Intelligent interactive enterprise management simulation system and method thereof

Abstract

The invention discloses an intelligent interactive enterprise management simulation system and a method thereof. The method comprises the following steps: constructing a multi-level state causal graph model with a multi-dimensional feature tag; fusing interaction data of the user, constructing a behavior intention tensor, and identifying a situation transition intention through a multi-modal fusion reasoning algorithm; a personalized interaction strategy is dynamically generated by a context-driven interaction optimization generation algorithm; a context semantic tensor is constructed based on a semantic context dynamic matching algorithm, a semantic compression response is realized in combination with a current intention and a historical information path, and a future strategy plan is actively generated through a causal relationship backstepping inference device; and establishing a situation feedback learning and weight updating mechanism, and dynamically adjusting a state causal graph, an intention tensor and an interaction strategy parameter to realize self-evolution closed-loop optimization of the model. Context changes can be perceived in real time, and the decision intention of the user can be deduced deeply.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise management, and in particular to an intelligent interactive enterprise management simulation system and method thereof. Background Art

[0002] In the process of modern enterprise management education and decision-making training, simulation systems, as an effective auxiliary tool, are gradually evolving from static teaching aids to highly interactive, personalized and intelligent simulation platforms. Enterprise management simulation not only covers multi-dimensional management subsystems such as marketing, financial operations, and human resource deployment, but also requires participants to make flexible and forward-looking management decisions in complex situations. With the development of artificial intelligence and data perception technology, building an intelligent interactive management simulation system has become a key path to improve the efficiency of enterprise management training and enhance resilience. However, most current simulation systems are still mainly rule-driven and pre-set scripts, lacking the ability to deeply understand user behavior and situational evolution, and are difficult to meet the highly dynamic, sudden and changeable interactive needs in real enterprise operations.

[0003] Especially in complex scenarios, such as simulating a sudden market crisis, supply chain disruption or major policy changes, users' operational behaviors, information focus, language instructions and decision-making intentions are often significantly different from daily management situations. This requires the system to not only identify behavioral changes, but also understand the management intentions behind the behavior and dynamically adjust the information presentation method and interaction strategy based on the current situation. However, in reality, existing systems generally rely on static templates and a single interaction channel, making it difficult to capture users' fine-grained decision intentions under multimodal data; at the same time, they lack the ability to perceive the semantics of the current enterprise situation, resulting in the output information feedback being either redundant and complicated, or deviating from the user's real needs, affecting the realism and effectiveness of decision simulation.

[0004] Therefore, how to design an enterprise management simulation system that can perceive context changes in real time, deeply infer user decision-making intentions, and dynamically generate optimal interaction strategies and accurate information feedback is a core technical problem that needs to be solved urgently. Summary of the invention

[0005] In order to solve the above problems, the present invention provides an intelligent interactive enterprise management simulation system and method thereof.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows: On the one hand, the present invention discloses an intelligent interactive enterprise management simulation method, comprising: Step 1: By constructing a multi-level state causal graph model with multi-dimensional feature labels, a structured situation modeling foundation supporting context recognition is formed; Step 2: Based on the state causality graph modeling, fuse the user's interaction data, construct a behavior intention tensor, and identify its context transition intention through a multi-modal fusion inference algorithm; Step 3: Jointly input the multi-dimensional feature labels and the behavior intention tensor into the graph neural decision network, and dynamically generate personalized interaction strategies by the context-driven interaction optimization generation algorithm; Step 4: Rely on the semantic context dynamic matching algorithm to construct a context semantic tensor, combine the current intention with the historical information path to achieve semantic compression response, and actively generate future strategy plans through the causal relationship inverse inference engine; Step 5: Establish a context feedback learning and weight update mechanism, take the policy execution effect, user feedback, and behavior deviation as inputs, and dynamically adjust the state causality graph, intention tensor, and interaction policy parameters to achieve self-evolution closed-loop optimization of the model.

[0007] Furthermore: The said Step 1 includes: Pre-define three types of context state nodes, namely macro management state, meso business state, and micro operation state, to form a hierarchical state system; Connect each hierarchical state node through a state mapping function, construct the mapping function based on the causal trigger rule set, and achieve the causal association between states; Assign multi-dimensional feature labels such as suddenness weight, controllability coefficient, and external dependence degree to each state node to form a node feature vector; According to specific causal trigger conditions and influence intensity parameters, define the edge weight causal relationship between state nodes by using a composite function based on logical rules and historical data regression; Define a state transition record matrix, and update the migration frequency and path strength between each state in real time, as the basis for the self-evolution of the state causality graph; Construct a state graph heat function, generate a state graph visualization interface in real time, display key state clusters, and assist the system to perceive context mutations in advance.

