Complex scene-oriented distributed agent interaction architecture method and system

Abstract

The application relates to an artificial intelligence technology and proposes a complex-scene-oriented distributed intelligent agent interaction architecture method and system, which comprises the following steps: reading interaction request data, determining a corresponding interaction scene type, calling a target intelligent agent according to the interaction scene type to form an initial interaction group, performing feature mapping on the interaction request data, performing confidence evaluation on joint features obtained after the feature mapping to obtain a credibility parameter, performing collaborative measurement on a historical interaction process between the target intelligent agents to obtain an interaction efficiency parameter, performing group cognitive consistency evaluation on the joint features to obtain counterfactual correction data, performing weighted fusion on the joint features and the counterfactual correction data based on the credibility parameter, and performing comprehensive evaluation on collaborative decision data obtained after the weighted fusion and the interaction efficiency parameter to output a target interaction group. The application can improve the collaborative decision efficiency, accuracy and reliability of the distributed intelligent agent system in a complex dynamic scene.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system based on a distributed intelligent agent interaction architecture for complex scenarios. Background Technology

[0002] Existing distributed multi-agent collaborative decision-making methods typically employ pre-defined agent invocation strategies and fixed collaborative topologies. Before task execution, several agents are selected from an agent resource pool to form a temporary group according to pre-configured rules. The composition and internal communication relationships of this group are fixed after team formation. However, in complex dynamic scenarios, the semantic features of interaction requests often exhibit high diversity and uncertainty. Fixed team formation strategies struggle to adaptively adjust agent combinations based on dynamic changes in task semantics. Furthermore, existing methods, when processing interaction requests collaboratively, often simply weight or aggregate the outputs of each agent through voting, lacking multi-dimensional considerations regarding data source quality, topological stability of semantic relationships between agents, and the loss of projection information during feature mapping. This makes it difficult to guarantee the reliability of the fusion results. Moreover, since multi-agent systems often share similar data distributions or optimization objectives during training, the judgment tendencies of each agent during collaborative decision-making tend to converge. Existing methods lack effective monitoring methods for group cognitive consistency and have not established proactive intervention mechanisms for excessive convergence, increasing the risk of collective decision-making errors.

[0003] Existing distributed multi-agent systems, after completing collaborative decision-making, typically lack quantitative evaluation and feedback optimization mechanisms for the collaborative process. This makes it difficult to comprehensively evaluate collaborative performance from multiple dimensions, such as response timeliness, task load balancing, and communication path rationality, and also hinders the objective assessment of the efficiency of the agent group's collaborative mode. After prolonged system operation, the initially formed agent group may experience problems such as uneven load distribution or communication congestion. However, existing methods struggle to accurately pinpoint the root cause of these problems, and updating group configurations often employs simple performance ranking or random replacement strategies, lacking refined analysis based on individual marginal contributions. This can easily lead to the accidental removal of potentially valuable agents or the inability to effectively identify performance bottlenecks. These shortcomings prevent multi-agent systems from forming an adaptive closed loop between decision execution, performance evaluation, and structural optimization. Therefore, there is an urgent need to develop a distributed agent interaction architecture method with capabilities such as scenario-adaptive team formation, multi-dimensional confidence assessment, proactive identification and counterfactual correction of group cognitive biases, and collaborative performance-driven topology dynamic optimization. This aims to address these issues and improve the efficiency and reliability of collaborative decision-making in complex dynamic scenarios. Summary of the Invention

[0004] This invention provides a method and system for a distributed intelligent agent interaction architecture for complex scenarios, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for a distributed intelligent agent interaction architecture for complex scenarios, comprising: S01: Read the interaction request data from the distributed data space and determine the interaction scenario type corresponding to the interaction request data; S02: Based on the interaction scenario type, call the target intelligent agent from the intelligent agent registration center to form an initial interaction group; S03: Based on the semantic association between the target intelligent agents, perform feature mapping on the interaction request data, and evaluate the confidence of the joint features obtained after feature mapping to obtain a confidence parameter. Based on the historical interaction records between the target intelligent agents, perform collaborative measurement on the historical interaction process between the target intelligent agents to obtain an interaction performance parameter. S04: Perform a group cognitive consistency assessment on the joint features. When the assessment index reaches or exceeds a preset cognitive convergence threshold, construct a cognitive perturbation signal. Based on the cognitive perturbation signal, perform counterfactual reasoning on the target agent to obtain counterfactual correction data. S05: Based on the credibility parameter, the joint features and the counterfactual correction data are weighted and fused. The collaborative decision data obtained after weighted fusion is comprehensively evaluated with the interaction performance parameter. Based on the evaluation result, the topology of the initial interaction group is updated, and the target interaction group is output.

[0006] In a preferred embodiment, the step of reading interaction request data from the distributed data space and determining the interaction scenario type corresponding to the interaction request data includes: The protocol decapsulation of interactive request data in the distributed data space is performed to obtain the original request message; The original request message is parsed to obtain the request semantic features; The semantic features of the request are matched with a preset scenario library to obtain the interaction scenario type corresponding to the interaction request data.

[0007] In a preferred embodiment, the step of calling the target intelligent agent from the intelligent agent registration center to form an initial interaction group according to the interaction scenario type includes: Based on the interaction scenario type, target intelligent agents matching the capability requirements are retrieved from the intelligent agent registration center; The target intelligent agent is initialized and arranged to obtain the initial interaction group.

[0008] In a preferred embodiment, the step of performing feature mapping on the interaction request data based on the semantic associations between the target agents, and evaluating the confidence level of the joint features obtained after feature mapping to obtain a confidence parameter, includes: The similarity between the semantic feature vectors of the target intelligent agent is calculated, and a semantic association matrix is ​​constructed based on the similarity. The semantic association matrix is ​​sparsified to obtain the association graph between the target intelligent agents; Based on the adjacency relationship of the association graph, the request semantic features in the interaction request data are projected onto a preset shared semantic space to obtain joint features; The joint features are decomposed into confidence scores, and the confidence scores of the data source quality, the confidence scores of the association graph structure, and the information fidelity during the feature projection process are calculated respectively. The confidence level of the data source quality, the confidence level of the relationship graph structure, and the information fidelity are subjected to nonlinear transformation to output a confidence parameter.

[0009] In a preferred embodiment, the step of performing collaborative measurement on the historical interaction process between the target intelligent agents based on the historical interaction records between the target intelligent agents to obtain interaction performance parameters includes: The interaction records of the target agent within a preset historical time window are obtained, the interaction records are time-series aligned, and the response delay sequence, message routing path, and task load distribution data of the target agent are extracted. Based on the response delay sequence, the average response delay and task throughput of the target agent are calculated. The reciprocal of the average response delay is normalized by min-max, the task throughput is normalized by Z-score, and the normalized results are weighted geometrically averaged to obtain the collaborative timeliness index. Based on the message routing path and the task load distribution data, the path redundancy coefficient in the message routing path and the load balancing Gini coefficient of the task load distribution data among the target agents are calculated. The path redundancy coefficient and the load balancing Gini coefficient are weighted and summed to obtain the collaborative balance index. The collaborative timeliness index and the collaborative balance index are nonlinearly fused to output the interaction efficiency parameter.

