of the world model. The real execution results include at least real transaction results, real user satisfaction feedback, real complaint records, real violation judgment results, and real refund records. Simultaneously, the system reads the predicted result data corresponding to the marketing strategy from the digital twin marketing environment, compares the real execution results with the predicted results state by state, and calculates the prediction deviation value for each state node. The prediction deviation value includes state transition direction deviation and state parameter deviation.

[0084] Based on the comprehensive analysis results of the prediction deviation, a hierarchical adaptive correction strategy is implemented. The system first inputs the calculated prediction deviation sequence into a deviation classifier, decomposing the deviation into short-term random fluctuation components and long-term systematic deviation components: short-term random fluctuations are characterized by the deviation value oscillating randomly around the mean without a sustained trend, while long-term systematic deviations are characterized by the deviation value continuously shifting in the same direction and increasing cumulatively over time. If it is determined that the current deviation is mainly caused by short-term random fluctuations, the system only performs a Bayesian online update on the noise variance parameter in the marketing state prediction parameters. That is, it uses the deviation information from this observation to perform a posterior update on the prior distribution of the noise variance, fine-tuning the confidence interval of the prediction parameters without changing the core structure of the state transition matrix. If the current deviation is determined to be a long-term systematic deviation (such as a fundamental change in user behavior patterns due to policy and regulatory adjustments or changes in the market environment), the system locks the target state transition rows in the marketing state transition matrix affected by this systematic deviation. Using a deviation tracing algorithm, it identifies the matrix rows containing the state transition relationships with the greatest contribution to the deviation. For these target rows, it performs an incremental graph learning algorithm using the latest collected real marketing results data to rewrite the transition probability values ​​and revenue parameter values ​​in those rows, ensuring that the locked target rows reflect the latest environmental change trends. Simultaneously, it maintains the stability of other data rows in the marketing state transition matrix that are not affected by the systematic deviation, avoiding the overwriting of existing correct knowledge due to global updates. After correction, the system synchronizes the updated state transition matrix, state prediction parameters, and state transition relationships to the core logic layer of the education marketing world model, completing a full self-evolutionary iteration of the world model.

[0085] Further, step S43 includes the following steps: S431: Read the initial marketing state corresponding to each marketing scenario to be simulated, and use each initial marketing state as the root node of the marketing decision search tree. Initialize the marketing decision search tree according to the marketing environment information corresponding to each initial marketing state to form multiple independent marketing simulation branches. Specifically, the system reads the initial marketing state corresponding to each of the multiple marketing scenarios to be simulated. The initial marketing state includes the complete user profile, course matching status, marketing stage identifier, transaction intention level, and environmental constraints for that scenario. The system treats each initial marketing state as the root node of an independent search tree, assigns a unique simulation branch identifier to each root node, and initializes the corresponding marketing decision search tree data structure starting from that root node. During initialization, the system allocates an independent simulation context cache space for each search tree to store all state nodes and transition relationships generated by that branch in subsequent simulations.

[0086] S432: Based on the marketing state transition matrix in the education marketing world model, calculate multiple candidate subsequent marketing states corresponding to the current marketing state, and generate a set of state transitions corresponding to each candidate marketing state by combining the marketing constraints corresponding to the current marketing state. For each current marketing state node, the system reads the row vector of the marketing state transition matrix corresponding to that state from the education marketing world model. This row vector records the conditional transition probabilities of reaching all possible subsequent states from the current state. The system filters out subsequent states with transition probabilities lower than a preset filtering threshold (e.g., 0.05) from the candidates, retaining multiple candidate subsequent states with transition probabilities higher than the threshold. Subsequently, the system reads the set of constraints in the current marketing environment layer, including policy and regulatory constraints, time constraints, and resource constraints. The system applies these constraints one by one to each candidate subsequent state: if the marketing behavior or state attribute corresponding to a candidate subsequent state violates any constraint, the system adds a violation mark to that state in the state transition set and sets its transition probability to zero or reduces its weight in the subsequent behavior selection.

