Modern office affair processing practical training system and method based on digital twinning

Abstract

The application relates to the technical field of digital twinning and office affairs training, and discloses a modern office affairs processing training system and method based on digital twinning, wherein the system comprises a retrieval track collection module, an information state transition graph generation module, a role intelligent agent double-layer state representation generation module, an information disclosure strategy generation module, a probabilistic utility weight calculation module, an initial retrieval path planning module, an interactive response and belief updating module, a dynamic path re-planning module and a deviation diagnosis module. The system drives the role intelligent agent to perform strategic information disclosure through a Bayesian game framework, converts retrieval edges from a deterministic increment into a probabilistic utility weight, and dynamically re-plans an optimal retrieval path based on actual disclosure behavior observation, and further performs multi-dimensional diagnosis on the ability defects of students.

Description

Technical Field

[0001] This invention relates to the field of digital twin and office administration training technology, and more specifically, to a modern office administration training system and method based on digital twins. Background Technology

[0002] In modern office processing training scenarios, trainees are required to complete complex office tasks involving multi-source information retrieval and comprehensive decision-making, such as cross-departmental project coordination, data collection from multiple channels, and solution formulation. Completing these tasks requires obtaining necessary information from multiple information sources within the organization, with a significant amount of critical information needing to be acquired through interpersonal interactions such as inquiring about project progress from colleagues and understanding the decision-making background from supervisors. However, in actual office scenarios, information holders may selectively disclose or obscure information based on their own interests, leading to uncertainty in the actual information increment of each retrieval operation.

[0003] Existing training assessment methods typically use information state transition diagrams to model the retrieval process of office tasks and employ sequence alignment methods to evaluate the quality of trainees' retrieval paths. However, these methods have the following drawbacks: First, the information increments of each retrieval operation edge in the information state transition diagram are encoded as deterministic values, failing to reflect the strategic disclosure behavior of personnel role information sources. This renders the optimal retrieval path calculated based on deterministic increments ineffective when faced with strategic information sources. Second, the sequence alignment assessment method misjudges reasonable alternative paths as deviations when multiple equivalent optimal retrieval paths exist for a task, and it cannot quantify trainees' ability to dynamically adjust retrieval paths based on the disclosure behavior of information holders. Consequently, the intertwined defects of information game skills and retrieval path planning skills cannot be diagnosed. Summary of the Invention

[0004] This invention provides a modern office affairs processing training system and method based on digital twins, which solves the technical problems in related technologies that cannot model the uncertain disclosure behavior of personnel role information sources in office affairs training scenarios, resulting in a lack of consideration for information game factors in retrieval path planning, and the inability to conduct a combined diagnosis of trainees' information game ability and retrieval strategy ability.

[0005] This invention discloses a modern office transaction processing training system based on digital twins, including: an information state transition graph generation module, configured to generate an information state transition graph based on the information dependency relationship of office transaction tasks, distinguishing information sources into deterministic information sources and personnel role information sources, and assigning deterministic utility weights to the retrieval edges corresponding to deterministic information sources; The two-layer state representation generation module is configured to generate a two-layer state representation containing a public state space and a private state space for the role agents associated with each person's role information source. The information disclosure strategy generation module is configured to define the information disclosure action space of the role agents and train each role agent to obtain the equilibrium information disclosure strategy through a Bayesian game framework. The probabilistic utility weight calculation module is configured to calculate the probabilistic utility weight of the retrieval edge corresponding to each personnel role information source based on the balanced information disclosure strategy, and generate a probabilistic information state transition graph. The dynamic path replanning module is configured to perform Bayesian posterior updates based on the actual disclosure behavior after each visit by a student to the personnel role information source, and replan the optimal subsequent retrieval path. The deviation diagnosis module is configured to compare the student's actual retrieval decisions with the dynamic replanning path, classify deviations, and generate a composite diagnostic report that integrates information game capabilities and retrieval strategy capabilities.

[0006] Furthermore, the public state space includes the workload and collaboration commitments publicly announced by the role agent, and the information in the public state space is shared knowledge observable by all parties involved in the task; the private state space includes three parameters: actual resource surplus, internal priority sequence, and reserved buffer ratio, where the actual resource surplus is the difference between the role's currently available resources and the publicly announced workload, the internal priority sequence is the role's actual ranking of each pending task, and the reserved buffer ratio is the proportion of resources reserved by the role to cope with uncertain tasks; the information in the private state space is only visible to the role agent itself; the three parameters are mapped to a mean normalization based on the range. Interval.

[0007] Furthermore, the information disclosure action space includes four types of disclosure actions: full disclosure, providing the trainee with the true and complete content of the queried information; partial disclosure, providing only the part of the queried information that does not conflict with the interests of the role; vague expression, conveying information in an imprecise way; and false statement, providing misleading content that does not conform to the true information. Each disclosure action corresponds to a different amount of information transmission and a different degree of information distortion.

[0008] Furthermore, in the Bayesian game framework, the type of each agent is defined by a combination of three normalized parameters in its private state space, and each agent maintains a prior belief distribution regarding the type distribution of other participants; each agent optimizes by maximizing its own objective function, which is:

[0009] in, For intelligent agents In action and private state The objective function value is below. For the benefit of information protection, The costs associated with maintaining the collaborative relationship; the benefits of information protection are:

[0010] in, For private state number Importance weights of item parameters For in action and private state Next The probability estimate of the parameter being inferred by the other party; the maintenance cost of the cooperative relationship is:

[0011] in, For the role intelligent agent The intensity of collaboration and dependence between trainees For action The corresponding degree of information distortion; the equilibrium information disclosure strategy is expressed as a mapping from the belief distribution and private state to the probability distribution of the four disclosure actions. ,in For intelligent agents Current belief distribution for other participant types.

[0012] Furthermore, the probabilistic utility weight calculation module targets each person's role information source. and each type of exploration method Based on the balanced information disclosure strategy, the probability of each of the four disclosure actions being selected and the information transmission amount corresponding to each action are obtained, and the expected information disclosure amount is obtained by weighted summation:

[0013] in, Personnel role information source In the exploration method The expected amount of information disclosure It is a set of actions consisting of four disclosure actions. For any disclosure action in the action set, For action The corresponding information transmission volume; the exploration methods include three categories: direct inquiry, indirect inference, and conditional information exchange proposal; the expected information disclosure volume of the same personnel role information source under the three exploration methods is combined into the probabilistic utility weight of the retrieval edge corresponding to the information source, replacing the original deterministic information increment encoding of the retrieval edge in the information state transition diagram.