[0010] In a preferred embodiment, the step of assessing the group cognitive consistency of the joint features, and constructing a cognitive perturbation signal when the assessment index reaches or exceeds a preset cognitive convergence threshold, includes: The confidence distribution generated internally by the target agent when processing the joint features is obtained, and the confidence distribution is discretized and encoded to obtain the cognitive state vector of the target agent. The Euclidean distance between the cognitive state vectors is calculated. Based on the Euclidean distance, a cognitive divergence matrix is ​​constructed. The cognitive divergence matrix is ​​then aggregated into a fully connected graph to obtain the group cognitive dispersion. The group cognitive dispersion is used as an evaluation index. The formula for calculating the dispersion of group cognition is as follows: ; In the formula, The dispersion of the group's cognition. The total number of target agents in the initial interaction group. For the first The cognitive state vector of a target intelligent agent For the first The cognitive state vector of a target intelligent agent For the first Confidence level of a target intelligent agent For the first The confidence level of each target agent; The group cognitive dispersion is compared with a preset cognitive convergence threshold. When the group cognitive dispersion is less than the cognitive convergence threshold, the group cognition is determined to be in an excessive convergence state, triggering a cognitive perturbation mechanism. When the dispersion of group cognition is greater than or equal to the cognitive convergence threshold, the group cognition is determined to be in a state of sufficient divergence, and the cognitive perturbation mechanism is not triggered. When the cognitive perturbation mechanism is triggered, based on the cognitive divergence matrix, at least one subgroup with the smallest cognitive divergence is extracted from the target agent as a converging subgroup. The centroid vectors of the cognitive state vectors of all agents in the converging subgroup are counted, and the cognitive perturbation signal is constructed with the opposite direction of the centroid vector as the perturbation direction.

[0011] In a preferred embodiment, the step of performing counterfactual reasoning on the target agent based on the cognitive perturbation signal to obtain counterfactual correction data includes: The cognitive disturbance signal is decomposed to extract the disturbance direction component and disturbance intensity component. Extract the corresponding original premises from the interaction request data, and construct alternative premises that contradict the original premises based on the perturbation direction component and the perturbation intensity component, to obtain a set of counterfactual premises; The set of counterfactual premises is broadcast to the target agent, driving the target agent to perform counterfactual deduction, and the counterfactual deduction result and corresponding confidence level of the target agent are obtained. The counterfactual inference results are semantically aligned, mapped to the shared semantic space, and weighted voting aggregation is performed on the counterfactual inference results based on the confidence level to output counterfactual correction data.

[0012] In a preferred embodiment, the step of weightedly fusing the joint features and the counterfactual correction data based on the credibility parameter, comprehensively evaluating the collaborative decision data obtained after weighted fusion with the interaction performance parameter, and updating the topology of the initial interaction group according to the evaluation result to output the target interaction group includes: The credibility parameter is used as the fusion weight of the joint features, and the difference between the unit weight and the credibility parameter is used as the fusion weight of the counterfactual correction data. The joint features and the counterfactual correction data are weighted and summed to obtain collaborative decision data. The collaborative decision-making data and the interaction efficiency parameters are jointly evaluated to obtain a comprehensive efficiency score, which is then used as the evaluation result. Obtain the topology of the initial interaction group, wherein the nodes in the topology correspond to the target agents, and the edges in the topology are used to represent the communication links and their weights between the corresponding two target agents; The overall performance score is compared with a preset performance achievement threshold. When the overall performance score is greater than or equal to the performance achievement threshold, the topology remains unchanged, and the initial interaction group is output as the target interaction group. When the overall performance score is less than the performance threshold, a topology optimization operation is performed on the topology to obtain an optimized topology. Based on the optimized topology, the initial interaction group is updated, and the target interaction group is output.

[0013] In a preferred embodiment, the topology optimization operation includes: The marginal contribution of the target intelligent agent to the collaborative decision-making data is statistically analyzed. Based on the marginal contribution, the link weights of the edges in the topology are positively incentivized or negatively attenuated. Based on the marginal contribution, replace or supplement the intelligent agent from the intelligent agent registration center, and replace or expand the nodes in the topology. Repeat the topology optimization operation until the overall performance score is greater than or equal to the performance threshold.

[0014] To address the above problems, this invention also provides a distributed intelligent agent interaction architecture system for complex scenarios, the system comprising: The scene perception module is used to read interaction request data from the distributed data space and determine the interaction scene type corresponding to the interaction request data; The group initialization module is used to call the target intelligent agent from the intelligent agent registration center according to the interaction scenario type to form an initial interaction group; The collaborative processing module is used to perform feature mapping on the interaction request data based on the semantic association between the target intelligent agents, and to evaluate the confidence of the joint features obtained after feature mapping to obtain a confidence parameter. Based on the historical interaction records between the target intelligent agents, it performs collaborative measurement on the historical interaction process between the target intelligent agents to obtain an interaction performance parameter. The cognitive correction module is used to evaluate the group cognitive consistency of the joint features. When the evaluation index reaches or exceeds the preset cognitive convergence threshold, a cognitive perturbation signal is constructed. Based on the cognitive perturbation signal, counterfactual reasoning is performed on the target intelligent agent to obtain counterfactual correction data. The topology optimization module is used to perform weighted fusion of the joint features and the counterfactual correction data based on the credibility parameter, comprehensively evaluate the collaborative decision data obtained after weighted fusion and the interaction performance parameter, update the topology of the initial interaction group according to the evaluation result, and output the target interaction group.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves adaptive formation and dynamic orchestration of multi-agent groups through scene awareness and on-demand team formation mechanisms, solving the problems of fixed group composition and difficulty in adapting to dynamic changes in task semantics in existing methods. Regarding feature mapping and confidence evaluation, this invention constructs a semantic association graph between target agents, projects interactive request data onto a shared semantic space to obtain joint features, and performs three-dimensional confidence decomposition and nonlinear fusion of the joint features based on data source quality, association graph structure stability, and feature projection information fidelity. This achieves a comprehensive quantitative evaluation of the quality of feature mapping results, significantly improving the reliability and accuracy of the fusion results. In terms of assessing the consistency of group cognition, this invention proposes a confidence-weighted formula for calculating the dispersion of group cognition by statistically analyzing the differences between the cognitive state vectors of each agent and constructing a cognitive divergence matrix. This enables precise quantitative monitoring of the degree of convergence in group cognition. When excessive convergence in group cognition is detected, a cognitive perturbation signal is constructed and each agent is driven to perform independent counterfactual deduction. Alternative premises that contradict the original premises are proactively introduced, effectively breaking the collective cognitive bias, reducing the risk of collective decision-making errors caused by cognitive convergence, and improving the decision-making quality and robustness of multi-agent systems in complex scenarios.

[0016] 2. This invention utilizes a collaborative measurement mechanism based on historical interaction records to comprehensively evaluate the collaborative effectiveness among agents from multiple dimensions, including response timeliness, task load balancing, and communication path rationality, achieving a comprehensive quantitative evaluation of the collaborative process. Building upon this, the invention achieves dynamic adaptive adjustment of the communication topology of the agent group through topology optimization driven by comprehensive effectiveness scoring. This includes positively incentivizing or negatively attenuating communication link weights based on the individual marginal contribution of each agent to the collaborative decision-making results, and accurately calling replacement or supplementary agents from the agent registry based on marginal contribution. This mechanism enables the system to accurately locate inefficient nodes and avoid mistakenly removing potentially valuable agents, forming an adaptive closed loop between decision execution, effectiveness evaluation, and structural optimization. This allows the multi-agent system to dynamically approach the optimal collaborative state during continuous operation, significantly improving the system's long-term operational efficiency and collaborative decision-making capabilities in complex dynamic scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a distributed intelligent agent interaction architecture method for complex scenarios provided in an embodiment of the present invention. Figure 2 This is a functional block diagram of a distributed intelligent agent interaction architecture system for complex scenarios provided in an embodiment of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] This application provides a method for a distributed intelligent agent interaction architecture for complex scenarios. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0021] Reference Figure 1 The diagram shown is a flowchart illustrating a distributed intelligent agent interaction architecture method for complex scenarios provided in an embodiment of the present invention. In this embodiment, the distributed intelligent agent interaction architecture method for complex scenarios includes: S01: Read the interaction request data from the distributed data space and determine the interaction scenario type corresponding to the interaction request data; In this embodiment of the invention, the step of reading interaction request data from the distributed data space and determining the interaction scenario type corresponding to the interaction request data includes: The protocol decapsulation of interactive request data in the distributed data space is performed to obtain the original request message; The original request message is parsed to obtain the request semantic features; The semantic features of the request are matched with a preset scenario library to obtain the interaction scenario type corresponding to the interaction request data.