[0008] Furthermore: The said Step 2 includes: Real-time collect natural language input streams, mouse and touch trajectory data, keyboard input and shortcut operations, pause and delay feature modal behavior data, and map them to the standard vector space through a feature extraction function to construct a multi-modal interaction vector splicing body; Around the established multi-level context state graph, bind the user behavior to a specific context state to form a state-behavior time series tensor, define a three-dimensional behavior intention tensor, and dynamically update it through a sliding window mechanism to maintain the latest behavior context; Use an adaptive tensor compression mechanism based on behavior - focused weights to construct a behavior - focused weight function. Perform tensor clipping according to the focused weights, and extract the compressed key behavior intention vector sequence as the input basis for intention inference; Based on the compressed behavior sequence, construct an intention transition map, define the set of intention nodes, and use a classifier function to map the behavior sequence to the intention nodes. When the intention changes, add edges in the map, and determine that the user enters a new context mode by detecting intention node jumps and path mutations; Feed the recognized current intention and its associated path back to the state causal graph, drive it to dynamically label the current state as a high - attention - high - transition - risk node, and provide a context - driven basis for generating subsequent interaction strategies, and generate a behavior intention discrimination result structure in real - time.

[0009] Further: The step 3 includes: Fuse the multi - dimensional feature labels of the currently activated state node with the core compressed representation in the current behavior intention tensor to form a unified context representation vector, and complete the fusion in a weighted non - linear combination manner; Design a graph neural decision network, using the hierarchical state causal graph as the underlying graph structure basis, embed the context semantic fusion representation vector into the graph as a node injection signal, drive policy generation through the GNN update iteration formula, and the output layer maps the fusion representation to an interaction policy triple, including the recommended information display method, information granularity structure, and system response rhythm; Introduce a policy execution feedback record and reinforcement learning update mechanism, continuously record the user's response behavior to each round of interaction policies, construct a policy reward - punishment function to guide the parameter update of the GDN, and perform network update through the policy gradient method to achieve the closed - loop evolution of policy generation - feedback learning - policy optimization.

[0010] Further: The step 4 includes: Based on the state node causal graph structure and the fused context vector, construct a high - order tensor to represent the current semantic background and intention nesting relationship, and express the composite nesting relationship between three - dimensional interactions through the tensor product as the semantic basis for subsequent semantic matching; Design a semantic compression kernel to compress the information components in the user's input request that are not relevant to the current context, and perform structural reconstruction based on the semantic tensor context; Design a multi - layer semantic matching mechanism based on the semantic reconstruction vector and the enterprise knowledge base data, fuse the causal graph node weights as the priority retrieval factor, ensure that the information response focuses on the key state and its potential evolution path, finally select the content with a high score as the first - round information response set, and complete the content supplementation mechanism according to the user's behavior trigger; Based on the structural reversibility of the state causality graph, design a reverse path backtracking and forward condition expansion inference engine to automatically invert the possible antecedents and future evolution paths in the current situation and generate decision plan suggestions.

[0011] Furthermore, step 5 includes: Design a multi-source self-learning mechanism that integrates the recommended adoption situation of strategies, users' subjective feedback, and the degree of deviation of behavioral trajectories to provide a basis for subsequent model parameter updates; Based on the state transition record matrix, compare the actual state transition path and the predicted path after the execution of the strategy, and correct the conditional transition probability of the state edge to achieve the dynamic adjustment of the state causality graph; According to the degree of behavior deviation, perform gradient descent correction on the weight matrix of the behavior intention tensor to ensure that the behavior intention tensor more accurately reflects the user's behavior characteristics; Maintain a dynamically evolving strategy success rate matrix to record the adoption and execution success situations of strategies in different states; for strategies that continuously fail, automatically enter the "strategy cold storage pool" and trigger a new strategy combination training module for alternative generation; Take the user feedback and adoption behavior as reward signals and connect them to the strategy generation layer of the graph neural decision network to update the strategy generation weights, so that the system gradually tends to generate interactive strategies that are frequently adopted by users and have excellent execution effects, and complete the strategy evolution closed-loop.

[0012] On the other hand, the present invention discloses an intelligent interactive enterprise management simulation system, including: Multi-level enterprise management simulation scenario dynamic modeling module: By constructing a multi-level state causality graph model with multi-dimensional feature labels, form a structured scenario modeling basis that supports context recognition; Real-time multi-modal interaction data perception and behavior intention reasoning module: Based on the state causality graph modeling, fuse the interaction data of users, construct a behavior intention tensor, and identify its context transition intention through a multi-modal fusion reasoning algorithm; Interactive strategy dynamic generation module: Jointly input the multi-dimensional feature labels and the behavior intention tensor into the graph neural decision network, and dynamically generate personalized interactive strategies by the context-driven interactive optimization generation algorithm; Information response module: Rely on the semantic context dynamic matching algorithm to construct a context semantic tensor, combine the current intention and the historical information path to achieve semantic compression response, and actively generate future strategy plans through a causal relationship inverse reasoning engine; Simulation feedback and model self-evolution module: Establish a context feedback learning and weight update mechanism, take the strategy execution effect, user feedback, and behavior deviation as inputs, and dynamically adjust the state causality graph, intention tensor, and interactive strategy parameters to achieve the self-evolution closed-loop optimization of the model.

