A Multi-Agent Network Traffic Prediction Method and System Based on Failure Gating and Shared Cost Ratio
By using a multi-agent network traffic prediction method based on failure gating and shared cost ratio, this paper addresses issues such as insufficient binding of agent reasoning to the network environment and improper knowledge failure management. It achieves credibility and physical rationality in network traffic prediction and supports the accumulation and cross-scenario transfer of predicted knowledge.
Patent Information
- Application Number
- CN202611500620.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing network traffic prediction technologies suffer from problems such as insufficient binding between agent reasoning and network environment grounding, improper knowledge expiration management, uncontrollable sharing among multiple agents, insufficient data sufficiency assessment, and unstable prediction scheduling linkage, resulting in prediction results lacking verifiable evidence and exhibiting systematic bias.
A multi-agent network traffic prediction method based on failure gating and shared cost ratio is adopted. By collecting heterogeneous data from multiple sources in the network, a temporal knowledge state graph is constructed. Grounding encoding and multi-agent collaborative reasoning are performed to calculate the shared cost ratio and failure gating attention. This enables knowledge failure management and data sufficiency assessment, eliminates common errors, constructs a dual-loop stability criterion, and outputs the prediction validity period.
It achieves grounding binding between agent reasoning and network environment, effectively manages knowledge failure, enables multiple agents to share controllability, ensures the credibility and physical rationality of prediction results, supports the accumulation and cross-scenario transfer of prediction knowledge, and alleviates the challenges of traffic prediction in complex network environments.
Smart Images

Figure CN123123191A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a multi-agent network traffic prediction method and system based on failure gating and shared cost ratio, belonging to the fields of artificial intelligence multi-agent technology, knowledge graph technology, time series deep learning and intelligent operation and maintenance technology of communication networks. Background Technology
[0002] Network traffic prediction is a fundamental upstream capability in the intelligent operation and maintenance chain of communication networks, and its output directly affects the rationality of network resource scheduling strategies. With the large-scale deployment of 5G-A, cloud computing, edge computing, and the Industrial Internet, the scale of network access devices is rapidly expanding, and business models are becoming increasingly complex. Network traffic exhibits characteristics of periodicity, burstiness, non-stationarity, and strong coupling of multiple factors. Currently, mainstream traffic prediction technologies include statistical models, traditional machine learning, deep learning time series models, and large-scale intelligent agents. The first three rely on the assumption of time series stationarity, manual feature engineering, and numerical mapping learning, respectively, and all have shortcomings in network semantics, causal logic, and state transition modeling.
[0003] Large model agents have been gradually applied to network traffic prediction. Related solutions include: using time-series knowledge graphs based on frequency statistics to record only the time when facts occur and estimating state transition probabilities based on historical frequency; using multi-agent ensemble methods to treat error correlation as an exogenous parameter and weighting it through conventional attention mechanisms; and combining prediction-decision fixed point solving, sub-model retrieval and graph heat spread methods to obtain prediction results.
[0004] However, the above-mentioned solutions still have the following shortcomings: 1. Overall, existing solutions suffer from insufficient environmental grounding. There is a lack of structured association between agent inference and the data, state, and feedback of the real network, resulting in a lack of verifiable evidence for the prediction results. 2. There is a lack of assessment of data sufficiency. The system's own data collection behavior introduces additional traffic, causing a systematic shift in the observation results; the study of fixed points for prediction decisions also does not consider this observation interference. 3. Its temporal knowledge graph only records the moment when the fact occurs, and cannot characterize the effective lifespan of the fact. The state transition probability only reflects historical frequency and cannot determine whether the knowledge supporting the state transition is currently valid. When traditional survival analysis models characterize knowledge lifetime, their covariates are exogenous fixed attributes, which cannot reflect the rewriting effect of agent actions on knowledge lifetime. Even with the superimposed bias, the weights of the attention mechanism are difficult to strictly set to zero, making it difficult to completely eliminate invalid evidence. 4. Multi-agent integration usually uses error correlation as an exogenous parameter, lacking an executable criterion for controlling the start and stop of information sharing. As the number of agents increases, it is easy to cause error convergence and a decrease in prediction performance. 5. Knowledge retrieval often employs simple semantic ranking or general submodal retrieval, failing to consider scenarios of shared knowledge failure and ignoring the risks of knowledge redundancy and shared failure. 6. The lack of a dual-loop joint stability determination between prediction and scheduling decisions makes it prone to mutual chasing between prediction and scheduling after the scheduling changes the traffic distribution. 7. Graph heat diffusion methods set diffusion time as a hyperparameter, failing to output a prediction validity period with a clear physical meaning, and making it difficult to proactively downgrade and abandon predictions when they fail. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a multi-agent network traffic prediction method based on failure gating and sharing cost ratio that can realize the binding of agent reasoning with network environment grounding, knowledge failure management, data sufficiency assessment, multi-agent sharing and controllability, fact screening and credibility verification and prediction scheduling linkage, has the ability to output the prediction validity period and actively degrade, and supports prediction knowledge accumulation, abstraction and cross-scenario transfer.
[0006] The present invention adopts the following technical solution: This invention employs the following steps in its multi-agent network traffic prediction method based on failure gating and shared cost ratio: S1. Collect multi-source heterogeneous data from the network to eliminate the rise in acquisition observation; assess data sufficiency to support degraded output; calculate the amplification factor of the acquisition loop; and dynamically switch the acquisition strategy. S2. Construct a temporal knowledge state graph and initialize a hierarchical memory storage framework; configure a rewritable fact failure rate, calculate the failure gating transition probability and failure weight allocation, and implement the shielding of corresponding state transitions when all supporting facts fail. S3. Perform grounding encoding on the temporal knowledge state graph to generate the shared environment state and local state of multiple agents; construct observation resource constraints to solve for the optimal shared observation area length, and dynamically switch public and private observation configurations based on the cost coefficient; S4. Calculate the shared cost ratio, construct the failure gating factor to obtain the failure gating attention, generate the cooperative channel switch and dynamic cooperative weight, control the start and stop of the channel according to the shared cost ratio, eliminate common errors, and complete the credibility identifiable signal-to-noise ratio determination. S5. Construct a sub-modal filtering objective function to filter the fact set, verify the set survival constraints, calculate the fact suspicion degree to intercept abnormal facts, and complete the traffic conservation correction; S6. Iteratively solve for self-consistent conditions and construct training objectives; monitor and suppress dual-loop instability; solve the failure time spectrum equation to obtain node validity period, complete aggregate computing power selection and downgrade abandonment; calibrate module offline error characteristic curves to achieve module trial-free running; S7. Receive real feedback from network operation, filter structured modifications through counterfactual verification and write them back to the temporal knowledge state diagram and hierarchical memory; update memory entry failure rate and knowledge activity, complete knowledge prototype update and agent knowledge transfer, realize memory dormancy archiving and reactivation, monitor collaboration channels and reconfigure as needed.
[0007] The specific calculations for eliminating the elevation of the acquisition observation, calculating the data sufficiency, and calculating the amplification factor of the acquisition loop in method S1 of this invention are as follows: S101. Estimation and Subtraction of Acquisition Interference Coefficient: The data acquisition requests of this device are recorded as a sequence of acquisition request quantities. The acquisition request quantities are then averaged to obtain the averaged acquisition request quantities. A least-squares regression is performed between the averaged acquisition request quantities and the original observed flow rate to estimate the acquisition interference coefficient. The interference term is then subtracted from the original observed flow rate to obtain the observed flow rate after interference subtraction. The acquisition interference coefficient and the observed flow rate after interference subtraction are as follows: ; In the formula, To collect the number of requests without taking the mean, For the original observed flow rate, To observe the average flow rate, To collect interference coefficients; After removing interference; S102. Data Sufficiency Calculation: Calculate the Fisher information matrices of environmental parameters under the current acquisition strategy and the ideal acquisition strategy respectively. Take the ratio of the trace of the inverse Fisher information matrix of the ideal acquisition strategy to the trace of the inverse Fisher information matrix of the current acquisition strategy to obtain the observation gap. Then, take the reciprocal of the observation gap as the data sufficiency. ,in In the formula, This is the Fisher information matrix for the current data collection strategy. For an ideal data collection strategy, Represents the trace operation of a matrix; S103. Calculation of the amplification factor of the self-acquisition loop: Multiply the acquisition density sensitivity, the acquisition quantity sensitivity to network load, and the proportion of the prediction result's absorption of the observation bias to obtain the amplification factor of the self-acquisition loop. In the formula, To improve density sensitivity, To assess the sensitivity of the data collection volume to network load. The proportion of observation bias absorbed by the prediction results.