[0014] Furthermore, it also includes an initial retrieval path planning module, configured to calculate the initial retrieval path with the expected information gain maximized using a random shortest path algorithm on the probabilistic information state transition graph, starting from the initial empty information state and ending at the target information state required to complete the office task, as the benchmark for the optimal retrieval strategy based on the current game equilibrium state before the trainee performs the task.

[0015] Furthermore, the dynamic path replanning module includes: a probing behavior analysis submodule, configured to analyze the communication content of trainees, extracting the probing method category, the specific information dimension targeted by the probing, and the range of known information actively exposed by the trainees; and a Bayesian belief update submodule, configured to update the posterior belief distribution of the role agent regarding the trainees' intentions according to the Bayesian formula based on the extracted probing behavior features.

[0016] in, To observe the characteristics of the exploration behavior Next, the type of trainee intention The posterior belief probability, For the complete set of learner intention types, For the student's intent type Exploration behavior characteristics were observed at the time. The likelihood probability, The prior belief probability; the information disclosure decision submodule is configured to select the information disclosure decision for this round from four disclosure actions based on the updated posterior belief distribution and the private state of the role agent, and generate the actual disclosure information content according to the balanced information disclosure strategy.

[0017] Furthermore, the dynamic path replanning module also includes a disclosure tendency posterior update submodule, configured to update disclosure behavior sequences that have been observed within the same department or for similar functional roles. As indirect evidence, regarding sources of personnel role information that have not yet been accessed. Calculate the posterior probability distribution of the four disclosure actions:

[0018] in, Personnel role information source After observing the sequence Then choose to disclose the action The posterior probability, In order to disclose tendencies Observed sequence The likelihood probability, The utility weight recalculation submodule is configured to recalculate the probabilistic utility weights of the retrieval edges corresponding to the remaining personnel role information sources based on the updated posterior probability distribution, according to the following formula:

[0019] in, This represents the updated expected amount of information disclosure. For action The corresponding information transmission volume; the path replanning submodule is configured to, on the updated probabilistic information state transition graph, take the student's current information state node as the starting point and use the random shortest path algorithm to replan the optimal subsequent retrieval path that maximizes the expected information gain.

[0020] Furthermore, the deviation diagnosis module includes: a path deviation calculation submodule, configured to calculate the path deviation by comparing the actual search operation selected by the student with the optimal search operation indicated by the dynamic replanning path at the corresponding time after each search decision is made; a deviation cause feature extraction submodule, configured to extract state context features, unselected optimal alternative operation features, and path extension cost caused by the actual operation for decision nodes with deviation exceeding a preset threshold; a defect classification submodule, configured to classify the deviation points into four defect categories: inappropriate information exploration method selection, misinterpretation of disclosure signals, redundant retrieval, and global path planning deviation, using the deviation cause feature vector of each deviation point as input; and a diagnosis report generation submodule, configured to classify the deviation points of inappropriate information exploration method selection and misinterpretation of disclosure signals into the information game ability assessment dimension, and the deviation points of redundant retrieval and global path planning deviation into the retrieval strategy ability assessment dimension, and generate the composite diagnosis report based on the distribution of each deviation point and the cumulative path extension cost under the two dimensions.

[0021] This invention discloses a training method for modern office transaction processing based on digital twins, used to execute the aforementioned training system for modern office transaction processing based on digital twins, comprising: Step 1: Generate an information state transition graph based on the information dependencies of office tasks, distinguish information sources into deterministic information sources and personnel role information sources, and assign deterministic utility weights to the retrieval edges corresponding to deterministic information sources; Step 2: Generate a two-layer state representation containing a public state space and a private state space for the role agents associated with each personnel role information source; Step 3: Define the information disclosure action space of the role agents, and train each role agent to obtain the equilibrium information disclosure strategy through a Bayesian game framework; Step 4: Calculate the probabilistic utility weights of the retrieval edges corresponding to the information sources of each personnel role based on the balanced information disclosure strategy, and generate a probabilistic information state transition graph; Step 5: After each visit to the personnel role information source by a student, perform a Bayesian posterior update based on the actual disclosure behavior and replan the optimal subsequent retrieval path; Step 6: Compare the trainee's actual retrieval decisions with the dynamic replanning path, classify the deviations, and generate a composite diagnostic report that integrates information game ability and retrieval strategy ability.

[0022] The beneficial effects of this invention are as follows: Firstly, by generating a two-layer state representation containing public and private state spaces for the role agents associated with personnel role information sources, and defining an information disclosure action space containing four types of information disclosure actions, including full disclosure, partial disclosure, fuzzy representation, and false statement, the virtual role agents in the digital twin office environment can strategically disclose information based on their own private state and the distribution of their beliefs about the trainees' intentions. This reproduces the information control behavior of personnel information sources in real office scenarios and overcomes the deficiency that virtual roles cannot simulate interpersonal information games under the assumption of global observability.

[0023] Secondly, by transforming the retrieval edges corresponding to personnel role information sources from deterministic information increments into probabilistic utility weights that depend on the game equilibrium strategy of the role-based intelligent agents, and by performing Bayesian posterior updates based on actual disclosure behavior observations after each interaction between the trainee and the personnel role information source to recalculate the probabilistic utility weights of the remaining retrieval edges and replan the optimal subsequent path, the optimal retrieval path can adaptively evolve as information is gradually revealed during the interactive game process. This overcomes the defect that the optimal path is fixed when facing strategic information sources due to the deterministic increment assumption.