[0013] Compared with the prior art, the technical progress achieved by the present invention lies in: The present invention constructs a dynamic enterprise scenario representation system with a three - level structure of "macro - management state - meso - business state - micro - operation state" through designing a hierarchical state - causality graph modeling algorithm, and updates the state transition path in real - time during the user interaction process, which can faithfully reflect the context evolution process in enterprise management and solve the technical short - board of the current simulation system of "static context and rigid state transition".

[0014] By introducing an interactive behavior intention reasoning algorithm with high - dimensional modal fusion, the system comprehensively analyzes modal data such as the user's language input, interaction trajectory, and pause duration, constructs a behavior intention tensor and an intention transition map, and effectively captures the deep decision - making intention behind the change of the user's behavior pattern, thus breaking through the limitation of the traditional system that only relies on language text or single operation to judge the user's intention.

[0015] Based on the state - node features and intention tensors obtained in the first two steps, this method uses a context - driven interaction optimization generation algorithm to fuse multi - dimensional information through a graph neural network model, dynamically generate the optimal interaction mode, information granularity, and response rhythm, effectively coping with the significant differences in user interaction requirements under daily operations and emergency situations, and solving the drawbacks of the current system of "rigid feedback and fixed rhythm".

[0016] By introducing a semantic context dynamic matching algorithm, through constructing a context semantic tensor and a semantic compression core, the system can focus on the user's real needs, actively compress non - critical information, and generate future strategy plans through a causal reverse - inference module, realizing the transformation of information response from "surface matching" to "context deep coupling" and solving the persistent problems of information feedback such as "off - topic, redundant, and inaccurate".

[0017] By constructing a context - feedback learning and weight - updating mechanism, the system continuously learns the user's preferences and feedback results after each round of simulation, and dynamically optimizes the state modeling, intention reasoning, and strategy generation modules through a strategy success rate matrix, thus forming a trinity evolution closed - loop of "data - model - strategy" and realizing the fundamental transition of the model from "passive response" to "active adaptation".

[0018] In summary, the present invention not only accurately responds to the core interaction requirements of "highly dynamic, multi - dimensional, and deep - semantic" in the enterprise management context, but also effectively breaks the drawbacks of the traditional system with slow response and rigid strategies through a series of structured tensor modeling, causality graph optimization, and neural decision - making mechanisms. It is a brand - new enterprise management simulation technical path with high context adaptability, intention interpretability, and system self - evolution ability. Brief Description of the Drawings

[0019] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0020] In the accompanying drawings: Figure 1 is a flowchart of the method according to the present invention; Figure 2 is a schematic block diagram of the system according to the present invention. Detailed implementation manners

[0021] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0022] Embodiment 1 As Figure 1 shown, the present invention discloses an intelligent interactive enterprise management simulation method, including: Step 1: Construct a multi-level enterprise management simulation scenario dynamic modeling system Its purpose is to construct a model with dynamic state migration ability, providing a structured, traceable, and computable scenario basis for the subsequent context recognition module.

[0023] 1. Scenario state hierarchy definition module By predefining three types of context state nodes, a hierarchical state system of "macro management state - meso business state - micro operation state" is formed: The set of macro management states is denoted as:

[0024] For example: industry prosperity, external policies, macro supply and demand, etc.

[0025] The set of meso business states is denoted as:

[0026] For example: revenue trend, cost fluctuation, cash flow pressure, etc.

[0027] The set of micro operation states is denoted as:

[0028] For example: inventory level, production capacity utilization, employee scheduling, etc.

[0029] The connections between each level are made through state mapping functions:

[0030]

[0031] The mapping function is constructed based on a set of causal trigger rules. For example: when "the market demand plummets", it may trigger "the deterioration of the accounts receivable turnover".

[0032] 2. Implementation of the Hierarchical Scenario Causal Graph Modeling (HSCGM) algorithm This algorithm includes the following core components: 2.1 Node construction rules Each state node has the following multi-dimensional feature labels: : The suddenness weight, which is a probability index indicating its sudden nature.

[0033] : The controllability coefficient, which indicates whether its state can be changed through enterprise intervention.

[0034] : The external dependence degree, which depicts whether this state depends on macro external variables.

[0035] Construct the node feature vector:

[0036] 2.2 Edge weight causal definition For any two state nodes , , if the following causal trigger conditions are met, a directed edge is established :

[0037] Among them, represents the current external environmental variable, represents the influence intensity parameter of the edge, and the function is a composite function based on logical rules and historical data regression, and its form is as follows:

[0038] Among them is the sigmoid function.

[0039] 3. Dynamic update mechanism of the state transition record matrix During the simulation process, each operation by the user may trigger a change in the state. For this reason, this embodiment defines the state transition record matrix to update the migration frequency and path strength between each state in real time: Let: : It represents the cumulative transfer weight of state at time .