[0022] When decapsulating interactive request data in a distributed data space, the first step is to determine the corresponding protocol specification based on the communication protocol type used by the interactive request data. Then, the message structure of the interactive request data is parsed field by field according to this protocol specification. The protocol header in the message structure is stripped to remove protocol control fields containing information such as source address, destination address, message sequence number, and checksum, resulting in the original request message after protocol decapsulation.

[0023] When parsing the payload of the original request message, the payload portion of the original request message is decomposed into fields according to a preset data format specification. From the decomposed fields, the request intent field, target object field, time constraint field, and parameter condition field are identified and extracted. The contents of these fields are converted into a numerical sequence of a unified dimension according to preset encoding rules; this numerical sequence is the request semantic feature. If the original request message uses a structured query language (SCL) format, the SCL is parsed to identify operation type keywords, data table names, field names in conditional clauses, and their value ranges. These identification results are converted into request semantic features according to the same encoding rules. If the original request message uses a key-value pair format, all key-value pairs are directly traversed, and the values ​​corresponding to the preset key names are converted into request semantic features according to the encoding rules. If the original request message uses a natural language format, word segmentation and dependency parsing are performed on the natural language to extract core verbs, noun entities, and modifying phrases. The extracted results are converted into request semantic features according to the encoding rules.

[0024] When performing feature matching between the request semantic features and the preset scene library, the preset scene library is pre-built and stored in the system memory before system operation. The construction process of the preset scene library is as follows: collect sample data of various typical interaction scenarios, extract the corresponding sample semantic features for each sample data according to the above-mentioned load parsing method, associate and store the sample semantic features with the scene type label to which the sample belongs, forming a scene template. All scene templates constitute the preset scene library. During matching calculation, calculate the matching degree value between the request semantic features and the sample semantic features of each scene template in the preset scene library, obtaining the same number of matching degree values ​​as the number of scene templates. Sort all matching degree values ​​in descending order, select the scene template with the largest matching degree value, and output the scene type label corresponding to the scene template as the interaction scene type. If the difference between the scene template with the largest matching degree value and the scene template with the second largest matching degree value is less than the preset difference threshold, further compare the degree of match between the key fields in the request semantic features and the corresponding fields of the two scene templates, and output the scene type label of the scene template with the higher degree of match as the interaction scene type.

[0025] The beneficial effects are that through a three-layer processing approach—protocol decapsulation, payload parsing, and scene feature matching—the system can accurately identify the scene type corresponding to an interaction request from a distributed data space. Protocol decapsulation enables the system to be compatible with data access under multiple communication protocols; payload parsing allows the system to uniformly extract request semantic features from request messages of different formats; and scene matching enables the system to automatically match a preset scene template based on the request semantic features and output a scene type label. These mechanisms ensure that upon receiving an interaction request, the system can adaptively determine the scene category of the current task without manual intervention, providing accurate pre-input for subsequent on-demand invocation and targeted orchestration of agents based on scene type.

[0026] S02: Based on the interaction scenario type, call the target intelligent agent from the intelligent agent registration center to form an initial interaction group; In this embodiment of the invention, the step of calling the target intelligent agent from the intelligent agent registration center to form an initial interaction group according to the interaction scenario type includes: Based on the interaction scenario type, target intelligent agents matching the capability requirements are retrieved from the intelligent agent registration center; The target intelligent agent is initialized and arranged to obtain the initial interaction group.

[0027] When retrieving a target agent from the agent registry based on the interaction scenario type to match its capability requirements, the first step is to determine the set of capability requirements needed to complete the interaction request in that scenario. This set of capability requirements is determined as follows: a pre-defined scenario capability mapping table stores the correspondence between each interaction scenario type and the required capability types and minimum requirement levels for each capability. This mapping table is configured by human experts based on the task characteristics of various scenarios and stored in the system memory before system operation. The corresponding capability requirement set is obtained by searching this mapping table based on the interaction scenario type. The agent registry stores the registration information of all available agent instances in the current system. Each registration entry includes a unique agent identifier, the agent's capability type tag, the quantitative rating of each capability, the agent's current load status, and the agent's health status. During retrieval, agents with a normal health status and a current load status below a preset load limit are first selected from the agent registry to form a candidate agent pool. Then, for each agent in the candidate agent pool, its capability type tag and quantitative rating are compared item by item with the capability types and minimum requirement levels in the capability requirement set. An agent is excluded if its capability type label does not contain any capability type from the capability requirement set. An agent is marked as a capability-matching agent if its capability type label contains all capability types from the capability requirement set and the quantified rating of each capability meets or exceeds the corresponding minimum requirement level. After comparing all candidate agents item by item, all agents marked as capability-matching agents constitute the search results, and these agents are output as the target agents.

[0028] When initializing the target agents, all retrieved target agents are treated as member nodes of the initial interaction group, and each target agent is assigned a unique logical identifier within the group. Based on the interaction scenario type, the stages of the task processing flow and their dependencies are determined, and each stage is assigned to a target agent with the corresponding capabilities. For stages requiring collaboration from multiple target agents, one is designated as the primary processing node, and the rest as auxiliary nodes. The primary processing node is responsible for aggregating the processing results from the auxiliary nodes and generating the final output for that stage. The data flow direction between target agents is determined based on a pre-defined communication topology template for the interaction scenario type. This template specifies which agents need to establish communication links and their initial priorities in a particular scenario. Logical communication links are established between two target agents that need to communicate, following the data flow direction, and the initial priority of the communication link is used as its weight. In cases where multiple data streams converge on the same target agent, a data buffer is configured for that target agent, and the data is queued and processed sequentially according to the link weights of each data stream from highest to lowest. After completing the above logical identifier allocation, phase assignment, primary and secondary node designation and communication link establishment, all target intelligent agents and their communication relationships together constitute the initial interaction group.

[0029] The beneficial effects include enabling on-demand adaptive assembly of agent groups through scenario-driven capability requirement matching and dynamic retrieval from the agent registry center, avoiding capability mismatch issues associated with static configurations or fixed groups in different scenarios. Multi-level initialization orchestration, including logical identifier allocation, stage assignment, primary / secondary node designation, and communication link establishment for target agents, ensures that the initial interaction group possesses a structured collaborative relationship and a clear data flow direction, enabling each agent to execute its tasks orderly within the group according to a unified logical framework. This mechanism ensures that the system can quickly assemble agent groups that precisely match the capability requirements of the interaction scenario after acquiring the scenario type, providing a clear organizational foundation with division of labor and communication topology for subsequent feature mapping, confidence assessment, and collaborative decision-making.

[0030] S03: Based on the semantic association between the target intelligent agents, perform feature mapping on the interaction request data, and evaluate the confidence of the joint features obtained after feature mapping to obtain a confidence parameter. Based on the historical interaction records between the target intelligent agents, perform collaborative measurement on the historical interaction process between the target intelligent agents to obtain an interaction performance parameter. In this embodiment of the invention, the step of performing feature mapping on the interaction request data based on the semantic association between the target intelligent agents, and evaluating the confidence of the joint features obtained after feature mapping to obtain a confidence parameter, includes: The similarity between the semantic feature vectors of the target intelligent agent is calculated, and a semantic association matrix is ​​constructed based on the similarity. The semantic association matrix is ​​sparsified to obtain the association graph between the target intelligent agents; Based on the adjacency relationship of the association graph, the request semantic features in the interaction request data are projected onto a preset shared semantic space to obtain joint features; The joint features are decomposed into confidence scores, and the confidence scores of the data source quality, the confidence scores of the association graph structure, and the information fidelity during the feature projection process are calculated respectively. The confidence level of the data source quality, the confidence level of the relationship graph structure, and the information fidelity are subjected to nonlinear transformation to output a confidence parameter.