[0024] Third, by comparing the deviation between the optimal path of dynamic replanning and the actual retrieval decisions of trainees, and classifying the deviations into four predefined defect categories—inappropriate selection of information exploration methods, misinterpretation of disclosure signals, redundant retrieval, and global path planning deviations—and then categorizing each defect category into two evaluation dimensions—information game ability and retrieval strategy ability—it is possible to locate the defects in trainees' office affairs processing ability in both dimensions. This overcomes the limitations of sequence alignment methods, which misjudge equivalent optimal paths as deviations and cannot distinguish between defects in different dimensions of ability. Attached Figure Description

[0025] Figure 1 This is a flowchart of a modern office transaction processing training method based on digital twins provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall data flow architecture of the digital twin training platform provided in this embodiment of the invention; Figure 3 This is a schematic diagram comparing the objective function decomposition of Supply Chain Manager B under four disclosure actions according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the probability distribution of the equilibrium disclosure strategy of supply chain manager B provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the probabilistic utility weights of the three personnel role information sources under three types of exploration methods provided in this embodiment of the invention; Figure 6 This is a schematic diagram of the two-layer state representation structure of the supply chain manager B provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the expected information disclosure amount of remaining information sources before and after dynamic path replanning, as provided in an embodiment of the present invention. Figure 8 This is a schematic diagram comparing the initial optimal retrieval path and the subsequent path of dynamic replanning provided in the embodiments of the present invention; Figure 9 This is a schematic diagram showing the comparison of the probability distribution of intentions of student A and colleague A in the marketing department before and after the belief update, provided in an embodiment of the present invention. Detailed Implementation

[0026] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0027] At least one embodiment of the present invention discloses a modern office transaction processing training system based on digital twins, the method of which is described below. Figure 1 The system includes: The retrieval trajectory acquisition module is configured to record all retrieval operations of trainees in the digital twin office environment and generate a retrieval trajectory sequence with status annotations.

[0028] The retrieval trajectory acquisition module captures each retrieval operation initiated by the student in real time during the student's office tasks. Each retrieval operation record contains the following four data items: retrieval keyword vector, target information source identifier, a snapshot of the state of the information set already acquired at the time of retrieval, and the information increment code of the retrieval result. Specifically, the retrieval keyword vector is the vectorized representation of the query content entered by the student when initiating the retrieval; the target information source identifier marks the specific information source accessed by the student in this retrieval; the state snapshot records the combined state of all information acquired by the student at the time of retrieval initiation; and the information increment code represents the encoded representation of the newly added information content actually returned by this retrieval operation. The retrieval trajectory acquisition module organizes all retrieval operation records arranged chronologically into a retrieval trajectory sequence with status annotations.

[0029] The information state transition diagram generation module is configured to generate information state transition diagrams based on the information dependencies of office tasks and to classify and label information sources.

[0030] The information state transition graph generation module obtains all available information sources associated with the current office task and the dependencies between them, generating an information state transition graph for the current office task. Each node in the information state transition graph represents a combination state of acquired information, and each directed edge represents an executable retrieval operation and the state transition it triggers.

[0031] The information source classification submodule retrieves a list of information sources associated with the task and categorizes them into two types: deterministic information sources and personnel role information sources. Deterministic information sources refer to sources where the information content is fixed and can be obtained completely, such as database query interfaces and archived files in document management systems. Personnel role information sources refer to sources where the information is held by specific personnel roles and the content disclosed is influenced by the holder's strategic behavior, such as department colleagues and project managers.

[0032] The deterministic utility weight assignment submodule uses information gain measurement to assign deterministic utility weight values ​​to each retrieval edge corresponding to a deterministic information source. The deterministic utility weight value reflects the deterministic increment of the information state obtained after performing the corresponding retrieval operation.

[0033] The role agent two-layer state representation generation module is configured to acquire organizational function data and generate a two-layer state representation containing a public state space and a private state space for the role agent associated with each personnel role information source.

[0034] The role-agent two-layer state representation generation module acquires functional data and resource management data of roles in various departments of the organization, and generates a two-layer state representation for the role-agent associated with each person's role information source.

[0035] The public state space generation submodule extracts the public state space of each role agent from the publicly released functional data of the organization. The public state space includes the workload and collaboration commitments that the role has publicly announced. The workload represents the current task load declared by the role, and the collaboration commitments represent the cross-departmental collaboration matters and their schedules that the role has publicly confirmed. The information in the public state space is shared knowledge that can be observed by all parties involved in the task.

[0036] The private state space generation submodule extracts the private state space for each role agent from the organization's internal resource management data. The private state space includes three parameters: actual resource surplus, internal priority sequence, and reserved buffer ratio. Actual resource surplus represents the difference between the role's currently available resources and its publicly announced workload; the internal priority sequence represents the role's actual ranking of pending tasks, which may differ from the publicly declared priority; and the reserved buffer ratio represents the percentage of resources reserved by the role to handle uncertain tasks, and this ratio is not publicly disclosed. Information in the private state space is only visible to the role agent itself and cannot be directly observed by trainees or other agents.

[0037] To ensure that the actual resource surplus, internal priority sequence, and reserved buffer ratio can be uniformly weighted and summed in the subsequent objective function, mean normalization based on the range is applied to each of the three parameters, mapping each parameter to... Intervals are used to eliminate the influence of dimensional differences on weighted calculations.

[0038] The dual-state representation enables the agent to maintain both publicly visible information and private information visible only to itself. The difference between the public and private state spaces forms the information basis for the agent to disclose strategic information.

[0039] The information disclosure strategy generation module is configured to define the information disclosure action space of the role agents and train each agent to learn the equilibrium information disclosure strategy through a Bayesian game framework.

[0040] The Information Disclosure Action Space Definition submodule defines four types of information disclosure action spaces for each role-based agent: Full Disclosure, which provides the student with the true and complete content of the queried information; Partial Disclosure, which provides only the portion of the queried information that does not conflict with the role's interests; Vague Expression, which conveys information in an imprecise manner, preventing the student from accurately extracting key values ​​or conclusions; and False Statement, which provides misleading content that does not conform to the truth. Each disclosure action corresponds to a different amount of information delivered and a different degree of information distortion.

[0041] The Bayesian game training submodule utilizes a Bayesian game framework to train the policies of each agent. In this framework, each agent's type is defined by a combination of three parameters (normalized): actual resource reserves in its private state space, internal priority sequence, and reserved buffer ratio. Agents cannot directly observe the types of other participants, but maintain a prior belief distribution regarding the type distributions of others. During training, each agent optimizes by maximizing its own objective function, which comprehensively considers both information protection gains and the cost of maintaining cooperative relationships. Information protection gains measure the private state advantage maintained by the agent through strategic disclosure, specifically defined as: when the agent... Select Action At that time, its private state The proportion of information in the private state that has not been inferred by the other party, multiplied by the sum of the weighted importance coefficients of each parameter of the private state, is...

