An organization relationship differentiation management method based on reinforcement learning

Abstract

The application discloses a kind of organization relationship differentiation management methods based on reinforcement learning, it is related to data processing technical field, comprising: based on the time window of pre-set, the task flow data between each node in distributed collaboration network and interactive log data are collected;Utilize interactive semantic analysis model to the feature extraction of the interactive log data, the stability index and time delay jitter index between node pair are calculated in combination with timestamp metadata, interactive impedance coefficient is measured and obtained, and global interactive impedance matrix is constructed;The state space of reinforcement learning intelligent agent is constructed, global interactive impedance matrix is mapped as graph structure data, and the comprehensive state vector containing network topology features and load features is generated by fusing task flow data;Comprehensive state vector is input into interactive strategy learning model and inferences, and the dynamic regulation and control action instruction for node is output through strategy network;Dynamic regulation and control action instruction is executed, and distributed collaboration network is regulated and controlled.

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 fluctuating in real time; 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 interactive 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 interactive links at the micro level, and achieves the technical effect of differentiated and accurate management.

[0007] The present application provides an organization relationship differentiated management method based on reinforcement learning, comprising:

[0008] Based on a preset time window, task flow data and interaction log data between nodes in a distributed collaboration network are collected;

[0009] 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 in combination with timestamp metadata; the interaction impedance coefficient is calculated based on the stability index and time delay jitter index, and the global interaction impedance matrix of the current time step is generated based on the interaction impedance coefficient;

[0010] A state space of a reinforcement learning agent is constructed, a 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;

[0011] The comprehensive state vector is input into an interaction policy learning model for inference, and a dynamic regulation action instruction for the node is output through the policy network;

[0012] 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 to calculate the reward function value to update the parameters of the interaction policy learning model in reverse.

[0013] The task flow data includes node queue accumulation depth, task average residence time delay, flow transfer rollback rate, and task blocking state bit;

[0014] 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;

[0015] The task average residence time delay is calculated based on the timestamp log of the task object to obtain the average residence time of the task in the node, and quantifies the processing throughput of the node;

[0016] The flow transfer rollback rate is obtained by counting the frequency of sending and returning task instructions between adjacent nodes within a time window, and is mapped to a logical shock feature at the network level;

[0017] 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.

[0018] 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 trajectory chaos degree, and operation undo and redo frequency.

[0019] 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:

[0020] The unstructured text information is segmented and vectorized embedded to detect the negative vocabulary density, imperative sentence frequency, and high-intensity punctuation symbol usage frequency contained therein, and the semantic sharpness score of a single interaction is calculated;

[0021] The number of interaction rounds between node pairs within a preset period and the number of times the same task instruction is transferred back and forth between nodes are counted to calculate the interaction shock frequency;

[0022] The key stroke rate, the confusion degree of the mouse movement trajectory and the frequency of operation undo and redo in the input behavior feature data are encoded into a behavior feature vector by a long short-term memory network, and a behavior state score is obtained according to the behavior feature vector;

[0023] The average waiting time length of message reply is calculated based on timestamp metadata to obtain interaction time length data;

[0024] The stability index includes a semantic sharpness score and an interaction oscillation frequency; and the time delay jitter index includes a behavior state score and interaction time length data.

[0025] An interaction impedance coefficient is calculated based on the stability index and the time delay jitter index, and a global interaction impedance matrix of the current time step is generated based on the interaction impedance coefficient.

[0026] The semantic sharpness score and the interaction oscillation frequency in the stability index, and the behavior state score and the interaction time length data in the time delay jitter index are fused by a weighted normalization algorithm to obtain an interaction impedance coefficient of a node pair at the current time step, and the interaction impedance coefficients of all node pairs together constitute the global interaction impedance matrix.

[0027] A state space of a reinforcement learning intelligent agent is constructed, the global interaction impedance matrix is mapped into graph structure data, and task flow data is fused to generate a comprehensive state vector including network topology features and load features.

[0028] A graph structure data is constructed with nodes as vertices and interaction relationships as edges, and the global interaction impedance matrix is mapped into a dynamic weight attribute of an edge in the graph structure data.

[0029] 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.

[0030] Node load feature vectors of nodes are generated by numerically standardizing node queue accumulation depth, task average residence time delay, flow back rate and task blocking state bits in the task flow data.

[0031] The enhanced graph embedding vector and the node load feature vector are spliced and fused to generate a multi-dimensional comprehensive state vector.

[0032] The interaction strategy learning model includes an actor network and a critic network.

[0033] The actor network is responsible for generating a deterministic parameter adjustment strategy according to an input integrated 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;

[0034] The reward function value design includes a positive gain term and a negative penalty term;

[0035] The positive gain term is positively correlated with the number of task completions per unit time in the entire network.

[0036] The negative penalty term is negatively correlated with the average norm of the interaction impedance matrix of the entire network and the duration of the high impedance link; this guides the model to actively reduce semantic friction in the network while pursuing task processing speed.

[0037] The output dynamic regulation action instruction includes a traffic shaping operation for the interaction link:

[0038] The traffic shaping operation includes setting a cooling time threshold and a frequency limit for message delivery; when the agent identifies that there is an overheated interaction impedance between a node pair, the dynamic regulation action instruction is output to lock the instant messaging input interfaces of both parties and delay the push time of non-urgent messages.