[0031] The step of performing collaborative measurement on the historical interaction process between the target intelligent agents based on the historical interaction records between the target intelligent agents to obtain interaction performance parameters includes: The interaction records of the target agent within a preset historical time window are obtained, the interaction records are time-series aligned, and the response delay sequence, message routing path, and task load distribution data of the target agent are extracted. Based on the response delay sequence, the average response delay and task throughput of the target agent are calculated. The reciprocal of the average response delay is normalized by min-max, the task throughput is normalized by Z-score, and the normalized results are weighted geometrically averaged to obtain the collaborative timeliness index. Based on the message routing path and the task load distribution data, the path redundancy coefficient in the message routing path and the load balancing Gini coefficient of the task load distribution data among the target agents are calculated. The path redundancy coefficient and the load balancing Gini coefficient are weighted and summed to obtain the collaborative balance index. The collaborative timeliness index and the collaborative balance index are nonlinearly fused to output the interaction efficiency parameter.

[0032] When calculating the similarity between the semantic feature vectors of target agents, each target agent stores a corresponding functional description semantic vector in the agent registry center. This vector is a numerical representation of the agent's functional semantics. The functional description semantic vectors of all target agents are read from the agent registry center. For any two target agents, the cosine similarity between their functional description semantic vectors is calculated as the semantic similarity between the two target agents. After calculating the cosine similarity for each pair of all target agents, the semantic similarity between any two target agents is filled into the corresponding matrix positions using the target agents as row and column indices. This matrix is ​​the semantic association matrix.

[0033] When sparsifying the semantic association matrix, a preset similarity threshold is set. This threshold is configured and stored in the system memory before system operation based on the semantic association sensitivity requirements in multi-agent collaboration scenarios. The similarity values ​​at all off-diagonal positions in the semantic association matrix are traversed. Each similarity value is compared with the preset similarity threshold. If the similarity value at a position is less than the threshold, the similarity value at that position is set to zero. If the similarity value at a position is greater than or equal to the threshold, the similarity value at that position remains unchanged. After traversal, only similarity values ​​greater than or equal to the threshold are retained in the semantic association matrix; all similarity values ​​less than the threshold are set to zero, resulting in the sparsified semantic association matrix. Using the sparsified semantic association matrix as an adjacency matrix, with the target agents as nodes and the non-zero elements in the matrix as the weights of the edges between nodes, a relationship graph between the target agents is constructed.

[0034] When projecting the request semantic features in the interaction request data onto a preset shared semantic space based on the adjacency relationships in the association graph, the preset shared semantic space is defined by a pre-constructed semantic projection basis vector set. This semantic projection basis vector set is obtained through statistical analysis of a large-scale general semantic corpus and stored in the system memory before system operation. The adjacency relationships of each target agent are read from the association graph, indicating which other target agents have non-zero weighted edges with the current target agent. For each target agent, the semantic feature vectors of all neighboring nodes in its adjacency relationships are weighted and aggregated with the semantic feature vector of the current target agent itself, with the weights being the weight values ​​of the corresponding edges in the association graph, to obtain the local semantic context vector of that target agent. The local semantic context vectors of all target agents are arranged and combined according to the order of the target agents to form a joint feature. This joint feature simultaneously contains the semantic information of each target agent itself and the semantic information of neighboring agents introduced through the association graph.

[0035] When performing confidence decomposition on the joint features, the data source portion originating from the interactive request data is extracted from the joint features. The completeness, format compliance, and missing field status of this data source portion are checked item by item, and a data source quality confidence score is generated based on the check results. The structural portion originating from the association graph is extracted from the joint features. The total number of edges, the number of connected components, and the degree distribution of each node in the association graph are statistically analyzed. The structural stability of the association graph is judged based on the statistical results, and an association graph structure confidence score is generated. The mapping portion originating from the projection process is extracted from the joint features. By comparing the consistency of the request semantic features before and after projection with the original request semantic features, the information fidelity during the feature projection process is generated.

[0036] When performing nonlinear transformations on data source quality confidence, relational graph structure confidence, and information fidelity, these three values ​​are used as inputs, and a comprehensive calculation is performed according to a preset nonlinear transformation rule. This preset nonlinear transformation rule is configured and stored in the system memory before system operation, based on the accuracy requirements of multi-source confidence fusion. After the nonlinear transformation, a single value is output, which is the confidence parameter.

[0037] When acquiring interaction records of target agents within a preset historical time window, the length of the preset historical time window is dynamically determined by the system runtime based on the task type. This length value is configured and stored in the system memory during system initialization. Each interaction event record occurring between all target agents within the preset historical time window is read from the system log storage area. Each interaction event record includes an event timestamp, sender agent identifier, receiver agent identifier, request summary, response summary, response time, and the task identifier to which this interaction belongs. All read interaction event records are sorted according to their event timestamps to obtain a sequence of interaction events arranged chronologically. For each record in the interaction event sequence, the clock source upon which its timestamp depends is checked. When there is a discrepancy between the clock sources of different agents, the timestamp of each record is offset and corrected using the system master clock as the reference, ensuring that all interaction event records have a unified reference on the timeline. From the corrected interaction event sequence, the set of all response times for each target agent when acting as a responder is extracted, forming the response delay sequence for that agent. From the corrected sequence of interaction events, trace the flow of each request message from one agent to its final processing, recording all flow links to form a message routing path. From the corrected sequence of interaction events, count the number of tasks processed by each target agent within a preset historical time window, summarizing the task counts of all target agents to form task load distribution data.

[0038] When calculating the average response latency and task throughput of a target agent based on the response latency sequence, the average response latency of the target agent is obtained by arithmetically averaging all response times in the response latency sequence for each target agent. For each target agent, the total number of tasks processed by the agent within a preset historical time window is counted, and this total number of tasks is divided by the length of the preset historical time window to obtain the task throughput of the target agent.

[0039] When performing minimum-maximum normalization on the reciprocal of the average response latency, first, take the reciprocal of the average response latency for each target agent to obtain its response speed value. Find the minimum response speed value among all target agent response speed values ​​as the global minimum speed value, and the maximum response speed value as the global maximum speed value. For each target agent's response speed value, subtract the global minimum speed value to obtain the difference, then divide the difference by the difference between the global maximum and global minimum speed values; the resulting ratio is the normalized response speed value for that target agent. When performing Z-score normalization on task throughput, calculate the average task throughput of all target agents and calculate the dispersion value for all task throughput. For each target agent's task throughput, subtract the average to obtain the difference, then divide the difference by the dispersion value; the resulting ratio is the normalized task throughput value for that target agent. For each target agent, a weighted geometric mean is calculated using the normalized response speed and normalized task throughput values. Pre-defined weighting coefficients are assigned to each value during the weighted geometric mean process. These weighting coefficients are configured and stored in the system memory before system operation, based on the emphasis of the timeliness evaluation. The normalized response speed value and normalized task throughput value of each target agent are both raised to the power of their respective weighting coefficients. The results of these two power operations are multiplied, and the product is then taken as the square root of the sum of the weighting coefficients to obtain the individual timeliness score for that target agent. The arithmetic mean of the individual timeliness scores for all target agents is then calculated to obtain the collaborative timeliness index.

[0040] When calculating the path redundancy coefficient in message routing paths based on message routing paths and task load distribution data, for each message routing path, the number of agent nodes the message passes through from the starting agent to the final target agent is determined, and this number is subtracted by one to obtain the actual hop count. The theoretical minimum hop count for a message type in an interaction scenario is determined according to a preset shortest hop count table. This preset shortest hop count table is configured and stored in the system memory before system operation based on the direct communication reachability between any two agents in the system network. The ratio of the actual hop count to the theoretical minimum hop count is used as the path redundancy coefficient of the message routing path. The arithmetic mean of the path redundancy coefficients of all message routing paths is used to obtain the statistical value of the path redundancy coefficient. When calculating the load balancing Gini coefficient based on task load distribution data, the number of tasks for all target agents is arranged in ascending order. The cumulative proportion is calculated for each item in the sorted sequence. The area of ​​the polygon corresponding to the Lorenz curve is plotted based on the cumulative proportion. The area of ​​this polygon is divided by the area enclosed by the perfect balance line, and the ratio is used as the load balancing Gini coefficient.