[0008] The calculation methods for the factual failure rate, state transition probability, and post-failure weight allocation in method S2 of this invention employ the following steps: S201. Factual Failure Rate Calculation: The factual failure rate is obtained by modulating the baseline risk rate and agent actions. The agent actions rewrite the factual risk rate as a multiplicative factor, i.e. In the formula, For historical information; Actions of the intelligent agent; For the event; The baseline risk rate for facts is the inherent failure rate of facts without intervention. For the facts The learnable coefficient vector; This is an action feature mapping function that maps the original action to a feature vector with modified failure rate. S202. Failure Gated State Transition Probability Calculation: Integrate the failure rate of all facts supporting the current state transition over the prediction period using time, take the negative exponent to obtain the joint survival probability of the fact set, and then multiply it by the transition tendency constituted by resource utilization and state change to obtain the state transition probability. ; In the formula, For state To state Predicting step size The failure gating state transition probability; The current moment; To predict the field of view (predicting the step size backward). It is an integral dummy variable; The set of facts that state transitions depend on; For a moment The failure rate of the following fact f; The normalization function outputs... Interval probability; These are the state transition probability weighting coefficients; Let t be the network resource utilization rate at time t; Let t be the rate of change of flow rate at time t; when all the facts supporting a certain state transition fail, the joint survival probability of the fact set is zero, the state transition probability is set to zero accordingly, and the target state is determined to be unreachable within the prediction period; S203. Redistribution of knowledge weights after failure: The proportion of the product of a fact's failure rate and its relevance among all surviving related facts of the same type is used as the competitive allocation ratio, as follows:
[0009] The knowledge weight of invalid facts is transferred to surviving related facts according to this proportion, avoiding the direct discarding of evidence information; where, The knowledge weight competition ratio allocated to surviving fact g after fact f becomes invalid; The degree of relevance between fact f and fact g; For candidate survival facts that are associated with f, it is a summation and traversal of indices; The failure rate is the factual failure source fact f; For candidate survival facts associated with f The fact of failure; For fact f and candidate survival facts The degree of correlation between them.
[0010] The calculation of grounding encoding, observation resource constraints, and optimal shared observation area length in method S3 of this invention specifically adopts the following steps: S301. Generation of Global Environment State and Local State of a Single Agent: The global environment state is obtained by concatenating and encoding four types of information: traffic observation, graph knowledge, network state, and hierarchical memory. Furthermore, it is divided into four components: observation, knowledge, state, and hierarchical memory, to obtain the local state of a single agent. ; in, The environment state vector is shared globally by multiple agents; For feature fusion function; Traffic characteristics; The knowledge representation encoded as a temporal knowledge state diagram; This is the network state vector; The hierarchical memory structure is represented by a vector. Let be the local state vector of the i-th agent; These are respectively agent-private observation, graph knowledge, network environment, and hierarchical memory components; S302. Observation Resource Constraints and Solution for Optimal Shared Observation Area: The total observation budget is divided into a shared observation area (shared by all agents and collected only once) and private observation areas (not overlapping between agents), constituting the observation resource constraints. Construct a lower bound for the multi-agent fusion error based on this constraint, and find the extreme value of the shared observation area length to obtain the optimal shared observation area. ;in, To share the length of the observation area; The length of a single agent's private observation zone; The total number of intelligent agents; For the total budget of observation; The optimal shared observation area length; This is a truncation function that restricts the range of output values. The length related to the flow space; The variance of individual errors that can be eliminated by shared observations; To share the observation cost coefficient; S303, Three-phase observation configuration switching and distribution: Based on the interval of the shared observation cost coefficient, switch and distribute the configuration between three observation modes: full private configuration, partial shared configuration, and full shared configuration, and divide the shared observation area and the private observation area of each agent accordingly.
[0011] The calculation of shared cost ratio, failure-gated attention, dynamic collaborative weight, and credibility-identifiable signal-to-noise ratio in method S4 of this invention specifically adopts the following steps: S401. Calculation of the sharing cost ratio: Using the sharing intensity as the independent variable, calculate the derivatives of the common error variance and the individual error variance with respect to the sharing intensity, and use the ratio of the two as the sharing cost ratio. ;in, For the cost-sharing ratio; the criterion is Close all collaboration channels; The overall sharing strength of the system characterizes the degree of information sharing among multiple agents; The common error variance introduced for sharing does not diminish with the number of agents; Individual error variance can be reduced through information sharing; S402. Failure-gated attention calculation: Calculate the inner product of the query vector and key vector for each agent and normalize it according to the dimension. Add the failure-gated factor in the logarithmic field to the attention score to obtain the failure-gated attention. Among them, failure gating factor In the formula, For attention weights in failure gating; hour, ; Let i be the query vector for the i-th agent; Let J be the key vector of the j-th agent; The Q / K vector dimension; As a failure gating factor, supporting The joint survival probability of the set of facts related to state transitions within the prediction window; S403, Cooperative Channel Switching and Dynamic Cooperative Weight Calculation: Indicative factors are constructed using the difference between the temporal state semantic relevance and the channel admission threshold, and the difference between the inverse of the number of agents and the sharing cost ratio. Cooperative channels are allowed to open only when both factors are positive, thus obtaining the cooperative channel switching mechanism. Applying the channel switch as a multiplicative factor to the normalized weights yields dynamic collaborative weights.
[0012] in, This is an indicator function; it returns 1 if the condition is true and 0 if the condition is false. For collaborative channel switching; The temporal state semantic correlation between agents i and j; This is the semantic relevance threshold; Dynamic collaborative weights; Let the credibility of the j-th agent be denoted as 'j'. S404. Credibility Identifiability Determination: Subtract the mean of all errors from the prediction errors of each agent to obtain the individual difference after deducting the common error. The signal-to-noise ratio can be identified by calculating the reliability of the proportion of the largest eigenvalue in the residual correlation matrix and the number of agents. ;in, The original prediction error of agent i; This refers to the pure individual residual error after deducting the global common error component; For credibility, the signal-to-noise ratio can be identified; when Credibility is unverifiable; credibility updates are frozen. This represents the proportion of the largest eigenvalue obtained from the eigenvalue decomposition of the residual correlation matrix.
[0013] The calculation of the sub-modulus screening objective function, set survival constraint, fact doubt degree, and flow conservation correction in method S5 of this invention specifically adopts the following steps: S501. Solving for the optimal fact set: Using the marginal information content of the conditions multiplied by the survival probability of the facts as the payoff term, and the sum of the failure rates of facts and the sum of the co-mortality correlation coefficients of facts as penalty terms, a monotonic submodular objective function is constructed. Maximizing this objective function, while ensuring the set size does not exceed a given upper limit, yields the optimal fact set.
[0014] In the formula, To filter the optimal set of facts in the output; To select an upper limit for the number of facts; Given the historical context and selected facts, the conditional marginal information content of fact f; Let t be the set of historical facts. This is the fact that f has already been selected into the set; Let f be the individual survival probability of a single fact f. The coefficient for the failure rate penalty term; The coefficient of the co-mortality related penalty term; Let f be the co-mortality correlation coefficient between facts f and g, which characterizes the degree of correlation between the two factual failure events; S502, Set-level Joint Survival Constraint Verification: Multiply the joint survival probability of the fact set during the prediction period by the co-mortality correlation correction term to obtain the set-level joint survival probability:
[0015] In the formula, This represents the aggregate-level joint survival probability after co-mortality correction. To accommodate the survival probability tolerance, the joint survival probability of the set must be greater than or equal to 1. ; S503, Fact Suspicion Calculation and Interception: Fact suspicion is constructed based on the degree of insufficient precursor support, the amount of violation of flow conservation, and the strength of fact generation. ; In the formula, The degree of suspicion of fact f; if it exceeds the threshold, the fact is blocked and prohibited from entering the fusion process; The preceding support level measures whether there is supporting evidence for the fact. This violates the principle of flow conservation. Strength is born from facts; It is the fact importance penalty coefficient, a preset constant hyperparameter used to adjust the exponential decay caused by fact importance; S504, Flow Conservation Correction: Constructing an orthogonal projection operator for the conservation subspace from the node-link correlation matrix. A closed-loop projection is performed on the multi-agent fusion result to obtain a corrected output that satisfies flow conservation. No learnable parameters are introduced; among them, For flow conservation orthogonal projection operators; It is the identity matrix; Network topology node-link correlation matrix; This is the pseudoinverse of the Moore-Penrose matrix; This is the original prediction vector without conservation correction; This is the final predicted output after projection correction and strict satisfaction of flow conservation constraints; The specific calculations for the training target, failure time spectrum equation, and module offline error characteristic curve calibration in method S6 of this invention are performed using the following steps: S601. Training Objective Construction: Construct training objectives with decision regret as the main objective and prediction error as the regularization term. This allows for higher prediction errors in some samples while improving decision performance; among them, For multi-agent training loss function, Regret for the decision For prediction error regularization, The regularization coefficient; S602, Dual-loop stability verification and suppression: Estimate the amplification factor of the acquisition loop and the decision reaction loop respectively, and determine whether the sum of the two is less than 1: If it is not less than 1, perform three types of suppression actions: reduce the degree of acquisition adaptation, increase the output change rate limit, and increase the aggregation granularity; if it is less than 1, maintain the current configuration. S603. Solving the Failure Time Spectrum Equation and Calculating the Predicted Effective Period: Taking the prediction horizon as the evolution time, the error diffusion process on the graph is expressed as a diffusion equation. Expanding this equation according to the graph's characteristic decomposition yields the failure time spectrum equation. The inverse solution yields the predicted validity period for each network node. The prediction validity period for backbone low-spectrum nodes is longer, while that for edge burst nodes is shorter; among them, For network node indexing; To solve for the output, nodes The predicted validity period; The error spread rate on the network graph; The Laplace matrix of the network topology graph is the first... One eigenvalue; For nodes Spectral energy on the k-th order eigenvector; Coverage factor; The standard deviation of the flow rate observation at the initial time; This is the business-permissible error tolerance, i.e., the maximum prediction error that the business side can tolerate; The common error variance; S604, Spatial Aggregation and Computing Power Tier Selection: Generate spatial aggregation clusters based on frequency domain feature similarity, and select spatial aggregation granularity and computing power tier under the condition of meeting time constraints; when all candidate computing power tiers do not meet the time constraints, output a downgrade flag and actively abandon the selection, and do not output point prediction; S605. Module Offline Error Characteristic Curve Calibration and Trial-Free Selection: Offline fitting of error characteristic curves for various candidate prediction modules at spectral position and prediction step size. After deployment, the expected error is calculated directly from this curve to complete module selection, without the need for trial runs at the target node; among which... For module At the node Predicting step size The mean square error at that point. For module constant term, Frequency band sensitivity coefficient, The error diffusion rate is the module's own value, and the three are inherent offline multiplexing parameters of the module.