[0042] in, For intelligent agents In action and private state Information protection benefits For private state number Importance weights of item parameters These correspond to three parameters: the normalized actual resource surplus, the internal priority sequence, and the reserved buffer ratio. This represents the total number of private state parameters. For in action and private state Next The probability estimate of the parameters inferred by the other party is obtained by pre-calibrating the degree of information distortion corresponding to each disclosure action. The cost of maintaining the cooperative relationship measures the damage to the long-term cooperative relationship caused by information asymmetry, specifically defined as the current disclosure action. Information distortion Dependency of collaboration between the role and the trainees The product of, i.e.

[0043] in, For intelligent agents In action The cost of maintaining the collaborative relationship. For the role intelligent agent The strength of the collaborative dependency between the role and the trainee is obtained by normalizing the historical collaboration frequency between the role and the trainee's department in the organizational function data, with a value range of [value missing]. ; For action The corresponding degree of information distortion is as follows: full disclosure corresponds to zero distortion, while partial disclosure, vague statements, and false declarations correspond to progressively increasing distortion levels, provided by a pre-defined distortion level calibration table, with values ​​ranging from [value missing]. Set up a role-based intelligent agent. The objective function is

[0044] in, For intelligent agents In action and private state The objective function value is below. For intelligent agents The chosen information disclosure action For intelligent agents The private state (i.e., the combination of normalized actual resource reserves, internal priority sequence, and reserved buffer ratio). For the benefit of information protection, The cost of maintaining collaborative relationships. Due to... , , and All are dimensionless normalized values. and Since the units are consistent, they can be directly subtracted. Each agent selects the action with the highest expected utility from four information disclosure actions based on its current belief distribution regarding the other party's type and its own private state. Training employs a gradient ascent optimization strategy. The input consists of a two-layer state representation and an information disclosure action space for each agent, and the output is the equilibrium information disclosure strategy for each agent under given belief distribution and private state conditions. The equilibrium information disclosure strategy is represented as a mapping from the belief distribution and private state to the probability distributions of the four disclosure actions. ,in For intelligent agents The equilibrium information disclosure strategy function, For intelligent agents Current belief distribution for other participant types.

[0045] The probabilistic utility weight calculation module is configured to calculate the probabilistic utility weight of the retrieval edge corresponding to the personnel role information source based on the balanced information disclosure strategy of the role-based intelligent agent and generate a probabilistic information state transition graph.

[0046] The probabilistic utility weight calculation module, based on the current belief distribution and private objective function of each agent, pre-calculates the expected information disclosure distribution under three types of inquiry methods for each agent's information source. The three inquiry methods are: direct inquiry, where trainees directly make explicit information requests to the agent; indirect inference, where trainees indirectly obtain information by asking relevant questions or engaging in guided dialogue; and conditional information exchange proposal, where trainees exchange information they already possess for information disclosure from the other party.

[0047] For each person's role information source and each type of exploration method The probabilistic utility weight calculation module is based on the equilibrium information disclosure strategy of the role-based intelligent agent. Obtain the probability that the agent will choose one of the four actions: full disclosure, partial disclosure, obscuring the statement, and making a false statement. and the amount of information transmitted for each action. The expected information disclosure amount under this exploration method is obtained by weighting and summing the probabilities of each action with the corresponding information transmission amount:

[0048] in, Personnel role information source In the exploration method The expected amount of information disclosure It is a set of actions consisting of four types of disclosure actions: full disclosure, partial disclosure, vague statements, and false statements. For action sets Any disclosure action in the process, For intelligent agents In private state Belief Distribution and exploration methods Choose action under conditions The probability, For action The corresponding amount of information transmitted. The expected amount of information disclosed from the same personnel role information source under the three types of investigation methods. , , The combination forms the probabilistic utility weights of the retrieval edge corresponding to the information source, where , , Corresponding to three exploration methods—direct inquiry, indirect inference, and conditional information exchange proposal—the probabilistic utility weights characterize the distribution of information obtainable under different exploration methods in the form of conditional probability distributions. The original deterministic information increment encoding of the retrieval edge corresponding to the personnel role information source in the information state transition diagram is replaced with the probabilistic utility weights, while the deterministic utility weights of the retrieval edge corresponding to the deterministic information source remain unchanged, thus forming the probabilistic information state transition diagram.

[0049] The initial retrieval path planning module is configured to calculate the initial retrieval path that maximizes the expected information gain on the probabilistic information state transition graph.

[0050] The initial retrieval path planning module, based on the probabilistic information state transition graph, starts with an initial empty information state and ends with the target information state required to complete the office task. It uses a stochastic shortest path algorithm to calculate the initial retrieval path that maximizes the expected information gain. The input to the stochastic shortest path algorithm is the probabilistic utility weight of each edge in the probabilistic information state transition graph, and the output is the path sequence with the optimal expected cumulative utility from the start point to the end point. The initial retrieval path serves as the benchmark for the optimal retrieval strategy based on the current game equilibrium state before the trainee executes the task.

[0051] The Interaction Response and Belief Update module is configured to analyze the exploration behavior and drive the role agent to generate information disclosure decisions when trainees access the personnel role information source.

[0052] When trainees perform a search operation to access personnel role information sources in a digital twin office environment, the interaction response and belief update module executes the following process: The probing behavior analysis submodule analyzes the communication content of trainees and extracts information probing behavior characteristics, including the type of probing method (direct inquiry, indirect inference, or conditional information exchange proposal), the specific information dimension targeted by the probing, and the range of known information that trainees actively reveal in the communication.

[0053] The Bayesian belief update submodule inputs the information exploration behavior characteristics and the set of information already disclosed by the learner into the Bayesian belief update process of the corresponding role agent. Let the prior belief distribution of the role agent regarding the learner's intention be... The observed exploration behavior characteristics are as follows Then the posterior belief distribution Calculated using Bayes' theorem:

[0054] in, For the role intelligent agent After observing the characteristics of the exploration behavior Next, the type of trainee intention The posterior belief probability, For the type of student intent, For the complete set of learner intention types, To Iteration variables for each intent type in the text. For the student's intent type Exploration behavior characteristics were observed at the time. The likelihood probability, For the role intelligent agent Types of trainees' intentions The prior belief probability. The updated posterior belief distribution. This reflects the agent's judgment on the student's actual information needs and subsequent behavioral direction after observing the student's exploration behavior in this round.