[0041] When performing a weighted summation of the path redundancy coefficient and the load balancing Gini coefficient, preset balance weight coefficients are configured for both coefficients. These weight coefficients are configured and stored in the system memory before system operation, based on the emphasis of the collaborative balance evaluation. The path redundancy coefficient is multiplied by its corresponding balance weight coefficient to obtain the weighted path redundancy value, and the load balancing Gini coefficient is multiplied by its corresponding balance weight coefficient to obtain the weighted load balance value. The weighted path redundancy value and the weighted load balance value are then added together to obtain the collaborative balance index.

[0042] When performing nonlinear fusion of the collaborative timeliness index and the collaborative balance index, the two values ​​of the collaborative timeliness index and the collaborative balance index are used as inputs and are comprehensively calculated according to the preset nonlinear fusion rules. The preset nonlinear fusion rules are configured based on the comprehensive mapping relationship of performance evaluation and stored in the system memory before the system runs. After nonlinear fusion, a single value is output, which is the interaction performance parameter.

[0043] The beneficial effects are that, through semantic association-driven feature mapping and three-dimensional confidence decomposition, the system can comprehensively evaluate the data source quality, association graph structure stability, and projection information fidelity of joint features at the multi-agent collaboration level and output confidence parameters. Simultaneously, through temporal alignment of historical interaction records and multi-dimensional collaborative metrics, it comprehensively evaluates the collaborative efficiency of the agent group from the perspectives of response timeliness and load balancing, and outputs interaction performance parameters. This mechanism achieves a dual-track quantitative evaluation of the quality of collaborative decision-making results and the efficiency of the collaborative process, providing reliable data support for subsequent weighted fusion and topology optimization.

[0044] S04: Perform a group cognitive consistency assessment on the joint features. When the assessment index reaches or exceeds a preset cognitive convergence threshold, construct a cognitive perturbation signal. Based on the cognitive perturbation signal, perform counterfactual reasoning on the target agent to obtain counterfactual correction data. In this embodiment of the invention, the step of assessing the group cognitive consistency of the joint features, and constructing a cognitive perturbation signal when the assessment index reaches or exceeds a preset cognitive convergence threshold, includes: The confidence distribution generated internally by the target agent when processing the joint features is obtained, and the confidence distribution is discretized and encoded to obtain the cognitive state vector of the target agent. The Euclidean distance between the cognitive state vectors is calculated. Based on the Euclidean distance, a cognitive divergence matrix is ​​constructed. The cognitive divergence matrix is ​​then aggregated into a fully connected graph to obtain the group cognitive dispersion. The group cognitive dispersion is used as an evaluation index. The formula for calculating the dispersion of group cognition is as follows: ; In the formula, The dispersion of the group's cognition. The total number of target agents in the initial interaction group. For the first The cognitive state vector of a target intelligent agent For the first The cognitive state vector of a target intelligent agent For the first Confidence level of a target intelligent agent For the first The confidence level of each target agent; The group cognitive dispersion is compared with a preset cognitive convergence threshold. When the group cognitive dispersion is less than the cognitive convergence threshold, the group cognition is determined to be in an excessive convergence state, triggering a cognitive perturbation mechanism. When the dispersion of group cognition is greater than or equal to the cognitive convergence threshold, the group cognition is determined to be in a state of sufficient divergence, and the cognitive perturbation mechanism is not triggered. When the cognitive perturbation mechanism is triggered, based on the cognitive divergence matrix, at least one subgroup with the smallest cognitive divergence is extracted from the target agent as a converging subgroup. The centroid vectors of the cognitive state vectors of all agents in the converging subgroup are counted, and the cognitive perturbation signal is constructed with the opposite direction of the centroid vector as the perturbation direction.

[0045] The step of performing counterfactual reasoning on the target agent based on the cognitive perturbation signal to obtain counterfactual correction data includes: The cognitive disturbance signal is decomposed to extract the disturbance direction component and disturbance intensity component. Extract the corresponding original premises from the interaction request data, and construct alternative premises that contradict the original premises based on the perturbation direction component and the perturbation intensity component, to obtain a set of counterfactual premises; The set of counterfactual premises is broadcast to the target agent, driving the target agent to perform counterfactual deduction, and the counterfactual deduction result and corresponding confidence level of the target agent are obtained. The counterfactual inference results are semantically aligned, mapped to the shared semantic space, and weighted voting aggregation is performed on the counterfactual inference results based on the confidence level to output counterfactual correction data.

[0046] When acquiring the confidence distribution generated internally by the target agent when processing joint features, each target agent, upon receiving the joint features, analyzes the joint features based on its internal knowledge base and reasoning logic, and outputs the agent's probability judgment of the result corresponding to the joint features in different possible categories. The probability values ​​of all possible categories together constitute the agent's confidence distribution. When discretizing the confidence distribution, the probability value of each category in the confidence distribution is converted into a corresponding discrete coded value according to a preset encoding mapping table. This encoding mapping table is configured according to the feature dimensions of the confidence distribution and stored in the system memory before system operation. The discrete coded values ​​of all categories are arranged in category order to form the cognitive state vector of the target agent.

[0047] When calculating the Euclidean distance between cognitive state vectors, for any two target agents' cognitive state vectors, the discrete encoded values ​​at corresponding positions of the two vectors are subtracted to obtain the difference. Each difference is squared, summed, and the square root of the sum is taken. The resulting value is the Euclidean distance between the two cognitive state vectors. After calculating the Euclidean distance for all target agents pairwise, the Euclidean distance between any two target agents is filled into the corresponding matrix positions using the target agents as row and column indices. This matrix is ​​the cognitive divergence matrix. When performing fully connected graph aggregation on the cognitive divergence matrix, all off-diagonal elements in the cognitive divergence matrix are comprehensively calculated according to a preset aggregation rule. This preset aggregation rule is configured and stored in the system memory before system operation based on the sensitivity requirements of group cognitive evaluation. After fully connected graph aggregation, a single value is output, which is the group cognitive dispersion, and this group cognitive dispersion is used as the evaluation index. The group cognitive dispersion is calculated as follows: For each target agent in the initial interaction group, the cognitive state vector and confidence level of the target agent are obtained. For any two different target agents, the directional consistency between their cognitive state vectors is calculated. The directional consistency is calculated by dividing the inner product of the two cognitive state vectors by the product of their respective magnitudes. The directional divergence is obtained by subtracting the directional consistency from the constant 1. The directional divergence is multiplied by a weight value, which is calculated by dividing the product of the confidence levels of the two target agents by the product of the average confidence level of all target agents and the average confidence level of the remaining agents after removing the current reference agent. All different target agents are combined in pairs, and the results calculated in the above manner are accumulated. The accumulated result is divided by the product of the total number of target agents in the initial interaction group and the total number of target agents minus one to obtain the group cognitive dispersion. In group cognitive dispersion, the cognitive state vector originates from the output of the discretization encoding step, the confidence level comes from the maximum probability value corresponding to the final judgment result in the confidence level distribution generated within the target agent, and the total number of target agents comes from the number of members in the initial interaction group. A smaller group cognitive dispersion value indicates that the cognitive states of each target agent are more consistent in direction, and that each target agent has a higher degree of confidence in this consistent cognition. A larger value indicates greater divergence in the cognitive states of each target agent in direction, or a lower degree of confidence. The directional divergence component of group cognitive dispersion decreases as the directions of two cognitive state vectors tend to be consistent, and increases as the directions of two cognitive state vectors tend to be opposite. The confidence weight component increases as the confidence levels of both target agents are higher than the group average, and decreases as the confidence level of either target agent is lower than the group average.

[0048] When comparing the group's cognitive dispersion with a preset cognitive convergence threshold, the preset threshold is configured and stored in the system memory before system operation, based on the system's sensitivity requirements to excessive convergence. When the group's cognitive dispersion is less than the cognitive convergence threshold, the group's cognition is determined to be in an excessive convergence state, triggering the cognitive perturbation mechanism. When the group's cognitive dispersion is greater than or equal to the cognitive convergence threshold, the group's cognition is determined to be in a sufficiently divergent state, and the cognitive perturbation mechanism is not triggered.

