Multi-subject relation and intention collaborative closed-loop optimization method and device based on dynamic graph

Abstract

The invention relates to the intelligent technical field of graph model and machine learning, and provides a multi-subject relationship and intention collaborative closed-loop optimization method and device based on a dynamic graph in order to solve the problems that a multi-subject system in the prior art is split in relationship and intention deduction, large in result fluctuation and insufficient in interactive noise robustness. Comprising the following steps: performing time alignment on original interaction data, constructing a dynamic graph structure, and encoding to generate node and edge representation; estimating a cooperative relationship based on edge representation, and inferring a behavior intention based on node representation; a closed-loop feedback mechanism is introduced, the cooperation relation is corrected through gating, and the behavior intention is corrected; and iteratively optimizing the process until a convergence condition is met, and finally outputting a relationship and intention estimation result after collaborative optimization. Experiments show that the method can simultaneously improve the performance of relation prediction and intention inference on a real multi-subject interaction data set, has higher parameter sharing efficiency and stability, and has practical application value.

Description

A method and apparatus for multi-agent relationship and intention-based collaborative closed-loop optimization based on dynamic graphs Technical Field

[0001] This invention belongs to the field of intelligent technology of graph models and machine learning, specifically relating to a method and apparatus for multi-agent relationship and intent collaborative closed-loop optimization based on dynamic graphs. Background Technology

[0002] With the development of artificial intelligence technology and complex system modeling methods, multi-agent collaborative systems are widely used in applications such as multi-device collaborative operation, distributed task execution, group behavior analysis, and complex organizational operation management. In such systems, multiple agents form collaborative relationships through continuous information interaction and behavioral feedback, and their collaborative state and behavioral intentions often change dynamically over time. Accurately characterizing the collaborative relationships between agents and effectively inferring their behavioral intentions are key technical issues for improving the overall decision-making ability and operational stability of multi-agent systems.

[0003] In existing technologies, multi-agent collaboration analysis methods mainly focus on two directions: collaborative relationship modeling and behavioral intention inference. One type of method emphasizes characterizing the collaborative structure between agents through relationship modeling, such as representing the connections between agents based on graph or network models, and analyzing relationship changes through updates to graph node features or edge weights. This type of method can usually describe the strength of associations between agents, but often treats behavioral intention as a latent variable, lacking explicit modeling of the agents' behavioral motivations. The other type of method focuses on behavioral intention inference, predicting agents' behavioral tendencies or decision preferences by modeling agent behavioral sequences, communication information, or state characteristics. However, it usually ignores the dynamic influence of the collaborative relationship structure between agents and struggles to reflect the constraining effect of changes in collaborative relationships on behavioral intentions.

[0004] In recent years, some studies have attempted to introduce dynamic graph models to model multi-agent systems, characterizing the evolution of relationships by updating the representations of nodes or edges over time. While these methods have improved the expressive power of relationship modeling to some extent, they mostly treat collaborative relationship modeling and behavioral intention inference as independent tasks, lacking a unified joint modeling mechanism. Furthermore, existing methods often employ a unidirectional inference process, either inferring behavioral intention based on the relationship structure first, or predicting relationship changes based on behavioral characteristics first, making it difficult to characterize the mutual influence between the two.

[0005] Because behavioral intentions and collaborative relationships are significantly coupled in multi-agent systems, an agent's behavioral intentions often influence the evolution of their collaborative relationships, while changes in collaborative relationships, in turn, constrain and modify the agent's behavioral intentions. Without modeling this two-way influence, existing technologies are prone to problems such as unstable inference results, high misjudgment rates, or delayed responses to changes when facing complex scenarios involving interactive noise, abnormal behavior, or rapid changes in relationships.

[0006] Given the shortcomings of the existing technologies, there is an urgent need for a technical solution that can jointly infer the collaborative relationships and behavioral intentions of multiple subjects within a unified framework, and achieve mutual constraint and collaborative optimization between the two through a closed-loop feedback mechanism, thereby improving the accuracy, stability and adaptability of the inference results to complex interaction scenarios. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a method and apparatus for collaborative closed-loop optimization of multi-agent relationships and intentions based on dynamic graphs.