[0016] The hierarchical memory-related iterative update in method S7 of this invention specifically adopts the following steps: S701. Dynamic Failure Rate Update for Memory Items: An exponential factor is constructed from the offset between the mean of recent decision regret and the mean of historical decision regret. This factor modulates the baseline failure rate of memory items to obtain the dynamic failure rate of memory items. Increased regret over decisions leads to accelerated memory loss; among them, The dynamic failure rate of memory entry r; The initial baseline failure rate for memory entry r; I regret my recent decisions; The average regret rate for long-term historical decisions; The sensitivity coefficient of memory impairment rate to decision regret bias; S702, Knowledge Activity Iteration: Knowledge activity is updated jointly by the forgetting curve decay term, the hit gain term, and the error penalty term.
[0017] in, Let i be the knowledge activity of the i-th memory sample; The time step interval; This serves as the forgetting time constant, controlling the natural rate of forgetting. This is the hit gain coefficient; the sample was successfully invoked, and its activity was increased. This represents the number of times the sample was successfully invoked at time t. This is the error penalty coefficient; when samples involved in prediction produce large errors, the activity decreases. This is the normalized error of the sample participating in the prediction output; S703. Knowledge Prototype Update: When the similarity between a new experience sample and an existing knowledge prototype is higher than the third threshold, the prototype center vector is finely adjusted towards the new sample by adjusting the step size according to the confidence level. The step size If the similarity between the new experience sample and all prototypes is below the third threshold, then a new knowledge prototype is created; where, Let the k-th knowledge prototype center vector be denoted as 'k'. For the feature vector of the newly arrived experience sample; The actual update step size is aggregated for the prototype; Base learning rate; The sample-prototype similarity function; The confidence level of the knowledge prototype is used; the higher the confidence level, the smaller the update step size. S704, Added Agent Knowledge Transfer: New agents read transferable skills from the skill memory layer and initialize strategies by weighting and combining them according to transferability scores. Skills with scores below a preset threshold are excluded from the combination to avoid negative transfer; among them, The initial predicted skill vector is obtained for initializing the new agent; This represents the k-th transferable skill vector in the skill memory layer. The score is based on transferability; skills with low scores are not included in the weighting to avoid negative transfer.
[0018] The counterfactual verification judgment logic in method S7 of this invention is as follows: only when the difference between the decision regret of the adoption state and the non-adoption state of the modification is lower than the set negative threshold, and the direction of the difference remains consistent in the set number of repeated trials, will the structured modification output by the reflective agent be adopted; monitor the marginal value of the collaborative channel, and when its performance index is lower than the negative threshold for multiple consecutive times, issue an alarm signal and reconfigure the observation range and collaborative channel.
[0019] The present invention relates to a system for predicting network traffic based on a multi-agent network traffic using failure gating and shared cost ratio. The system includes a multi-source data acquisition module, a temporal knowledge state graph construction module, a grounding state encoding module, a multi-agent collaborative reasoning module, a state constraint prediction module, and a feedback update module. The multi-source data acquisition module is used for multi-source data acquisition and acquisition interference deduction, and switches the acquisition strategy according to the data sufficiency and the amplification factor of the acquisition loop. The temporal knowledge state graph construction module is used to maintain the temporal knowledge state graph and manage the fact failure rate, state transition probability, and knowledge weight redistribution after failure. The grounding state encoding module is used to fuse graph, traffic, state and memory information to output the shared environment state of multiple agents; The multi-agent collaborative reasoning module is used for observation configuration calculation, failure gating attention calculation, and control of collaborative channel on / off based on shared cost ratio. The state constraint prediction module is used to solve the decision self-consistent fixed point, solve the failure time spectrum equation to obtain the prediction validity period, select the aggregation granularity and computing power level, and support active abandonment and degradation output. The feedback update module is used to calculate decision regret, perform counterfactual verification on the amount of modification output by the reflective agent, and update the graph and the three-layer hierarchical memory.
[0020] The positive effects of this invention are as follows: This invention employs a fact-based failure rate to achieve end-to-end control of knowledge failure, enabling the identification of unreachable states within the prediction period; it proposes a shared cost ratio criterion to achieve controllable start and stop of multi-agent collaborative channels, suppressing the error convergence problem caused by the increase in the number of agents; it achieves system acquisition interference elimination and data sufficiency assessment, avoiding the output of unfounded prediction results; it adopts a submodal fact screening with co-death correlation penalties to avoid knowledge redundancy and the risk of common failure; it constructs a flow conservation hard constraint to ensure the physical rationality of network prediction results; it establishes a dual-loop stability criterion to solve the problem of mutual pursuit between prediction and scheduling; it outputs the node prediction validity period through the failure time spectrum equation, allowing for the selection of computing power and aggregation granularity according to timeliness requirements, supporting proactive degradation and abandonment of prediction failures; it builds a three-layer hierarchical memory architecture to achieve experience accumulation, abstraction, and cross-agent knowledge transfer, supporting memory dormancy archiving and reactivation, mitigating cold start and negative migration interference for new agents.
[0021] This invention can be applied to intelligent network traffic prediction in scenarios such as 5G and 5G-A mobile communication networks, industrial internet, metropolitan IP networks, data center networks, edge computing cluster networks and operator core networks. It provides predictive input for intelligent operation and maintenance services such as network resource pre-allocation, congestion warning, fault prediction and network slice scheduling, and improves the reliability and engineering practicality of traffic prediction in complex network environments. Attached Figure Description
[0022] Appendix Figure 1 This is a schematic diagram of the overall implementation process of the present invention; Appendix Figure 2 This is a flowchart of the multi-source network data acquisition module of the present invention; Appendix Figure 3 This is a flowchart of the timing knowledge state graph construction module of the present invention; Appendix Figure 4 This is a flowchart of the grounding state encoding module of the present invention; Appendix Figure 5 This is a diagram of the multi-agent failure-gated collaborative reasoning structure of the present invention; Appendix Figure 6 This is a flowchart of the state constraint prediction module of the present invention; Appendix Figure 7 This is a flowchart of the feedback update and hierarchical memory module of the present invention; Appendix Figure 8This is a schematic diagram of the temporal knowledge state diagram structure of the present invention; Appendix Figure 9 This is a schematic diagram of the multi-agent failure-gated collaborative reasoning structure of the present invention; Appendix Figure 10 This is a schematic diagram of the layered memory structure of the present invention; Appendix Figure 11 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0023] As attached Figure 1 , 11 As shown, the system of the multi-agent network traffic prediction method for failure gating and sharing cost ratio of the present invention includes a multi-source data acquisition module, a temporal knowledge state graph construction module, a grounding state encoding module, a multi-agent collaborative reasoning module, a state constraint prediction module, and a feedback update module. The multi-source data acquisition module is used for multi-source data acquisition and acquisition interference deduction, and switches the acquisition strategy according to the data sufficiency and the amplification factor of the acquisition loop. The temporal knowledge state graph construction module is used to maintain the temporal knowledge state graph and manage the fact failure rate, state transition probability, and knowledge weight redistribution after failure. The grounding state encoding module is used to fuse graph, traffic, state and memory information to output the shared environment state of multiple agents; The multi-agent collaborative reasoning module is used for observation configuration calculation, failure gating attention calculation, and control of collaborative channel on / off based on shared cost ratio. The state constraint prediction module is used to solve the decision self-consistent fixed point, solve the failure time spectrum equation to obtain the prediction validity period, select the aggregation granularity and computing power level, and support active abandonment and degradation output. The feedback update module is used to calculate decision regret, perform counterfactual verification on the amount of modification output by the reflective agent, and update the graph and the three-layer hierarchical memory.