[0055] The information disclosure decision-making submodule is based on the updated posterior belief distribution. Private state of the role's intelligent agent Based on the balanced information disclosure strategy obtained during training The system selects from four options for information disclosure in this round: full disclosure, partial disclosure, vague description, and false statement, and generates the corresponding actual disclosure information.

[0056] The information increment deviation calculation submodule encodes the actual disclosure information output by the role agent into the actual information increment. Calculate the corresponding exploration methods in the actual information increment and probabilistic information state transition diagram. Expected information increment Deviation values ​​between:

[0057] in, This represents the deviation between the actual information increment and the expected information increment. The information increment corresponding to the disclosure content actually output by the role's intelligent agent. Personnel role information source In the exploration method The expected amount of information disclosure. and All are dimensionless information quantity measurements that have been normalized, have consistent dimensions, and can be directly subtracted.

[0058] The dynamic path replanning module is configured to update the probabilistic information state transition graph based on actual information increments and disclosure behavior observations, and replan the optimal subsequent retrieval path.

[0059] The information status update submodule is based on the actual information increment obtained. This transfers the student's current information state node from the previous state to a new state node corresponding to the actual information increment.

[0060] The disclosure propensity posterior update submodule utilizes Bayesian inference, using observed historical disclosure behaviors of each role agent as evidence to update the posterior estimate of disclosure propensity for unvisited personnel role information sources. For each unvisited personnel role information source... Based on its prior disclosure tendency distribution Based on the observed sequence of disclosure behaviors within the same department or similar functional roles. As indirect evidence, the posterior probability distribution of the information source for the four actions of full disclosure, partial disclosure, obscuring statements, and false statements is calculated:

[0061] in, For personnel role information sources that have not yet been visited After observing the historical disclosure behavior sequence Then choose to disclose the action The posterior probability, For action sets Any disclosure action in the process, In order to disclose tendencies Historical disclosure behavior sequence observed The likelihood probability, Personnel role information source Choose disclosure action The prior probability.

[0062] The utility weight recalculation submodule is based on the updated disclosure propensity posterior estimate. ,by Replace the probabilistic utility weight calculation module The remaining personnel role information sources in the probabilistic information state transition diagram are recalculated using the same weighted summation method at their positions. The probabilistic utility weights of the corresponding search edges:

[0063] in, For personnel role information sources that have not yet been visited In the exploration method The expected amount of information disclosure after the update. For action sets Any disclosure action in the process, Personnel role information source After observing the historical disclosure behavior sequence Then choose to disclose the action The posterior probability, For action The corresponding information transmission volume is consistent with the definition in the probabilistic utility weight calculation module.

[0064] The path replanning submodule, on the updated probabilistic information state transition graph, uses the student's current information state node as the starting point and the target information state as the ending point, and employs a random shortest path algorithm to replan the optimal subsequent retrieval path that maximizes the expected information gain. This submodule is triggered after each visit to the personnel role information source by the student, ensuring that the optimal subsequent retrieval path is continuously updated as information is gradually revealed during the interactive game.

[0065] The deviation diagnosis module is configured to compare the student's actual retrieval decisions with the dynamic replanning path and generate a composite diagnostic report that integrates information game capabilities and retrieval strategy capabilities.

[0066] The path deviation calculation submodule calculates the path deviation between the student's actual search operation at the current decision node and the optimal search operation indicated by the dynamic replanning path at the corresponding moment after each search decision is made.

[0067] The deviation cause feature extraction submodule extracts the following three types of deviation cause features for decision nodes whose deviation exceeds a preset threshold: First, the state context features corresponding to the decision node, including the set of information currently acquired by the trainee and the observed disclosure behavior history of each role agent; Second, the features of the optimal alternative operation that was not selected, including the alternative retrieval operation indicated by the dynamic replanning path and its corresponding probabilistic utility weight; Third, the path extension cost caused by the actual operation, that is, the amount of loss in the total utility of the expected path compared to the optimal alternative operation in terms of the actual operation selected by the trainee.

[0068] The defect classification submodule utilizes a decision tree classifier, taking the feature vectors of deviation causes at each deviation point as input and four defect category labels as output. It employs Gini impurity as the splitting criterion, uses labeled historical deviation samples as training data, and uses a gradient descent optimization strategy for training. The loss function is cross-entropy loss. Predefined defect categories include the following four types: Inappropriate information exploration method selection, where trainees choose a low-return exploration method for a specific role, resulting in a significantly lower actual information increment at that retrieval edge than the expected information increment when using the optimal exploration method; Misinterpretation of disclosure signals, where trainees fail to adjust their subsequent retrieval direction after receiving clear strategic disclosure signals (such as a role agent continuously executing ambiguous statements or partial disclosure actions), continuing to access the same information source or similar highly defensive information sources along the original path; Redundant retrieval, where trainees repeatedly access information sources for which sufficient information has already been obtained, resulting in no effective information increment from the retrieval operation; and Global path planning deviation, where the trainee's retrieval order systematically deviates from the optimal path at multiple consecutive decision nodes, and this deviation is not caused by information game factors but by an unreasonable retrieval order arrangement.

[0069] The diagnostic report generation submodule summarizes and integrates the defect classification results, corresponding state context features, and path extension costs for each deviation point. Deviation points with defects of inappropriate selection of attribution information exploration methods and misinterpretation of disclosure signals are categorized under the information game capability assessment dimension; deviation points with defects of redundant attribution retrieval and global path planning deviation are categorized under the retrieval strategy capability assessment dimension. Based on the distribution of each deviation point and the cumulative path extension cost under these two dimensions, a composite diagnostic report integrating the information game capability and retrieval strategy capability assessment dimensions is generated.

[0070] This invention generates a two-layer state representation containing a public state space and a private state space for the role intelligent agent associated with the personnel role information source, and defines an information disclosure action space containing four types of information disclosure actions: full disclosure, partial disclosure, fuzzy representation, and false statement. This enables the virtual role intelligent agent in the digital twin office environment to strategically disclose information based on its own private state and the distribution of beliefs about the trainees' intentions, thereby reproducing the information control behavior of personnel information sources in real office scenarios and overcoming the defect that virtual roles cannot simulate interpersonal information games under the global observability assumption.