[0008] This invention provides a closed-loop optimization method for multi-agent relationships and intent collaboration based on dynamic graphs, comprising: step S100, performing data preprocessing and time alignment, including dividing the original multi-agent interaction data into continuous time steps according to time windows, and extracting the subject state features, interaction evidence, and dynamic graph structure of each time step; step S200, performing graph representation encoding: encoding the subject state features and interaction features between subjects based on the dynamic graph structure of the current time step, generating node representations and edge representations; step S300, estimating the collaboration relationship based on the edge representations, and outputting the initial collaboration relationship estimation result between subjects in the current time step; step S400, inferring the behavioral intent based on the node representations, and outputting... Step S500: Calculate the initial behavioral intent estimation results for each subject in the current time step; Step S600: Perform closed-loop feedback correction: Use the initial behavioral intent estimation results to perform gating correction on the initial cooperation relationship estimation results to obtain the corrected cooperation relationship; Use the corrected cooperation relationship to perform smoothing constraints on the initial behavioral intent estimation results to obtain the corrected behavioral intent; Step S600: Perform iterative optimization: Determine if the iteration termination condition is met; If the iteration termination condition is not met, update the graph representation based on the corrected cooperation relationship and behavioral intent, and return to step S200 or step S300 to continue the next iteration; If the iteration termination condition is met, output the final cooperation relationship and behavioral intent results.

[0009] On the other hand, the present invention also provides a multi-agent relationship and intent collaboration closed-loop optimization device based on dynamic graphs, comprising: a first module for data preprocessing and time alignment, including dividing the original multi-agent interaction data into continuous time steps according to time windows, and extracting the subject state features, interaction evidence, and dynamic graph structure of each time step; a second module for graph representation encoding: encoding the subject state features and interaction features between subjects based on the dynamic graph structure of the current time step, generating node representations and edge representations; a third module for performing collaboration relationship estimation based on the edge representations, and outputting the initial collaboration relationship estimation result between subjects in the current time step; and a fourth module for performing behavioral intent inference based on the node representations. The first module outputs the initial behavioral intent estimation results for each subject in the current time step; the second module performs closed-loop feedback correction: it uses the initial behavioral intent estimation results to perform gating correction on the initial cooperation relationship estimation results to obtain the corrected cooperation relationship; it uses the corrected cooperation relationship to perform smoothing constraints on the initial behavioral intent estimation results to obtain the corrected behavioral intent; the third module performs iterative optimization: it determines whether the iteration termination condition is met; if the iteration termination condition is not met, it updates the graph representation based on the corrected cooperation relationship and behavioral intent, and returns to the steps of the second module or the third module to continue the iteration; if the iteration termination condition is met, it outputs the final cooperation relationship and behavioral intent results.

[0010] Compared with the prior art, the beneficial effects obtained by the present invention include: (1) By representing the interaction process of the multi-agent system at discrete time steps as a dynamic graph structure that evolves over time, the subject state features and the interaction features of subject pairs (between subjects) are constructed, and the subject representation and subject pair representation are obtained based on graph representation encoding; on this basis, a cooperative relationship estimation branch and a behavioral intention inference branch are set up simultaneously, so that the cooperative relationship strength / type of subject pairs and the behavioral intention probability distribution of subjects or subject pairs can be output synchronously within the same framework, thereby avoiding the problem of "training two models separately and information not being shared" in the prior art. Experimental comparison shows that, under the same data conditions, the cooperative framework of the present invention improves the cooperative relationship prediction F1 from 0.78 of independent modeling to 0.86, and the intention inference F1 from 0.74 of independent modeling to 0.83, indicating that joint inference can achieve complementary enhancement by utilizing the structured context of cooperative relationships and the high-level semantic signals of intentions, thereby improving prediction accuracy and reducing the misjudgment rate.

