Organizational relationship differentiation management method based on reinforcement learning

Abstract

The invention discloses an organization relation differentiation management method based on reinforcement learning, and relates to the technical field of data processing, and the method comprises the steps: collecting task flow data and interaction log data between nodes in a distributed cooperative network based on a preset time window; performing feature extraction on the interaction log data by using an interaction semantic analysis model, calculating a stability index and a delay jitter index between node pairs in combination with timestamp metadata, performing measurement and calculation to obtain an interaction impedance coefficient, and constructing a global interaction impedance matrix; constructing a state space of the reinforcement learning agent, mapping the global interaction impedance matrix into graph structure data, and fusing task circulation data to generate a comprehensive state vector containing network topology features and load features; inputting the comprehensive state vector into an interactive strategy learning model for reasoning, and outputting a dynamic regulation and control action instruction aiming at the node through a strategy network; and executing the dynamic regulation and control action instruction, and regulating and controlling the distributed collaborative network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an organization relationship differentiated management method based on reinforcement learning. BACKGROUND

[0002] In a complex organization system such as a modern enterprise, a large R&D team or a supply chain system, the running efficiency of the organization relationship directly determines the overall output efficiency. The traditional organization relationship management mode is usually based on a static hierarchical architecture and fixed process rules, and lacks adaptability to dynamic interactive environment.

[0003] In actual operation, the interaction quality between different nodes is real-time fluctuation; for example, high-frequency interaction friction may occur between collaborative nodes due to poor communication, cognitive bias or resource contention, which is manifested in the data layer as an increase in task flow delay, an increase in information rollback rate and the accumulation of negative semantics in unstructured interaction text. The traditional technology lacks means to convert these implicit organization relationship differences into quantifiable calculation indicators, cannot accurately match the needs of different subjects, and leads to lagging management.

[0004] Most of the existing solutions rely on manual intervention to alleviate conflicts or redistribute tasks, which is not only inefficient, but also cannot achieve real-time and accurate regulation of thousands of interaction links at the micro level.

[0005] Therefore, there is an urgent need for a digital technology solution that can convert organization interaction behavior into a state space, automatically learn the optimal strategy using an intelligent agent, and thus achieve differentiated and accurate management. SUMMARY

[0006] The present application provides an organization relationship differentiated management method based on reinforcement learning, to solve the problem that the prior art cannot accurately regulate complex interaction links at the micro level, and achieve the technical effect of differentiated and accurate management.

[0007] The present application provides an organization relationship differentiated management method based on reinforcement learning, comprising: Based on a preset time window, task flow data and interaction log data between nodes in a distributed collaboration network are collected; An interaction semantic analysis model is used to extract features from the interaction log data, and a stability index and a time delay jitter index between node pairs are calculated in combination with timestamp metadata; an interaction impedance coefficient is calculated based on the stability index and the time delay jitter index, and a global interaction impedance matrix for the current time step is generated based on the interaction impedance coefficient; A state space of a reinforcement learning intelligent agent is constructed, the global interaction impedance matrix is mapped to graph structure data, and task flow data is fused to generate a comprehensive state vector containing network topology features and load features; The comprehensive state vector is input into the interaction policy learning model for inference, and a dynamic regulation action instruction for the node is output through the policy network; The dynamic regulation action instruction is executed to regulate the distributed collaboration network, and the change of the global interaction impedance matrix is detected in the next time window, and the reward function value is calculated to update the parameters of the interaction policy learning model in reverse.

[0008] The task flow data includes node queue accumulation depth, task average residence time delay, flow back rate, and task blocking state bit; The node queue accumulation depth is obtained by normalizing the number of task objects in the node to-do list, and is used as a numerical feature representing the instantaneous load pressure of the node; The task average residence time delay is calculated based on the timestamp log of the task object, and the average residence time of the task in the node is obtained, which quantifies the processing throughput of the node; The flow back rate is obtained by counting the frequency of sending and returning the task instruction between adjacent nodes within a time window, which is mapped to the logical oscillation feature of the network level; The task blocking state bit is a Boolean signal activated when the task residence time exceeds a preset threshold, indicating a potential local network deadlock point.

[0009] The interaction log data includes unstructured text information between nodes, input behavior feature data of node operation terminals, and corresponding timestamp metadata; The input behavior feature data includes keystroke rate, mouse movement track chaos degree, and operation undo and redo frequency.

[0010] The interaction semantic analysis model is used to extract features from the interaction log data, and the stability index and time delay jitter index between node pairs are calculated based on the timestamp metadata, including: The unstructured text information is segmented and vectorized embedded, and the negative vocabulary density, imperative sentence frequency and high-intensity punctuation symbol usage frequency contained therein are detected, and the semantic sharpness score of a single interaction is calculated; The number of interaction rounds between node pairs within a preset period and the number of times the same task instruction is transferred between nodes are counted, and the interaction oscillation frequency is calculated; The keystroke rate, mouse movement track chaos degree, and operation undo and redo frequency in the input behavior feature data are encoded into a behavior feature vector through a long short-term memory network, and a behavior state score is obtained according to the behavior feature vector; The average waiting time of message reply is calculated based on the timestamp metadata to obtain interaction time data; The stability indicators include semantic sharpness scores and interaction oscillation frequencies; and the time delay jitter indicators include behavior state scores and interaction duration data.

[0011] The interaction impedance coefficients are calculated based on the stability indicators and the time delay jitter indicators, and a specific process of generating a global interaction impedance matrix at a current time step based on the interaction impedance coefficients includes: The semantic sharpness scores and the interaction oscillation frequencies in the stability indicators and the behavior state scores and the interaction duration data in the time delay jitter indicators are fused by a weighted normalization algorithm to obtain interaction impedance coefficients of node pairs at the current time step, and the interaction impedance coefficients of all node pairs jointly constitute the global interaction impedance matrix.

[0012] A specific process of constructing a state space of a reinforcement learning intelligent agent, mapping the global interaction impedance matrix into graph structure data, and fusing task flow data to generate a comprehensive state vector including network topology features and load features includes: A graph structure data is constructed with nodes as vertices and interaction relationships as edges, and the global interaction impedance matrix is mapped into dynamic weight attributes of edges in the graph structure data. The graph structure data is input into a graph convolutional neural network layer, impedance weight features of first-order neighbors and second-order neighbors of a target node are aggregated by a convolution kernel, and an enhanced graph embedding vector representing a network topology risk propagation situation is output. Node load feature vectors of nodes are generated by numerically standardizing node queue accumulation depth, task average residence time delay, flow back-off rate, and task blocking state bits in the task flow data. The enhanced graph embedding vector and the node load feature vector are spliced and fused to generate the comprehensive state vector in multiple dimensions.