[0024] This invention discloses a multi-agent network traffic prediction method based on failure gating and sharing cost ratio, comprising the following steps: S1. Collect multi-source data and subtract observation self-interference: Collect multi-source heterogeneous data on network traffic, topology, device operating status, and events, and perform time synchronization and feature encoding; record the device's own data acquisition requests to eliminate the systematic increase in observation traffic caused by acquisition behavior; assess the data sufficiency under the current operating conditions and determine whether the data sufficiency is lower than the set threshold: if it is lower than the threshold, it is determined that the current observation is insufficient to support point prediction, and no point prediction result is output, only a data insufficiency mark is output and the output is degraded; if it is not lower than the threshold, it continues to calculate the amplification factor of its own acquisition loop and determines whether the amplification factor reaches the judgment threshold: if it reaches the threshold, it is determined that the positive feedback of "the more data is collected, the greater the observation deviation, and the more data needs to be collected" has been formed, and the adaptive acquisition density mode is switched to the fixed density acquisition mode or the random time acquisition mode; if it does not reach the threshold, the adaptive acquisition density mode is maintained. S2. Construct a temporal knowledge state graph and calculate failure-gated state transitions: Construct a temporal knowledge state graph and initialize a corresponding hierarchical memory storage framework; the hierarchical memory storage framework includes a context memory layer, a semantic memory layer, and a skill memory layer. During the framework initialization phase, only storage space is allocated, and historical working condition samples and knowledge entries are not filled in; configure a corresponding rewritable fact failure rate for each fact in the temporal knowledge state graph. This failure rate can be dynamically modified by the counterfactual verification mechanism in the subsequent feedback phase; calculate the state transition probability based on the fact failure rate; when a fact fails, the knowledge weight carried by that fact is assigned to the remaining surviving related facts; determine whether all the facts supporting a certain state transition have failed: if all have failed, set the state transition probability to zero, determine that the target state is unreachable within the prediction period, and exclude it from subsequent prediction reasoning; if not all have failed, continue to participate in reasoning according to the state transition probability after being gated by the survival probability; S3. Grounding Encoding and Observation Range Division: Grounding encoding is performed on the temporal knowledge state graph to generate an environment state representation that can be shared by multiple agents, and at the same time, a local state representation corresponding to each agent is generated; a multi-agent system including a traffic prediction agent, a state perception agent, a knowledge reasoning agent, a decision-making agent, and a reflection agent is built; a total constraint on observation resources is set, and the optimal shared observation zone length is obtained by solving the problem; the interval of the shared observation cost coefficient is determined: if the shared observation cost coefficient is higher than the first critical value, all private configurations are issued, and the shared observation zone length is zero; if the shared observation cost coefficient is lower than the second critical value, all shared configurations are issued, and the shared observation zone length is the total observation budget; if the shared observation cost coefficient is between the two critical values, some shared configurations are issued, and the shared observation zone and the private observation zones of each agent are divided according to the optimal shared observation zone length. S4. Shared Cost Ratio Determination and Failure-Gated Collaborative Reasoning: Calculate the shared cost ratio among agents, construct a failure gating factor based on the joint survival probability of the fact set, and generate failure-gated attention weights by combining the gating factor; generate collaborative channel switches by combining temporal state semantic relevance and shared cost ratio, and further obtain dynamic collaborative weights; determine whether the shared cost ratio reaches the reciprocal of the number of agents: if it reaches this reciprocal, it is determined that the marginal benefit of continuing to share is negative, all collaborative channels are closed, and each agent completes its prediction independently before merging the results; if it does not reach this reciprocal, dynamic collaborative weights are allocated according to relevance and credibility, and intermediate results are continued to be exchanged; remove the common error component from the prediction error of each agent, calculate the credibility identifiable signal-to-noise ratio; determine whether the credibility is identifiable: if it is not identifiable, freeze the credibility update process and re-evaluate the weights of each agent using the counterfactual marginal contribution method; if it is identifiable, continue to update the credibility of each agent according to the individual residuals after deducting the common error component. S5. Submodal Fact Filtering, Suspicion Interception, and Flow Conservation Correction: A monotonic submodal optimization method is used to filter the fact set from the temporal knowledge state graph, setting a set-level joint survival constraint. It checks whether the selected fact set satisfies the set-level joint survival constraint: if not, it iteratively removes the fact with the smallest marginal information content until the constraint is satisfied; if satisfied, it outputs the filtered fact set. The suspicion of each fact to be fused is calculated, and it is determined whether the suspicion exceeds a set threshold: if it exceeds the threshold, the fact is intercepted, preventing it from entering multi-agent fusion; if it does not exceed the threshold, the fact is allowed to participate in fusion. The fusion result after filtering is corrected for flow conservation and then output. S6. State Constraint Prediction, Prediction Validity Period, and Computational Power Level Selection: Iteratively solve the state transition self-consistency condition and decision self-consistency condition, setting a training objective with decision regret as the main factor and prediction error as the regularization term; monitor the stability of the dual-loop consisting of the acquisition loop and the decision reaction loop, and determine whether the sum of the amplification factors of the two loops is less than 1: if it is not less than 1, the dual-loop is determined to be unstable, and three types of suppression processing are implemented: reducing the degree of acquisition adaptation, increasing the output change rate limit, and increasing the aggregation granularity; if it is less than 1, the current configuration is maintained; solve the failure time spectrum equation to obtain the prediction validity period corresponding to each network node; determine the overall... If the network prediction validity period is 0: if it is 0, the prediction at the current moment is determined to be unreliable, and a data insufficiency flag is output, and no point prediction is output; if it is not 0, a spatial cluster is generated according to the frequency domain feature similarity, and the spatial cluster granularity and computing power level are selected under the condition of meeting the time constraint; if the feasible set of candidate computing power levels is empty: if it is empty, the prediction value is determined to be unable to be delivered within the validity period, and it is actively abandoned and a downgrade flag is output; if it is not empty, the error characteristic curves of various candidate prediction modules are calibrated offline, and the expected error is directly calculated from the curve to complete the module selection during the selection stage, without on-site trial running at the target node; S7. Feedback Update and Hierarchical Memory Maintenance: Utilizing real-world network feedback, predictive errors and decision regret are statistically analyzed; the agent outputs two types of structured modifications: factual failure rate and range of selectable actions; a counterfactual verification mechanism is employed to determine whether the difference in decision regret before and after the modification meets a set threshold, and whether the direction of this difference remains consistent across multiple repeated trials; if both are satisfied, the modification is accepted and written back to the temporal knowledge state graph and the hierarchical memory structure; if not, the modification is discarded and not written to the graph or memory; the hierarchical memory includes a contextual memory layer, a semantic memory layer, and a skill memory layer, which respectively complete... Experience sample storage, knowledge prototype updating, and cross-agent knowledge transfer; updating the dynamic failure rate and knowledge activity of memory entries; determining whether the knowledge activity is lower than the first threshold: if it is lower than the first threshold, the memory sample is moved to the dormant archive instead of being directly deleted; if it is not lower than the first threshold, it is retained in the active area; determining whether the similarity between the new working condition and the dormant memory is higher than the second threshold: if it is higher than the second threshold, the memory is reactivated and the dormant sample is recalled to the active area; continuously monitoring the marginal value of the collaborative channel and determining whether its performance indicators are lower than the negative threshold multiple times in a row: if so, an alarm signal is issued, and the observation range and collaborative channel are reconfigured.
[0025] As attached Figure 2As shown, this is the program execution process of step S1 above; after collecting multi-source heterogeneous data, time synchronization and feature encoding are completed, the observation rise caused by the collection itself is deducted, data sufficiency is calculated, the collection loop amplification factor is calculated, and the processed flow feature tensor is output and sent to the time series knowledge state graph construction module; the method for eliminating the observation rise caused by the collection, calculating data sufficiency, and calculating the amplification factor of the collection loop itself adopts the following steps: S101. Estimation and Subtraction of Acquisition Interference Coefficient: The data acquisition requests of this device are recorded as a sequence of acquisition request quantities. The acquisition request quantities are then averaged to obtain the averaged acquisition request quantities. A least-squares regression is performed between the averaged acquisition request quantities and the original observed flow rate to estimate the acquisition interference coefficient. The interference term is then subtracted from the original observed flow rate to obtain the observed flow rate after interference subtraction. The acquisition interference coefficient and the observed flow rate after interference subtraction are respectively... and In the formula, To collect the number of requests without taking the mean, For the original observed flow rate, To observe the average flow rate, To collect interference coefficients; After removing interference; S102. Data Sufficiency Calculation: Calculate the Fisher information matrices of environmental parameters under the current acquisition strategy and the ideal acquisition strategy respectively. Take the ratio of the trace of the inverse Fisher information matrix of the ideal acquisition strategy to the trace of the inverse Fisher information matrix of the current acquisition strategy to obtain the observation gap. Then, take the reciprocal of the observation gap as the data sufficiency. ,in The data sufficiency is no greater than 1; the closer it is to 1, the more sufficient the current observations are. In the formula, This is the Fisher information matrix for the current data collection strategy. For an ideal data collection strategy, Represents the trace operation of a matrix; S103. Calculation of the amplification factor of the self-acquisition loop: Multiply the acquisition density sensitivity, the acquisition quantity sensitivity to network load, and the proportion of the prediction result's absorption of the observation bias to obtain the amplification factor of the self-acquisition loop. When the magnification reaches 1, it is determined that positive feedback has been formed, and the system switches to either fixed-density sampling or random-time sampling; where, To improve density sensitivity, To assess the sensitivity of the data collection volume to network load. To predict the proportion of observation bias absorbed by the results; when the amplification factor of its own acquisition loop reaches the set threshold, the adaptive acquisition density mode will be switched to the fixed density acquisition mode or the random time acquisition mode.