[0039] The beneficial effects of the present application are:

[0040] 1. The present application has significant technical advantages through multi-dimensional feature processing of task flow data, solving the subjective problem of node load assessment in traditional management; by introducing node queue accumulation depth and normalization processing, it can cross different types of business nodes and 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 entire network.

[0041] 2. The present application can capture the real impedance hidden under silence or surface politeness by introducing behavior state scores and interaction duration data; in addition, the calculation of interaction shock frequency can effectively identify the invalid communication trap; even if the communication is semantic friendly, if it falls into an infinite loop of discussion without output, it will still consume a large amount of system resources; this index 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 and achieving accurate identification.

[0042] 3、The application utilizes the actor-critic framework to realize the adaptive evolution of management strategies; through the guidance of the composite reward function, the model can automatically learn extremely complex management strategies in continuous trial and error and iteration; in the load peak period, slight interactive friction is appropriately tolerated to ensure speed, while in the load valley period, aggressive impedance regulation is carried out to optimize the network structure, so that the distributed collaborative network has self-regulation and evolution ability and can cope with changing external demand and internal environment. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 A flowchart of the organizational relationship differentiation management method based on reinforcement learning.

[0044] Figure 2 A data logic diagram of the organizational relationship differentiation management method based on reinforcement learning. DETAILED DESCRIPTION

[0045] The above technical solutions will be described in detail below in combination with the 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 skilled in the art without creative labor fall within the scope of protection of the present application. It should be noted that, for convenience of description, only parts related to the present application are shown in the drawings, not all. EMBODIMENT

[0046] The present application proposes an organizational relationship differentiation management method based on reinforcement learning, the flow of the method is as shown in Figure 1 The data logic of the method is as shown in Figure 2

[0047] Based on the preset time window, the task flow data and the interaction log data between the nodes in the distributed collaborative network are collected;

[0048] The interaction semantic analysis model is used to extract features from the interaction log data, and the stability index and the time delay jitter index between the node pairs are calculated in combination with the timestamp metadata; the interaction impedance coefficient is calculated 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;

[0049] The state space of the reinforcement learning agent is constructed, the global interaction impedance matrix is mapped to the graph structure data, and the task flow data is fused to generate a comprehensive state vector containing network topology features and load features; ​

[0050] 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;

[0051] 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.

[0052] The task flow data includes node queue accumulation depth, task average residence time delay, flow transfer rollback rate, and task blocking state bit;

[0053] 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;

[0054] The task average residence time delay is calculated based on the timestamp log of the task object to obtain the average residence time of the task in the node, and quantifies the processing throughput of the node;

[0055] The flow transfer rollback rate is obtained by counting the frequency of sending task instructions between adjacent nodes in a time window, and is mapped to a logical oscillation feature at the network level;

[0056] 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.

[0057] In this embodiment, the task flow data flowing in the distributed collaboration network is deeply structured and processed to construct a digital portrait reflecting the node's work efficiency and pressure. The specific data processing procedure includes the following four rigorous calculation steps:

[0058] First, regarding the calculation of the node queue accumulation depth, the system backend service scans the to-do task buffer of each processing node at a frequency of seconds; 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 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 number is divided by the maximum parallel processing capacity to obtain a normalized ratio between 0 and 1. When the ratio approaches 1, it indicates that the node is in a full-load state.

[0059] Secondly, regarding the acquisition of the average task residence delay, the timestamp of each task data packet entering the node and the timestamp marked as completed or transferred out are extracted by using log mining technology. In a preset time window (for example, the past one hour), the residence time of all task data packets passing through the node is calculated by arithmetic mean; the index can dynamically reflect the computing throughput efficiency of the node, and if the residence delay presents exponential rise, it usually means that processing blockage occurs inside the node.

[0060] Thirdly, regarding the monitoring of the transfer back-off 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 the task instruction is sent from node A to node B and then rejected or reassigned back to node A by node B in a very short time, such an event is marked as a logical shock; the proportion of such shock path in the total transfer path is calculated to quantify the shirking phenomenon between nodes or the incompatibility of the cooperation interface.

[0061] 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 the running state; once it is found that the residence time of a task in a single node exceeds the safety threshold set by the system, for example, 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.

[0062] The present application has significant technical advantages through multi-dimensional feature processing of task transfer data, solving the subjectivity problem of node load evaluation in traditional management; by introducing the node queue accumulation depth and normalization processing, it can cross different types of business nodes and compare the 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.

[0063] The introduction of the transfer back-off rate is one of the key innovative points of the present application; traditional load balancing algorithms often only pay attention to whether the task is sent out, but ignore whether the task is effectively received; high back-off rate often implies cognitive conflict or interface mismatch between nodes, and this index can make the system aware of the invalid flow consumption, thereby avoiding the waste of idle computing resources.

[0064] The task blocking state bit provides a deterministic risk warning mechanism; before the reinforcement learning model converges or reacts, this hard threshold determination can serve as a bottom protection mechanism to prevent deadlocked tasks from occupying node memory and computing slots indefinitely, significantly improving the robustness and safety of the system.

[0065] The interaction log data includes unstructured text information between nodes, input behavior feature data of a node operation terminal, and corresponding timestamp metadata; the input behavior feature data includes a key stroke rate, a mouse movement trajectory chaos degree, and a frequency of operation undo and redo.

