Hierarchical agent traffic prediction information fusion method
By calculating the credibility, timeliness, and relevance indices of lower-level intelligent agents and dynamically adjusting the fusion weights, the adaptability and accuracy issues of information fusion in hierarchical intelligent agent systems are resolved, thereby improving the accuracy of traffic prediction and decision-making transparency in vehicle-road cooperative scenarios.
Patent Information
- Application Number
- CN202511808243.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-03
- Publication Date
- 2026-03-03
AI Technical Summary
In existing technologies, hierarchical multi-agent systems cannot adapt to dynamically changing environments and task requirements when integrating prediction information from different lower-level agents, and lack effective conflict detection and resolution mechanisms, resulting in information loss.
By calculating the credibility, timeliness, and relevance indices of the lower-level intelligent agents, the fusion weights are dynamically adjusted, and the credibility index is optimized using a Bayesian update mechanism. The decision-making scenarios are divided into two categories: ordinary and critical, and information fusion is performed.
It has improved the accuracy of traffic prediction information in vehicle-road cooperative scenarios, especially in accident early warning, with strong adaptability and improved decision-making accuracy and transparency.
Smart Images
Figure CN121600714A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and intelligent agent systems, and specifically relates to a hierarchical intelligent agent traffic prediction information fusion method. Background Technology
[0002] With the rapid development of artificial intelligence technology, hierarchical multi-agent systems are being increasingly widely used in complex decision-making scenarios. In a hierarchical architecture, upper-layer agents need to comprehensively process predictive information from multiple lower-layer agents to make accurate decisions. Specifically, in vehicle-road-cloud collaborative perception scenarios, the cloud-based decision-making system needs to integrate identification and judgment data (such as road congestion information) transmitted from multiple lower-layer agents, such as vehicle-side and roadside devices, to make corresponding decisions (such as vehicle dispatching decisions).
[0003] However, current methods for predicting information from multiple lower-level agents typically employ only a simple weighted average fusion approach, assigning each agent a pre-defined fixed weight; or a voting decision-making mechanism. The former cannot adapt to dynamically changing environments and task requirements, while the latter easily overlooks minority but correct opinions. Furthermore, when lower-level agents provide contradictory information, existing systems lack effective conflict detection and resolution mechanisms, often resorting to simple averaging or discarding strategies, resulting in information loss. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a hierarchical intelligent agent traffic prediction information fusion method to solve the problem of reasonable fusion of prediction information from different lower-level intelligent agents.
[0005] The present invention achieves the above-mentioned technical objectives through the following technical means.
[0006] A hierarchical intelligent agent traffic prediction information fusion method:
[0007] For the output information of multiple lower-level intelligent agents, the fusion weight of each piece of information is calculated based on credibility, timeliness, and relevance respectively;
[0008] Based on the fusion weight, the output information of each lower-level agent is fused.
[0009] Furthermore, the fusion weights are calculated based on the decision-making scenario to which the decision-making task belongs, including:
[0010] For the first type of decision-making scenario, the weighted calculation is as follows:
[0011]
[0012] For the second type of decision-making scenario, the weight calculation is as follows:
[0013]
[0014] In the formula, For lower-level intelligent agents The fusion weight, , , They are the lower-level intelligent agents. Credibility index, timeliness index, and relevance index.
[0015] Furthermore, the credibility index The calculation is as follows:
[0016]
[0017] In the formula, , , , They are the lower-level intelligent agents. Historical accuracy, professionalism score, consistency score, and peer validation score. , , , These are the weighting coefficients.
[0018] Furthermore, the consistency score is calculated as follows:
[0019]
[0020] In the formula, The variance of the information output by the lower-level agent within the sliding window range.
[0021] Furthermore, the same-layer verification score is calculated by first calculating the degree of deviation between the output information of each lower-layer agent and the output information of other lower-layer agents, and then taking a value based on the degree of deviation, wherein the greater the degree of deviation, the lower the same-layer verification score.