[0013] The interaction strategy learning model includes an actor network and a critic network. The actor network is responsible for generating a deterministic parameter adjustment strategy according to the input comprehensive state vector, and the critic network is responsible for predicting a future cumulative reward value according to a current state and an action output by the actor network. The reward function value design includes a positive gain term and a negative penalty term. The positive gain term is positively correlated with the number of task completions per unit time in the entire network. The negative penalty term is negatively correlated with the average norm of the global interaction impedance matrix and the duration of high impedance links, thereby guiding the model to actively reduce semantic friction in the network while pursuing task processing speed.

[0014] The output dynamic regulation and control action instruction includes a traffic shaping operation for an interaction link. The traffic shaping operation includes setting a cooling time threshold and a frequency limit of message delivery; when the intelligent agent identifies that there is an overheated interaction impedance between the node pairs, a dynamic regulation action instruction is output to lock the instant messaging input interfaces of both sides and delay the push time of non-urgent messages.

[0015] The beneficial effects of the present application are: 1、The present application has significant technical advantages through multi-dimensional feature processing of task flow data, solves the subjectivity problem of node load assessment in traditional management, and through the introduction of node queue accumulation depth and normalization processing, can cross different types of business nodes, compare load pressure horizontally on a unified mathematical benchmark, so that the resource scheduling algorithm can accurately identify the short board and bottleneck of the whole network.

[0016] 2、The present application can capture the real resistance hidden under silence or surface politeness by introducing behavior state score and interaction duration data; in addition, the calculation of interaction shock frequency can effectively identify the invalid communication trap; even if the communication is friendly in semantics, if it falls into an endless cycle of discussion without output, it will still consume a lot of system resources; this index enables the reinforcement learning model to learn to suppress this inefficient high-frequency interaction, thereby greatly improving the effective information throughput rate of the network and achieving accurate identification.

[0017] 3、The present application uses the actor-critic architecture to realize the adaptive evolution of management strategy; through the guidance of the composite reward function, the model can automatically learn extremely complex management strategies in continuous trial and error and iteration; it tolerates slight interaction friction appropriately during load peak period to ensure speed, and aggressively regulates resistance during load trough period to optimize network structure, so that the distributed collaboration network has self-regulation and evolution ability and can cope with changing external demands and internal environment. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 It is a flowchart of the organizational relationship differentiated management method based on reinforcement learning of the present application.

[0019] Figure 2 It is a data logic diagram of the organizational relationship differentiated management method based on reinforcement learning of the present application. DETAILED DESCRIPTION

[0020] The above technical solutions will be described in detail below in combination with the accompanying drawings and specific embodiments, so that the above technical solutions can be better understood. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments of the present application, and it should be understood that the present application is not limited to the example embodiments for explaining the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. It should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, not all. Embodiments

[0021] The present application proposes a reinforcement learning-based organization relationship differentiated management method, the flow of the method is as shown in Figure 1 The data logic of the method is as shown in Figure 2 The method comprises the following steps: Based on a preset time window, task flow data and interaction log data between nodes in a distributed collaboration network are collected; The interaction log data is feature-extracted by using an interaction semantic analysis model, and the stability index and the time delay jitter index between node pairs are calculated in combination with timestamp metadata; the interaction impedance coefficient is measured based on the stability index and the time delay jitter index, and the global interaction impedance matrix of the current time step is generated based on the interaction impedance coefficient; The state space of the reinforcement learning intelligent agent is constructed, the global interaction impedance matrix is mapped into graph structure data, and the comprehensive state vector containing network topology features and load features is generated by fusing the task flow data; The comprehensive state vector is input into the interaction policy learning model for reasoning, and the dynamic regulation and control action instruction for the node is output through the policy network; The dynamic regulation and control action instruction is executed to regulate and control the distributed collaboration network, and the change of the global interaction impedance matrix is detected in the next time window, and the reward function value is calculated to update the parameters of the interaction policy learning model in reverse.

[0022] The task flow data includes node queue accumulation depth, task average residence time delay, flow transfer rollback rate, and task blocking state bit; The node queue accumulation depth is obtained by normalizing the number of task objects in the node to-do list, as a numerical feature representing the instantaneous load pressure of the node; The task average residence time delay is obtained by calculating the average residence time of the task in the node based on the timestamp log of the task object, and quantifying the processing throughput of the node; The flow transfer rollback rate is obtained by counting the frequency of sending and returning the task instruction between adjacent nodes in the time window, and is mapped to the logical oscillation feature at the network level; The task blocking state bit, a Boolean signal activated when the task stays for a duration exceeding a preset threshold, identifies a potential local network deadlock point.

[0023] In this embodiment, the task flow data flowing in the distributed collaborative network is deeply structured and processed, and a digital portrait reflecting the node work efficiency and pressure is constructed. The specific data processing procedure includes the following four rigorous calculation steps: First, regarding the calculation of the node queue accumulation depth, the system backend service scans the pending task buffer of each processing node at a second-level frequency; it does not focus on the specific content of the task, but counts the total number of task objects accumulated in the buffer. In order to eliminate the differences in processing capacity benchmarks of different nodes, the maximum parallel processing capacity of the node in the historical best state is read from the configuration database, and the current accumulated total is divided by the maximum parallel processing capacity to obtain a normalized ratio between 0 and 1. When this ratio tends to 1, it indicates that the node is in a full-load carrying state.

[0024] Second, regarding the acquisition of the average task residence time delay, the log mining technology is used to extract the timestamp of each task data packet entering the node and the timestamp marked as completed or transferred out. In a preset time window (e.g. the past hour), the residence time of all task data packets flowing through the node is calculated by arithmetic mean; this indicator can dynamically reflect the computing throughput efficiency of the node, and if the residence time delay shows exponential growth, it usually means that processing congestion has occurred within the node.

[0025] Third, regarding the monitoring of the transfer rollback rate, a directed path graph of task transfer is constructed, and in the time window, the complete hop path of each task instruction is tracked; if it is detected that a task instruction is sent from node A to node B and then rejected or reassigned back to node A within a very short time, such an event is marked as a logical shock; the proportion of such shock paths in the total transfer paths is calculated to quantify the shirking phenomenon between nodes or the incompatibility of the collaboration interface.

[0026] Finally, regarding the determination of the task blocking state bit, it is a binary fuse signal; a monitoring daemon is built in, which periodically checks all tasks in progress; once it finds that a task has been staying in a single node for more than the system's safety threshold, e.g. 48 hours without any state update, the daemon immediately marks the blocking state bit associated with the task as true, which is equivalent to identifying a potential deadlock point at the network level.