[0066] Feature extraction is performed on the interaction log data by using an interaction semantic analysis model, and a stability index and a time delay jitter index between node pairs are calculated in combination with timestamp metadata, including:

[0067] Segmentation and vectorization embedding are performed on the unstructured text information, a negative word density, an imperative sentence frequency, and a use frequency of a high-intensity punctuation symbol contained in the unstructured text information are detected, and a semantic sharpness score of a single interaction is calculated;

[0068] The number of interaction rounds between node pairs in a preset period and the number of times a same task instruction is circulated between nodes are counted, and an interaction oscillation frequency is calculated.

[0069] The key stroke rate, the mouse movement trajectory chaos degree, and the frequency of operation undo and redo in the input behavior feature data are encoded into a behavior feature vector by using a long short-term memory network, and a behavior state score is recognized according to the behavior feature vector; the mouse movement trajectory chaos degree is defined as a variance of a curvature change rate of a mouse cursor coordinate point in a unit time.

[0070] Further, the process of obtaining the mouse movement trajectory chaos degree is as follows:

[0071] The coordinate sequence of a mouse cursor in office software operation is collected at a frequency of 100 Hz, and the sampling duration is not less than 1 second; for three continuous coordinate points, the curvature of an arc formed by the three points is calculated through vector cross multiplication and 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 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 preset maximum variance threshold is set, and if the threshold is exceeded, 1 is counted. The higher the score is, the more disordered the trajectory is, such as high-frequency shaking when anxious or repeated turning back when entangled, and the lower the score is, the more stable the trajectory is, such as straight-line movement when focused, and the implicit operation state of the cooperation subject is accurately quantified, thereby providing core data for behavior feature analysis.

[0072] Interaction duration data is obtained by calculating an average message reply waiting time based on timestamp metadata.

[0073] The stability index includes the semantic sharpness score and the interaction oscillation frequency; and the time delay jitter index includes the behavior state score and the interaction duration data.

[0074] The interactive semantic analysis model is used to extract accurate impedance indicators from a large amount of chaotic logs. The model adopts a parallel pipeline architecture and is divided into four independent data processing channels.

[0075] In the first branch, unstructured text information is processed. First, the text is segmented into basic word units using a word segmentation tool and mapped into a high-dimensional word vector. Then, a pre-set negative dictionary and sentiment dictionary are matched and 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 exclamation marks, question marks, and other high-intensity punctuation marks is counted. These statistical features are input into a logistic regression layer to output a semantic sharpness score for a single interaction. The higher the score, the stronger the antagonistic and conflict emotions contained in the text.

[0076] In the second branch, a statistical-based interactive oscillation frequency calculation is performed. A node pair interaction frequency table is maintained in memory. If there are more than dozens of message round trips between two nodes within a very short preset period, and the same task instruction is repeatedly transferred between them, the system determines that the link is in a high-frequency and inefficient oscillation state, and the oscillation frequency is normalized and output.

[0077] 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 sent to a long short-term memory network for encoding. The long short-term memory network captures the time dependence in operation behavior through a gating mechanism and can identify typical anxious behavior patterns such as rapid tapping and disordered sliding, outputting a behavior state score between 0 and 1.

[0078] In the fourth branch, the interaction duration data is calculated based on 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 response speed.

[0079] The interactive semantic analysis model adopts a four-branch parallel pipeline architecture, uniformly uses an Adam optimizer, and finally outputs four-dimensional core indicators of semantic sharpness score, interaction oscillation frequency, behavior state score and interaction duration data after 30000 steps of total iteration training. The first branch processes unstructured text, and after jieba word segmentation, Word2Vec word embedding and CNN feature extraction, outputs the semantic sharpness score in the interval [0, 1] through logistic regression, and fuses negative / sentiment dictionary matching and imperative sentence, high-intensity punctuation statistics; the second branch is based on a sliding window, and the number of interaction rounds and the number of task round trips between nodes are counted, and the interaction oscillation frequency is output after Min-Max normalization, without independent training; the third branch receives 3-dimensional time series data such as key interval and mouse chaos degree, and outputs the behavior state score through the LayerNorm layer and single-layer unidirectional LSTM coding, and the full connection 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 normalize it.

[0080] The 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 collaborative networks, many high-impedance antagonistic behaviors are implicit, for example, although nodes do not use dirty words or obvious rejection words in text, they express antagonism through deliberately delaying replies or violent keyboard operations with emotions. Breaking through the limitations of traditional management that 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 intervenes in implicit conflicts in advance.

[0081] The present scheme 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 the 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.

[0082] The specific process of generating a global interaction impedance matrix of the current time step based on the stability index and the delay jitter index includes:

[0083] The semantic sharpness score and the interaction oscillation frequency in the stability index, and the behavior state score and the interaction duration data in the delay jitter index are fused by a weighted normalization algorithm to obtain the interaction impedance coefficient of the node pair at the current time step, and the global interaction impedance matrix is composed of the interaction impedance coefficients of all node pairs.

[0084] First, a set of dynamic weight parameters are defined, corresponding to the four dimensions of semantics, oscillation, behavior and timeliness respectively. These weights are not fixed, but can be pre-configured according to business scenarios.

[0085] The weighted normalization algorithm is adopted to perform weighted summation on the semantic sharpness score output by the first branch and the interaction oscillation frequency output by the second branch to generate a stability index representing explicit conflict, wherein the weights are both defaulted to 0.5; at the same time, 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 defaulted to 0.5.