[0026] As attached Figure 3 As shown, the program execution process of step S2 above is as follows: constructing sets of entities, relations, and facts; assigning a failure rate to each fact; calculating the state transition probability; handling the knowledge weight redistribution after fact failure; determining unreachable states; and outputting a temporal knowledge state graph carrying the failure rate and sending it to the grounding state encoding module. The calculation method for the fact failure rate, state transition probability, and knowledge weight redistribution after failure adopts the following steps: S201. Factual Failure Rate Calculation: The factual failure rate is obtained by modulating the baseline risk rate and agent actions. The agent actions rewrite the factual risk rate as a multiplicative factor, i.e. In the formula, For historical information; Actions of the intelligent agent; For the event; The baseline risk rate for facts is the inherent failure rate of facts without intervention. For the facts The learnable coefficient vector; This is an action feature mapping function that maps the original action to a feature vector with modified failure rate. S202. Failure Gated State Transition Probability Calculation: Integrate the failure rate of all facts supporting the current state transition over the prediction period using time, take the negative exponent to obtain the joint survival probability of the fact set, and then multiply it by the transition tendency constituted by resource utilization and state change to obtain the state transition probability. ; In the formula, For state To state Predicting step size The failure gating state transition probability; The current moment; To predict the field of view; It is an integral dummy variable; The set of facts that state transitions depend on; For a moment The failure rate of the following fact f; The normalization function outputs... Interval probability; These are the state transition probability weighting coefficients; Let t be the network resource utilization rate at time t; Let t be the rate of change of flow rate at time t; when all the facts supporting a certain state transition fail, the joint survival probability of the fact set is zero, the state transition probability is set to zero accordingly, and the target state is determined to be unreachable within the prediction period; S203. Redistribution of knowledge weights after failure: The proportion of the product of a fact's failure rate and its relevance among all surviving related facts of the same type is used as the competitive allocation ratio, as follows:
[0027] The knowledge weight of invalid facts is transferred to surviving related facts according to this proportion, avoiding the direct discarding of evidence information; where, The knowledge weight competition ratio allocated to surviving fact g after fact f becomes invalid; The degree of relevance between fact f and fact g; For candidate survival facts that are associated with f, it is a summation and traversal of indices; The failure rate is the factual failure source fact f; For candidate survival facts associated with f The fact of failure; For fact f and candidate survival facts The degree of correlation between them.
[0028] In this embodiment, the fact failure rate is modulated by both the baseline risk rate and the agent's actions. When all facts supporting "high load → congestion" fail, the state transition probability is directly set to zero, and the congestion state is marked as unreachable during the prediction period and no longer participates in subsequent prediction and inference. When a fact fails, the knowledge weight is transferred to the remaining surviving related facts according to the competition allocation ratio.
[0029] As attached Figure 4 As shown, this is the program execution process of step S3 above; the graph encoding is completed, the traffic, network state and hierarchical memory information are fused, the global shared environment state vector is output, the local state of each agent is obtained by decomposition, and it is distributed to each agent; the grounding encoding, observation resource constraints and the calculation of the optimal shared observation area length are specifically carried out through the following steps: S301. Generation of Global Environment State and Local State of a Single Agent: The global environment state is obtained by concatenating and encoding four types of information: traffic observation, graph knowledge, network state, and hierarchical memory. Furthermore, it is divided into four components: observation, knowledge, state, and hierarchical memory, to obtain the local state of a single agent. ;in, The environment state vector is shared globally by multiple agents; For feature fusion function; Traffic characteristics; The knowledge representation encoded as a temporal knowledge state diagram; This is the network state vector; The hierarchical memory structure is represented by a vector. Let be the local state vector of the i-th agent; These are respectively agent-private observation, graph knowledge, network environment, and hierarchical memory components; S302. Observation Resource Constraints and Solution for Optimal Shared Observation Area: The total observation budget is divided into a shared observation area (shared by all agents and collected only once) and private observation areas (not overlapping between agents), constituting the observation resource constraints. Construct a lower bound for the multi-agent fusion error based on this constraint, and find the extreme value of the shared observation area length to obtain the optimal shared observation area. And a truncation function is used to restrict it to the range of 0 to the total observation budget; among which, To share the length of the observation area; The length of a single agent's private observation zone; The total number of intelligent agents; For the total budget of observation; The optimal shared observation area length; This is a truncation function that restricts the range of output values. The length related to the flow space; The variance of individual errors that can be eliminated by shared observations; To share the observation cost coefficient; S303, Three-phase observation configuration switching and distribution: Based on the interval of the shared observation cost coefficient, switch and distribute the configuration between three observation modes: full private configuration, partial shared configuration, and full shared configuration, and divide the shared observation area and the private observation area of each agent accordingly. The number of agents in this embodiment Total budget for observation 200 data collection points are selected; the optimal shared observation area is limited to 0 to 1 using the clip truncation function. Interval; when sharing observation cost coefficient If the coefficient is too large, switch to all private observations; if the coefficient is too small, switch to all shared observations; in the intermediate range, execute partial shared configuration. The traffic prediction agent focuses on observation and state components, the knowledge reasoning agent focuses on graph and memory components, and the decision-making agent uses all four components.
[0030] As attached Figure 5 As shown, the program execution process of step S4 above is as follows: The environmental state is read, the optimal shared observation area is calculated to complete the observation configuration; the shared cost ratio, failure-gated attention, and dynamic collaborative weights are calculated; the collaborative channel is switched on and off based on the shared cost ratio threshold; the credibility-identifiable signal-to-noise ratio is calculated, credibility update or counterfactual reassessment weights are selected, and the multi-agent fusion intermediate results are output and sent to the state constraint prediction module; the calculation methods for the shared cost ratio, failure-gated attention, dynamic collaborative weights, and credibility-identifiable signal-to-noise ratio are as follows: S401. Calculation of the sharing cost ratio: Using the sharing intensity as the independent variable, calculate the derivatives of the common error variance and the individual error variance with respect to the sharing intensity, and use the ratio of the two as the sharing cost ratio. This ratio characterizes the common error variance injected for each unit of individual error variance eliminated; where, For the cost-sharing ratio; the criterion is Close all collaboration channels; The overall sharing strength of the system characterizes the degree of information sharing among multiple agents; The common error variance introduced for sharing does not diminish with the number of agents; Individual error variance can be reduced through information sharing; S402. Failure-gated attention calculation: Calculate the inner product of the query vector and key vector for each agent and normalize it according to the dimension. Add the failure-gated factor in the logarithmic field to the attention score to obtain the failure-gated attention. Among them, failure gating factor When all supporting facts fail, the gating factor is zero, corresponding to a strictly zero attention level; where, For attention weights in failure gating; hour, ; Let i be the query vector for the i-th agent; Let J be the key vector of the j-th agent; The Q / K vector dimension; As a failure gating factor, supporting The joint survival probability of the set of facts related to state transitions within the prediction window; S403. Cooperative Channel Switching and Dynamic Cooperative Weight Calculation: Indicative factors are constructed using the difference between the temporal state semantic relevance and the channel admission threshold, and the difference between the inverse of the number of agents and the sharing cost ratio. Cooperative channels are allowed to open only when both factors are positive, thus obtaining the cooperative channel switching mechanism. The channel switch is applied as a multiplicative factor to the normalized weights to obtain dynamic collaborative weights. This allows the weights to be strictly zero; where, This is an indicator function; it returns 1 if the condition is true and 0 if the condition is false. For collaborative channel switching; The temporal state semantic correlation between agents i and j; This is the semantic relevance threshold; Dynamic collaborative weights; Let the credibility of the j-th agent be denoted as 'j'. S404. Credibility Identification Determination: Subtract the mean of all errors from the prediction errors of each agent to obtain the individual residuals after deducting the common error. The signal-to-noise ratio can be identified by calculating the reliability of the proportion of the largest eigenvalue in the residual correlation matrix and the number of agents. When the signal-to-noise ratio is less than 1, the reliability is deemed indistinguishable; among which, The original prediction error of agent i; This refers to the pure individual residual error after deducting the global common error component; For credibility, the signal-to-noise ratio can be identified; when When credibility is unrecognizable, credibility updates are frozen. The percentage of the largest eigenvalue obtained from the eigenvalue decomposition of the residual correlation matrix of the decision agent.
[0031] In this embodiment, the threshold When the cost of sharing is higher than When the signal-to-noise ratio reaches 0.2, all collaborative channel switches are set to zero, and the agents no longer exchange intermediate results; instead, their individual predictions are combined and output. If the value is less than 1, the credibility iteration update is frozen, and the weights of each agent are re-evaluated using the counterfactual marginal contribution method. The logarithm of the failure gating factor is added to the attention. When all supporting facts fail, the attention of the corresponding channel is strictly set to zero, completely blocking the output of that agent.
[0032] As attached Figure 6 The above steps S5 and S6 are illustrated in the program execution process: sub-modal fact screening, suspicion interception, and flow conservation correction are performed; the decision self-consistent fixed point is iteratively solved; the failure time spectrum equation is solved to obtain the node prediction validity period; aggregated clusters are generated according to frequency domain similarity, and the aggregation granularity and computing power level are selected; if the conditions are not met, downgrading or abandonment is performed, and the prediction results are output and sent to the feedback update and hierarchical memory module. The calculation methods for the sub-modal screening objective function, set survival constraint, fact suspicion, flow conservation correction, training objective, failure time spectrum equation, and module offline error characteristic curve are as follows: S501. Solving for the optimal fact set: Using the marginal information content of the conditions multiplied by the survival probability of the facts as the payoff term, and the sum of the failure rates of facts and the sum of the co-mortality correlation coefficients of facts as penalty terms, a monotonic submodular objective function is constructed. Maximizing this objective function, while ensuring the set size does not exceed a given upper limit, yields the optimal fact set.