[0027] The application has obvious technical advantages by multi-dimensional feature processing of task flow data, solves the subjectivity problem of node load evaluation in traditional management, and through the introduction of node queue accumulation depth and normalization processing, can cross different types of business nodes, compare load pressure horizontally on the unified mathematical benchmark, so that the resource scheduling algorithm can accurately identify the short board and bottleneck of the whole network.

[0028] The introduction of flow back rate is one of the key innovations of the application, and the traditional load balancing algorithm often only pays attention to whether the task is sent out, but ignores whether the task is effectively received; high back rate often implies cognitive conflict or interface mismatch between nodes, and this index can make the system perceive invalid traffic consumption, thereby avoiding the waste of idle computing resources.

[0029] The task blocking state bit provides a deterministic risk warning mechanism, which can be used as a bottom protection mechanism to prevent deadlocked tasks from occupying node memory and computing slots indefinitely before the reinforcement learning model converges or responds, significantly improving the robustness and safety of the system.

[0030] The interaction log data includes unstructured text information between nodes, input behavior feature data of node operation terminal and corresponding timestamp metadata; the input behavior feature data includes key stroke rate, mouse movement track chaos degree and operation undo and redo frequency.

[0031] The interaction semantic analysis model is used for feature extraction of the interaction log data, and the stability index and time delay jitter index between node pairs are calculated combined with the timestamp metadata. The unstructured text information is segmented and vectorized embedded, the negative word density, imperative sentence frequency and high-intensity punctuation symbol usage frequency contained therein are detected, and the semantic sharpness score of single interaction is calculated; The number of interaction rounds between node pairs in a preset period and the number of times the same task instruction is transferred back and forth between nodes are counted to obtain the interaction oscillation frequency. The key stroke rate, mouse movement track chaos degree and operation undo and redo frequency in the input behavior feature data are encoded into behavior feature vectors by a long short-term memory network, and a behavior state score is identified according to the behavior feature vector; the mouse movement track chaos degree is defined as the variance of the curvature change rate of the mouse cursor coordinate point per unit time.

[0032] Further, the mouse movement track chaos degree is obtained as follows: The coordinate sequence of the mouse cursor in the operation of office software is collected at a frequency of 100 Hz, and the sampling duration is not less than 1 second; for the continuous three coordinate points, the curvature of the arc formed by the three points is calculated through the vector cross product and the module length, and then the ratio of the continuous curvature value to the sampling interval is obtained to obtain the curvature change rate; the variance of all the curvature change rates is counted to reflect the instability degree of the trajectory corner change; finally, the variance is normalized to the interval [0, 1], a maximum variance threshold is preset, and if the threshold is exceeded, it is counted as 1. The higher the score, the more disordered the trajectory, such as high-frequency shaking when anxious, repeated U-turns when tangled, the lower the score, the more stable, such as straight-line movement when focused, and the more accurate quantification of the implicit operation state of the collaborative subject, providing core data for behavior feature analysis.

[0033] The average waiting time of message reply is calculated based on timestamp metadata to obtain interaction time length data; The stability index includes semantic sharpness score and interaction oscillation frequency; the time delay jitter index includes behavior state score and interaction time length data.

[0034] An interaction semantic analysis model is used to extract accurate resistance indicators from a large amount of chaotic logs; the model uses a parallel pipeline architecture and is divided into four independent data processing channels: In the first branch, unstructured text information is processed. First, the text is divided into basic word units using a word segmentation tool and mapped to high-dimensional word vectors. Then, a pre-set negative dictionary and sentiment dictionary are scanned to calculate the density of negative words (such as errors, rejections, etc.) in the text; at the same time, imperative sentence structures are identified through syntax analysis trees, and the continuous occurrence frequency of high-intensity punctuation marks such as exclamation marks and question marks is counted; these statistical features are input into a logistic regression layer to output the semantic sharpness score of a single interaction, the higher the score, the stronger the antagonistic and conflict emotions contained in the text.

[0035] In the second branch, a statistical-based interaction oscillation frequency calculation is performed. A node pair interaction frequency table is maintained in memory; if dozens of message exchanges occur between two nodes within a very short preset period, and the same task instruction is repeatedly transferred between the two nodes with state changes, the system determines that the link is in a high-frequency and low-efficiency oscillation state, and the oscillation frequency is normalized and output.

[0036] In the third branch, behavior feature analysis is introduced; the input event stream of the node operation terminal is collected, including millisecond-level key interval sequences, mouse movement trajectories in the screen pixel coordinate system, and trigger logs of the undo combination key; these time series data are input into a long short-term memory network for encoding, which can capture the time dependence of operation behavior through a gating mechanism, recognize typical anxiety behavior patterns such as rapid tapping and disordered sliding, and output a behavior state score between 0 and 1.

[0037] In the fourth branch, the interaction duration data is calculated based on the timestamp metadata. Specifically, the time difference between the message sending time and the reply receiving time, i.e., the round-trip delay, is calculated; the historical average delay benchmark of the node pair is maintained, and the current actual delay is compared with the benchmark to quantify the jitter of the response speed.

[0038] The interaction semantic analysis model adopts a four-branch parallel pipeline architecture, uniformly uses the Adam optimizer, and iteratively trains for 30000 steps. The final output is four-dimensional core indicators, including semantic sharpness score, interaction oscillation frequency, behavior state score, and interaction duration data. The first branch processes unstructured text. After jieba word segmentation, Word2Vec word embedding, and CNN feature extraction, the semantic sharpness score in the [0, 1] interval is output by logistic regression, and negative / sentiment dictionary matching and imperative sentences / high-intensity punctuation statistics are fused. The second branch is based on a sliding window, and the number of interaction rounds and the number of task round-trip flows of the node pair are counted. After Min-Max normalization, the interaction oscillation frequency is output without independent training. The third branch receives three-dimensional time series data such as key interval and mouse chaos degree, encodes them through the LayerNorm layer and single-layer unidirectional LSTM, and outputs the behavior state score through the fully connected layer with a learning rate of 0.0008 and a Batch Size of 32. The fourth branch calculates the message round-trip delay, compares it with the historical benchmark to obtain the deviation rate and normalizes it.