[0086] 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.

[0087] The calculation formula of the interaction impedance coefficient is:

[0088] ;

[0089] ;

[0090] ;

[0091] 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 the 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, defaulted to 0.5, and and are 1;

[0092] is the stability index; is the semantic sharpness score, the text conflict probability extracted by NLP, and the normalized value; is the interaction oscillation frequency, the normalized value of the number of task state changes per unit time; , is the weighted coefficient of the stability index, both defaulted to 0.5, and the sum is 1;

[0093] is the time delay jitter index; Score the behavior state, score the anxiety of the input device operation behavior based on the LSTM, and normalize the data; For the interaction duration data, the jitter deviation value of the message response time delay relative to the historical benchmark, the normalized data; 、 The weighting coefficient of the time delay jitter index is 0.5 by default, and the sum is 1.

[0094] Finally, all node pairs in the network are traversed, the impedance coefficient is calculated, 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.

[0095] In the present application, the interaction impedance coefficient is analogous to the impedance in the circuit system, i.e. the 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;

[0096] Saturation characteristic: when the explicit conflict and implicit friction reach a critical value, the growth rate of the coefficient slows down significantly, i.e. in actual collaboration, the degree of hindrance decreases when the conflict escalates to a certain extent.

[0097] 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.

[0098] 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 this relationship 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 is sensitive to extreme outliers while being robust to low-intensity noise. This means that a high impedance will not be misjudged 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 the management interference caused by false positives. By converting complex organizational relationship networks into standardized mathematical matrices, we achieve relationship visualization and conflict quantification, enabling management decisions to accurately locate high-friction collaboration links and avoid blind regulation, thereby improving management targeting.

[0099] The state space of the reinforcement learning 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 specific process includes:

[0100] The construction node is a graph structure data with vertices and edges of interactive relationship, and the global interactive impedance matrix is mapped as dynamic weight attributes of the edges in the graph structure data;

[0101] The graph structure data is input into a graph convolutional neural network layer, the impedance weight features of the first-order neighbors and the second-order neighbors of the target node are aggregated by using a convolution kernel, and an enhanced graph embedding vector representing a network topology risk propagation situation is output.

[0102] The node queue accumulation depth, the task average residence time delay, the flow transfer rollback rate and the task blocking state bit in the task flow transfer data are subjected to numerical standardization processing to generate a node load feature vector of the node.

[0103] The enhanced graph embedding vector and the node load feature vector are subjected to feature splicing and fusion to generate a multi-dimensional comprehensive state vector.

[0104] On the topology-aware route, the global interactive impedance matrix is regarded as a weighted adjacency matrix of a graph, and a graph structure data with processing nodes as vertices and interactive relationships as edges is constructed. The graph structure is input into a graph convolutional neural network layer. The graph convolutional network uses the spectral graph convolution principle to define a convolution kernel, which slides on the graph to aggregate the impedance weights of each node, its first-order neighbors (directly connected nodes) and second-order neighbors (indirectly connected nodes). The activation function uses ReLU, and a BatchNorm layer is added before output. The optimizer is an Adam optimizer, and the training hyperparameters are: learning rate 0.0012, Batch Size 64, Dropout rate 0.2, 35,000 iterations, and cross-entropy loss function. After multiple convolution operations, a high-dimensional enhanced graph embedding vector is output. This vector not only contains the impedance information of the node itself, but also implicitly contains the potential trend of impedance propagation in the network.

[0105] On the load-aware route, the obtained node queue accumulation depth and other four task flow transfer data are subjected to numerical standardization, such as Z-score standardization, to generate a node load feature vector.

[0106] Specifically, the node queue accumulation depth, the task average residence time delay and the flow transfer rollback rate are processed by the Z-score standardization algorithm to convert the feature values into standardized numerical values with a mean of 0 and a variance of 1.

[0107] The task blocking state bit (Boolean signal) is subjected to numerical mapping: the inactivated state is mapped to the numerical value 0, indicating no potential deadlock point; the activated state is mapped to the numerical value 1, indicating the existence of a potential deadlock point; since the mapping is in the 0-1 interval, the binary numerical feature is kept unchanged to retain the strong signal characteristics of the deadlock warning.

[0108] Finally, the feature splicing and fusion operation is performed; the enhanced graph embedding vector and the node load feature vector are concatenated in the channel dimension to form a multi-dimensional comprehensive state vector. The vector contains not only the topological features of where the relationship network is blocked, but also the load features of who has too much work, forming a complete context benchmark for the reinforcement learning model to reason.

[0109] The advanced graph neural network technology is adopted in the application to give the system the ability to predict risk propagation. Traditional management methods can only see the current conflict, i.e. point thinking, while the present scheme can enable the agent to perceive the risk transmission by aggregating the features of first-order and second-order neighbors. For example, when node A and node B have a high impedance interaction, although node C is currently in a normal state, since C is closely connected to B, the graph convolution network will calculate the potential risk faced by C, which is reflected in the comprehensive state vector. At the same time, the topological features and load features are spliced and fused to solve the problem of relationship and task fragmentation. The decision made by the agent based on the vector will neither simply relieve the relationship to cause task backlog, nor ignore the relationship deterioration to speed up the progress, thus finding the best balance point between multiple objectives.