[0040] The update rule is as follows:

[0041] Where: : The smoothing factor that controls the weight of historical memory; If is triggered by the current interaction, otherwise it is 0.

[0042] This matrix serves as the basis for the self-evolution of the state causal graph and can dynamically enhance or weaken specific state paths according to the user's behavior in different simulation scenarios, thereby achieving the "soft adjustment" of the situation graph.

[0043] 4. State graph visualization and heat dynamic perception interface Construct a state graph heat function to reflect the current trigger frequency, interaction activity, and prediction risk level of a certain state node:

[0044] Where: : The trigger frequency of the edges with this node as the starting point; : The operation frequency of the user related to this node; : The risk assessment value of this node in the simulation (from subsequent model predictions); , , : The weighting coefficient is adjustable.

[0045] Generate a state graph visualization interface in real time through the heat function to display the state clusters of "high risk - high frequency operation - high dependence", and assist the system to perceive context mutations in advance.

[0046] The above are four closely related technical steps of hierarchical node modeling, state causal graph construction, real-time update of the state transfer record matrix, and visualization of state heat feedback, forming a structured and dynamically adaptive situation engine, which solves the problem of insufficient context perception basis in the enterprise simulation process.

[0047] Step 2: Design a real-time multi-modal interaction data perception and behavior intention reasoning module Its purpose is to achieve real-time, multimodal perception and high-dimensional representation of users' ever-changing interactive behavior characteristics in enterprise management simulation systems, and on this basis to infer their potential decision-making intentions and information demand preferences to support the subsequent dynamic generation process of interactive strategies.

[0048] 1. Multimodal interaction data collection and structured mapping mechanism construction During the interaction with the user, the system continuously collects the following four types of main modal behavior data in real time to build a standardized "interaction feature dictionary": : Natural language input stream (including query terms, commands, and semantic structures) : Mouse and touch trajectory data (including movement path, click hotspot) : Keyboard input and shortcut operations (including key frequency and combination commands) : Pause and delay characteristics (including lag time before operation and rollback frequency) Each type of modal data is extracted through the feature extraction function , uniformly mapped to the standard vector space:

[0049] And construct a multimodal interaction vector mosaic:

[0050] This vector is the user's time slice in the simulation scene. The interactive behavior representation vector under .

[0051] 2. Build Behavioral Intent Tensor (BIT) Based on the multi-level situational state diagram established in the above step 1, user behavior is bound to specific situational states to form a "state-behavior time series tensor".

[0052] Define a three-dimensional behavior intention tensor:

[0053] in: Indicates the state node number (from the state diagram ); represents the time step; Represents feature dimension; Indicates in status Previous The moment One behavioral feature value.

[0054] This tensor can be dynamically updated through a sliding window mechanism to maintain the latest behavioral context.

[0055] 3. Tensor Compression and Key Interaction Feature Extraction Mechanism To avoid the problems of overly sparse or dimensional explosion of the behavioral intention tensor, an innovative adaptive tensor compression mechanism based on behavior-focused weights (Adaptive Attention-Guided Tensor Reduction, AAGTR) is used. Its core idea is to extract local behavioral segments that contribute most to intention recognition in the state dimension and time dimension.

[0056] Construct the behavior-focused weight function:

[0057] Where: : Represents the weight vector of the th state; : Is the historical behavioral hidden state (which can be generated by GRU / RNN); : Represents the behavioral weight of a certain state at a certain moment.

[0058] Perform tensor pruning according to the focus weights:

[0059] Thus, the compressed key behavioral intention vector sequence is extracted as the input basis for intention inference.

[0060] 4. Construct an Intent Transition Graph to Identify User Context Switching Based on the compressed behavioral sequence, the system constructs an Intent Transition Graph (ITG) to describe the decision-making intention migration pattern of the user in different context states.

[0061] The construction rules are as follows: Define the set of intention nodes:

[0062] Each intention node represents a potential goal, such as "stable operation", "emergency stop loss", "market expansion".

[0063] Use the classifier function , and map the behavioral sequence to the intention node:

[0064] If the intention changes between time steps i.e., , then an edge is added to the ITG:

[0065] where represents the transition time, and represents the change amplitude of the behavior tensor.

[0066] When the system detects frequent jumps of intention nodes, path mutations, or a transition path that is significantly different from previous trajectories, it can be determined that the user has entered a new situational mode (e.g., transitioning from "steady-state operation" to "crisis response").

[0067] 5. Output of the behavior intention discrimination result and linkage with the state diagram Finally, the currently identified intention and its associated path are fed back to the state causality diagram in step 1, driving its dynamic annotation of the current state as a "high-concern - high-transition-risk" node and providing a context-driven basis for generating subsequent interaction strategies.