[0011] (2) This invention introduces a closed-loop feedback correction mechanism in the joint inference framework: on the one hand, the intention inference result is used as a gating signal to correct the estimation of the cooperation relationship, so that the cooperation relationship is no longer determined unidirectionally by local interaction evidence, but can be corrected by the long-term tendency and consistency constraints at the intention level; on the other hand, the change in the cooperation relationship in adjacent time steps is used to construct a smooth constraint on the intention update, so that the intention output remains continuous when the relationship is stable and allows rapid adjustment when the relationship changes rapidly, thereby suppressing the "output jitter" and "abnormal trigger drift" common in the prior art. Experimental results further show that after adding end-to-end fine-tuning on the basis of alternating optimization, the cooperation relationship F1 of this invention is improved from 0.84 to 0.86, the intention F1 is improved from 0.81 to 0.83, and the stability error (MAE) is reduced from 0.067 to 0.061 and the total loss is reduced from 0.73 to 0.69, indicating that closed-loop feedback and joint fine-tuning can effectively reduce residual mismatch between modules and improve overall stability and reliability.

[0012] (3) By using a shared representation space alignment mechanism, the collaborative relationship branches and the intention branches gradually learn a consistent representation space, reducing the convergence difficulties caused by the inconsistency of feature spaces in multi-module collaborative training. Training process statistics show that the mean similarity of shared representations increased from 0.42 to 0.81, the standard deviation of similarity decreased from 0.31 to 0.12, and the alignment loss decreased from 1.87 to 0.28, indicating that the collaborative closed-loop training of the present invention has clear convergence characteristics and can stably obtain a usable joint inference model, thus improving the controllability of the training process from an engineering implementation perspective.

[0013] (4) This invention introduces structural embedding of dynamic graphs into intent inference, enabling intent inference to comprehensively utilize the subject's position, association strength, and relationship evolution trend in the collaborative network, thereby overcoming the insufficient intent recognition ability caused by relying solely on local features. Experimental results show that the F1 score of intent inference with complete configuration including structural embedding is 0.85, while the F1 score drops to 0.72 after removing structural embedding, indicating that structural context has a significant gain for intent inference and can improve the recognition accuracy in complex interaction scenarios.

[0014] (5) From the perspective of engineering cost, the present invention adopts a collaborative framework for parameter sharing and joint training, which reduces the repetitive feature extraction and model redundancy during independent modeling, thereby improving performance while reducing training costs. Experimental comparison shows that the number of parameters of the collaborative framework of the present invention is 12.1M, which is less than the sum of the number of parameters of the two models trained independently (15.1M); the collaborative training time is 4.5 hours, which is less than the total training time of 6.0 hours for independent training, indicating that the present invention has better deployment efficiency and engineering usability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0016] Figure 1 is a flowchart of a multi-agent relationship and intent collaboration closed-loop optimization method based on dynamic graphs in one embodiment of the present invention; Figure 2 is a schematic diagram of an interaction evidence example and a complete multi-agent relationship and intent collaboration closed-loop optimization method flowchart in one embodiment of the present invention, taking a game scene as an example; Figure 3 is a schematic diagram of the multi-level abstract construction of dynamic graph structure in one embodiment of the present invention, wherein Figure 3(a) is a schematic diagram of the geographical scope of the study area, Figure 3(b) is a schematic diagram of the spatial characteristics of the network space between countries, and Figure 3(c) is a schematic diagram of defining spatial resolution; Figure 4 is a schematic diagram of the basic processing unit in the graph convolutional network in one embodiment of the present invention; Figure 5 is a schematic diagram of the edge representation construction and collaboration relationship estimation process in one embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] In one embodiment, as shown in Figure 1, the present invention provides a multi-agent relationship and intent collaboration closed-loop optimization method based on dynamic graphs, including: step S100, performing data preprocessing and time alignment, including dividing the original multi-agent interaction data into continuous time steps according to time windows, and extracting the subject state features, interaction evidence, and dynamic graph structure of each time step; step S200, performing graph representation encoding: based on the dynamic graph structure of the current time step, encoding the subject state features and interaction features between subjects to generate node representations and edge representations; step S300, performing collaboration relationship estimation based on the edge representations, and outputting the initial collaboration relationship estimation results between subjects in the current time step; step S400, performing line representation estimation based on the node representations. For intent inference, output the initial behavioral intent estimation results of each subject in the current time step; Step S500, perform closed-loop feedback correction: use the initial behavioral intent estimation results to perform gating correction on the initial cooperation relationship estimation results to obtain the corrected cooperation relationship; use the corrected cooperation relationship to perform smoothing constraint on the initial behavioral intent estimation results to obtain the corrected behavioral intent; Step S600, perform iterative optimization: determine whether the iteration termination condition is met; if the iteration termination condition is not met, update the graph representation based on the corrected cooperation relationship and behavioral intent, and return to step S200 or step S300 to continue the next round of iteration; if the iteration termination condition is met, output the final cooperation relationship and behavioral intent results.