[0049] When the cognitive perturbation mechanism is triggered, at least one subgroup with the smallest cognitive divergence is extracted from the target agents based on the cognitive divergence matrix as a converging subgroup. Specifically, the agent combination with the smallest sum of Euclidean distances among them is found from the cognitive divergence matrix, and all agents in this combination constitute the converging subgroup. The average value of the cognitive state vector of all agents in the converging subgroup is calculated in each dimension, and the average values ​​of all dimensions are combined into a vector, which is the centroid vector. The inverse value of each dimension of the centroid vector is taken to obtain the inverse vector, and the cognitive perturbation signal is constructed using this inverse vector as the perturbation direction.

[0050] When decomposing the cognitive disturbance signal, it is broken down into two independent components: the first component is a vector indicating the direction of the disturbance, and the second component is a scalar indicating the magnitude of the disturbance. This scalar is configured and stored in the system memory based on historical disturbance response statistics before system operation. The corresponding original preconditions are extracted from the interaction request data. Specifically, the conditional clauses, time constraints, and hypothetical statements in the interaction request data are identified and extracted, and the extracted results are used as the original preconditions. Alternative preconditions that contradict the original preconditions are constructed based on the disturbance direction and intensity components. Specifically, the structure of the original preconditions remains unchanged, and the key attribute values ​​in the original preconditions are adjusted according to the direction indicated by the disturbance direction component. The adjustment magnitude is determined by the intensity indicated by the disturbance intensity component. The adjusted preconditions are used as alternative preconditions, and all alternative preconditions constitute a counterfactual precondition set.

[0051] When broadcasting the counterfactual premise set to the target agents, each alternative premise in the counterfactual premise set is sent to each target agent individually, ensuring that each target agent receives the complete counterfactual premise set. When driving the target agents to perform counterfactual deduction, each target agent takes each alternative premise in the counterfactual premise set as input, replaces the original premise with the alternative premise, and re-executes its reasoning process under the alternative premise, obtaining the deduction result under the alternative premise and the corresponding confidence level. The deduction results under all alternative premises constitute the counterfactual deduction result of the target agent, and the confidence levels under all alternative premises constitute the corresponding confidence level distribution.

[0052] When semantically aligning the counterfactual inference results, the counterfactual inference results output by all target agents are formatted according to a unified semantic expression standard, ensuring that the counterfactual inference results output by different target agents have a consistent semantic representation. When mapping the semantically aligned counterfactual inference results to a shared semantic space, the shared semantic space is the same shared semantic space used when projecting the request semantic features in the interaction request data to a preset shared semantic space. For each semantically aligned counterfactual inference result, it is converted into a vector representation in the shared semantic space according to the same projection rules. When performing weighted voting aggregation of the counterfactual inference results based on confidence levels, for each counterfactual inference result, its corresponding confidence level is used as the voting weight. All counterfactual inference results are weighted according to their voting weights, and the counterfactual inference result with the highest weighted score is selected as the aggregated result output. This aggregated result is the counterfactual correction data.

[0053] The beneficial effect is that by measuring the pairwise differences and aggregating fully connected graphs of the cognitive state vectors of the target agents, the quantitative calculation of the dispersion of group cognition is achieved. This enables the system to automatically identify excessive convergence states and trigger a cognitive perturbation mechanism when group cognition reaches or exceeds the cognitive convergence threshold. By decomposing the perturbation signal, constructing alternative preconditions that contradict the original preconditions, driving each agent to independently perform counterfactual inference under the alternative preconditions, and performing semantic alignment and weighted voting aggregation on the inference results, counterfactual correction data with cognitive correction function is generated. The above mechanism enables the system to proactively introduce alternative hypotheses from multiple perspectives when group cognition faces the risk of collective bias, effectively breaking the decision-making deadlock caused by cognitive convergence, and providing a source of corrective data with adversarial and diverse characteristics for subsequent weighted fusion.

[0054] S05: Based on the credibility parameter, the joint features and the counterfactual correction data are weighted and fused. The collaborative decision data obtained after weighted fusion is comprehensively evaluated with the interaction performance parameter. Based on the evaluation result, the topology of the initial interaction group is updated, and the target interaction group is output.

[0055] In this embodiment of the invention, the step of weightedly fusing the joint features and the counterfactual correction data based on the credibility parameter, comprehensively evaluating the collaborative decision data obtained after weighted fusion with the interaction performance parameter, and updating the topology of the initial interaction group according to the evaluation result to output the target interaction group includes: The credibility parameter is used as the fusion weight of the joint features, and the difference between the unit weight and the credibility parameter is used as the fusion weight of the counterfactual correction data. The joint features and the counterfactual correction data are weighted and summed to obtain collaborative decision data. The collaborative decision-making data and the interaction efficiency parameters are jointly evaluated to obtain a comprehensive efficiency score, which is then used as the evaluation result. Obtain the topology of the initial interaction group, wherein the nodes in the topology correspond to the target agents, and the edges in the topology are used to represent the communication links and their weights between the corresponding two target agents; The overall performance score is compared with a preset performance achievement threshold. When the overall performance score is greater than or equal to the performance achievement threshold, the topology remains unchanged, and the initial interaction group is output as the target interaction group. When the overall performance score is less than the performance threshold, a topology optimization operation is performed on the topology to obtain an optimized topology. Based on the optimized topology, the initial interaction group is updated, and the target interaction group is output.

[0056] The topology optimization operation includes: The marginal contribution of the target intelligent agent to the collaborative decision-making data is statistically analyzed. Based on the marginal contribution, the link weights of the edges in the topology are positively incentivized or negatively attenuated. Based on the marginal contribution, replace or supplement the intelligent agent from the intelligent agent registration center, and replace or expand the nodes in the topology. Repeat the topology optimization operation until the overall performance score is greater than or equal to the performance threshold.

[0057] When using the credibility parameter as the fusion weight of the joint features, the value is directly read from the output of the credibility parameter and used as the weight coefficient of the joint features in the weighted summation. The difference between the unit weight and the credibility parameter is used as the weight coefficient of the counterfactual correction data in the weighted summation, and the unit weight is a fixed value of one. When performing a weighted summation of the joint features and the counterfactual correction data, the value of each dimension of the joint features is multiplied by the credibility parameter to obtain the weighted joint feature value, and the value of each dimension of the counterfactual correction data is multiplied by the difference between the unit weight and the credibility parameter to obtain the weighted counterfactual correction value. The weighted joint feature value and the weighted counterfactual correction value are added one by one in the corresponding dimensions to obtain the weighted summation result for each dimension. The weighted summation results of all dimensions together constitute the collaborative decision data.

[0058] When jointly evaluating collaborative decision-making data and interaction performance parameters, the collaborative decision-making data is parsed according to its data format, extracting the decision result field, decision confidence field, and decision completeness field. These fields are then comprehensively calculated using preset evaluation weights to obtain a decision quality score. These preset evaluation weights are configured and stored in the system memory before system operation, based on the emphasis of the decision quality evaluation. Interaction performance parameters are treated as an independent evaluation dimension and compared with preset performance level classification standards to obtain an interaction performance level score. These preset performance level classification standards are defined and stored in the system memory before system operation, based on historical statistical data of collaborative performance. The decision quality score and interaction performance level score are then fused according to preset comprehensive evaluation rules. These rules are configured and stored in the system memory before system operation, based on the balance requirements of comprehensive performance evaluation. After fusion, a single numerical value is output, which is the comprehensive performance score, and this score serves as the evaluation result.