[0022] Furthermore, a Bayesian update mechanism is employed to dynamically update and adjust the credibility index of each lower-level agent:
[0023]
[0024] In the formula, For lower-level intelligent agents The updated credibility index Its prior confidence index. To achieve a given credibility index Data observed under conditions The likelihood probability, For data The marginal probability.
[0025] Furthermore, the timeliness index is calculated as follows:
[0026]
[0027] In the formula, The attenuation rate, The duration of the message transmission.
[0028] Furthermore, the correlation index is calculated as follows:
[0029]
[0030] In the formula, , , They are the lower-level intelligent agents. Location relevance, context relevance, task relevance, , , These are the weighting coefficients;
[0031] Location Relevance The degree of correlation between the geographic location of the information perceived by the lower-level intelligent agent and the target area of the current decision-making task;
[0032] Contextual relevance The degree of matching between the dynamic situation in which the lower-level agent is located and the situational pattern that the current decision-making task is concerned with;
[0033] Task relevance This represents the strength of the association between the inherent attributes of the lower-level intelligent agent and the current decision task type.
[0034] Furthermore, the merged information for:
[0035]
[0036] In the formula, For lower-level intelligent agents The output information, This represents the number of lower-level intelligent agents.
[0037] Furthermore, it also includes calculating the uncertainty of the fused information. :
[0038]
[0039] In the formula, For lower-level intelligent agents The confidence level of the output information.
[0040] The beneficial effects of this invention are as follows:
[0041] (1) This invention provides a hierarchical intelligent agent traffic prediction information fusion method, which significantly improves decision accuracy by comprehensively considering the characteristics of multi-dimensional information, especially for accident early warning in vehicle-road cooperative scenarios.
[0042] (2) The present invention adaptively and dynamically sets the fusion weight of each lower-level agent, which can adapt to different decision-making task scenarios and environments; wherein the decision-making scenarios are divided into two categories: ordinary (first category) and critical (second category). Ordinary decision-making scenarios take into account the three dimensions of credibility, timeliness and relevance in a balanced manner, while critical decision-making scenarios emphasize the credibility of the lower-level agents.
[0043] (3) When calculating the credibility index, this invention comprehensively considers the historical accuracy, professionalism score, consistency score and peer verification score of the corresponding lower-level intelligent agent. The calculated credibility index is more comprehensive and more accurate and objective. In addition, a Bayesian update mechanism is used to dynamically update the credibility index to ensure its timeliness. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of a traffic prediction information fusion scenario according to the present invention;
[0045] Figure 2 This is a flowchart of the credibility assessment process in this invention;
[0046] Figure 3 This is a functional module architecture diagram of the upper-layer intelligent body in the traffic prediction information fusion of this invention. Detailed Implementation
[0047] Embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein similar or identical reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0048] like Figure 1 As shown, the system consists of an onboard perception layer, a roadside perception layer, and a cloud-based perception layer. The onboard perception layer includes intelligent connected vehicles equipped with various sensors, with each vehicle acting as a mobile lower-level intelligent agent. The roadside perception layer includes fixed facilities such as roadside units (RSUs), surveillance cameras, and millimeter-wave radars deployed along the road, with each facility acting as a fixed lower-level intelligent agent. The cloud-based decision layer deploys the information fusion system of this invention, acting as an upper-level intelligent agent to process and fuse lower-level information.
[0049] Each lower-level agent reports information in a unified data format. Each piece of information includes the identifier, timestamp, location information, and prediction information (such as traffic flow, congestion index, etc.) and confidence level of each lower-level agent for the corresponding event.