[0071] By transforming the retrieval edges corresponding to personnel role information sources from deterministic information increments into probabilistic utility weights dependent on the game equilibrium strategy of role-based intelligent agents, and performing Bayesian posterior updates based on actual disclosure behavior observations after each interaction between the trainee and the personnel role information source to recalculate the probabilistic utility weights of the remaining retrieval edges and replan the optimal subsequent path, the optimal retrieval path can adaptively evolve as information is gradually revealed during the interactive game process. This overcomes the defect that the optimal path remains fixed when facing strategic information sources due to the deterministic increment assumption.

[0072] By comparing the deviation between the optimal path derived from dynamic replanning and the trainees' actual retrieval decisions, and classifying the deviations into four predefined defect categories—inappropriate selection of information exploration methods, misinterpretation of disclosure signals, redundant retrieval, and global path planning deviations—and then categorizing each defect category into two assessment dimensions—information game ability and retrieval strategy ability—it is possible to pinpoint the trainees' deficiencies in office affairs processing ability on both dimensions. This overcomes the limitations of sequence alignment methods, which misjudge equivalent optimal paths as deviations and cannot distinguish between defects in different dimensions of ability.

[0073] See Figures 2-9 The following application examples are proposed: In a company's digital twin training platform, trainee A is executing a cross-departmental new product launch coordination task. The task requires trainee A, within a digital twin office environment, to retrieve information from multiple sources and summarize three core information categories: product pricing strategy, supply chain inventory preparation progress, and marketing schedule, ultimately forming a launch coordination plan. The information sources associated with the task include two types: deterministic information sources are the product database and document management system; personnel role information sources include colleague A from the marketing department, supply chain manager B, and colleague C from the brand department. At the start of the training, trainee A's initial information state is an empty set, and the target information state is that all three core information categories have been acquired.

[0074] Retrieval trajectory acquisition module: During the task execution, the retrieval trajectory acquisition module captured each of Trainee A's retrieval operations in real time and organized them into a retrieval trajectory sequence with status annotations. The following shows the record of Trainee A's first three retrieval operations.

[0075] Table 1. Student A's First Three Steps of Search Operation Record

[0076] Among them, the increment of deterministic information sources can be determined during collection, while the increment of personnel role information sources is marked as "to be parsed" and will be filled in after being processed by the interaction response and belief update module.

[0077] Role-based intelligent agent two-layer state representation generation module: The system generates two-tiered state representations for three personnel roles—Marketing Department Colleague A, Supply Chain Supervisor B, and Brand Department Colleague C—from organizational functional data and internal resource management data. Taking Supply Chain Supervisor B as an example, their public state space comes from publicly released organizational functional data, while their private state space comes from internal resource management data. These three private parameters are mapped to a mean-normalized value based on the range. Interval.

[0078] Table 2. Two-tiered status representation of Supply Chain Supervisor B

[0079] The workload declaration in the public state is 75%, while the normalized value of the actual resource surplus in the private state is 0.71, reflecting that the actual available resources are higher than the external declaration. The difference between the two forms the information basis for Supply Chain Manager B to make strategic information disclosures.

[0080] Information disclosure strategy generation module: Taking supply chain manager B as an example, their private status... The importance weights of the three parameters are preset to be: , , Collaborative Dependency Strength (Obtained by normalization of historical collaboration frequency).

[0081] During Bayesian game training, the system calculates the objective function value for Supply Chain Manager B under four different disclosure actions to determine the equilibrium disclosure strategy. The "partial disclosure" action is used as an example. For example, preset distortion level The three parameters are pre-calibrated with inferred probabilities as follows: , , ,but:

[0082]

[0083]

[0084] Table 3. Objective function values ​​of Supply Chain Manager B under four disclosure actions

[0085] The action that maximizes the objective function value is partial disclosure. Therefore, under the current private state and belief distribution, Supply Chain Manager B's equilibrium information disclosure strategy will select partial disclosure with a higher probability. The equilibrium strategy output from the training is mapped as follows: The probability of full disclosure is 0.12, the probability of partial disclosure is 0.58, the probability of vague statements is 0.24, and the probability of false statements is 0.06.

[0086] Probabilistic utility weight calculation module: Taking supply chain manager B as an example, the system calculates the expected information disclosure volume for each of the three exploration methods. The information transmission volume corresponding to the four disclosure actions is preset as follows: full disclosure Partial disclosure Vague expression False statement By asking directly For example, the probability distribution of the equilibrium strategy in this method follows the training results mentioned above:

[0087] Table 4. Probabilistic utility weights of the three personnel role information sources under three types of exploration methods.

[0088] The probabilistic utility weights replace the original deterministic incremental codes of the three personnel role retrieval edges in the information state transition graph. The deterministic utility weights of the corresponding retrieval edges for the product database and document management system remain unchanged, forming a probabilistic information state transition graph. The initial retrieval path planning module runs a random shortest path algorithm on this graph, outputting the initial optimal retrieval path as: Product Database (direct query) → Marketing Department Colleague A (conditional information exchange proposal) → Brand Department Colleague C (direct inquiry) → Supply Chain Supervisor B (conditional information exchange proposal), with an expected cumulative utility of 2.98.

[0089] Interactive Response and Belief Update Module: In step 2 of the search, student A directly inquired with colleague A in the marketing department. The probing behavior analysis submodule extracted the probing method as direct inquiry, the probing dimension as pricing scheme, and the known information proactively revealed by the student as historical pricing benchmarks. The Bayesian belief update submodule then applied these probing behavior characteristics. Input the belief update process of colleague A from the marketing department. Assume the prior beliefs that the trainee's intention type "only needs pricing confirmation" has a probability of 0.40 and "needs full pricing details" has a probability of 0.60. After observing the direct inquiry behavior, the likelihood probability of the "needs full pricing details" intention is higher, and the posterior beliefs are updated to "only needs pricing confirmation" with a probability of 0.22 and "needs full pricing details" with a probability of 0.78. The information disclosure decision submodule then updates the posterior beliefs accordingly. And a balanced strategy, selectively disclosing certain actions, resulting in an actual increase in information. .