[0039] The present application realizes holographic perception of node interaction behavior by constructing a multi-modal, four-branch feature extraction system, breaking through the limitations of single text analysis. In actual collaboration networks, many high-impedance antagonistic behaviors are implicit, such as nodes not using dirty words or explicit rejection words in text, but expressing antagonism through deliberate delay of reply or emotional violent keyboard operation. Breaking through the limitations of traditional management, which only focuses on explicit collaboration results and ignores implicit interaction processes, capturing emotional conflicts in text communication and state abnormalities in operation behavior, enables management decisions to penetrate surface collaboration and touch real collaboration quality, and intervene in implicit conflicts in advance.

[0040] The present application can capture the real impedance hidden under silence or surface politeness by introducing the behavior state score and interaction duration data. In addition, the calculation of interaction oscillation frequency can effectively identify the invalid communication trap. Even if the communication is friendly in semantics, if it falls into an infinite loop of discussion without output, it will still consume a large amount of system resources. This indicator enables the reinforcement learning model to learn to suppress this inefficient high-frequency interaction, thereby significantly improving the effective information throughput rate of the network.

[0041] An interaction impedance coefficient is calculated based on the stability index and the time delay jitter index. A specific process of generating a global interaction impedance matrix of a current time step based on the interaction impedance coefficient includes: By using a weighted normalization algorithm, the semantic sharpness score in the stability index, the interaction oscillation frequency, and the behavior state score in the time delay jitter index, and the interaction time length data are fused to obtain an interaction impedance coefficient of the node pair at the current time step. The interaction impedance coefficients of all node pairs jointly constitute the global interaction impedance matrix.

[0042] First, a set of dynamic weight parameters corresponding to the four dimensions of semantics, oscillation, behavior, and timeliness are defined. These weights are not fixed but can be pre-configured according to business scenarios.

[0043] By using a weighted normalization algorithm, the semantic sharpness score output by the first branch and the interaction oscillation frequency output by the second branch are weighted and summed to generate a stability index representing explicit conflict, wherein the weights are both 0.5 by default. Meanwhile, the behavior state score of the third branch and the interaction time length data of the fourth branch are fused to generate a time delay jitter index representing implicit friction, wherein the weights are both 0.5 by default.

[0044] Subsequently, a nonlinear activation function, such as a hyperbolic tangent function, is introduced to map the stability index and the time delay jitter index into the interval of 0 to 1 to obtain the interaction impedance coefficient of the node pair at the current time. The coefficient is a scalar, and the larger the value, the more congested the channel between the two nodes and the greater the logical friction.

[0045] The calculation formula of the interaction impedance coefficient is as follows: ; ; ; wherein, is the current time window, representing the time step index; is the interaction impedance coefficient, representing the logical friction degree of the node i and j at t, and the value range is (0, 1); is a hyperbolic tangent activation function, used to nonlinearly map the linear combination result to the interval of (0, 1) to simulate the saturation effect of impedance; , is a hyperparameter for balancing the influence of the stability index and the jitter index, and the default value is 0.5, and and are 1; is the stability index; is the semantic sharpness score, which is the text conflict probability extracted by NLP, and the normalized value; is the interaction oscillation frequency, which is the normalized value of the number of task state changes per unit time; , The weighted coefficients of the stability indicators are both 0.5 by default and add up to 1; is a time delay jitter indicator; is a behavior state score, which is based on the anxiety score of the input device operation behavior by LSTM, and the data is normalized; is interaction duration data, a jitter deviation value of message response time delay relative to historical benchmarks, and the data is normalized; , The weighted coefficients of the time delay jitter indicators are both 0.5 by default and add up to 1.

[0046] Finally, all node pairs in the network are traversed to calculate impedance coefficients, and these coefficients are filled into a two-dimensional matrix of N rows and N columns, where N is the total number of nodes in the network. This matrix is the global interaction impedance matrix, which constitutes the heat map of the entire collaborative network at the current time step.

[0047] In the present application, the interaction impedance coefficient is analogous to the impedance in the circuit system, i.e., a physical quantity that hinders current transmission. Specifically, the interaction impedance coefficient in the present application represents the comprehensive hindrance to the transmission of task, information, and other collaborative energy between nodes due to explicit conflicts and implicit friction, which is a quantitative measure of non-efficiency loss in organizational collaboration. Saturation characteristic: When the explicit conflict and the implicit friction reach a critical value, the growth rate of the coefficient significantly slows down, i.e., in actual collaboration, the degree of hindrance decreases when the conflict escalates to a certain extent.

[0048] The present application provides a standardized mathematical model for complex organizational relationship management by generating a global interaction impedance matrix, realizing the digitization and computability of relationships.

[0049] In traditional management, organizational relationships are described in a fuzzy and qualitative manner, which cannot be directly processed by computers. However, the present application quantifies such relationships into an accurate matrix, enabling subsequent deep learning algorithms and graph algorithms to directly perform convolution, decomposition, or eigenvalue calculation on this matrix. In addition, through a nonlinear fusion strategy, this method remains sensitive to extreme outliers while being robust to low-intensity noise. This means that a high impedance will not be falsely determined due to occasional reply delays. Only when multiple indicators simultaneously show negative trends will the coefficient in the matrix significantly increase, thereby ensuring the accuracy of system decisions and avoiding management interference caused by false positives. By converting complex organizational relationship networks into standardized mathematical matrices, the present application realizes relationship visualization and conflict quantification, enabling management decisions to accurately locate high-friction collaboration links and avoid blind regulation, thereby improving management targeting.

[0050] The specific process of constructing the state space of a reinforcement learning agent, mapping the global interaction impedance matrix to graph-structured data, and fusing task flow data to generate a comprehensive state vector containing network topology features and load features includes: Construct a graph structure data with nodes as vertices and interaction relationships as edges, and map the global interaction impedance matrix to the dynamic weight attribute of the edges in the graph structure data; The graph structure data is input into the graph convolutional neural network layer, and the impedance weight features of the first-order and second-order neighbors of the target node are aggregated by the convolution kernel to output an enhanced graph embedding vector that represents the propagation status of network topology risk. The node queue stacking depth, average task dwell time, flow rollback rate, and task blocking status bits in the task flow data are numerically standardized to generate node load feature vectors. The enhanced graph embedding vector and the node load feature vector are concatenated and fused to generate the multi-dimensional comprehensive state vector.