[0110] The interaction strategy learning model includes an actor network and a critic network:

[0111] 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 the future cumulative reward value according to the current state and the action output by the actor network;

[0112] The design of the reward function value 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, and the negative penalty term is negatively correlated with the average norm of the interaction impedance matrix in the entire network and the duration of the high impedance link, so as to guide the model to actively reduce semantic friction in the network while pursuing task processing speed.

[0113] The embodiment adopts the actor-critic algorithm architecture in deep reinforcement learning to construct the interaction strategy learning model.

[0114] The actor network is a multi-layer perceptron that receives the comprehensive state vector as input, and finally outputs a deterministic action strategy distribution after weighted processing by layer after layer of neurons. The strategy determines which nodes should be applied with what degree of regulation parameters.

[0115] The critic network plays the role of a referee, which also receives the comprehensive state vector, but its output is a value scalar, which is used to predict the expected cumulative reward value that can be obtained in the future after executing the strategy of the actor network.

[0116] The actor network adopts an MLP (Multi-Layer Perceptron) architecture, the input layer receives a comprehensive state vector, the number of neurons is consistent with the input dimension; 2 fully connected hidden layers are set, the number of neurons is 256 and 128 respectively, the activation function is ReLU, and the BatchNorm layer is added to accelerate convergence. The number of neurons in the output layer matches the dimension of the regulation action parameter, and is mapped to the [0, 1] interval through the Sigmoid activation. The hyperparameter configuration is: learning rate 0.0005, Adam optimizer, weight decay 1e-5, Batch Size=32, hidden layer Dropout rate 0.1, iterative training 50000 steps, and the deterministic policy output requirement of the collaboration scene is adapted.

[0117] The critic network is also based on MLP, the input layer is consistent with the actor network, and receives a comprehensive state vector; the hidden layer structure is the same, 2 fully connected, 256 and 128 neurons respectively; ReLU activation and BatchNorm layer, the output layer has only one neuron, which outputs a value scalar through Linear activation. In terms of hyperparameters, the learning rate is 0.001, the same Adam optimizer configuration is used, Batch Size=32, and the Dropout rate is 0.1. The target network is updated every 100 steps, the output value scalar range is [-10, 10], which is used to accurately predict the expected value of cumulative rewards after executing the policy in the current state, and to provide stable evaluation feedback for the actor network.

[0118] During the model training and updating process, a composite reward function is designed; the function includes 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 in the entire network, that is, the throughput rate; the negative penalty term is negatively related to the average norm of the global interaction impedance matrix, that is, the overall friction level, and the duration of the high impedance link. Every time a time window passes, the actual reward obtained is calculated, and the prediction error of the critic network is calculated, and the error is used to update the weight parameters of the actor network and the critic network through the backpropagation algorithm.

[0119] The formula of the reward function is:

[0120] ;

[0121] Wherein, is the reward function value; is the number of tasks successfully completed in the entire network in the current time window, that is, the positive gain core; is the preset time window length; is the Frobenius norm of the global interaction impedance matrix, which is used to punish the overall friction level of the entire network; is the global interaction impedance matrix; is a preset high impedance warning threshold, which can be set to 0.8 for example; is a preset high impedance warning threshold, which can be set to 0.8 for example; is the normalized data of the duration of the high impedance state between nodes i and j; is the task output gain coefficient; is the network impedance penalty coefficient; is the malignant link duration penalty coefficient.

[0122] The determination process of the task output gain coefficient, the network impedance penalty coefficient, and the malignant link duration penalty coefficient is summarized as follows:

[0123] The first stage: based on the magnitude normalization purpose of statistics, eliminate the huge difference in numerical size of the three indexes, task quantity, impedance norm, and duration, to prevent the item with larger value from dominating the gradient descent direction. Collect historical interaction logs, calculate the standard deviation of the above three indexes in statistics. Adopt the principle of inverse weighting for initialization, set each coefficient to the inverse of the corresponding index standard deviation, that is, the greater the index fluctuation, the smaller the coefficient weight, to ensure that the initial contributions of the three to the reward function are in the same order of magnitude.

[0124] Specifically, the core of this process is to eliminate the magnitude difference of the three indexes (task quantity, impedance norm, and duration) to ensure balanced contribution to the reward function. First, collect the organization collaboration historical interaction logs in the past 30 days, filter the effective data, cover normal collaboration, high load, and high conflict scenarios, and extract the original data of the three indexes; then calculate the statistical standard deviation of each index, quantify the data fluctuation degree through the formula, and obtain the task quantity standard deviation σ1, the impedance norm standard deviation σ2, and the duration standard deviation σ3. Finally, initialize the coefficients according to the principle of inverse weighting: the task output gain coefficient w1 = 1 / σ1, the network impedance penalty coefficient w2 = 1 / σ2, and the malignant link duration penalty coefficient w3 = 1 / σ3; this way makes the larger the index fluctuation, the smaller the corresponding coefficient weight, and finally makes the initial contributions of the three indexes to the reward function in the same order of magnitude, avoiding the dominance of a single index in the gradient descent direction.

[0125] The second stage: based on the simulation optimization purpose of grid search, find the best balance point between the two conflicting goals of pursuing efficiency and reducing friction, which is the Nash equilibrium point; operation: construct a three-dimensional parameter search grid centered on the normalized values, and run multiple parameter combinations in parallel in a virtual simulation environment. By drawing the Pareto frontier curve, select the parameter combination that achieves the highest task throughput rate under the premise that the network impedance is no longer significantly increased as the benchmark fixed value for model deployment.