[0050] like Figure 3 The diagram shows the functional architecture of the upper-layer intelligent body, including a multi-dimensional evaluation module, a weight calculation module, a fusion decision module, an uncertainty quantification module, a learning optimization module, and an interpretability report module. Specifically: ① The multi-dimensional evaluation module evaluates the reported information based on three dimensions: credibility, timeliness, and relevance; ② The weight calculation module then generates fusion weights for the reported information from each lower-layer intelligent body; ③ The fusion decision module fuses the reported information, and the uncertainty quantification module outputs the uncertainty of the fused information, which serves as the basis for the upper-layer intelligent body's decision-making; ④ Finally, the interpretability report module outputs a corresponding decision report (i.e., the basis for the fusion decision, the weight distribution of each intelligent body's contribution, and the decision confidence analysis, helping users understand the decision-making process and improving system transparency and user trust); and the learning optimization module continuously adjusts and optimizes model parameters, including but not limited to credibility weights, timeliness decay rates, and relevance weights, based on the actual execution effect of the fusion decision and feedback from environmental changes, thereby improving the system's adaptability and long-term performance. The specific steps are as follows:
[0051] Step 1: For information fed back by multiple different lower-level agents, perform ① credibility assessment, ② timeliness assessment, and ③ relevance assessment respectively, where:
[0052] 1. Credibility Assessment
[0053] See Figure 2 As shown, the credibility assessment is:
[0054] 1.1 The credibility assessment is specifically divided into four aspects: ① historical accuracy, ② professionalism, ③ consistency, and ④ peer verification.
[0055] 1) Historical accuracy The value range is [0,1], which is used to measure the accuracy of the past prediction results of the corresponding lower-level agent. It is obtained by statistically analyzing the degree of consistency between the past prediction results and the actual situation.
[0056] 2) Professionalism rating The value range is [0,1], which is used to reflect the performance capability of the corresponding lower-level intelligent agent in a specific application field. It can be set in advance by manual scoring. For example, the professional score of a vehicle intelligent agent equipped with LiDAR is higher than that of a vehicle intelligent agent equipped with only a visual sensor in road congestion recognition. Similarly, the professional score of a roadside unit equipped with more comprehensive road information sensing elements is higher than that of a conventional surveillance camera. For example, in this embodiment, the professional score of a vehicle equipped with 64-line LiDAR is 0.9, the professional score of a vehicle equipped with 16-line LiDAR is 0.7, and the professional score of a vehicle with only a camera is 0.5.
[0057] 3) Consistency score The value range is [0,1], which is used to measure the stability of the output of the corresponding lower-level agent. In this embodiment, the variance of the output information of each lower-level agent over a period of time is first calculated by using a sliding window (the window size can be adjusted according to the actual situation, the default is 20). Then the consistency score of the corresponding lower-level agent is:
[0058]
[0059] The smaller the variance mentioned above, the better the consistency.
[0060] 4) Same-layer verification scoring The value range is [0,1]. It is verified by comparing the information about the same thing output by other agents in the same layer at the same time. If the output of a certain lower-level agent is similar to that of most lower-level agents, the corresponding peer verification score of that lower-level agent is higher; conversely, if the output of a certain lower-level agent deviates more from the output of most lower-level agents, the peer verification score of that lower-level agent is lower. Specifically, the deviation between the output information of each lower-level agent and the overall population can be quantified by calculating the standard deviation.
[0061] 1.2, For lower-level intelligent agents The credibility index of its output information The calculation is as follows:
[0062]
[0063] In the formula, , , , These are the weighting coefficients for historical accuracy, professionalism score, consistency score, and peer validation score, respectively. subscript This is the ID number of the lower-level intelligent agent.
[0064] 1.3 A Bayesian update mechanism is adopted to dynamically update and adjust the credibility index of each lower-level agent; that is, whenever new observation data is obtained, the credibility index of the agent is updated according to the likelihood probability.
[0065]
[0066] In the formula, For lower-level intelligent agents The updated credibility index Its prior confidence index. To achieve a given credibility index Data observed under conditions The likelihood probability, For data The marginal probability.