[0051] On topology-aware paths, the global interaction impedance matrix is ​​treated as a weighted adjacency matrix of the graph, constructing a graph structure with processing nodes as vertices and interaction relationships as edges. This graph structure is then fed into a graph convolutional neural network layer. The graph convolutional network utilizes the principle of spectral graph convolution, defining a convolutional kernel that slides across the graph, aggregating the impedance weights of each node and its first-order neighbors (direct connectors) and second-order neighbors (indirect connectors). The activation function is ReLU, and a BatchNorm layer is added before the output. The optimizer is the Adam optimizer; training hyperparameters include a learning rate of 0.0012, a batch size of 64, a dropout rate of 0.2, 35,000 iterations, and cross-entropy as the loss function. After multiple convolutional operations, a high-dimensional enhanced graph embedding vector is output. This vector not only contains the impedance information of the nodes themselves but also implicitly contains the potential trend of impedance propagation in the network.

[0052] On the load-aware path, the acquired task flow data, including the node queue stacking depth, are numerically standardized, such as Z-score standardization, to generate node load feature vectors.

[0053] Specifically, the Z-score standardization algorithm is used to process the node queue backlog depth, average task dwell time, and flow rollback rate, converting the feature values ​​into standardized values ​​with a mean of 0 and a variance of 1. The task blocking status bit (Boolean signal) is numerically mapped: the inactive state is mapped to the value 0, indicating that there is no potential deadlock point; the active state is mapped to the value 1, indicating that there is a potential deadlock point; since it is already in the 0-1 range after mapping, the binary numerical characteristics are kept unchanged to retain the strong signal characteristics of deadlock warning.

[0054] Finally, feature concatenation and fusion operations are performed; the enhanced graph embedding vector and the node load feature vector are concatenated along the channel dimension to form a multi-dimensional comprehensive state vector. This vector contains both topological features indicating where the relationship network is blocked and load features indicating who has too much work, forming a complete contextual basis for the reinforcement learning model to perform inference.

[0055] This invention employs advanced graph neural network technology, endowing the system with the ability to predict the propagation of risks. Traditional management methods often only see the current conflict, i.e., point-based thinking; while this solution, by aggregating the features of first-order and second-order neighbors, enables the agent to perceive the transmissibility of risks. For example, when node A and node B have a high-impedance interaction, although node C is currently in a normal state, because C is closely connected to B, the graph convolutional network will calculate the potential risk faced by C, which is then reflected in the comprehensive state vector. At the same time, by splicing and fusing topological features and load features, the problem of the separation between relationships and tasks is solved. The decisions made by the agent based on this vector will neither cause task backlog simply to alleviate relationships, nor ignore relationship deterioration in order to rush the schedule, thus finding the optimal balance among multiple objectives.

[0056] The interactive policy learning model includes an actor network and a critic network: The actor network is responsible for generating deterministic parameter adjustment strategies based on the input comprehensive state vector, and the critic network is responsible for predicting future cumulative reward values ​​based on the current state and the actions output by the actor network. The design of the reward function includes a positive gain term and a negative penalty term. The positive gain term is positively correlated with the number of tasks completed per unit time in the entire network, while the negative penalty term is negatively correlated with the norm mean of the network's interactive impedance matrix and the duration of high-impedance links. This guides the model to actively reduce semantic friction within the network while pursuing task processing speed.

[0057] This embodiment uses the actor-commentator algorithm architecture in deep reinforcement learning to construct an interactive policy learning model.

[0058] An actor network is a multilayer perceptron that receives a comprehensive state vector as input, processes it through weighted layers of neurons, and ultimately outputs a deterministic action policy distribution. This policy determines which nodes should be applied with what degree of control parameters.

[0059] The commentator network acts as the referee. It also receives the composite state vector, but its output is a value scalar that predicts the expected cumulative reward that can be obtained in the future after following the agent network's policy in the current state.

[0060] The actor network adopts an MLP (Multilayer Perceptron) architecture. The input layer receives the comprehensive state vector, and the number of neurons is consistent with the input dimension. Two fully connected hidden layers are set, with 256 and 128 neurons respectively, both using ReLU activation functions. A BatchNorm layer is added to accelerate convergence. The number of neurons in the output layer matches and adjusts the dimension of the action parameters, which are mapped to the [0,1] interval through Sigmoid activation. The hyperparameters are configured as follows: learning rate 0.0005, Adam optimizer, weight decay 1e-5, batch size = 32, hidden layer dropout rate 0.1, and 50,000 iterations of training, adapting to the deterministic policy output requirements of collaborative scenarios.

[0061] The critic network is also built on an MLP, with the same input layer as the actor network, receiving a combined state vector. The hidden layer structure is identical, consisting of two fully connected layers with 256 and 128 neurons respectively. ReLU activation is used, along with a BatchNorm layer. The output layer has only one neuron, outputting a scalar value through Linear activation. Hyperparameters include a learning rate of 0.001, the same Adam optimizer configuration, a batch size of 32, and a dropout rate of 0.1. The target network performs a soft update every 100 steps, outputting a scalar value in the range [-10, 10], used to accurately predict the expected cumulative reward after implementing the policy in the current state, providing stable evaluation feedback for the actor network.

[0062] During model training and updating, a composite reward function was designed, comprising a positive gain term and a negative penalty term. The positive gain term is directly linked to the number of tasks completed per unit time across the entire network, i.e., throughput. The negative penalty term is negatively correlated with the norm mean of the global interaction impedance matrix, i.e., the overall friction level, and the duration of high-impedance links. After each time window, the actual reward is calculated, and the prediction error of the critic network is also calculated. This error is then used to update the weight parameters of both the actor network and the critic network simultaneously via backpropagation.

[0063] The formula for the reward function is: ; in, The reward function value; This represents the number of tasks successfully completed across the entire network within the current time window, i.e., the positive gain core. The preset time window length; The Frobenius norm of the global interactive impedance matrix is ​​used to penalize the overall friction level of the entire network; This is the global interactive impedance matrix; This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise. The preset high impedance warning threshold can be set to, for example, 0.8; The duration of the high impedance state between nodes i and j is normalized data. This is the task output gain coefficient; This is the network impedance penalty coefficient; This represents the penalty coefficient for the duration of malicious links.

[0064] The process for determining the task output gain coefficient, network impedance penalty coefficient, and malicious link duration penalty coefficient is summarized as follows: Phase 1: Based on the statistical goal of normalization, eliminate significant differences in the numerical values ​​of the three indicators—task quantity, impedance norm, and duration—to prevent larger values ​​from dominating the gradient descent direction. Historical interaction logs are collected, and the standard deviations of the three indicators are calculated. Initialization is performed using a reciprocal weighting principle, setting each coefficient to the reciprocal of the corresponding indicator's standard deviation; that is, the greater the indicator's volatility, the smaller the coefficient's weight, ensuring that the initial contributions of the three to the reward function are of the same order of magnitude.