[0126] Specifically, the simulation optimization process aims to find a balanced parameter combination of efficiency and low friction, as follows: the search range is based on the normalized initial coefficient, covering three types of strategy scenarios: efficiency-oriented, low-friction-oriented, and balanced. The search step is set to 0.1 times the initial coefficient, and the corresponding candidate values are generated for each coefficient dimension. The three-dimensional grid forms multiple parameter combinations, taking into account optimization accuracy and computational efficiency. The stopping condition is met when any of the following conditions is met: 1) all parameter combinations have completed virtual simulation testing; 2) performance converges, and the optimal solution on the Pareto frontier does not change for 10 consecutive simulation rounds, i.e., the network impedance does not significantly increase, and the task throughput is the highest, with a reward function value fluctuation of ≤1%. The optimal parameter combination is output after stopping, serving as the baseline fixed value for model deployment.

[0127] The third stage: adaptive dynamic adjustment based on system load. According to the real-time urgency, dynamically adjust the focus of the strategy. Operation: Real-time monitoring of node queue depth, in high load mode, automatically adjust the task output gain coefficient, guide the agent to tolerate slight friction, and prioritize system stability; in low load mode, automatically adjust the network impedance penalty coefficient, guide the agent to perform strict traffic shaping, and deeply manage potential stubborn high impedance links.

[0128] This embodiment uses the actor-critic architecture to achieve adaptive evolution of management strategies. Unlike traditional rule-based systems, the model of the invention does not require pre-written if-then rules. Through the guidance of the 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 appropriately during system load peaks to ensure speed, while aggressively regulating impedance during load valleys to optimize network structure; enabling the distributed collaboration network to have self-regulating and evolutionary capabilities, capable of responding to changing external demands and internal environments.

[0129] The output dynamic regulation action instruction includes traffic shaping operations for interactive links:

[0130] The traffic shaping operation includes setting the cooling time threshold and frequency limit for message delivery; when the agent identifies that there is an overheated interaction impedance between the node pair, the dynamic regulation action instruction is output to lock the instant messaging input interface of both parties and delay the push time of non-emergency messages.

[0131] This embodiment describes how the agent's output decision instructions are executed at the physical level.

[0132] When the reinforcement learning model outputs traffic shaping instructions for a specific node pair, such as node A and node B, the control interface layer resets the parameters of the underlying instant messaging service and task distribution engine.

[0133] Specific operations include: first, set the cooling time threshold.

[0134] Inserting delay middleware in the message sending queue of node A, and forcibly introducing minute-level delivery delay for non-high-priority text messages.

[0135] Second, frequency limit and interface lock. When detecting extremely overheated interaction impedance, instructions are sent to the front-end client to temporarily gray out the chat input box of node A to node B, and a prompt in the system flow control is displayed, forcing both parties to enter a physical silent period. This locked state will continue until the impedance monitoring value of the next time window falls within the safe interval.

[0136] The traffic shaping operation provided by the application can quickly block malignant interaction cycles. In organizational collaboration, antagonistic emotions often spiral upward with immediate refutation and quarrel, leading to complete stop of effective work by both parties. The present application cuts off the immediate feedback loop of negative emotions by forcibly inserting a silent period, giving nodes, i.e. collaboration workers, time to cool down. This non-intrusive but mandatory intervention method can prevent local high impedance oscillation from evolving into a full network paralysis at minimal system cost, effectively protecting the stability of the collaboration network and ensuring the smooth flow of core task channels. By non-intrusive intervention to block malignant interaction cycles, the lag of traditional management in resolving conflicts is avoided, and communication conflicts are resolved at minimal cost.

[0137] Further, on the basis of the traffic shaping operation, a routing reconstruction operation for task distribution is also included; the routing reconstruction operation includes dynamically adjusting the connection weight between nodes, when the interaction impedance coefficient of a specific link is identified to exceed the threshold, the agent outputs an instruction to reduce the task allocation weight of the link, and searches for a third-party node with a lower impedance coefficient in the network as a relay platform, to construct a detour routing path, thereby isolating the high impedance node pair at the logical topology level.

[0138] Embodiment 2:

[0139] The application provides an organization relationship differentiated management method based on reinforcement learning, comprising:

[0140] Based on the preset time window, task flow data and interaction log data between nodes in the distributed collaboration network are collected;

[0141] The interaction semantic analysis model is used for feature extraction of the interaction log data, the stability index and the time delay jitter index between node pairs are calculated in combination with the timestamp metadata, the interaction impedance coefficient is calculated 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;

[0142] A state space of a reinforcement learning agent is constructed, a 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;

[0143] The comprehensive state vector is input into an interaction policy learning model for inference, and a dynamic regulation action instruction for the node is output through the policy network;

[0144] 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.

[0145] The node refers to the collaboration subject in the organization, such as people, departments, teams, and computer processing units that support their work, such as terminals, servers, and business system modules.

[0146] The task in the task flow data refers to the work task in the organization, such as project division, approval process, and business demand docking. Such tasks record the flow trajectory and store related data, such as to-do lists and processing time, through computer systems.

[0147] For example, the method described in the present application is adapted to enterprise WeChat, Dingding and other office software that can realize instant communication and collaboration. The nodes of the distributed collaboration network are set as follows:

[0148] The department node, such as the product department and the technology development department, is a non-leaf node with independent permission configuration, such as allowing only internal members to view department tasks.