[0067] 2. Timeliness assessment
[0068] The timeliness of dynamically updating information using an exponential decay model:
[0069]
[0070] In the formula, For lower-level intelligent agents The timeliness index It is a natural exponential function. The attenuation rate, The duration of information transmission (i.e., the duration of the lower-level intelligent agent) The time difference between when information is output and when the current upper-level agent makes a decision.
[0071] Attenuation rate The value of is set according to the specific type of decision-making task. For urgent decision-making tasks such as collision pre-tightening, the information value decays rapidly, therefore... Take the larger value; for non-urgent decision-making tasks such as route planning, the information value decays more slowly. Take the smaller value. For example, in this embodiment, the attenuation rate values are given in Table 1 below:
[0072] Table 1: Examples of Attenuation Rate Values
[0073]
[0074] 3. Correlation assessment
[0075] The relevance of information to the current decision-making task is assessed from three perspectives: location relevance, context relevance, and task relevance. Among these, location relevance... This measures the degree of correlation between the geographic location of the lower-level agent's perceived information and the target area of the current decision-making task; contextual relevance. This measures the degree of match between the dynamic situation in which the lower-level agent exists and the situational patterns that the current decision-making task focuses on; task relevance. It measures the strength of the correlation between the inherent attributes of the lower-level agent and the current decision-making task type.
[0076] Specifically in this embodiment:
[0077] 1) Location Relevance The calculation method is as follows: it is measured by the geographical intersection-union ratio (IoU) between the target region and the agent's perceived region. First, the target polygon region is defined according to the decision focus. Secondly, based on the sensor parameters and real-time pose of the intelligent agent, its perceptual polygonal region is determined. Finally, the intersection and union areas of the two polygons are calculated using a geometry library, and their ratio IoU is used as... The output parameters are:
[0078]
[0079] in, For area, It is a mapping function, which can be a linear function or a nonlinear function.
[0080] Preferably, given that the confidence or accuracy of the sensor drops significantly at the edge of the sensing area, only sensing within the "high-confidence core region (IoU ≥ T)" is valuable. Therefore, a piecewise linear function with an activation threshold T is used, i.e., when IoU is below the threshold T, a decision is made... When IoU is greater than or equal to T, :
[0081]
[0082] 2) The method for calculating context relevance is as follows: compare the current velocity and heading angle motion state perceived by the agent with the target state of the task, and calculate the degree of deviation, i.e.:
[0083]
[0084] In the formula, , For intelligent agents The report includes the vehicle's speed and heading. , The target speed and heading are of interest to the mission. For the velocity normalization parameter, and For weights.
[0085] 3) The method for calculating task relevance is as follows: Based on the inherent attributes of the lower-level agent, the driving mode (autonomous driving, manual driving), and the current decision task type, different relevance scores are assigned in advance. Table 2 shows an example of obtaining the driving task relevance through triple mapping:
[0086] Table 2: Examples of Driving Task Relevance
[0087]
[0088] Then the lower-level intelligent agent correlation index for:
[0089]
[0090] In the formula, , , Location relevance Contextual relevance Task relevance The weighting coefficients, .
[0091] Step 2, based on the credibility index Timeliness Index Correlation index Calculate the lower-level intelligent agent Fusion weights It is divided into:
[0092] 1) Conventional decision-making scenarios (Category I decision-making scenarios):
[0093]
[0094] A balanced consideration should be made across the three dimensions to ensure that a significant deficiency in any one dimension will substantially reduce the influence of the information source.
[0095] 2) Key Decision-Making Scenarios (Second Category of Decision-Making Scenarios):
[0096]
[0097] The impact of credibility is a key consideration.
[0098] The aforementioned routine decision-making scenarios and key decision-making scenarios can be pre-classified by humans into different types of decision-making tasks.
[0099] Step 3, based on the fusion weights of each lower-layer agent The prediction information is then fused.
[0100]
[0101]
[0102] In the formula, For the fused prediction information, For lower-level intelligent agents Predictive information, This represents the number of lower-level intelligent agents; This represents the uncertainty of the fused prediction information; the smaller the value, the lower the uncertainty of the prediction information. For lower-level intelligent agents The confidence level of the prediction information.