[0065] Specifically, the core of this process is to eliminate the magnitude differences among the three types of indicators (task quantity, impedance norm, and duration) to ensure a balanced contribution to the reward function. First, historical interaction logs of organizational collaboration over the past 30 days are collected, and valid data is filtered to cover scenarios such as normal collaboration, high load, and high conflict, extracting the raw data for the three types of indicators. Next, the statistical standard deviation is calculated for each type of indicator, quantifying the degree of data fluctuation through formulas to obtain the standard deviation σ1 for task quantity, σ2 for impedance norm, and σ3 for duration. Finally, coefficients are initialized according to the reciprocal weighting principle: task output gain coefficient w1 = 1 / σ1, network impedance penalty coefficient w2 = 1 / σ2, and malicious link duration penalty coefficient w3 = 1 / σ3. This method ensures that the greater the fluctuation of the indicator, the smaller the corresponding coefficient weight, ultimately making the initial contributions of the three types of indicators to the reward function of the same magnitude, avoiding a single indicator dominating the gradient descent direction.

[0066] The second stage: Simulation optimization based on grid search aims to find the optimal balance point, or Nash equilibrium point, between the conflicting goals of pursuing efficiency and reducing friction. Operation: A three-dimensional parameter search grid is constructed centered on the normalized values, and multiple parameter combinations are run in parallel within a virtual simulation environment. By plotting the Pareto front curve, the coefficient combination that achieves the highest task throughput without a significant increase in network impedance is selected as the baseline fixed value for model deployment.

[0067] Specifically, the simulation optimization process aims to find a balanced parameter combination that balances efficiency and low friction. The search range is based on normalized initial coefficients, covering three strategy scenarios: efficiency-focused, low-friction-focused, and balanced. The search step size is set to 0.1 times the initial coefficients. Candidate values ​​are generated for each coefficient dimension, and a three-dimensional mesh forms multiple parameter combinations, balancing optimization accuracy and computational efficiency. A stopping condition is met if either of the following is satisfied: 1) all parameter combinations have completed virtual simulation testing; or 2) performance convergence occurs, with the Pareto front optimal solution remaining unchanged for 10 consecutive simulation rounds (i.e., the highest task throughput without a significant increase in network impedance, and the corresponding reward function value fluctuating by ≤1%). After stopping, the optimal parameter combination is output as the baseline fixed value for model deployment.

[0068] Phase 3: Adaptive dynamic adjustment based on system load. The strategy's focus is dynamically adjusted according to the real-time urgency. Operation: Real-time monitoring of node queue backlog depth. In high-load mode, the task output gain coefficient is automatically increased to guide the agent to tolerate minor friction and prioritize preventing system crashes. In low-load mode, the network impedance penalty coefficient is automatically increased to guide the agent to perform strict traffic shaping and deeply manage potentially stubborn high-impedance links.

[0069] This embodiment utilizes an actor-critic architecture to achieve adaptive evolution of management strategies. Unlike traditional rule-based systems, the model of this invention does not require pre-written if-then rules. Guided by a composite reward function, the model can automatically learn extremely complex management strategies through continuous trial and error and iteration. For example, the model may automatically learn to tolerate slight interaction friction during peak system load periods to ensure speed, while aggressively performing impedance regulation during off-peak periods to optimize the network structure; this enables the distributed collaborative network to possess self-regulation and evolutionary capabilities, and to cope with constantly changing external demands and internal environments.

[0070] The output dynamic control action commands include traffic shaping operations for the interaction link: The traffic shaping operation includes setting a cooling-off time threshold and frequency limit for message delivery; when the agent detects overheating interaction impedance between node pairs, it outputs a dynamic adjustment action command to lock the instant communication input interface of both parties and delay the push time of non-urgent messages.

[0071] This embodiment describes how decision-making instructions output by an intelligent agent are executed at the physical level.

[0072] When the reinforcement learning model outputs a traffic shaping instruction for a specific node pair, such as node A and node B, the parameters of the underlying real-time communication service and task distribution engine are reset through the control interface layer.

[0073] The specific steps include: First, setting a cooldown time threshold.

[0074] Insert a delay middleware into the message sending queue of node A to forcibly introduce a minute-level delivery delay for non-high-priority text messages.

[0075] Second, frequency limits and interface locking. When extremely overheated interactive impedance is detected, a command is sent to the front-end client to temporarily gray out the chat input box of node A for node B and display a prompt from the system flow control, forcing both parties into a physical silence period. This locked state will continue until the impedance monitoring value in the next time window falls back to a safe range.

[0076] The traffic shaping operation provided by this invention can quickly interrupt vicious cycles of interaction. In organizational collaboration, antagonistic emotions often spiral upwards with immediate rebuttals and arguments, causing both parties to completely cease effective work. This solution, by forcibly inserting a silent period, cuts off the immediate feedback loop of negative emotions, giving nodes—that is, collaborators—time to cool down cognitively. This non-invasive yet mandatory intervention method can prevent local high-impedance oscillations from evolving into network-wide paralysis with minimal system cost, effectively protecting the stability of the collaborative network and ensuring the smooth flow of core tasks. By interrupting vicious cycles of interaction through non-invasive intervention, it avoids the lag of manual mediation after conflict escalation in traditional management, resolving communication conflicts at minimal cost.

[0077] Furthermore, in addition to traffic shaping, the system also includes route reconstruction for task distribution. The route reconstruction includes dynamically adjusting the connection weights between nodes. When the interaction impedance coefficient of a specific link is found to exceed a threshold, the agent outputs an instruction to reduce the task allocation weight of that link and searches for a third-party node with a lower impedance coefficient in the network as a relay stepping stone to construct a bypass route, thereby isolating high-impedance node pairs at the logical topology level.

[0078] Example 2: This invention proposes a method for differentiated management of organizational relationships based on reinforcement learning, comprising: Based on a preset time window, task flow data and interaction log data between nodes in the distributed collaboration network are collected. The interaction log data is feature extracted using an interaction semantic analysis model, and the stability index and latency jitter index between node pairs are calculated in combination with timestamp metadata. The interaction impedance coefficient is calculated based on the stability index and latency jitter index, and the global interaction impedance matrix for the current time step is generated based on the interaction impedance coefficient. Construct the state space of the reinforcement learning agent, map the global interaction impedance matrix into graph structure data, and fuse task flow data to generate a comprehensive state vector containing network topology features and load features. The integrated state vector is input into the interactive policy learning model for inference, and the policy network outputs dynamic control action instructions for the nodes. Execute dynamic control action instructions to regulate the actions of the distributed cooperative network, and detect changes in the global interaction impedance matrix in the next time window to calculate the reward function value to update the parameters of the interaction policy learning model in reverse.