[0087] S433: For each candidate marketing state, call each functional agent in the dynamic graph agent collaborative network to generate corresponding marketing behaviors. Each functional agent jointly completes marketing strategy negotiation based on the current marketing state, and determines the target marketing behavior for the corresponding candidate marketing state based on the negotiation results, so as to realize the joint deduction between marketing state and marketing behavior. Specifically, the system invokes various functional agents in the dynamic graph agent collaborative network, using the current state and candidate subsequent states as joint inputs to generate the marketing actions required to reach the candidate state. The specific collaborative decision-making process is as follows: the system inputs the attribute differences between the current state node and the candidate subsequent state node into each functional agent, enabling the marketing generation agent to generate multiple candidate action plans; the compliance review agent to conduct compliance reviews of each candidate action plan; the sales prediction agent to estimate the probability of a sale after each action plan reaches the target state; and the risk assessment agent to score the potential risks of each action plan.

[0088] S434: Based on the state change results after the target marketing behavior acts on the education marketing world model, generate corresponding child nodes for each candidate marketing state, and establish the marketing state transition relationship between the parent node and the child node to complete one round of path expansion of the marketing decision search tree; The target marketing behavior determined in step S433 is input into the behavior execution interface of the education marketing world model. The behavior is simulated and executed in the world model's simulation environment, and the resulting state changes are obtained. These state changes include changes in user demand characteristics, updates to the level of purchase intention, reassessment of risk level, and shifts in user psychological state. The system comprehensively compares the state change results with the candidate subsequent states predicted in step S432, fusing the two information to generate complete attribute information for the child node. Subsequently, the system creates a corresponding child node under the current parent node, stores the complete state attributes of the child node, and establishes a directed connection edge between the parent node and the child node. This connection edge records the target marketing behavior identifier corresponding to the transfer, the corrected transfer probability, the change in expected revenue, and the change in expected risk.

[0089] S435: For each newly added sub-node, repeatedly execute the marketing status prediction, agent collaborative decision-making and marketing status expansion process, and recursively expand the marketing decision search tree in multiple rounds until the preset inference depth, marketing goal completion status or termination constraint is reached, forming a multi-layer marketing decision search tree. Using the newly added child nodes in step S434 as the current node, the operation sequence described in steps S432 to S434 is repeated, namely, state expansion prediction, agent collaborative decision-making to generate target marketing behavior, child node creation and transfer relationship establishment. This recursive expansion process continues until any of the following termination conditions are met: (1) the current inference depth has reached the preset maximum inference depth; (2) the current state node is in the marketing goal completion state; (3) the marketing risk level of the current state node exceeds the preset acceptable upper limit; (4) the transaction intention level of the current state node has not changed for several consecutive rounds and is at a low level. When any termination condition is triggered, the branch stops expanding and the current node is marked as a leaf node. After all branches have completed expansion, the system outputs a complete multi-layer marketing decision search tree, which contains all reachable paths from the initial state to each termination state and the state node sequence and transfer behavior sequence on each path.

[0090] S436: Based on the predicted transaction revenue, marketing risk, user satisfaction, course renewal, and marketing compliance evaluation results corresponding to each marketing status node, comprehensively evaluate each marketing path in the marketing decision search tree and identify the target marketing path that meets the preset revenue conditions. Traverse each complete path from the root node to a leaf node in the search tree, and accumulate the prediction parameters corresponding to all state nodes on the path; read the transaction revenue prediction value, marketing risk prediction value, user satisfaction prediction value, course renewal prediction value, and marketing compliance evaluation result for each state node on the path.