[0059] When obtaining the topology of the initial interaction group, the topology data generated and stored during the initial orchestration phase is read from the system storage area. This topology data includes a set of nodes and a set of edges. Each node in the node set corresponds to a target agent in the initial interaction group, and the node information includes the target agent's unique logical identifier and capability type label. Each edge in the edge set corresponds to a communication link between two target agents, and the edge information includes the unique logical identifiers of the two target agents connected by the communication link and the link weight of the communication link.

[0060] When comparing the overall performance score with a preset performance threshold, the preset performance threshold is configured and stored in the system memory before system operation based on the system's requirements for collaborative performance. When the overall performance score is greater than or equal to the performance threshold, the link weights of all nodes and edges in the current topology remain unchanged, and the current initial interaction group is directly output as the target interaction group. When the overall performance score is less than the performance threshold, a topology optimization operation is performed on the current topology. After the topology optimization operation is completed, the optimized topology is obtained. Based on the optimized topology, the node correspondences, edge connections, and link weights in the initial interaction group are updated, and the updated initial interaction group is output as the target interaction group.

[0061] When calculating the marginal contribution of a target agent to the collaborative decision-making data, for each target agent, it is temporarily removed from the current initial interaction group. Under the new group configuration after removing the agent, the process of acquiring joint features and generating collaborative decision-making data is re-executed to obtain new collaborative decision-making data. The original collaborative decision-making data is compared with the new collaborative decision-making data, and the change in the decision result field of the new collaborative decision-making data relative to the original collaborative decision-making data is calculated. This change is the marginal contribution of the target agent. This temporary removal and change calculation operation is performed sequentially for all target agents to obtain the marginal contribution of each target agent.

[0062] When applying positive incentives or negative attenuations to the link weights of edges in the topology based on marginal contribution, the average marginal contribution of all target agents is calculated. For each target agent, its marginal contribution is compared to the average. If the marginal contribution of a target agent is greater than the average, the link weights of all communication links associated with that agent are increased according to a preset incentive ratio. This preset incentive ratio is configured and stored in the system memory before system operation, based on the required adjustment step size for the link weights. If the marginal contribution of a target agent is less than the average, the link weights of all communication links associated with that agent are decreased according to a preset attenuation ratio. This preset attenuation ratio is configured and stored in the system memory before system operation, based on the required adjustment step size for the link weights. If the marginal contribution of a target agent equals the average, the link weights of all communication links associated with that agent remain unchanged.

[0063] When calling a replacement or supplementary agent from the agent registry based on marginal contribution, the target agent with the smallest marginal contribution is determined. A replacement agent with the same capability type label as the target agent with the smallest marginal contribution and a lower current load state is retrieved from the agent registry and called as the replacement agent. If no replacement agent meeting the above conditions exists in the agent registry, a supplementary capability agent with a current load state lower than a preset load limit and whose capability type label does not belong to the existing capability type set in the current initial interaction group is retrieved from the agent registry and called as the supplementary agent. When replacing nodes in the topology, the node corresponding to the target agent with the smallest marginal contribution is removed from the node set, the node corresponding to the replacement agent is added to the node set, and all edges connected to the removed node are transferred to the replacement agent node, maintaining the edge connections unchanged during the transfer. When expanding the nodes in the topology, the nodes corresponding to the supplementary agents are added to the node set. Based on the semantic similarity between the capability type labels of the supplementary agents and the capability type labels of the existing nodes, a communication link is established between the supplementary agent nodes and the existing nodes, and an initial link weight is assigned.

[0064] When repeatedly performing topology optimization operations, after each node replacement or expansion and corresponding link weight adjustment, the comprehensive performance score of the current topology is recalculated using the aforementioned method of jointly evaluating collaborative decision-making data and interaction performance parameters. The recalculated comprehensive performance score is then compared to a preset performance threshold. If the recalculated comprehensive performance score is still less than the performance threshold, the next round of topology optimization is performed, and marginal contribution statistics, link weight adjustments, and node replacement or expansion are repeated. When the recalculated comprehensive performance score is greater than or equal to the performance threshold, the topology optimization operation is stopped, and the current topology is output as the optimized topology.

[0065] The beneficial effects include: adaptive integration of joint features and counterfactual correction data through weighted fusion driven by credibility parameters; precise determination of topology optimization trigger conditions through comparison of comprehensive performance scores and performance achievement thresholds; refined adaptive optimization of the agent group structure through marginal contribution statistics, dynamic adjustment of link weights, and node replacement and expansion; and a complete closed-loop iterative mechanism formed by repeatedly executing topology optimization operations until the comprehensive performance score is achieved. These mechanisms enable the system to automatically adjust the initial interaction group's membership and communication topology based on real-time feedback from credibility and interaction performance parameters, ultimately outputting a performance-verified target interaction group. This ensures the collaborative decision-making quality and long-term operational stability of the multi-agent system in complex dynamic scenarios.

[0066] like Figure 2 The diagram shown is a functional block diagram of a distributed intelligent agent interaction architecture system for complex scenarios provided in an embodiment of the present invention.

[0067] The distributed intelligent agent interaction architecture system for complex scenarios described in this invention can be installed in electronic devices. Depending on the functions implemented, the distributed intelligent agent interaction architecture system for complex scenarios may include a scene perception module, a group initialization module, a collaborative processing module, a cognitive correction module, and a topology optimization module. The modules described in this invention can also be called units, referring to a series of computer program segments that can be executed by the processor of an electronic device and perform a fixed function, stored in the memory of the electronic device.

[0068] In this embodiment, the functions of the joint feature module / unit are as follows: The scene perception module is used to read interaction request data from the distributed data space and determine the interaction scene type corresponding to the interaction request data. The group initialization module is used to call the target intelligent agent from the intelligent agent registration center according to the interaction scenario type to form an initial interaction group; The collaborative processing module is used to perform feature mapping on the interaction request data according to the semantic association between the target intelligent agents, and to evaluate the confidence of the joint features obtained after feature mapping to obtain a confidence parameter. Based on the historical interaction records between the target intelligent agents, it performs collaborative measurement on the historical interaction process between the target intelligent agents to obtain an interaction performance parameter. The cognitive correction module is used to evaluate the group cognitive consistency of the joint features. When the evaluation index reaches or exceeds the preset cognitive convergence threshold, a cognitive perturbation signal is constructed. Based on the cognitive perturbation signal, counterfactual reasoning is performed on the target intelligent agent to obtain counterfactual correction data. The topology optimization module is used to perform weighted fusion of the joint features and the counterfactual correction data based on the credibility parameter, comprehensively evaluate the collaborative decision data obtained after weighted fusion and the interaction performance parameter, and update the topology of the initial interaction group according to the evaluation result, and output the target interaction group.