[0079] Among them, nodes refer to the collaborating entities such as people, departments, and teams in an organization, as well as the computer processing units that support their work, such as terminals, servers, and business system modules. Tasks in task flow data refer to work tasks within an organization, such as project division of labor, approval processes, and business requirement coordination. These tasks are recorded and their flow trajectory is stored through computer systems, including to-do lists and processing times.

[0080] For example, the method described in this invention is adapted to office software such as WeChat Work and DingTalk, which enable instant communication and collaboration. The node settings of the distributed collaboration network are as follows: Departmental nodes, such as the product department and the technology development department, are non-leaf nodes and have independent permission configurations, such as allowing only internal members to view departmental tasks. Member node: All employees can bind this member node using their authorized login accounts.

[0081] Instant communication: one-on-one chat and departmental group chat in office software, where members send text and files through their accounts, forming the unstructured text information and operational behavior characteristics required for patents; Distributed collaboration: This includes document collaboration, spreadsheet collaboration, etc., which can enable collaboration among different member nodes.

[0082] The task flow data includes node queue backlog depth, average task dwell time, flow rollback rate, and task blocking status bit. The node queue stacking depth is obtained by counting the number of task objects in the node's to-do list and then normalizing it, serving as a numerical feature characterizing the instantaneous load pressure of the node. The average dwell time of the task is calculated based on the timestamp log of the task object, and the average dwell time of the task within the node is obtained to quantify the processing throughput of the node. The back-off rate is obtained by the frequency of the task instruction sending and back-off loop path between adjacent nodes within a statistical time window, and is mapped to the logical oscillation characteristics at the network level. The task blocking status bit is a Boolean signal activated when the task dwell time exceeds a preset threshold, identifying potential local network deadlock points.

[0083] The interaction log data includes unstructured text information between nodes, input behavior feature data of node operation terminals, and corresponding timestamp metadata; the input behavior feature data includes key press rate, mouse movement trajectory disorder, and operation undo / redo frequency.

[0084] The interactive semantic analysis model extracts and identifies stability and latency jitter metrics, including: Unstructured text information is segmented and vectorized for embedding, and the density of negative words, frequency of imperative sentences, and frequency of high-intensity punctuation are detected. The semantic sharpness score of a single interaction is calculated. The interaction oscillation frequency is calculated by statistically analyzing the number of interaction rounds between node pairs and the number of times the same task instruction travels back and forth between nodes within a preset period. Based on the key-click rate, mouse movement trajectory disorder, and operation undo / redo frequency in the input behavior feature data, the data is encoded into behavior feature vectors through a long short-term memory network. Behavior state scores are obtained based on the behavior feature vectors. The average message reply waiting time is calculated based on the timestamp to obtain the interaction duration data. Specifically, this embodiment focuses on two specific processing nodes in a distributed collaborative network: node A is the upstream business demand publishing unit, and node B is the downstream technology development execution unit; it describes the microscopic operating logic of this method when high-impedance interaction occurs between the two.

[0085] Scenario: During a high-priority project sprint, node A is responsible for issuing specific parameter configuration tasks to node B.

[0086] Step 1: Real-time capture of abnormal data; Interaction log capture: Node A was detected to have sent five text messages to Node B within two minutes; Input behavior capture: At the same time, the operation terminal of node B was found to be abnormal through the monitoring of the embedded data: when replying to messages, the click frequency of the backspace key reached 40 times per minute, far exceeding the average value, and the mouse trajectory showed high-frequency, small-amplitude disordered circular motion.

[0087] Step 2: Parallel computation of multimodal features; Semantic Branch: The interactive semantic analysis model identifies words such as "not right," "immediately," and "wrong" in text messages as high-weight negation and imperative words. Combined with high-frequency sending behavior in a short period of time, the semantic sharpness score is calculated.

[0088] Interactive oscillation frequency: Statistical analysis revealed that the status of the same task instruction changed three times between pending confirmation and rejection. This frequency was compared with historical benchmarks and normalized to calculate the interactive oscillation frequency.

[0089] Behavioral branch: The Long Short-Term Memory network processes the key sequence of node B, identifies the entanglement pattern of "input-delete-re-input", and calculates the behavioral state score.

[0090] Interaction duration data: Calculates the average response latency of node B to the last five messages. Historical benchmarks show that B's normal response time is three minutes. Combining this with the response time displayed in the current monitoring data, the interaction duration data is quantified.

[0091] Impedance generation: Combining the above indicators, the cross-impedance coefficient between node A and node B is calculated. At this point, the corresponding element in the global cross-impedance matrix turns red.

[0092] Step 3: Reinforce the reasoning and decision-making of the intelligent agent; State input: The agent reads the high impedance between A and B, and combined with the task flow data, it finds that there are three high-priority tasks waiting to be processed in the task queue of node B, and these tasks depend on a third-party node C.

[0093] Risk prediction: Enhanced graph embedding vectors show that if node B continues to be blocked by node A's entanglement, it will cause node C connected to it to cascade and wait, thus triggering a local network deadlock.

[0094] Policy Output: The agent outputs compound action instructions: Traffic Shaping: Apply a message cooling policy to node A, setting the delivery delay of its non-urgent messages to five minutes.

[0095] Routing Reordering: Temporarily suspend new tasks issued by node A in node B's to-do queue, lower their priority, and promote tasks from node C to the head of the queue for priority execution.

[0096] Step Four: Execution and Feedback; Execution result: Due to the message delay, node A could not interrupt node B by continuously urging it. Node B used this quiet period to concentrate on completing node C's task.

[0097] State recovery: Five minutes later, the keyboard typing rate of node B was detected to have returned to a stable level, indicating that the adjustment was effective and the agent was given a positive reward.

[0098] To verify the objective technical effectiveness of this method, a comparative test was conducted in a collaborative network environment containing one hundred processing nodes and an average daily processing volume of five thousand tasks.

[0099] Control group: Employs a traditional first-in-first-out (FIFO) task allocation strategy and lacks cross-impedance sensing capabilities. Communication between nodes is not interfered with. Experimental group: Deploy a reinforcement learning-based interactive impedance dynamic adjustment system and enable traffic shaping and route reconstruction functions.

[0100] The testing period was seven consecutive working days, and the statistical data are shown in Table 1.

[0101] Table 1 Comparison of experimental index data

[0102] As can be seen from the table data, the average task flow cycle of the experimental group was shortened by nearly 30%. This is because when the invention detects an increase in interaction impedance, it actively avoids congested nodes through route rearrangement, enabling tasks to flow through a smoother path.