[0068] The system generates the following behavior intention discrimination result structure in real time: { "current_intent": "Response to market drawdown", "intent_confidence": 0.94, "recent_intent_transitions": {"from": "Steady operation", "to": "Alert to revenue decline", "Δt": 5s}, {"from": "Alert to revenue decline", "to": "Response to market drawdown", "Δt": 3s} , "associated_state_node": "Macro management state s_2", "triggered_edge": "s_2, s_5 (high intensity)" } This step, through multi-modal data collection, construction of behavioral intention tensors, extraction of tensor compression features, identification of intention transition graphs, and annotation of result-linked status icons, not only achieves high-precision, low-latency, and multi-modal real-time capture of the user's behavioral changes in the enterprise simulation, but also further constructs a user state understanding mechanism centered on intention recognition, laying a solid semantic and structural foundation for the subsequent dynamic adjustment of interaction strategies and information feedback by the system.

[0069] Step 3: Construct a dynamic generation mechanism for context-aware guided interaction strategies Its purpose is to automatically generate a set of optimal interaction strategies in three dimensions: information presentation mode, content granularity, and response rhythm, based on the user's current situational state and potential behavioral intentions, and continuously optimize them through a feedback mechanism.

[0070] 1. Construct a Contextual Semantic Fusion Vector (CSFV) Structurally fuse the multi-dimensional feature labels of the currently activated state nodes in Step 1 with the core compressed representation in the current behavioral intention tensor in Step 2 to form a unified context representation vector , which is used to drive the policy generation network.

[0071] Suppose: : is the multi-dimensional structural semantic vector of the currently activated state node (including suddenness, controllability, hierarchy, etc.); : is the compressed behavioral intention representation at the current time step (from the tensor compression module); : is the fused context vector.

[0072] The fusion method adopts weighted non-linear combination:

[0073] where , is the trainable feature fusion weight matrix, is the bias term.

[0074] 2. Construct a Graph Decision Net (GDN) to generate interaction strategy triples Design an innovative Graph Decision Net (GDN) structure, using the hierarchical state causal graph in Step 1 as the underlying graph structure basis, and at the same time embedding the above fusion vector into the graph as a node injection signal to drive policy generation.

[0075] The graph structure is defined as follows: Graph , where each node corresponds to a state node; Edge weight represents the state transition probability and causal strength; The node features are initialized as the semantic vectors of the state nodes ; Update the iteration formula through GNN:

[0076] where: is the initial state node feature; is the attention weight; is used to regulate to make it biased towards the current context.

[0077] The output layer will fuse the representation and map it to an interaction policy triple:

[0078] where: represents the recommended information presentation method; represents the recommended information granularity structure; represents the system response rhythm.

[0079] These output categories are completed by three parallel classifier head networks , , :

[0080] 3. Build an interaction policy feedback loop mechanism (Policy Reinforcement Loop) To enhance the self-correction ability of the policy generation mechanism, a set of policy execution feedback recording and reinforcement learning update mechanisms are introduced. This mechanism continuously records the user's response behavior to each round of interaction policy and constructs a policy reward and punishment function to guide the parameter update of GDN.

[0081] Define the reward function:

[0082] where: : the proportion of user click feedback information; : Whether an effective decision is made based on system information; : Whether the user actively skips the system's suggestions or requests to change the display method; is an adjustable weight parameter.

[0083] The GDN network is updated by the policy gradient method:

[0084] In this way, after each policy execution, the network continuously optimizes its interaction policy preference according to the user feedback signal, thus realizing the closed-loop evolution of "policy generation - feedback learning - policy optimization".

[0085] This step constructs a complete mechanism chain integrating multi-modal context fusion, graph neural policy decision-making, and reinforcement feedback learning around the core goal of "from situation-intention to policy generation". By introducing the fusion vector , the policy generation graph network GDN, and the policy optimization mechanism based on behavioral reward and punishment feedback, the system can adapt to dynamic situation changes with high precision, dynamically select the optimal information display and interaction path, and continuously self-iterate and optimize through actual effects, thus effectively solving the "situation dynamics - behavior change - policy response" fault problem in enterprise management simulation.

[0086] Step 4: Implement an intelligent information response engine based on semantic tensor matching Its purpose is to realize the dynamic semantic reconstruction and high-precision response to user requests based on the current context semantic state and the inferred behavioral intention, combined with the historical interaction trajectory, so as to provide decision-making assistance content highly aligned with the scenario and avoid redundant or misleading information output.

[0087] 1. Construct a Contextual Semantic Tensor (CST) Based on the state node causal graph structure constructed in Step 1 and the fusion context vectors generated in Steps 2 and 3 , construct a high-order tensor to represent the current semantic background and intention nesting relationship.

[0088] Let: : The multi-dimensional feature label of the currently activated state node; : The behavioral intention tensor, from Step 2; : The historical interaction path tensor, recording the content features that the user has requested, ignored, or repeatedly called; Then the contextual semantic tensor is defined as:

[0089] Among them represents the outer product, which is used to express the complex nested relationship of the context-intention-history path in three-dimensional interaction and serves as the semantic basis for subsequent semantic matching.

[0090] 2. Construct a Semantic Compression Kernel (SCK) to achieve information reconstruction Design an attention-guided module called "Semantic Compression Kernel" to compress the information components in the user input request (natural language, click intention, etc.) that are not relevant to the current context, and perform structural reconstruction based on the semantic tensor context, making the response content more context-focused.