[0091] S437: For marketing paths whose comprehensive evaluation results are lower than the preset evaluation threshold, dynamic pruning is performed on the corresponding marketing paths based on the marketing constraints, marketing risk constraints and collaborative evaluation results corresponding to the dynamic graph Agent collaborative network in the education marketing world model. Marketing paths that meet the marketing objectives are retained, and an optimized marketing decision search tree is generated. Each path to be pruned is read sequentially, and the following attributes are obtained: whether the path meets the current marketing constraints in the world model, whether the cumulative marketing risk value of the path exceeds the preset risk limit, and whether the collaborative evaluation result of the path is too low. If a path meets any of the above pruning conditions, the path is marked as a path to be pruned. Before formally executing pruning, the system verifies the legality of the pruning operation in the simulation environment to ensure that the number of remaining paths after pruning is not less than the minimum number of retained paths, and that the integrity of the search tree is not damaged after pruning. After successful verification, the system executes the pruning operation, removing all nodes and connecting edges of the marked paths from the search tree. After pruning, the system retains all target marketing paths that meet the marketing objectives, and these retained paths together constitute the optimized marketing decision search tree.

[0092] S438: Based on the optimized marketing decision search tree, extract the order of marketing state nodes corresponding to each target marketing path, and generate a future marketing state sequence according to the order in which the marketing states occur. The future marketing state sequence is used for subsequent marketing revenue prediction, digital twin simulation and marketing strategy optimization. Traverse all retained target marketing paths in the optimized search tree. For each target marketing path, sequentially read all attribute information of each marketing state node along the path from the root node to the leaf node, including the state identifier, user demand characteristics, transaction intention level, marketing risk level, and user psychological state of each state node. Arrange the read state nodes according to the order in which they occur during the actual deduction process to generate the future marketing state sequence corresponding to the path. Each future marketing state sequence is an ordered list of state nodes, where the first element is the initial state node and the last element is the terminating state node. Adjacent state nodes in the sequence form a state transition relationship through the directed connection edge established in step S434.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart quality inspection effect optimization and evaluation system based on multi-agent collaboration, characterized in that, include: S1: Construct a world model for education marketing, integrating user profiles, course information, historical transaction cases, marketing rules and regulatory policies to form a predictable marketing state space; S2: Generate a dynamic graph agent collaboration network corresponding to the current consulting task based on the world model, and adaptively determine the type, number, connection relationship and execution order of the participating agents; S3: Based on the marketing generation agent, marketing quality inspection agent, regulatory constraint agent and transaction prediction agent, multiple sets of candidate marketing strategies are generated iteratively in a dynamic graph network through a self-game mechanism, and the output of each agent is adjusted in real time. S4: Using a world model, conduct multiple rounds of future consultations for each candidate marketing strategy to predict the probability of closing a deal, complaint risk, violation risk, refund risk, and long-term customer value. S5: Based on the comprehensive multi-objective optimization results of the decision agent, select the target marketing strategy and output real-time marketing guidance, quality inspection results and evidence chain.

2. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 1, characterized in that, S1 includes the following steps: S11: Collect multi-source heterogeneous data in the education marketing process, and standardize the data according to a unified data format to form the original marketing dataset; S12: Perform semantic parsing on the original marketing dataset, identify the corresponding marketing behavior events during the consultation process, and generate corresponding marketing event nodes according to the preset marketing event types; S13: Establish an event association network based on the time sequence, contextual reference relationship, and marketing goal association relationship between each marketing event, calculate the state transition relationship between each marketing event, and generate an education marketing state diagram; S14: Based on the marketing state diagram, perform statistical analysis on the transition probabilities between each state node, establish a marketing state transition matrix, and combine historical transaction cases to calculate the transaction probability, violation probability, complaint probability and user churn probability corresponding to each state, forming marketing state prediction parameters. S15: Based on the marketing state diagram, marketing state transition matrix and marketing state prediction parameters, construct an education marketing world model, and use each marketing state in the world model as the operating environment of the dynamic graph agent network.

3. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 2, characterized in that, The process of establishing an event association network based on the chronological order, contextual references, and marketing objective relationships among various marketing events, calculating the state transition relationships between these events, and generating an educational marketing state diagram includes the following steps: S131: Reconstruct the time sequence of all marketing events according to their occurrence time, consultation session, and marketing stage to generate a marketing event time sequence. Marketing events within the same consultation session are arranged in the actual order of occurrence, and continuous time relationships are established between different consultation sessions based on the user's historical consultation trajectory. S132: Based on the semantic content, historical reference information and contextual semantic similarity of each marketing behavior event, perform reference association analysis on marketing behavior events with cross-session reference relationships, establish contextual reference edges between marketing behavior events, and form a marketing event association structure across consultation cycles; S133: Based on the marketing objective type, course recommendation objective, sales promotion objective, user trust building objective, and marketing risk control objective corresponding to each marketing activity event, calculate the degree of objective correlation between the marketing activity event and each marketing objective, and establish a marketing objective correlation network; S134: Based on the marketing event time sequence, context reference edges, and marketing goal association network, aggregate the states of multiple marketing behavior events with the same marketing semantics, the same user intent, and the same marketing goal to generate corresponding marketing state nodes; S135: Statistically analyze the continuous transitions between marketing state nodes in the historical marketing process, calculate the frequency of state transitions, duration of state transitions, and stability of state transitions between marketing state nodes, and establish marketing state transition relationships based on the state transition results. S136: Establish an education marketing state diagram that can be used for digital twin simulation based on marketing state nodes and marketing state transition relationships.

4. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 2, characterized in that, S15 includes the following steps: S151: Read each marketing state node and marketing state transition relationship in the marketing state diagram, and establish a marketing state space based on the state transition relationship between marketing state nodes; S152: Collect course information, enrollment policies, marketing rules, education supervision requirements, market competition information and user profile information corresponding to the current education marketing task, and construct a marketing environment layer based on the environmental information to form a set of environmental states describing the current education marketing task operation environment; S153: Based on the marketing state nodes in the marketing state diagram, standardize and encode marketing states with the same user demand characteristics, marketing goals, user psychological states and transaction stages, construct a marketing state layer, and establish environmental constraint relationships between the marketing state layer and the marketing environment layer to determine the effective range of each marketing state under different marketing environments. S154: Based on the marketing objectives corresponding to each marketing state, configure a set of corresponding marketing behaviors for each marketing state, and establish the state interaction relationship between marketing behaviors and marketing states to form a marketing behavior layer; S155: Establish a marketing results layer based on historical marketing results, link the course transaction results, user satisfaction, refund results, complaint results, course renewal results and marketing violation results corresponding to each marketing behavior to the corresponding marketing behavior, and establish a result mapping relationship between the marketing behavior layer and the marketing results layer; S156: Based on the marketing state transition matrix and marketing state prediction parameters, perform state deduction on the marketing state change process after the marketing behavior takes effect, generate prediction results of the future marketing state under the corresponding marketing behavior, and establish dynamic evolution relationship between the marketing environment layer, marketing state layer, marketing behavior layer and marketing result layer to form an educational marketing world model. S157: Establish an operational interface between the education marketing world model and the dynamic graph agent collaborative network. Open marketing status reading interface, marketing behavior execution interface, and marketing result feedback interface to each functional agent in the dynamic graph agent collaborative network, so that each functional agent can use marketing status as input, marketing behavior as decision output, and marketing result as feedback information to complete collaborative operation in the unified education marketing world model.

5. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 1, characterized in that, S2 includes the following steps: S21: Read the state node corresponding to the current marketing state in the education marketing world model, perform state complexity analysis on the current marketing task based on the state node, and identify multiple decision factors that affect the current marketing decision. S22: Based on the state complexity analysis results, decouple the current marketing task and divide the current marketing goal into multiple interrelated marketing sub-tasks; S23: Read the capability description information corresponding to each Agent in the Agent capability resource pool, calculate the capability matching degree between each Agent and the marketing sub-task according to the capability requirements of each marketing sub-task, and complete the mapping relationship between each marketing sub-task and the corresponding Agent based on the capability matching degree. S24: Based on the data dependencies, temporal constraints, and capability synergy among the various marketing sub-tasks, establish dynamic connection relationships between multiple agents and generate a dynamic graph agent collaboration network corresponding to the current marketing state; S25: Based on the generated dynamic graph Agent collaborative network, configure the shared context, running priority, input / output interface, long-term memory access permission and collaborative communication strategy for each Agent, and complete the initialization of the shared state space between Agents so that Agents can complete collaborative operation under the unified education marketing world model; S26: After the dynamic graph agent collaborative network is initialized, the dynamic graph agent collaborative network is topologically adjusted according to the historical collaboration success rate, historical decision contribution and current marketing status changes of each agent. The connection relationship between corresponding agents is dynamically added, deleted or reconnected to form a self-evolving agent collaborative network that adapts to the current marketing status changes.

6. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 1, characterized in that, The topology self-adjustment of the dynamic graph agent collaborative network includes the following steps: S261: Map Agent capabilities and subtask requirements to high-dimensional embedding vectors respectively, calculate the dynamic capability matching degree between each Agent and the corresponding marketing subtask based on cosine similarity, and generate an initial Agent collaboration graph, where the node weights correspond to the Agent's current confidence score. S262: Continuously collect real-time interaction data between agents within a preset time window and calculate the cooperation entropy value; when the cooperation entropy value exceeds the preset chaos threshold, determine that the current topology is in an inefficient state and trigger a topology evolution event. S263: In response to the topology evolution event, calculate the marginal contribution value of each connection edge in the current collaboration graph, that is, the change in the overall decision benefit of the system after removing the corresponding connection edge from the graph; mark the connection edge with a marginal contribution value lower than the first preset threshold as a redundant edge and delete it; at the same time, according to the capability vector space model, select the enhanced agent with a matching degree exceeding the second preset threshold from the candidate agents that have never participated in the collaboration, establish a new connection edge between it and the associated agent in the existing collaboration graph, and assign the weight of the new connection edge according to the historical collaboration success rate and capability complementarity of the agent; S264: During topology mutation, atomic snapshots are performed on the shared state space, long-term memory call pointers, and intermediate decision results of each Agent. After mutation, state backfilling and consistency verification are performed on the shared context of each Agent based on the snapshots, so that the dynamic graph Agent collaborative network can complete the seamless hot migration and self-evolution of the topology without interrupting the current marketing service process.

7. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 1, characterized in that, S3 includes the following steps: S31: Based on the functional responsibilities of each agent in the dynamic graph agent collaborative network, the marketing strategy generation agent generates multiple candidate marketing strategies by combining the current marketing status in the education marketing world model. S32: In the control dynamic graph agent collaborative network, each functional agent independently analyzes each candidate marketing strategy, and each functional agent outputs the evaluation results for the candidate marketing strategy based on its corresponding capabilities; S33: Establish a negotiation and interaction mechanism among the functional agents, generate corresponding strategy correction suggestions based on the evaluation results output by each functional agent, and feed the strategy correction suggestions back to the marketing strategy generation agent to drive the marketing strategy generation agent to readjust the corresponding candidate marketing strategies. S34: The marketing strategy generation agent regenerates new candidate marketing strategies based on the strategy modification suggestions returned by each functional agent, and sends them again to each functional agent in the dynamic graph agent collaboration network for re-evaluation. The strategy generation, strategy evaluation and strategy modification process is repeated until the preset collaboration convergence condition is met. S35: Calculate the current degree of collaborative convergence based on the consistency of the output results of each functional agent, the magnitude of evaluation changes, and the changes in overall benefits during each round of negotiation. When the degree of collaborative convergence reaches the preset convergence threshold, stop the negotiation interaction between each functional agent and determine the target marketing strategy. S36: Record the information interaction content, strategy correction trajectory, evaluation change trajectory and final negotiation result between each functional agent during each round of negotiation, and generate the corresponding agent collaborative decision trajectory, and store the agent collaborative decision trajectory in the long-term memory module.

8. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 7, characterized in that, S4 includes the following steps: S41: Based on the stated target marketing strategy, replicate the current marketing state in the education marketing world model and construct a digital twin marketing environment that corresponds to the real marketing environment; S42: Using the target marketing strategy as the initial marketing path, and combining it with the education marketing world model to automatically generate multiple counterfactual marketing strategies. Each counterfactual marketing strategy adopts different course recommendation methods, marketing script combinations, preferential schemes, trial class arrangements and user communication paths to form multiple marketing scenarios to be deduced. S43: Using each marketing scenario to be simulated as the initial node, perform multi-step state simulation of the subsequent marketing process based on the marketing state transition matrix, and construct a marketing decision search tree based on the transition relationship between each marketing state. Perform multiple rounds of path expansion and state search for each marketing path to obtain the future marketing state sequence corresponding to each marketing path. S44: Calculate the comprehensive marketing revenue of the corresponding marketing path according to each future marketing state sequence. The comprehensive marketing revenue includes at least the course transaction probability, violation risk, complaint probability, refund probability, long-term customer value, brand influence value and user satisfaction, and establish the comprehensive revenue evaluation result of the corresponding marketing path. S45: Compare the comprehensive revenue evaluation results corresponding to each marketing path, determine the target marketing path according to the preset revenue optimization rules, and output the marketing strategy execution order, marketing risk level and marketing revenue prediction results corresponding to the target marketing path. S46: Compare the consistency between the actual marketing execution results and the predicted results in the digital twin marketing environment, calculate the prediction deviation between each marketing state, and adaptively correct the state transition relationship, state prediction parameters and marketing state transition matrix in the education marketing world model based on the prediction deviation, so as to update the education marketing world model.

9. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 8, characterized in that, S43 includes the following steps: S431: Read the initial marketing state corresponding to each marketing scenario to be simulated, and use each initial marketing state as the root node of the marketing decision search tree. Initialize the marketing decision search tree according to the marketing environment information corresponding to each initial marketing state to form multiple independent marketing simulation branches. S432: Based on the marketing state transition matrix in the education marketing world model, calculate multiple candidate subsequent marketing states corresponding to the current marketing state, and generate a state transition set corresponding to each candidate marketing state by combining the marketing constraints corresponding to the current marketing state. S433: For each candidate marketing state, call each functional agent in the dynamic graph agent collaborative network to generate corresponding marketing behaviors. Each functional agent jointly completes marketing strategy negotiation based on the current marketing state, and determines the target marketing behavior for the corresponding candidate marketing state based on the negotiation results, so as to realize the joint deduction between marketing state and marketing behavior. S434: Based on the state change results after the target marketing behavior acts on the education marketing world model, generate corresponding child nodes for each candidate marketing state, and establish the marketing state transition relationship between the parent node and the child node to complete one round of path expansion of the marketing decision search tree; S435: For each newly added sub-node, repeatedly execute the marketing status prediction, agent collaborative decision-making and marketing status expansion process, and recursively expand the marketing decision search tree in multiple rounds until the preset inference depth, marketing goal completion status or termination constraint is reached, forming a multi-layer marketing decision search tree. S436: Based on the predicted transaction revenue, marketing risk, user satisfaction, course renewal, and marketing compliance evaluation results corresponding to each marketing status node, comprehensively evaluate each marketing path in the marketing decision search tree and identify the target marketing path that meets the preset revenue conditions. S437: For marketing paths whose comprehensive evaluation results are lower than the preset evaluation threshold, dynamic pruning is performed on the corresponding marketing paths based on the marketing constraints, marketing risk constraints and collaborative evaluation results corresponding to the dynamic graph Agent collaborative network in the education marketing world model. Marketing paths that meet the marketing objectives are retained, and an optimized marketing decision search tree is generated. S438: Based on the optimized marketing decision search tree, extract the order of marketing state nodes corresponding to each target marketing path, and generate a future marketing state sequence according to the order in which marketing states occur.

10. The intelligent quality inspection effect optimization and evaluation system based on multi-agent collaboration according to claim 8, characterized in that, The following is included before S43: Before expanding the marketing decision search tree, based on the environmental constraints corresponding to the initial marketing state, an orthogonal experimental design method is applied to decouple and perturb the key decision variables in the candidate marketing strategies, generating a set of mutually orthogonal counterfactual perturbation factors. These counterfactual perturbation factors include at least price sensitivity perturbation factors, course matching perturbation factors, and emotional communication style perturbation factors. Each perturbation factor is then superimposed onto the current candidate marketing strategy to generate multiple counterfactual marketing strategies covering different decision dimensions.

Images