[0033] The objective function does not have a universal scalar score that can achieve global optimum, therefore a simple sorting method cannot be used to filter facts; where, To filter the optimal set of facts in the output; To select an upper limit for the number of facts; Given the historical context and selected facts, the conditional marginal information content of fact f; Let t be the set of historical facts. This is the fact that f has already been selected into the set; Let f be the individual survival probability of a single fact f. The coefficient for the failure rate penalty term; The coefficient of the co-mortality related penalty term; Let f be the co-mortality correlation coefficient between facts f and g, which characterizes the degree of correlation between the two factual failure events; S502, Set-level Joint Survival Constraint Verification: Multiply the joint survival probability of the fact set during the prediction period by the co-mortality correlation correction term to obtain the set-level joint survival probability. And require it to be no lower than the set level. When the condition is not met, the fact that minimizes marginal information is removed cyclically; where, This represents the aggregate-level joint survival probability after co-mortality correction. To accommodate the survival probability tolerance, the joint survival probability of the set must be greater than or equal to 1. ; S503. Fact Suspicion Calculation and Interception: Fact suspicion is constructed based on the degree of insufficient precursor support, the amount of violation of flow conservation, and the strength of fact generation. If the suspicion level exceeds a set threshold, the fact will be blocked, preventing it from entering the multi-agent fusion process; where, The degree of suspicion of fact f; if it exceeds the threshold, the fact is blocked and prohibited from entering the fusion process; The preceding support level measures whether there is supporting evidence for the fact. This violates the principle of flow conservation. Strength is born from facts; It is the fact importance penalty coefficient, a preset constant hyperparameter used to adjust the exponential decay caused by fact importance; S504, Flow Conservation Correction: Constructing an orthogonal projection operator for the conservation subspace from the node-link correlation matrix. A closed-loop projection is performed on the multi-agent fusion result to obtain a corrected output that satisfies flow conservation. No learnable parameters are introduced; among them, For flow conservation orthogonal projection operators; It is the identity matrix; Network topology node-link correlation matrix; This is the pseudoinverse of the Moore-Penrose matrix; This is the original prediction vector without conservation correction; This is the final predicted output after projection correction and strict satisfaction of flow conservation constraints; S505. Training Objective Construction: Construct training objectives with decision regret as the primary objective and prediction error as the regularization term. This allows for higher prediction errors in some samples while improving decision performance; among them, For multi-agent training loss function, Regret for the decision For prediction error regularization, The regularization coefficient; S506, Dual-loop stability verification and suppression: Estimate the amplification factor of the acquisition loop and the decision reaction loop respectively, and determine whether the sum of the two is less than 1: If it is not less than 1, perform three types of suppression actions: reduce the degree of acquisition adaptation, increase the output change rate limit, and increase the aggregation granularity; if it is less than 1, maintain the current configuration. S507. Solving the Failure Time Spectrum Equation and Calculating the Predicted Effective Period: Taking the prediction horizon as the evolution time, the error diffusion process on the graph is expressed as a diffusion equation. Expanding this equation according to the graph's characteristic decomposition yields the failure time spectrum equation. The inverse solution yields the predicted validity period for each network node. The prediction validity period for backbone low-spectrum nodes is longer, while that for edge burst nodes is shorter; among them, For network node indexing; To solve for the output, nodes The predicted validity period; The error spread rate on the network graph; The Laplace matrix of the network topology graph is the first... One eigenvalue; For nodes Spectral energy on the k-th order eigenvector; Coverage factor; The standard deviation of the flow rate observation at the initial time; This is the business-permissible error tolerance, i.e., the maximum prediction error that the business side can tolerate; The common error variance; S508, Spatial Aggregation and Computing Power Tier Selection: Spatial aggregation clusters are generated based on frequency domain feature similarity. Spatial aggregation granularity and computing power tier are selected under the condition of meeting time constraints. When all candidate computing power tiers do not meet the time constraints, a downgrade flag is output and the candidate is actively abandoned. No point prediction is output. S509. Module Offline Error Characteristic Curve Calibration and Trial-Free Selection: Offline fitting of error characteristic curves for various candidate prediction modules at spectral position and prediction step size. After deployment, the expected error is calculated directly from this curve to complete module selection, without the need for trial runs at the target node; among which... For module At the node Predicting step size The mean square error at that point. For module constant term, Frequency band sensitivity coefficient, The error diffusion rate is the module's own value, and the three are inherent offline multiplexing parameters of the module.
[0034] In this embodiment, the tolerance of the aggregate survival probability Set to 0.9; when the overall joint survival probability of the set does not meet the standard, the fact with the smallest marginal information content is deleted in a loop; intermediate facts with a suspicion level exceeding the threshold are directly intercepted and do not enter the multi-agent fusion; orthogonal projection operators are used to perform closed-loop flow conservation correction to ensure that the output strictly meets the network flow conservation constraint. When the sum of the dual-loop amplification factors is not less than 1, the degree of adaptive acquisition is reduced, the output change rate limit is increased, and the aggregation granularity is improved; when all candidate computing power levels do not meet the time delay constraint, a downgrade mark is output and the candidate is actively abandoned; each prediction module pre-fits the error characteristic curve offline, and after going online, the expected error is directly calculated by looking up the table to complete the module selection.
[0035] As attached Figure 7 As shown, the execution process of step S7 above is as follows: receiving real feedback from the network, calculating prediction error and decision regret; reflecting on the structured modification amount output by the agent, and performing counterfactual verification; updating the temporal knowledge state graph; updating hierarchical memory, including memory entry failure rate, knowledge activity, knowledge prototype and knowledge transfer; monitoring the marginal value of the collaborative channel, and if a deterioration alarm occurs, reverting to reconfigure the observation range and collaborative channel, and entering the next prediction cycle. The method of hierarchical memory related iterative update and counterfactual verification adopts the following steps: S701. Dynamic Failure Rate Update for Memory Items: An exponential factor is constructed from the offset between the mean of recent decision regret and the mean of historical decision regret. This factor modulates the baseline failure rate of memory items to obtain the dynamic failure rate of memory items. Increased regret over decisions leads to accelerated memory loss; among them, The dynamic failure rate of memory entry r; The initial baseline failure rate for memory entry r; I regret my recent decisions; The average regret rate for long-term historical decisions; The sensitivity coefficient of memory impairment rate to decision regret bias; S702, Knowledge Activity Iteration: Knowledge activity is updated jointly by the forgetting curve decay term, the hit gain term, and the error penalty term. ;in, Let i be the knowledge activity of the i-th memory sample; The time step interval; This serves as the forgetting time constant, controlling the natural rate of forgetting. This is the hit gain coefficient; the sample was successfully invoked, and its activity was increased. This represents the number of times the sample was successfully invoked at time t. This is the error penalty coefficient; when samples involved in prediction produce large errors, the activity decreases. This is the normalized error of the sample participating in the prediction output; S703. Knowledge Prototype Update: When the similarity between a new experience sample and an existing knowledge prototype is higher than the third threshold, the prototype center vector is finely adjusted towards the new sample by adjusting the step size according to the confidence level. The step size If the similarity between the new experience sample and all prototypes is below the third threshold, then a new knowledge prototype is created; where, Let the k-th knowledge prototype center vector be denoted as 'k'. For the feature vector of the newly arrived experience sample; The actual update step size is aggregated for the prototype; Base learning rate; The sample-prototype similarity function; The confidence level of the knowledge prototype is used; the higher the confidence level, the smaller the update step size. S704, Added Agent Knowledge Transfer: New agents read transferable skills from the skill memory layer and initialize strategies by weighting and combining them according to transferability scores. Skills with scores below a preset threshold are excluded from the combination to avoid negative transfer; among them, The initial predicted skill vector is obtained for initializing the new agent; This represents the k-th transferable skill vector in the skill memory layer. The score is based on transferability; skills with low scores are excluded from weighting to avoid negative transfer. S705, Counterfactual Verification: Only when the difference between the decision regret of the adoption state and the non-adoption state is lower than the set negative threshold, and the direction of this difference remains consistent in the set number of repeated trials, will the structured modification output by the reflective agent be adopted. In this embodiment, the reflective agent only outputs the edited modifications to the structured graph, not the free text; the modifications must be verified by counterfactual verification before being written into the graph and memory; the memory entry failure rate dynamically changes with decision regret, and memory failure accelerates as decision regret increases; when the knowledge activity is below the first threshold, the sample is moved into dormant archiving instead of being directly deleted; when the similarity between the new working condition and the dormant memory is higher than the second threshold, the memory is reactivated; newly launched agents read the weighted combination initialization strategy of transferable skills from the skill memory layer, and skills with scores below the threshold are discarded; when the marginal value of the collaborative channel continues to deteriorate, an alarm is triggered, and the system reverts to S3 to reconfigure the observation range and collaborative channel.
[0036] As attached Figure 8 As shown, the nodes in the diagram contain entities such as network devices, regions, and events, and the edges represent various business relationships. Each fact is bound to a corresponding fact failure rate parameter, and the birth time, duration, and failure rate of the fact are recorded. Based on this, it supports fact failure determination, knowledge weight redistribution after failure, and failure gate state transition probability calculation. Both entity relationship facts and state transition facts carry fact failure rates, and the state chain of the network state layer is connected sequentially in the temporal knowledge state diagram.
[0037] As attached Figure 9 As shown, the diagram illustrates the structure in which traffic prediction agents, state perception agents, knowledge reasoning agents, decision-making agents, and reflection agents share the same environmental state representation and each holds a private observation area. It also illustrates the control logic of failure-gated attention and collaborative channel switching. Specifically, the overall start and stop of all collaborative channels is controlled by comparing the shared cost ratio with the reciprocal of the number of agents, and the freeze and recovery of the credibility update process is controlled by the credibility-identifiable signal-to-noise ratio.
[0038] As attached Figure 10 As shown, the diagram illustrates a three-layer progressive structure from the contextual memory layer, the semantic memory layer to the skill memory layer, showing the processing path of experience sample storage, knowledge prototype abstraction, and cross-agent knowledge transfer. It also shows the evolution curves of knowledge activity decreasing with the forgetting curve, increasing with hit rate, and decreasing with error suppression, as well as the positions of the first threshold, the second threshold, the dormant zone, and the re-consolidation activation.