[0103] For example, three lower-level agents output traffic flow information of 300 pcu / h, 320 pcu / h, and 310 pcu / h, respectively, with corresponding confidence levels of 0.91, 0.82, and 0.75. After steps 1 to 2, the fusion weights of the three lower-level agents are 0.648, 0.072, and 0.216, respectively. Then, the traffic flow information after fusion by the upper-level agent is 303.846 pcu / h, with an uncertainty of 0.134.
[0104] This invention is not limited to the above-described embodiments. Any obvious improvements, substitutions, or modifications that can be made by those skilled in the art without departing from the essence of this invention are within the scope of protection of this invention.
Claims
1. A hierarchical intelligent agent traffic prediction information fusion method, characterized in that: For the output information of multiple lower-level intelligent agents, the fusion weight of each piece of information is calculated based on credibility, timeliness, and relevance. Based on the fusion weight, the output information of each lower-level agent is fused.
2. The hierarchical intelligent agent traffic prediction information fusion method according to claim 1, characterized in that: The fusion weights are calculated based on the decision-making scenario to which the decision-making task belongs, including: For the first type of decision-making scenario, the weighted calculation is as follows: For the second type of decision-making scenario, the weight calculation is as follows: In the formula, For lower-level intelligent agents The fusion weight, , , They are the lower-level intelligent agents. Credibility index, timeliness index, and relevance index.
3. The hierarchical intelligent agent traffic prediction information fusion method according to claim 2, characterized in that: The credibility index The calculation is as follows: In the formula, , , , They are the lower-level intelligent agents. Historical accuracy, professionalism score, consistency score, and peer validation score. , , , These are the weighting coefficients.
4. The hierarchical intelligent agent traffic prediction information fusion method according to claim 3, characterized in that: The consistency score is calculated as follows: In the formula, The variance of the information output by the lower-level agent within the sliding window range.
5. The hierarchical intelligent agent traffic prediction information fusion method according to claim 3, characterized in that: The same-layer verification score is calculated by first calculating the degree of deviation between the output information of each lower-layer agent and the output information of other lower-layer agents, and then taking a value based on the degree of deviation. The greater the degree of deviation, the lower the same-layer verification score.
6. The hierarchical intelligent agent traffic prediction information fusion method according to claim 3, characterized in that: A Bayesian update mechanism is used to dynamically update and adjust the credibility index of each lower-level agent: In the formula, For lower-level intelligent agents The updated credibility index Its prior confidence index. To achieve a given credibility index Data observed under conditions The likelihood probability, For data The marginal probability.
7. The hierarchical intelligent agent traffic prediction information fusion method according to claim 2, characterized in that: The timeliness index is calculated as follows: In the formula, The attenuation rate, The duration of the message transmission.
8. The hierarchical intelligent agent traffic prediction information fusion method according to claim 2, characterized in that: The correlation index is calculated as follows: In the formula, , , They are the lower-level intelligent agents. Location relevance, context relevance, task relevance, , , These are the weighting coefficients; Location Relevance The degree of correlation between the geographic location of the information perceived by the lower-level intelligent agent and the target area of the current decision-making task; Contextual relevance The degree of matching between the dynamic situation in which the lower-level agent is located and the situational pattern that the current decision-making task is concerned with; Task relevance This represents the strength of the association between the inherent attributes of the lower-level intelligent agent and the current decision task type.
9. The hierarchical intelligent agent traffic prediction information fusion method according to claim 2, characterized in that: Information after fusion for: In the formula, For lower-level intelligent agents The output information, This represents the number of lower-level intelligent agents.
10. The hierarchical intelligent agent traffic prediction information fusion method according to claim 9, characterized in that: It also includes calculating the uncertainty of the fused information. : In the formula, For lower-level intelligent agents The confidence level of the output information.