[0090] The information increment deviation calculation submodule calculates the deviation value:

[0091] Deviation value A negative result indicates that the actual information gain obtained by Trainee A from directly inquiring with Marketing Department Colleague A was lower than the expected value under this probing method, suggesting that Trainee A did not choose the probing method with the highest information return (conditional information exchange proposal). ).

[0092] Dynamic path replanning module: After student A completes step 2 of the search, the dynamic path replanning module is triggered. The information status update submodule updates the student's current information status from {historical pricing benchmark} to {historical pricing benchmark, pricing scheme (partial)}. The disclosure propensity posterior update submodule uses some disclosure behaviors already observed by marketing colleague A as indirect evidence to update the posterior estimate of disclosure propensity for supply chain manager B and brand department colleague C, who have not yet been visited, based on the historical observations of colleagues in the same department or with similar functional roles. This increases the posterior probability of partial disclosure for both. The utility weight recalculation submodule recalculates the expected information disclosure amounts for the remaining two personnel role information sources based on the updated posterior distribution.

[0093] Table 5 Comparison of Expected Information Disclosure Amount of Remaining Personnel Role Information Sources Before and After Dynamic Updates

[0094] The path replanning submodule replans the optimal subsequent path on the updated probabilistic information state transition graph, starting from the current information state. The output is: Brand Department colleague C (conditional information exchange proposal) → Supply Chain Supervisor B (conditional information exchange proposal). The expected cumulative utility of the dynamic optimal subsequent path is 1.39.

[0095] Deviation diagnosis module: In step 2 of the search, student A directly asked for a request instead of offering a conditional information exchange proposal to colleague A in the marketing department. The path deviation calculation submodule calculated that the path deviation of this decision node exceeded a preset threshold. The deviation cause feature extraction submodule extracted three types of features: the current state context is that historical pricing benchmarks have been obtained, and colleague A in the marketing department has not yet been visited; the optimal alternative operation that was not selected was the conditional information exchange proposal (probabilistic utility weight 0.74); and the path extension cost caused by the actual operation is... (Expected information loss). The defect classification submodule inputs this feature vector into the decision tree classifier and outputs the defect category as "inappropriate selection of information exploration method", which is included in the information game ability assessment dimension.

[0096] Table 6 Summary of Deviation Point Diagnosis for Student A

[0097] The diagnostic report generation submodule summarizes the classification results of the two deviations: the cumulative path extension cost in the information game ability dimension is 0.38, and there are no deviations in the retrieval strategy ability dimension. The composite diagnostic report points out that student A's main deficiency is concentrated in the information game ability dimension, specifically manifested in the continuous use of direct inquiry methods with low information return rate for strategic information sources, and failure to use conditional information exchange proposals to improve information acquisition efficiency.

[0098] Throughout the implementation process, data starts from the initial empty information state of trainee A. The retrieval trajectory acquisition module generates a sequence of operations with state annotations. The role agent's two-layer state representation generation module establishes a hierarchical information structure of public and private information for the three personnel role information sources. The information disclosure strategy generation module quantifies the private state into an equilibrium disclosure probability distribution through Bayesian game training. The probabilistic utility weight calculation module transforms the equilibrium strategy into the expected information disclosure amount on the retrieval edge and replaces the original deterministic increment, forming a probabilistic information state transition graph for initial path planning. Subsequently, each time the trainee interacts with a personnel role information source, the interaction response and belief update module updates the role agent's belief in the trainee's intention using Bayesian formula and generates the actual disclosure decision. The dynamic path replanning module uses the observed disclosure behavior sequence to update the posterior disclosure tendency of the remaining information sources and replans the optimal subsequent path, allowing the optimal path to continuously evolve throughout the game process. The deviation diagnosis module ultimately compares the trainee's actual decisions with the dynamic optimal path, classifies deviations into four types of defects and categorizes them into two dimensions: information game ability and retrieval strategy ability. It generates a composite diagnostic report with diagnostic positioning function, realizing a complete data flow from raw retrieval operation data to multi-dimensional capability defect diagnostic conclusions.

[0099] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A modern office transaction processing training system based on digital twins, characterized in that, include: The information state transition graph generation module is configured to generate information state transition graphs based on the information dependencies of office tasks, distinguishing information sources into deterministic information sources and personnel role information sources, and assigning deterministic utility weights to the retrieval edges corresponding to deterministic information sources. The two-layer state representation generation module is configured to generate a two-layer state representation containing a public state space and a private state space for the role agents associated with each person's role information source. The information disclosure strategy generation module is configured to define the information disclosure action space of the role agents and train each role agent to obtain the equilibrium information disclosure strategy through a Bayesian game framework. The probabilistic utility weight calculation module is configured to calculate the probabilistic utility weight of the retrieval edge corresponding to each personnel role information source based on the balanced information disclosure strategy, and generate a probabilistic information state transition graph. The dynamic path replanning module is configured to perform Bayesian posterior updates based on the actual disclosure behavior after each visit by a student to the personnel role information source, and replan the optimal subsequent retrieval path. The deviation diagnosis module is configured to compare the student's actual retrieval decisions with the dynamic replanning path, classify deviations, and generate a composite diagnostic report that integrates information game capabilities and retrieval strategy capabilities.

2. The modern office transaction processing training system based on digital twins according to claim 1, characterized in that, The public state space includes the workload and collaboration commitments publicly announced by the role agent. The information in the public state space is shared knowledge observable by all parties involved in the task. The private state space includes three parameters: actual resource surplus, internal priority sequence, and reserved buffer ratio. The actual resource surplus is the difference between the role's currently available resources and the publicly announced workload. The internal priority sequence is the role's actual ranking of pending tasks. The reserved buffer ratio is the percentage of resources reserved by the role to handle uncertain tasks. The information in the private state space is only visible to the role agent itself. These three parameters are mapped to a mean normalization method based on the range. Interval.

3. The modern office transaction processing training system based on digital twins according to claim 1, characterized in that, The information disclosure action space includes four types of disclosure actions: full disclosure, which provides the trainees with the true and complete content of the queried information; partial disclosure, which provides only the part of the queried information that does not conflict with the interests of the role; vague expression, which conveys the information content in an imprecise way; and false statement, which provides misleading content that does not conform to the true information. Each disclosure action corresponds to a different amount of information transmission and a different degree of information distortion.