[0091] Define the original input semantic representation as , then the compression and reconstruction process is as follows:

[0092] Among them, is the semantic attention function, which is essentially selective compression and reconstruction based on context attention weights in the tensor space:

[0093] This mechanism ensures that the generated compressed query vector can accurately express the user's true intention and scene key points, pointing the way for subsequent information extraction.

[0094] 3. Implement a semantic matching-based information response module Design a multi-layer semantic matching mechanism based on the semantic reconstruction vector and the structured and unstructured text data in the enterprise knowledge base, and fuse the causal graph node weights from step 1 as the priority retrieval factor to ensure that the information response always focuses on the key state and its potential evolution path.

[0095] The matching score function is defined as:

[0096] Among them: : represents the semantic vector of the knowledge base document or metric; : represents the node activation weight of this document or metric in the causal graph of the current state; , : is the adjustable fusion weight.

[0097] Finally, select the top The high-priority content is used as the first-round information response set and trigger the content completion mechanism according to user behavior when necessary.

[0098] 4. Introduce a causal relationship reverse inference engine to realize future scenario simulation Based on the structural reversibility of the state causal graph in step 1, design a reverse path backtracking and forward condition expansion inference engine. When the user faces a highly uncertain scenario (such as a market crash), automatically invert the possible antecedents and future evolution paths in the current situation, and generate decision plan suggestions.

[0099] The implementation method is as follows: 1. Reverse infer antecedent nodes: Trace all sets of parent nodes with high causal weights backward from the currently activated state node ; ; 2. Simulate future state sequences: Starting from the current node, generate the path sequence with the highest probability through the multi-round state transition prediction model ; ; 3. Trigger the plan generator: Use the state combination under this path as the conditional input, and call the strategy simulation module to generate candidate decision paths ;

[0100] The formal expression is as follows:

[0101] Finally, the plan is actively pushed in the form of a strategy recommendation card, which constitutes a part of the information response and has high forward-looking.

[0102] Based on the system logic of "scenario construction - intention recognition - strategy generation" in the previous three steps, this step proposes a closed-loop information response system around "semantic tensor reconstruction - compression-guided matching - causal path simulation". Through tensor-level context expression, compression kernel-controlled semantic redirection, structured graph-driven information recall and future plan recommendation, it truly realizes "delivering the right content in the right way at the right time" in the enterprise management simulation system.

[0103] Step 5: Build a closed-loop mechanism for simulation feedback and model self-evolution Its purpose is that after each round of enterprise management simulation interaction is completed, the system should automatically adjust its core model parameters (including the state causal graph structure, behavior intention tensor, interaction strategy generation module) based on the user's behavior adoption, feedback preference, and strategy execution effect, so as to achieve enhanced dynamic capabilities of personalization, adaptability, and progressive evolution.

[0104] 1. Construct a Contextual Feedback Learning with Weight Updating (CFLWU) mechanism Design a multi-source self-learning mechanism that integrates three types of feedback metrics. The core components include: : The current policy recommendation set; : The actual adoption behavior of the user (click, execute, modify, etc.); : The subjective feedback of the user (satisfaction score, dwell time, frequent return, etc.); : The degree of deviation of the behavior trajectory and the difference in the matching degree with the previous intention tensor.

[0105] 2. Update the node transition probability and causal edge weight in the state causal graph Trace back to the "state transition record matrix" defined in step 1, and compare each state transition path after each policy execution as the "actual state transition path": Let the current state be , and after the policy execution, enter the state ; The original predicted path in the system is ; Then, correct the conditional transition probability of each state edge :

[0106] Where is the Kronecker function (1 for match, 0 for non-match), is the learning rate hyperparameter that controls the feedback adjustment amplitude.

[0107] 3. Adjust the behavior feature weights in the behavior intention tensor For the behavior intention tensor constructed in step 2 , the system needs to correct the tensor axis weights according to the behavior deviation degree .

[0108] Let the behavior deviation distance be:

[0109] Where represents the Frobenius norm. If the deviation amount is higher than the preset threshold , then perform a gradient descent correction on the tensor weight matrix :

[0110] where is the behavior learning rate, and the loss function is the difference function between the actual behavior and the expected intention.

[0111] 4. Construct a policy success rate matrix and a policy elimination mechanism Maintain a dynamically evolving policy success rate matrix within the system , where: Rows represent different state contexts ; Columns represent candidate policy combinations ; The element represents the number of times the policy is adopted and executed successfully in the state .

[0112] Whenever a policy is executed, based on user feedback judge whether the policy is successful and update the matrix:

[0113] If a certain policy fails to execute continuously rounds in a certain state (such as the adoption rate is lower than a certain threshold ), it will automatically enter the "policy cold storage pool" and the system will trigger a new policy combination training module for replacement generation.