[0103] Friction is significantly reduced: The duration of high-impedance links is significantly reduced, thanks to the traffic shaping mechanism, which effectively cools down the conflict in its early stages and prevents the semantic adversarial process from escalating indefinitely.

[0104] Reduced rework rate: A significant reduction in task rollback rate, achieved by pausing the dispatch of complex tasks to nodes when they are in a high behavioral state score, avoids low-quality delivery and subsequent rework.

[0105] In summary, this invention achieves a qualitative leap in organizational collaboration efficiency at the macro level through millisecond-level perception and control at the micro level, effectively solving the problems of network congestion and resource waste caused by interpersonal interaction friction under the traditional static management model.

[0106] It should be understood that the embodiments disclosed in this invention and the above description enable those skilled in the art to implement this invention. However, this invention is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this invention.

Claims

1. A method for differentiated management of organizational relationships based on reinforcement learning, characterized in that, include: Based on a preset time window, task flow data and interaction log data between nodes in the distributed collaboration network are collected. The interaction log data is feature extracted using an interaction semantic analysis model, and the stability index and latency jitter index between node pairs are calculated in combination with timestamp metadata. The interaction impedance coefficient is calculated based on the stability index and latency jitter index, and the global interaction impedance matrix for the current time step is generated based on the interaction impedance coefficient. Construct the state space of the reinforcement learning agent, map the global interaction impedance matrix into graph structure data, and fuse task flow data to generate a comprehensive state vector containing network topology features and load features. The integrated state vector is input into the interactive policy learning model for inference, and the policy network outputs dynamic control action instructions for the nodes. Execute dynamic control action instructions to regulate the actions of the distributed cooperative network, and detect changes in the global interaction impedance matrix in the next time window to calculate the reward function value to update the parameters of the interaction policy learning model in reverse.

2. The organizational relationship differentiation management method based on reinforcement learning as described in claim 1, characterized in that, The task flow data includes node queue backlog depth, average task dwell time, flow rollback rate, and task blocking status bit. The node queue stacking depth is obtained by counting the number of task objects in the node's to-do list and then normalizing it, serving as a numerical feature characterizing the instantaneous load pressure of the node. The average dwell time of the task is calculated based on the timestamp log of the task object, and the average dwell time of the task within the node is obtained to quantify the processing throughput of the node. The back-off rate is obtained by the frequency of the task instruction sending and back-off loop path between adjacent nodes within a statistical time window, and is mapped to the logical oscillation characteristics at the network level. The task blocking status bit is a Boolean signal activated when the task dwell time exceeds a preset threshold, identifying potential local network deadlock points.

3. The organizational relationship differentiation management method based on reinforcement learning as described in claim 1, characterized in that, The interaction log data includes unstructured text information between nodes, input behavior feature data of node operation terminals, and corresponding timestamp metadata; the input behavior feature data includes key press rate, mouse movement trajectory disorder, and operation undo / redo frequency.

4. The method for differentiated organizational relationship management based on reinforcement learning as described in claim 3, characterized in that, The interaction log data is used to extract features using an interaction semantic analysis model, and the stability indicators and latency jitter indicators between node pairs are calculated by combining timestamp metadata, including: Unstructured text information is segmented and vectorized for embedding, and the density of negative words, frequency of imperative sentences, and frequency of high-intensity punctuation are detected. The semantic sharpness score of a single interaction is calculated. The interaction oscillation frequency is calculated by counting the number of interaction rounds between node pairs and the number of times the same task instruction flows back and forth between nodes within a preset period. The keystroke rate, mouse movement trajectory disorder, and operation undo / redo frequency in the input behavioral feature data are encoded into behavioral feature vectors through a long short-term memory network, and behavioral state scores are obtained based on the behavioral feature vectors. Calculate the average message response waiting time based on timestamp metadata to obtain interaction duration data; The stability indicators include semantic sharpness score and interaction oscillation frequency; the latency jitter indicators include behavioral state score and interaction duration data.

5. The organizational relationship differentiation management method based on reinforcement learning as described in claim 4, characterized in that, The specific process of calculating the interaction impedance coefficient based on the stability index and the time delay jitter index, and generating the global interaction impedance matrix for the current time step based on the interaction impedance coefficient includes: By using a weighted normalization algorithm, the semantic sharpness score and interaction oscillation frequency in the stability index, as well as the behavioral state score and interaction duration data in the latency jitter index, are fused to obtain the interaction impedance coefficient of the node pair at the current time step. The interaction impedance coefficients of all node pairs together constitute the global interaction impedance matrix.

6. The method for differentiated organizational relationship management based on reinforcement learning as described in claim 1, characterized in that, The specific process of constructing the state space of a reinforcement learning agent, mapping the global interaction impedance matrix to graph-structured data, and fusing task flow data to generate a comprehensive state vector containing network topology features and load features includes: Construct a graph structure data with nodes as vertices and interaction relationships as edges, and map the global interaction impedance matrix to the dynamic weight attribute of the edges in the graph structure data; The graph structure data is input into the graph convolutional neural network layer, and the impedance weight features of the first-order and second-order neighbors of the target node are aggregated by the convolution kernel to output an enhanced graph embedding vector that represents the propagation status of network topology risk. The node queue stacking depth, average task dwell time, flow rollback rate, and task blocking status bits in the task flow data are numerically standardized to generate node load feature vectors. The enhanced graph embedding vector and the node load feature vector are concatenated and fused to generate the multi-dimensional comprehensive state vector.

7. The organizational relationship differentiation management method based on reinforcement learning as described in claim 1, characterized in that, The interactive policy learning model includes an actor network and a critic network: The actor network is responsible for generating deterministic parameter adjustment strategies based on the input comprehensive state vector, while the critic network is responsible for predicting future cumulative reward values ​​based on the current state and the actions output by the actor network. The reward function is designed to include a positive gain term and a negative penalty term. The positive gain term is positively correlated with the number of tasks completed per unit time across the entire network, while the negative penalty term is negatively correlated with the norm mean of the network's interaction impedance matrix and the duration of high-impedance links. This guides the model to actively reduce semantic friction within the network while pursuing task processing speed.

8. The organizational relationship differentiation management method based on reinforcement learning as described in claim 1, characterized in that, The output dynamic control action commands include traffic shaping operations for the interaction link: The traffic shaping operation includes setting a cooling-off time threshold and frequency limit for message delivery; when the agent detects overheating interaction impedance between node pairs, it outputs a dynamic adjustment action command to lock the instant communication input interface of both parties and delay the push time of non-urgent messages.

Images