4. The modern office transaction processing training system based on digital twins according to claim 2, characterized in that, In the Bayesian game framework, the type of each agent is defined by a combination of three normalized parameters in its private state space. Each agent maintains a prior belief distribution regarding the type distributions of other participants. Each agent optimizes by maximizing its own objective function, which is: ； in, For intelligent agents In action and private state The objective function value is below. For the benefit of information protection, The costs associated with maintaining the collaborative relationship; the benefits of information protection are: ； in, For private state number Importance weights of item parameters For in action and private state Next The probability estimate of the parameter being inferred by the other party; the maintenance cost of the cooperative relationship is: ； in, For the role intelligent agent The intensity of collaboration and dependence between trainees For action The corresponding degree of information distortion; the equilibrium information disclosure strategy is expressed as a mapping from the belief distribution and private state to the probability distribution of the four disclosure actions. ,in For intelligent agents Current belief distribution for other participant types.

5. The modern office transaction processing training system based on digital twins according to claim 4, characterized in that, The probabilistic utility weight calculation module targets each person's role information source. and each type of exploration method Based on the balanced information disclosure strategy, the probability of each of the four disclosure actions being selected and the information transmission amount corresponding to each action are obtained, and the expected information disclosure amount is obtained by weighted summation: ； in, Personnel role information source In the exploration method The expected amount of information disclosure It is a set of actions consisting of four disclosure actions. For any disclosure action in the action set, For action The corresponding information transmission volume; the exploration methods include three categories: direct inquiry, indirect inference, and conditional information exchange proposal; the expected information disclosure volume of the same personnel role information source under the three exploration methods is combined into the probabilistic utility weight of the retrieval edge corresponding to the information source, replacing the original deterministic information incremental encoding of the retrieval edge in the information state transition diagram.

6. The modern office transaction processing training system based on digital twins according to claim 1, characterized in that, It also includes an initial retrieval path planning module, configured to calculate the initial retrieval path with the expected information gain maximized using a random shortest path algorithm on the probabilistic information state transition graph, starting from the initial empty information state and ending at the target information state required to complete the office task, as the benchmark for the optimal retrieval strategy based on the current game equilibrium state before the trainee performs the task.

7. The modern office transaction processing training system based on digital twins according to claim 5, characterized in that, The dynamic path replanning module includes: The probing behavior analysis submodule is configured to analyze the communication content of trainees, extract the probing method categories, the specific information dimensions targeted by the probing, and the range of known information actively revealed by the trainees; The Bayesian belief update submodule is configured to update the posterior belief distribution of the agent regarding the trainee's intentions according to the Bayesian formula based on the extracted probing behavior features. ； in, To observe the characteristics of the exploration behavior Next, the type of trainee intention The posterior belief probability, For the complete set of learner intention types, For the student's intent type Exploration behavior characteristics were observed at the time. The likelihood probability, For prior belief probability; The information disclosure decision submodule is configured to select the information disclosure decision for this round from four disclosure actions based on the updated posterior belief distribution and the private state of the role agent, and generate the actual disclosure information content according to the balanced information disclosure strategy.

8. The modern office transaction processing training system based on digital twins according to claim 7, characterized in that, The dynamic path replanning module also includes: The disclosure tendency post-hoc update submodule is configured to be based on a sequence of disclosure behaviors that have been observed within the same department or with similar functional roles. As indirect evidence, regarding sources of personnel role information that have not yet been accessed. Calculate the posterior probability distribution of the four disclosure actions: ； in, Personnel role information source After observing the sequence Then choose to disclose the action The posterior probability, In order to disclose tendencies Observed sequence The likelihood probability, This is the prior probability; The utility weight recalculation submodule is configured to recalculate the probabilistic utility weights of the retrieval edges corresponding to the remaining personnel role information sources based on the updated posterior probability distribution, according to the following formula: ； in, This represents the updated expected amount of information disclosure. For action The corresponding amount of information transmitted; The path replanning submodule is configured to, on the updated probabilistic information state transition graph, use the student's current information state node as the starting point and employ the random shortest path algorithm to replan the optimal subsequent retrieval path that maximizes the expected information gain.

9. The modern office transaction processing training system based on digital twins according to claim 1, characterized in that, The deviation diagnosis module includes: The path deviation calculation submodule is configured to calculate the path deviation by comparing the actual search operation selected by the student with the optimal search operation indicated by the dynamic replanning path at the corresponding time after each search decision is made. The deviation cause feature extraction submodule is configured to extract state context features, the features of the optimal alternative operation that was not selected, and the path extension cost caused by the actual operation for decision nodes whose deviation exceeds a preset threshold. The defect classification submodule is configured to take the feature vector of the deviation cause of each deviation point as input and classify the deviation points into four defect categories: inappropriate selection of information exploration method, misinterpretation of disclosure signal, redundant retrieval, and global path planning deviation. The diagnostic report generation submodule is configured to include deviations such as improper selection of attribution information exploration methods and misinterpretation of disclosure signals in the information game capability assessment dimension, and deviations such as attribution redundancy retrieval and global path planning deviations in the retrieval strategy capability assessment dimension, and generate the composite diagnostic report based on the distribution of each deviation point under the two dimensions and the cumulative path extension cost.

10. A training method for modern office transaction processing based on digital twins, used to execute the training system for modern office transaction processing based on digital twins as described in any one of claims 1 to 9, characterized in that, include: Step 1: Generate an information state transition graph based on the information dependencies of office tasks, distinguish information sources into deterministic information sources and personnel role information sources, and assign deterministic utility weights to the retrieval edges corresponding to deterministic information sources; Step 2: Generate a two-layer state representation containing a public state space and a private state space for the role agents associated with each personnel role information source; Step 3: Define the information disclosure action space of the role agents, and train each role agent to obtain the equilibrium information disclosure strategy through a Bayesian game framework; Step 4: Calculate the probabilistic utility weights of the retrieval edges corresponding to the information sources of each personnel role based on the balanced information disclosure strategy, and generate a probabilistic information state transition graph; Step 5: After each visit to the personnel role information source by a student, perform a Bayesian posterior update based on the actual disclosure behavior and replan the optimal subsequent retrieval path; Step 6: Compare the trainee's actual retrieval decisions with the dynamic replanning path, classify the deviations, and generate a composite diagnostic report that integrates information game ability and retrieval strategy ability.

Images