[0114] 5. Introduce a reinforcement learning factor to optimize policy generation Take user feedback and adoption behavior as reward signals and connect them to the policy generation layer of the Graph Decision Net constructed in step 3 to update the policy generation weights:

[0115] where represents the feedback round reward, and is the learning rate.

[0116] Through the reinforcement learning mechanism, the system gradually tends to generate interaction policies that are frequently adopted by users and have excellent execution effects, thus completing the policy evolution closed-loop.

[0117] This step constitutes a complete self - learning closed - loop architecture that can dynamically evolve the structure, weights, and strategies based on real - user behavior feedback through four major mechanisms: "policy adoption rate modeling - state causal structure fine - tuning - behavior tensor weight correction - policy elimination / reinforcement learning", enabling the system to have the evolutionary capabilities of "continuous adaptation, gradual optimization, and accurate prediction", ensuring that the entire intelligent enterprise management simulation platform becomes smarter and more accurate with each round of interaction.

[0118] Embodiment 2 As Figure 2 shown, this embodiment discloses an intelligent interactive enterprise management simulation system, including: Multi - level enterprise management simulation scenario dynamic modeling module: By constructing a multi - level state - causal graph model with multi - dimensional feature tags, a structured situation modeling foundation that supports context recognition is formed; Real - time multi - modal interaction data perception and behavior intention reasoning module: Based on the state - causal graph modeling, integrating the user's interaction data, constructing a behavior intention tensor, and identifying its context transition intention through a multi - modal fusion reasoning algorithm; Interactive policy dynamic generation module: Jointly inputting the multi - dimensional feature tags and the behavior intention tensor into a graph neural decision network, and dynamically generating personalized interaction policies by a context - driven interaction optimization generation algorithm; Information response module: Relying on a semantic context dynamic matching algorithm to construct a context semantic tensor, realizing semantic compression response by combining the current intention and the historical information path, and actively generating future policy plans through a causal relationship reverse inference engine; Simulation feedback and model self - evolution module: Establishing a context feedback learning and weight update mechanism, taking the policy execution effect, user feedback, and behavior deviation as inputs, and dynamically adjusting the state - causal graph, intention tensor, and interaction policy parameters to achieve self - evolution closed - loop optimization of the model.

[0119] The modules in Embodiment 2 are used to implement the functions in Embodiment 1. This embodiment can be implemented by a system including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent interactive enterprise management simulation method described in Embodiment 1 of the present application is realized. The system also includes other components well - known to those skilled in the art such as a communication bus and a communication interface, and their settings and functions are known in the art, so they will not be elaborated here.

[0120] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. An intelligent interactive enterprise management simulation method, characterized in that, Including: Step 1: Form a structured situation modeling basis for supporting context recognition by constructing a multi-level state causal graph model with multi-dimensional feature labels; Step 2: On the basis of state causal graph modeling, fuse the user's interaction data, construct a behavior intention tensor, and identify its context transition intention through a multi-modal fusion inference algorithm; Step 3: Jointly input the multi-dimensional feature labels and the behavior intention tensor into a graph neural decision network, and dynamically generate a personalized interaction strategy by a context-driven interaction optimization generation algorithm; Step 4: Rely on a semantic context dynamic matching algorithm to construct a context semantic tensor, combine the current intention with the historical information path to achieve semantic compression response, and actively generate a future strategy plan through a causal relationship inverse reasoning engine; Step 5: Establish a context feedback learning and weight update mechanism, take the strategy execution effect, user feedback and behavior deviation as inputs, and dynamically adjust the state causal graph, intention tensor and interaction strategy parameters to achieve self-evolution closed-loop optimization of the model.

2. The intelligent interactive enterprise management simulation method according to claim 1, wherein The said Step 1 includes: Pre-define three types of context state nodes: macro management state, middle-level business state and micro operation state, and form a hierarchical state system; Connect each hierarchical state node through a state mapping function, construct the mapping function based on a set of causal trigger rules, and realize the causal association between states; Assign multi-dimensional feature labels of suddenness weight, controllability coefficient and external dependence degree to each state node to form a node feature vector; Define the edge weight causal relationship between state nodes by using a composite function based on logical rules and historical data regression according to specific causal trigger conditions and influence intensity parameters; Define a state transition record matrix, and update the migration frequency and path strength between each state in real time as the basis for the self-evolution of the state causal graph; Construct a state graph heat function, generate a state graph visualization interface in real time, display key state clusters, and assist the system to perceive context mutations in advance.