[0149] Member node: all employees, through authorized login accounts, bind the member node.

[0150] Instant messaging: single chat and department group chat of office software, member nodes send text and files through accounts to form unstructured text information and operation behavior features required for patents;

[0151] Distributed collaboration: including document collaboration, table collaboration, etc., which can realize the collaboration of different member nodes.

[0152] The task flow data includes node queue accumulation depth, task average residence time delay, flow back-off rate, and task blocking state bit.

[0153] 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.

[0154] The task average residence time delay is calculated based on the timestamp log of the task object, the average residence time of the task in the node is obtained, and the processing throughput of the node is quantified;

[0155] The flow back-off rate is obtained by counting the frequency of sending the task instruction back and forth between adjacent nodes in a time window, and is mapped to the logical oscillation feature at the network level;

[0156] 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.

[0157] 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 keystroke rate, mouse movement trajectory confusion degree and operation undo and redo frequency.

[0158] The interaction semantic analysis model extracts and identifies stability indicators and time delay jitter indicators, including:

[0159] 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 a single interaction is calculated;

[0160] The number of interaction rounds between nodes in a preset period and the number of times the same task instruction is transferred back and forth between nodes are counted, and the interaction oscillation frequency is calculated; According to the keystroke rate, mouse movement trajectory confusion degree and operation undo and redo frequency in the input behavior feature data, the behavior feature vector is encoded by a long short-term memory network, and the behavior state score is identified according to the behavior feature vector; The message reply average waiting time calculated based on the timestamp is obtained, and the interaction time length data is obtained;

[0161] Specifically, the embodiment focuses on two specific processing nodes in a distributed collaborative network, node A is an upstream business demand publishing unit, and node B is a downstream technical development execution unit; When high impedance interaction occurs between the two, the micro operation logic of the method is described.

[0162] Scenario background: In a high-priority project sprint cycle, node A is responsible for issuing specific parameter configuration tasks to node B.

[0163] Step 1: Real-time capture of abnormal data;

[0164] Interaction log capture: It is monitored that node A sends five text messages to node B within two minutes;

[0165] Input behavior capture: At the same time, through the monitoring of the buried point data, it is found that the operation terminal of node B is abnormal: when replying to the message, the click frequency of the backspace key reaches 40 times per minute, far exceeding the average value, and the mouse trajectory presents high-frequency, small-amplitude disordered circular motion.

[0166] Step two: parallel calculation of multi-modal features;

[0167] Semantic branch: The interactive semantic analysis model identifies "not" "immediately" "wrong" and other high-weight negative and imperative words in the text message, and calculates the semantic sharpness score in combination with the high-frequency sending behavior in a short period of time.

[0168] Interaction oscillation frequency: Statistics show that the state involving the same task instruction has changed between pending confirmation and rejection three times. By comparing this frequency with the historical benchmark, the interaction oscillation frequency is calculated by normalization.

[0169] Behavior branch: The long short-term memory network processes the key sequence of node B and identifies the "input-delete-reinput" entanglement mode, and calculates the behavior state score.

[0170] Interaction duration data: Calculate the average response time of node B to the last five messages. The historical benchmark shows that the normal response time of B is three minutes, combined with the response time shown by the current monitoring data, the interaction duration data is quantified.

[0171] Resistance generation: Fusion of the above indicators, the interaction resistance coefficient between node A and node B is calculated. At this time, the element corresponding to the position in the global interaction resistance matrix turns red.

[0172] Step three: reasoning and decision-making of reinforcement learning agent;

[0173] State input: The agent reads the high resistance between A and B, and at the same time, combined with the task flow data, it is found that there are still three high-priority tasks in the task queue of node B waiting for processing, and these tasks depend on the third-party node C.

[0174] Risk prediction: Enhanced graph embedding vector shows that if node B continues to be entangled by node A, it will cause a cascade waiting of node C connected to it, and further cause a local network deadlock.

[0175] Strategy output: The agent outputs a composite action instruction:

[0176] Traffic shaping: Apply message cooling strategy to node A, set the delivery delay of its non-urgent message to five minutes.

[0177] Routing rearrangement: The new task sent by node A is temporarily suspended in the to-do queue of node B, the priority is reduced, and the task of node C is promoted to the top priority for execution.

[0178] Step four: execute and feedback;

[0179] Effect: Because the message is delayed, node A cannot interrupt node B by continuously urging, and node B uses this period of silence to concentrate on completing the task of node C.

[0180] State rollback: After five minutes, it is monitored that the keyboard typing rate of node B returns to normal, and it is determined that this regulation is effective, and the agent is given a positive reward.

[0181] In order to verify the objective technical effect of the method, comparative tests were carried out in a cooperative network environment containing one hundred processing nodes and five thousand daily processing tasks.

[0182] Control group: Adopt the traditional first-in-first-out task allocation strategy, and do not have the interactive impedance sensing function. The communication between nodes is not intervened;

[0183] Experimental group: Deploy an interactive impedance dynamic adjustment system based on reinforcement learning, and enable traffic shaping and route reconstruction functions.

[0184] The test period is seven consecutive working days, and the index data is shown in Table 1.