[0069] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0070] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0071] Furthermore, in the embodiments of the joint features of the present invention, the joint feature functional modules can be integrated into one processing unit, or the joint feature units can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0072] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0073] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A distributed intelligent agent interaction architecture method for complex scenarios, characterized in that, The method includes: S01: Read the interaction request data from the distributed data space and determine the interaction scenario type corresponding to the interaction request data; S02: Based on the interaction scenario type, call the target intelligent agent from the intelligent agent registration center to form an initial interaction group; S03: Based on the semantic association between the target intelligent agents, perform feature mapping on the interaction request data, and evaluate the confidence of the joint features obtained after feature mapping to obtain a confidence parameter. Based on the historical interaction records between the target intelligent agents, perform collaborative measurement on the historical interaction process between the target intelligent agents to obtain an interaction performance parameter. S04: Perform a group cognitive consistency assessment on the joint features. When the assessment index reaches or exceeds a preset cognitive convergence threshold, construct a cognitive perturbation signal. Based on the cognitive perturbation signal, perform counterfactual reasoning on the target agent to obtain counterfactual correction data. S05: Based on the credibility parameter, the joint features and the counterfactual correction data are weighted and fused, the collaborative decision data obtained after weighted fusion is comprehensively evaluated with the interaction performance parameter, and the topology of the initial interaction group is updated according to the evaluation result, and the target interaction group is output. A group cognitive consistency assessment is performed on the joint features. When the assessment index reaches or exceeds a preset cognitive convergence threshold, a cognitive perturbation signal is constructed, including: The confidence distribution generated internally by the target agent when processing the joint features is obtained, and the confidence distribution is discretized and encoded to obtain the cognitive state vector of the target agent. The Euclidean distance between the cognitive state vectors is calculated. Based on the Euclidean distance, a cognitive divergence matrix is ​​constructed. The cognitive divergence matrix is ​​then aggregated into a fully connected graph to obtain the group cognitive dispersion. The group cognitive dispersion is used as an evaluation index. The formula for calculating the dispersion of group cognition is as follows: ； In the formula, The dispersion of the group's cognition. The total number of target agents in the initial interaction group. For the first The cognitive state vector of a target intelligent agent For the first The cognitive state vector of a target agent For the first Confidence level of a target intelligent agent For the first The confidence level of each target agent; The group cognitive dispersion is compared with a preset cognitive convergence threshold. When the group cognitive dispersion is less than the cognitive convergence threshold, the group cognition is determined to be in an excessive convergence state, triggering a cognitive perturbation mechanism. When the dispersion of group cognition is greater than or equal to the cognitive convergence threshold, the group cognition is determined to be in a state of sufficient divergence, and the cognitive perturbation mechanism is not triggered. When the cognitive perturbation mechanism is triggered, based on the cognitive divergence matrix, at least one subgroup with the smallest cognitive divergence is extracted from the target agent as a converging subgroup. The centroid vectors of the cognitive state vectors of all agents in the converging subgroup are counted, and the cognitive perturbation signal is constructed with the opposite direction of the centroid vector as the perturbation direction.

2. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 1, characterized in that, The step of reading interaction request data from the distributed data space and determining the interaction scenario type corresponding to the interaction request data includes: The protocol decapsulation of interactive request data in the distributed data space is performed to obtain the original request message; The original request message is parsed to obtain the request semantic features; The semantic features of the request are matched with a preset scenario library to obtain the interaction scenario type corresponding to the interaction request data.

3. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 1, characterized in that, The step of calling the target intelligent agent from the intelligent agent registration center to form an initial interaction group according to the interaction scenario type includes: Based on the interaction scenario type, target intelligent agents matching the capability requirements are retrieved from the intelligent agent registration center; The target intelligent agent is initialized and arranged to obtain the initial interaction group.

4. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 1, characterized in that, The step of performing feature mapping on the interaction request data based on the semantic relationships between the target agents, and evaluating the confidence level of the joint features obtained after feature mapping to obtain a confidence parameter includes: The similarity between the semantic feature vectors of the target intelligent agent is calculated, and a semantic association matrix is ​​constructed based on the similarity. The semantic association matrix is ​​sparsified to obtain the association graph between the target intelligent agents; Based on the adjacency relationship of the association graph, the request semantic features in the interaction request data are projected onto a preset shared semantic space to obtain joint features; The joint features are decomposed into confidence scores, and the confidence scores of the data source quality, the confidence scores of the association graph structure, and the information fidelity during the feature projection process are calculated respectively. The confidence level of the data source quality, the confidence level of the relationship graph structure, and the information fidelity are subjected to nonlinear transformation to output a confidence parameter.

5. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 1, characterized in that, The step of performing collaborative measurement on the historical interaction process between the target intelligent agents based on the historical interaction records between the target intelligent agents to obtain interaction performance parameters includes: The interaction records of the target agent within a preset historical time window are obtained, the interaction records are time-series aligned, and the response delay sequence, message routing path, and task load distribution data of the target agent are extracted. Based on the response delay sequence, the average response delay and task throughput of the target agent are calculated. The reciprocal of the average response delay is normalized by min-max, the task throughput is normalized by Z-score, and the normalized results are weighted geometrically averaged to obtain the collaborative timeliness index. Based on the message routing path and the task load distribution data, the path redundancy coefficient in the message routing path and the load balancing Gini coefficient of the task load distribution data among the target agents are calculated. The path redundancy coefficient and the load balancing Gini coefficient are weighted and summed to obtain the collaborative balance index. The collaborative timeliness index and the collaborative balance index are nonlinearly fused to output the interaction efficiency parameter.

6. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 4, characterized in that, The step of performing counterfactual reasoning on the target agent based on the cognitive perturbation signal to obtain counterfactual correction data includes: The cognitive disturbance signal is decomposed to extract the disturbance direction component and disturbance intensity component. Extract the corresponding original premises from the interaction request data, and construct alternative premises that contradict the original premises based on the perturbation direction component and the perturbation intensity component, to obtain a set of counterfactual premises; The set of counterfactual premises is broadcast to the target agent, driving the target agent to perform counterfactual deduction, and the counterfactual deduction result and corresponding confidence level of the target agent are obtained. The counterfactual inference results are semantically aligned, mapped to the shared semantic space, and weighted voting aggregation is performed on the counterfactual inference results based on the confidence level to output counterfactual correction data.

7. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 1, characterized in that, Based on the credibility parameter, the joint features and the counterfactual correction data are weighted and fused. The resulting collaborative decision data and the interaction performance parameter are comprehensively evaluated. Based on the evaluation result, the topology of the initial interaction group is updated, and the target interaction group is output, including: The credibility parameter is used as the fusion weight of the joint features, and the difference between the unit weight and the credibility parameter is used as the fusion weight of the counterfactual correction data. The joint features and the counterfactual correction data are weighted and summed to obtain collaborative decision data. The collaborative decision-making data and the interaction efficiency parameters are jointly evaluated to obtain a comprehensive efficiency score, which is then used as the evaluation result. Obtain the topology of the initial interaction group, wherein the nodes in the topology correspond to the target agents, and the edges in the topology are used to represent the communication links and their weights between the corresponding two target agents; The overall performance score is compared with a preset performance achievement threshold. When the overall performance score is greater than or equal to the performance achievement threshold, the topology remains unchanged, and the initial interaction group is output as the target interaction group. When the overall performance score is less than the performance threshold, a topology optimization operation is performed on the topology to obtain an optimized topology. Based on the optimized topology, the initial interaction group is updated, and the target interaction group is output.

8. The method for distributed intelligent agent interaction architecture based on complex scenarios as described in claim 7, characterized in that, The topology optimization operation includes: The marginal contribution of the target intelligent agent to the collaborative decision-making data is statistically analyzed. Based on the marginal contribution, the link weights of the edges in the topology are positively incentivized or negatively attenuated. Based on the marginal contribution, replace or supplement the intelligent agent from the intelligent agent registration center, and replace or expand the nodes in the topology. Repeat the topology optimization operation until the overall performance score is greater than or equal to the performance threshold.

9. A distributed intelligent agent interaction architecture system for complex scenarios, characterized in that: The system is used to implement the distributed intelligent agent interaction architecture method for complex scenarios as described in claim 1, the system comprising: The scene perception module is used to read interaction request data from the distributed data space and determine the interaction scene type corresponding to the interaction request data; The group initialization module is used to call the target intelligent agent from the intelligent agent registration center according to the interaction scenario type to form an initial interaction group; The collaborative processing module is used to perform feature mapping on the interaction request data based on the semantic association between the target intelligent agents, and to evaluate the confidence of the joint features obtained after feature mapping to obtain a confidence parameter. Based on the historical interaction records between the target intelligent agents, it performs collaborative measurement on the historical interaction process between the target intelligent agents to obtain an interaction performance parameter. The cognitive correction module is used to evaluate the group cognitive consistency of the joint features. When the evaluation index reaches or exceeds the preset cognitive convergence threshold, a cognitive perturbation signal is constructed. Based on the cognitive perturbation signal, counterfactual reasoning is performed on the target intelligent agent to obtain counterfactual correction data. The topology optimization module is used to perform weighted fusion of the joint features and the counterfactual correction data based on the credibility parameter, comprehensively evaluate the collaborative decision data obtained after weighted fusion and the interaction performance parameter, update the topology of the initial interaction group according to the evaluation result, and output the target interaction group.

Images