[0039] As attached Figure 11 As shown, the data input flows vertically through the multi-source data acquisition module, the temporal knowledge state graph construction module, the grounding state encoding module, the multi-agent collaborative reasoning module, the state constraint prediction module, and the prediction result output, as well as two feedback loops from the feedback update and hierarchical memory module to the multi-agent collaborative reasoning module and the temporal knowledge state graph construction module.
[0040] This embodiment can run on a 5G-A mobile communication network management server and edge computing node hardware device. The processor executes the computer program stored in the memory to implement all the above-mentioned method processes.
[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some 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 the present invention.
Claims
1. A multi-agent network traffic prediction method based on failure gating and shared cost ratio, characterized in that, The following steps are used: S1. Collect multi-source heterogeneous data from the network to eliminate the rise in acquisition observation; assess data sufficiency to support degraded output; calculate the amplification factor of the acquisition loop; and dynamically switch the acquisition strategy. S2. Construct a temporal knowledge state diagram and initialize a hierarchical memory storage framework; The configuration can rewrite the failure rate of facts, calculate the failure gating transition probability and failure weight allocation, and realize the shielding of the corresponding state transition when all supporting facts fail. S3. Perform grounding encoding on the temporal knowledge state graph to generate the shared environment state and local state of multiple agents; Construct observation resource constraints to solve for the optimal shared observation area length, and dynamically switch public and private observation configurations based on cost coefficients; S4. Calculate the shared cost ratio, construct the failure gating factor to obtain the failure gating attention, generate the cooperative channel switch and dynamic cooperative weight, control the start and stop of the channel according to the shared cost ratio, eliminate common errors, and complete the credibility identifiable signal-to-noise ratio determination. S5. Construct a sub-modal filtering objective function to filter the fact set, verify the set survival constraints, calculate the fact suspicion degree to intercept abnormal facts, and complete the traffic conservation correction; S6. Iteratively solve for self-consistent conditions and construct training objectives; monitor and suppress dual-loop instability; solve the failure time spectrum equation to obtain node validity period, complete aggregate computing power selection and downgrade abandonment; calibrate module offline error characteristic curves to achieve module trial-free running; S7. Receive real feedback from network operation, filter the structured modification amount through counterfactual verification, and write it back to the time-series knowledge state diagram and hierarchical memory; Update memory entry failure rate and knowledge activity, complete knowledge prototype update and agent knowledge transfer, realize memory dormancy archiving and reactivation, monitor collaboration channels and reconfigure as needed.
2. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The specific calculations for eliminating the acquisition observation elevation, calculating data sufficiency, and calculating the acquisition loop amplification factor in S1 are performed using the following steps: S101. Estimation and subtraction of interference coefficient: Record the data acquisition requests of this device as a sequence of acquisition request quantities, and perform mean-reduction processing on the acquisition request quantities to obtain the mean-reduction acquisition request quantities. The mean-removed acquisition request quantity and the original observation flow are subjected to least squares regression to estimate the acquisition interference coefficient. The interference term is then subtracted from the original observation flow to obtain the observation flow after interference is deducted. The interference coefficient and the observed flow rate after interference subtraction are respectively: ; In the formula, To collect the number of requests without taking the mean, For the original observed flow rate, To observe the average flow rate, To collect interference coefficients; After removing interference; S102. Data Sufficiency Calculation: Calculate the Fisher information matrices of environmental parameters under the current acquisition strategy and the ideal acquisition strategy respectively. Take the ratio of the trace of the inverse Fisher information matrix of the ideal acquisition strategy to the trace of the inverse Fisher information matrix of the current acquisition strategy to obtain the observation gap. Then, take the reciprocal of the observation gap as the data sufficiency. ,in In the formula, This is the Fisher information matrix for the current data collection strategy. For an ideal data collection strategy, Represents the trace operation of a matrix; S103. Calculation of the amplification factor of the self-acquisition loop: Multiply the acquisition density sensitivity, the acquisition quantity sensitivity to network load, and the proportion of the prediction result's absorption of the observation bias to obtain the amplification factor of the self-acquisition loop. In the formula, To improve density sensitivity, To assess the sensitivity of the data collection volume to network load. The proportion of observation bias absorbed by the prediction results.
3. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The calculation methods for the factual failure rate, state transition probability, and post-failure weight allocation in S2 are as follows: S201. Factual Failure Rate Calculation: The factual failure rate is obtained by modulating the baseline risk rate and agent actions. The agent actions rewrite the factual risk rate as a multiplicative factor, i.e. In the formula, For historical information; Actions of the intelligent agent; For the event; The baseline risk rate for facts is the inherent failure rate of facts without intervention. For the facts The learnable coefficient vector; This is an action feature mapping function that maps the original action to a feature vector with modified failure rate. S202. Failure Gated State Transition Probability Calculation: Integrate the failure rate of all facts supporting the current state transition over the prediction period using time, take the negative exponent to obtain the joint survival probability of the fact set, and then multiply it by the transition tendency constituted by resource utilization and state change to obtain the state transition probability. ; In the formula, For state To state Predicting step size The failure gating state transition probability; The current moment; To predict the field of view (predicting the step size backward). It is an integral dummy variable; The set of facts that state transitions depend on; For a moment The failure rate of the following fact f; The normalization function outputs... Interval probability; These are the state transition probability weighting coefficients; Let t be the network resource utilization rate at time t; Let t be the rate of change of flow rate at time t; when all the facts supporting a certain state transition fail, the joint survival probability of the fact set is zero, the state transition probability is set to zero accordingly, and the target state is determined to be unreachable within the prediction period; S203. Redistribution of knowledge weights after failure: The proportion of the product of a fact's failure rate and its relevance among all surviving related facts of the same type is used as the competitive allocation ratio, as follows: The knowledge weight of invalid facts is transferred to surviving related facts according to this proportion, avoiding the direct discarding of evidence information; where, The knowledge weight competition ratio allocated to surviving fact g after fact f becomes invalid; The degree of relevance between fact f and fact g; For candidate survival facts that are associated with f, it is a summation and traversal of indices; The failure rate is the factual failure source fact f; For candidate survival facts associated with f The fact of failure; For fact f and candidate survival facts The degree of correlation between them.
4. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The calculation of grounding encoding, observation resource constraints, and optimal shared observation area length in S3 follows these steps: S301. Generation of Global Environment State and Local State of a Single Agent: The global environment state is obtained by concatenating and encoding four types of information: traffic observation, graph knowledge, network state, and hierarchical memory. Furthermore, it is divided into four components: observation, knowledge, state, and hierarchical memory, to obtain the local state of a single agent. ;in, The environment state vector is shared globally by multiple agents; For feature fusion function; Traffic characteristics; The knowledge representation encoded as a temporal knowledge state diagram; This is the network state vector; The hierarchical memory structure is represented by a vector. Let be the local state vector of the i-th agent; These are respectively agent-private observation, graph knowledge, network environment, and hierarchical memory components; S302. Observation Resource Constraints and Solution for Optimal Shared Observation Area: The total observation budget is divided into a shared observation area (shared by all agents and collected only once) and private observation areas (not overlapping between agents), constituting the observation resource constraints. Construct a lower bound for the multi-agent fusion error based on this constraint, and find the extreme value of the shared observation area length to obtain the optimal shared observation area. ;in, To share the length of the observation area; The length of a single agent's private observation zone; The total number of intelligent agents; For the total budget of observation; The optimal shared observation area length; This is a truncation function that restricts the range of output values. The length related to the flow space; The variance of individual errors that can be eliminated by shared observations; To share the observation cost coefficient; S303, Three-phase observation configuration switching and distribution: Based on the interval of the shared observation cost coefficient, switch and distribute the configuration between three observation modes: full private configuration, partial shared configuration, and full shared configuration, and divide the shared observation area and the private observation area of each agent accordingly.
5. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The calculation of shared cost ratio, failure-gated attention, dynamic collaborative weights, and credibility-identifiable signal-to-noise ratio in S4 specifically adopts the following steps: S401. Calculation of the sharing cost ratio: Using the sharing intensity as the independent variable, calculate the derivatives of the common error variance and the individual error variance with respect to the sharing intensity, and use the ratio of the two as the sharing cost ratio. ;in, For the cost-sharing ratio; the criterion is Close all collaboration channels; The overall sharing strength of the system characterizes the degree of information sharing among multiple agents; The common error variance introduced for sharing does not diminish with the number of agents; Individual error variance can be reduced through information sharing; S402. Failure-gated attention calculation: Calculate the inner product of the query vector and key vector for each agent and normalize it according to the dimension. Add the failure-gated factor in the logarithmic field to the attention score to obtain the failure-gated attention. Among them, failure gating factor In the formula, For attention weights in failure gating; hour, ; Let i be the query vector for the i-th agent; Let J be the key vector of the j-th agent; The Q / K vector dimension; As a failure gating factor, supporting The joint survival probability of the set of facts related to state transitions within the prediction window; S403, Cooperative Channel Switching and Dynamic Cooperative Weight Calculation: Indicative factors are constructed using the difference between the temporal state semantic relevance and the channel admission threshold, and the difference between the inverse of the number of agents and the sharing cost ratio. Cooperative channels are allowed to open only when both factors are positive, thus obtaining the cooperative channel switching mechanism. Applying the channel switch as a multiplicative factor to the normalized weights yields dynamic collaborative weights. in, This is an indicator function; it returns 1 if the condition is true and 0 if the condition is false. For collaborative channel switching; The temporal state semantic correlation between agents i and j; This is the semantic relevance threshold; Dynamic collaborative weights; Let the credibility of the j-th agent be denoted as 'j'. S404. Credibility Identifiability Determination: Subtract the mean of all errors from the prediction errors of each agent to obtain the individual difference after deducting the common error. The signal-to-noise ratio can be identified by calculating the reliability of the proportion of the largest eigenvalue in the residual correlation matrix and the number of agents. ;in, The original prediction error of agent i; This refers to the pure individual residual error after deducting the global common error component; For credibility, the signal-to-noise ratio can be identified; when Credibility is unverifiable; credibility updates are frozen. This represents the proportion of the largest eigenvalue obtained from the eigenvalue decomposition of the residual correlation matrix.
6. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The calculation of the submodal screening objective function, set survival constraint, fact doubtfulness, and flow conservation correction in S5 adopts the following steps: S501. Solving for the optimal fact set: Using the marginal information content of the conditions multiplied by the survival probability of the facts as the payoff term, and the sum of the failure rates of facts and the sum of the co-mortality correlation coefficients of facts as penalty terms, a monotonic submodular objective function is constructed. Maximizing this objective function, while ensuring the set size does not exceed a given upper limit, yields the optimal fact set. In the formula, To filter the optimal set of facts in the output; To select an upper limit for the number of facts; Given the historical context and selected facts, the conditional marginal information content of fact f; Let t be the set of historical facts. This is the fact that f has already been selected into the set; Let f be the individual survival probability of a single fact f. The coefficient for the failure rate penalty term; The coefficient of the co-mortality related penalty term; Let f be the co-mortality correlation coefficient between facts f and g, which characterizes the degree of correlation between the two factual failure events; S502, Set-level Joint Survival Constraint Verification: Multiply the joint survival probability of the fact set during the prediction period by the co-mortality correlation correction term to obtain the set-level joint survival probability: In the formula, This represents the aggregate-level joint survival probability after co-mortality correction. To accommodate the survival probability tolerance, the joint survival probability of the set must be greater than or equal to 1. ; S503, Fact Suspicion Calculation and Interception: Fact suspicion is constructed based on the degree of insufficient precursor support, the amount of violation of flow conservation, and the strength of fact generation. ; In the formula, The degree of suspicion of fact f; if it exceeds the threshold, the fact is blocked and prohibited from entering the fusion process; The preceding support level measures whether there is supporting evidence for the fact. This violates the principle of flow conservation. Strength is born from facts; It is the fact importance penalty coefficient, a preset constant hyperparameter used to adjust the exponential decay caused by fact importance; S504, Flow Conservation Correction: Constructing an orthogonal projection operator for the conservation subspace from the node-link correlation matrix. A closed-loop projection is performed on the multi-agent fusion result to obtain a corrected output that satisfies flow conservation. No learnable parameters are introduced; among them, For flow conservation orthogonal projection operators; It is the identity matrix; Network topology node-link correlation matrix; This is the pseudoinverse of the Moore-Penrose matrix; This is the original prediction vector without conservation correction; This is the final predicted output after projection correction and strict satisfaction of flow conservation constraints.
7. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The specific calculations for training objectives, failure time spectrum equations, and module offline error characteristic curve calibration in S6 follow these steps: S601. Training Objective Construction: Construct training objectives with decision regret as the main objective and prediction error as the regularization term. This allows for higher prediction errors in some samples while improving decision performance; among them, For multi-agent training loss function, Regret for the decision For prediction error regularization, The regularization coefficient; S602, Dual-loop stability verification and suppression: Estimate the amplification factor of the acquisition loop and the decision reaction loop respectively, and determine whether the sum of the two is less than 1: If it is not less than 1, perform three types of suppression actions: reduce the degree of acquisition adaptation, increase the output change rate limit, and increase the aggregation granularity; if it is less than 1, maintain the current configuration. S603. Solving the Failure Time Spectrum Equation and Calculating the Predicted Effective Period: Taking the prediction horizon as the evolution time, the error diffusion process on the graph is expressed as a diffusion equation. Expanding this equation according to the graph's characteristic decomposition yields the failure time spectrum equation. The inverse solution yields the predicted validity period for each network node. The prediction validity period for backbone low-spectrum nodes is longer, while that for edge burst nodes is shorter; among them, For network node indexing; To solve for the output, nodes The predicted validity period; The error spread rate on the network graph; The Laplace matrix of the network topology graph is the first... One eigenvalue; For nodes Spectral energy on the k-th order eigenvector; Coverage factor; The standard deviation of the flow rate observation at the initial time; This is the business-permissible error tolerance, i.e., the maximum prediction error that the business side can tolerate; The common error variance; S604, Spatial Aggregation and Computing Power Tier Selection: Generate spatial aggregation clusters based on frequency domain feature similarity, and select spatial aggregation granularity and computing power tier under the condition of meeting time constraints; when all candidate computing power tiers do not meet the time constraints, output a downgrade flag and actively abandon the selection, and do not output point prediction; S605. Module Offline Error Characteristic Curve Calibration and Trial-Free Selection: Offline fitting of error characteristic curves for various candidate prediction modules at spectral position and prediction step size. After deployment, the expected error is calculated directly from this curve to complete module selection, without the need for trial runs at the target node; among which... For module At the node Predicting step size The mean square error at that point. For module constant term, Frequency band sensitivity coefficient, The error diffusion rate is the module's own value, and the three are inherent offline multiplexing parameters of the module.
8. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to claim 1, characterized in that, The specific steps for iterative updates related to hierarchical memory in S7 are as follows: S701. Dynamic Failure Rate Update for Memory Items: An exponential factor is constructed from the offset between the mean of recent decision regret and the mean of historical decision regret. This factor modulates the baseline failure rate of memory items to obtain the dynamic failure rate of memory items. Increased regret over decisions leads to accelerated memory loss; among them, The dynamic failure rate of memory entry r; The initial baseline failure rate for memory entry r; I regret my recent decisions; The average regret rate for long-term historical decisions; The sensitivity coefficient of memory impairment rate to decision regret bias; S702, Knowledge Activity Iteration: Knowledge activity is updated jointly by the forgetting curve decay term, the hit gain term, and the error penalty term. in, Let i be the knowledge activity of the i-th memory sample; The time step interval; This serves as the forgetting time constant, controlling the natural rate of forgetting. This is the hit gain coefficient; the sample was successfully invoked, and its activity was increased. This represents the number of times the sample was successfully invoked at time t. This is the error penalty coefficient; when samples involved in prediction produce large errors, the activity decreases. This is the normalized error of the sample participating in the prediction output; S703. Knowledge Prototype Update: When the similarity between a new experience sample and an existing knowledge prototype is higher than the third threshold, the prototype center vector is finely adjusted towards the new sample by adjusting the step size according to the confidence level. The step size If the similarity between the new experience sample and all prototypes is below the third threshold, then a new knowledge prototype is created; where, Let the k-th knowledge prototype center vector be denoted as 'k'. For the feature vector of the newly arrived experience sample; The actual update step size is aggregated for the prototype; Base learning rate; The sample-prototype similarity function; The confidence level of the knowledge prototype is used; the higher the confidence level, the smaller the update step size. S704, Added Agent Knowledge Transfer: New agents read transferable skills from the skill memory layer and initialize strategies by weighting and combining them according to transferability scores. Skills with scores below a preset threshold are excluded from the combination to avoid negative transfer; among them, The initial predicted skill vector is obtained for initializing the new agent; This represents the k-th transferable skill vector in the skill memory layer. The score is based on transferability; skills with low scores are not included in the weighting to avoid negative transfer.
9. The multi-agent network traffic prediction method based on failure gating and sharing cost ratio according to any one of claims 1-8, characterized in that, The counterfactual verification logic in S7 is as follows: only when the difference between the decision regret of the adoption state and the non-adoption state is lower than the set negative threshold, and the direction of this difference remains consistent in the set number of repeated trials, will the structured modification output by the reflective agent be adopted; monitor the marginal value of the collaborative channel, and when its performance index is lower than the negative threshold for multiple consecutive times, an alarm signal will be issued, and the observation range and collaborative channel will be reconfigured.
10. A system for implementing the multi-agent network traffic prediction method based on failure gating and sharing cost ratio as described in any one of claims 1-8, characterized in that, It includes a multi-source data acquisition module, a time-series knowledge state graph construction module, a grounding state encoding module, a multi-agent collaborative reasoning module, a state constraint prediction module, and a feedback update module; The multi-source data acquisition module is used for multi-source data acquisition and acquisition interference deduction, and switches the acquisition strategy according to the data sufficiency and the amplification factor of the acquisition loop. The temporal knowledge state graph construction module is used to maintain the temporal knowledge state graph and manage the fact failure rate, state transition probability, and knowledge weight redistribution after failure. The grounding state encoding module is used to fuse graph, traffic, state and memory information to output the shared environment state of multiple agents; The multi-agent collaborative reasoning module is used for observation configuration calculation, failure gating attention calculation, and control of collaborative channel on / off based on shared cost ratio. The state constraint prediction module is used to solve the decision self-consistent fixed point, solve the failure time spectrum equation to obtain the prediction validity period, select the aggregation granularity and computing power level, and support active abandonment and degradation output. The feedback update module is used to calculate decision regret, perform counterfactual verification on the amount of modification output by the reflective agent, and update the graph and the three-layer hierarchical memory.