[0185] Table 1 Comparison of experimental index data

[0186]

[0187] From the table data, it can be seen that the average task flow cycle of the experimental group is shortened by nearly thirty percent; this is because the invention actively avoids congested nodes by route rearrangement when detecting an increase in interactive impedance, so that tasks can flow through a smoother path.

[0188] Friction is greatly reduced: the duration of high impedance link is greatly reduced, which is due to the traffic shaping mechanism, which effectively cools down at the beginning of the conflict, preventing the escalation of semantic confrontation.

[0189] Re-work rate is reduced: the significant reduction of task flow back rate, by pausing the dispatch of complex tasks to nodes in high behavior state score, avoids low-quality delivery and subsequent rework.

[0190] In summary, the invention realizes a qualitative leap in organizational collaboration efficiency at the macro level through millisecond-level perception and regulation at the micro level, effectively solving the problem of network congestion and resource waste caused by interpersonal friction in the traditional static management mode.

[0191] It should be understood that the embodiments disclosed herein and the foregoing description thereof are merely exemplary in nature and, thus, that various changes in the details thereof can be implemented by those skilled in the art without departing from the spirit and scope of the present application. Such changes are intended to fall within the scope of the present application as defined by the appended claims.

Claims

1. A method for differentiating management of organizational relationships based on reinforcement learning, characterized by, The method comprises the following steps: Based on the preset time window, collect the task flow data and interaction log data between nodes in the distributed collaboration network; Use the interaction semantic analysis model to extract features from the interaction log data, calculate the stability index and time delay jitter index between node pairs combined with timestamp metadata; calculate the interaction impedance coefficient based on the stability index and time delay jitter index; generate the global interaction impedance matrix of the current time step based on the interaction impedance coefficient; The stability index and time delay jitter index between node pairs include: Carry out word segmentation and vector embedding on unstructured text information, detect the negative vocabulary density, imperative sentence frequency and high-intensity punctuation symbol usage frequency contained therein, and calculate the semantic sharpness score of single interaction; Statistical 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, and calculate the interaction oscillation frequency; Encode the key stroke rate, mouse movement track chaos degree and operation undo and redo frequency in the input behavior feature data into a behavior feature vector through a long short-term memory network, and identify the behavior state score according to the behavior feature vector; Calculate the average waiting time of message reply 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; The specific process of calculating the interaction impedance coefficient based on the stability index and time delay jitter index and generating the global interaction impedance matrix of the current time step based on the interaction impedance coefficient includes: through a weighted normalization algorithm, the semantic sharpness score and interaction oscillation frequency in the stability index, and the behavior state score and interaction time length data in the time delay jitter index are fused to obtain the interaction impedance coefficient of the node pair at the current time step, and the global interaction impedance matrix is composed of the interaction impedance coefficients of all node pairs; Map the global interaction impedance matrix to graph structure data and fuse the task flow data to generate a comprehensive state vector containing network topology features and load features by constructing a state space of a reinforcement learning agent; Input the comprehensive state vector into the interaction policy learning model for reasoning, and output the dynamic regulation action instruction for the node through the policy network; Execute the dynamic regulation action instruction to regulate the distributed collaboration network, and detect the change of the global interaction impedance matrix in the next time window, calculate the reward function value to update the parameters of the interaction policy learning model in reverse.

2. The method of claim 1, wherein the method is based on reinforcement learning. 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, which is 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 to obtain the average residence time of the task in the node, which quantifies the processing throughput of the node; The flow back rate is obtained by counting the frequency of task instructions sending back and forth in a cycle path between adjacent nodes within a time window, and is mapped to the logical shock feature at the network level. The task blocking state bit is a Boolean signal activated when the task stays for a length of time exceeding a preset threshold, indicating a potential local network deadlock point.

3. The method of claim 1, wherein the method is based on reinforcement learning. 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 trajectory chaos degree, and operation undo and redo frequency.

4. The method of claim 1, wherein the method is based on reinforcement learning. The state space of the reinforcement learning 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 specific process includes: A graph structure data is constructed with nodes as vertices and interaction relationships as edges, and the global interaction impedance matrix is mapped to the dynamic weight attribute of the edges in the graph structure data; The graph structure data is input into a graph convolutional neural network layer, which aggregates the impedance weight features of the first-order neighbors and second-order neighbors of the target node using convolution kernels, and outputs an enhanced graph embedding vector representing the network topology risk propagation situation; The node queue accumulation depth, task average residence time delay, flow back rate, and task blocking state bit in the task flow data are numerically standardized to generate a node load feature vector of the node; The enhanced graph embedding vector and the node load feature vector are concatenated and fused to generate a multi-dimensional comprehensive state vector.

5. The method of claim 1, wherein the method is based on reinforcement learning. 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 based on the input comprehensive state vector, and the critic network is responsible for predicting the future cumulative reward value based on the current state and the 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, and the negative penalty term is negatively correlated with the average norm of the global interaction impedance matrix and the duration of high impedance links, so as to guide the model to pursue task processing speed while actively reducing semantic friction in the network.

6. The method of claim 1, wherein the method is based on reinforcement learning. The output dynamic regulation and control action instruction includes a traffic shaping operation for the interaction link: The traffic shaping operation includes setting the cooling time threshold and frequency limit of message delivery; when the agent identifies that there is an overheated interaction impedance between node pairs, the dynamic regulation and control action instruction is output to lock the instant messaging input interfaces of both parties and delay the push time of non-urgent messages.

Images