3. An intelligent interactive enterprise management simulation method according to claim 2, characterized in that, The said Step 2 includes: Collect natural language input streams, mouse and touch trajectory data, keyboard input and shortcut operations, pause and delay feature modal behavior data in real time, map them to a standard vector space through a feature extraction function, and construct a multi-modal interaction vector splicing body; Around the established multi-level context state graph, bind the user behavior to a specific context state to form a state-behavior time series tensor, define a three-dimensional behavior intention tensor, and dynamically update it through a sliding window mechanism to maintain the latest behavior context; Use an adaptive tensor compression mechanism based on behavior focus weight to construct a behavior focus weight function, perform tensor clipping according to the focus weight, and extract the compressed key behavior intention vector sequence as the input basis for intention inference; Based on the compressed behavior sequence, construct an intention transition map, define an intention node set, map the behavior sequence to the intention node by using a classifier function, add an edge to the map when the intention changes, and determine that the user enters a new context mode by detecting intention node jumps and path mutations; Feed the recognized current intention and its associated path back into the state causal graph, driving it to dynamically label the current state as a high-concern - high-transition-risk node, and providing a context-driven basis for generating subsequent interaction strategies, and generating a behavior intention discrimination result structure in real time.

4. An intelligent interactive enterprise management simulation method according to claim 3, characterized in that Step 3 includes: Fuse the multi-dimensional feature labels of the currently activated state node with the core compressed representation in the current behavior intention tensor to form a unified context representation vector, and complete the fusion in a weighted non-linear combination manner; Design a graph neural decision network, using the hierarchical state causal graph as the underlying graph structure basis, embed the context semantic fusion representation vector into the graph as a node injection signal, drive the generation of strategies through the GNN update iteration formula, and the output layer maps the fusion representation into an interaction strategy triple, including the recommended information display method, information granularity structure, and system response rhythm; Introduce a policy execution feedback record and reinforcement learning update mechanism, continuously record the user's reaction behavior to each round of interaction strategies, construct a policy reward and punishment function to guide the parameter update of the GDN, and perform network update through the policy gradient method to achieve a closed-loop evolution of policy generation - feedback learning - policy optimization.

5. An intelligent interactive enterprise management simulation method according to claim 4, characterized in that, Step 4 includes: Based on the state node causal graph structure and the fused context vector, construct a high-order tensor to represent the current semantic background and intention nesting relationship, and express the composite nesting relationship between three-dimensional interactions through the tensor product, serving as the semantic basis for subsequent semantic matching; Design a semantic compression kernel to compress the information components in the user input request that are irrelevant to the current context, and perform structural reconstruction based on the semantic tensor context; Design a multi-layer semantic matching mechanism based on the semantic reconstruction vector and the enterprise knowledge base data, fuse the causal graph node weights as the priority retrieval factor, ensure that the information response focuses on the key states and their potential evolution paths, finally select the content with high scores as the first-round information response set, and trigger the content completion mechanism according to the user behavior; Based on the structural reversibility of the state causal graph, design a reverse path backtracking and forward conditional expansion inference engine to automatically invert the possible antecedents and future evolution paths in the current situation, and generate decision plan suggestions.

6. An intelligent interactive enterprise management simulation method according to claim 5, characterized in that, Step 5 includes: Design a multi-source self-learning mechanism that integrates the adoption situation of the fusion strategy recommendation, the user's subjective feedback, and the degree of deviation of the behavior trajectory, providing a basis for subsequent model parameter updates; Based on the state transition record matrix, compare the actual state transition path and the predicted path after the policy execution, and correct the conditional transition probability of the state edge to achieve the dynamic adjustment of the state causal graph; According to the degree of behavior deviation, perform gradient descent correction on the weight matrix of the behavior intention tensor to ensure that the behavior intention tensor more accurately reflects the user's behavior characteristics; Maintain a dynamically evolving policy success rate matrix to record the adoption and execution success situations of the policy in different states; for continuously failed policies, automatically enter the "policy cold storage pool", and trigger a new policy combination training module for alternative generation; Taking user feedback and adoption behavior as reward signals, they are connected to the policy generation layer of the graph neural decision network to update the policy generation weights, enabling the system to gradually tend to generate interaction policies that are frequently adopted by users and have excellent execution effects, thus completing the policy evolution closed-loop.

7. An intelligent interactive enterprise management simulation system, characterized in that, It includes: Multi-level enterprise management simulation scenario dynamic modeling module: By constructing a multi-level state causal graph model with multi-dimensional feature labels, a structured situation modeling foundation for supporting context recognition is formed; Real-time multi-modal interaction data perception and behavior intention reasoning module: Based on the state causal graph modeling, the interaction data of users is fused to construct a behavior intention tensor, and its context transition intention is identified through a multi-modal fusion reasoning algorithm; Interaction policy dynamic generation module: The multi-dimensional feature labels and the behavior intention tensor are jointly input into the graph neural decision network, and personalized interaction policies are dynamically generated by the context-driven interaction optimization generation algorithm; Information response module: Relying on the semantic context dynamic matching algorithm to construct a context semantic tensor, semantic compression response is realized by combining the current intention and the historical information path, and future policy plans are actively generated through the causal relationship inverse reasoning engine; Simulation feedback and model self-evolution module: A context feedback learning and weight update mechanism is established. Taking the policy execution effect, user feedback and behavior deviation as inputs, the state causal graph, intention tensor and interaction policy parameters are dynamically adjusted to realize the self-evolution closed-loop optimization of the model.

Images