[0019] Specifically, step S100, which involves data preprocessing and time alignment, includes: dividing the original multi-subject interaction data into continuous time windows. Each time step; at each time step , The extracted information includes: (1) Subject state characteristics: (2) Interactive evidence: including but not limited to original data such as communication records, collaborative behavior records, resource exchange records, and conflict event records, used for feature supplementation in subsequent steps; (3) Dynamic graph structure Adjacency matrix or candidate edge set The edge can be directly input or provided externally; where the candidate edge set is... Generate based on interaction evidence or prior rules in the original multi-agent interaction data.

[0020] In this invention, the following concepts need to be clarified: an interaction event refers to an objective interaction fact that occurs between subjects, including but not limited to communication, collaboration, resource exchange, conflict, etc.; interaction evidence refers to the original data carrier that records the interaction event; in addition, interaction features are feature vectors that can be directly input into a neural network after feature extraction of the interaction evidence.

[0021] Taking a game scenario in a multi-agent interaction system as an example, as shown in Figure 2, the raw data includes the historical game states and collaborative behavior records of the agents (such as location coordinates, supply center status, movement commands, attack actions, etc.). After data preprocessing and time alignment in step S100, interaction (event) evidence is first extracted from this raw data, including but not limited to communication records, collaborative behavior records, resource exchange records, etc. Figure 2 shows a typical interaction evidence extracted in this step—the communication record between agent A and agent B: at a certain time step, agent A sends a request to agent B, "Can you help me complete the current task?", and agent B replies, "Okay, I agree." Based on the extracted interaction evidence and combined with the agent state characteristics, the time step shown in the middle of Figure 2 is constructed. The graph structure is a dynamic graph. In this graph structure, nodes represent interacting entities (e.g., entity A, entity B, entity C, etc.), and the nodes implicitly contain entity state characteristics (e.g., location, supply center information). The edges between nodes represent interaction relationships supported by evidence of interaction between entities (e.g., the communication edge between entity A and entity B). This dynamic graph structure will serve as the unified data foundation for subsequent steps and will be input into the graph representation encoding module for further processing.

[0022] Furthermore, the construction of dynamic graph structures needs to be rooted in the physical or logical structure of the environment in which the subject resides. As shown in Figure 3, in this invention, the original space of the interaction environment, such as the geographical range of the study area shown in Figure 3(a), is abstracted into a network topology composed of nodes (subjects) and edges (interaction possibilities), as shown in Figure 3(b) for the spatial characteristics of the network space between countries. Based on this, a graph structure abstraction granularity for feature extraction is defined, as shown in Figure 3(c) for defining the spatial resolution. This is based on the adjacency matrix. or candidate edge set Dynamic graph structure abstracted from form This forms the basis for all subsequent steps.

[0023] Further, in step S200, graph representation encoding is performed to generate node representations and edge representations, including: (1) Node encoding to generate node representations: using a node encoding function to map the subject state features to the latent space, specifically, the nodes / subjects extracted in step S100 are used to generate node representations. Subjective state characteristics The initial features of the graph nodes are mapped to node representations through a node encoding function: ;in, It is a node encoding function, which can be a linear transform or a multilayer perceptron.

[0024] In this invention, multiple steps (including feature extraction, graph representation encoding, relation estimation, intent inference, etc.) involve graph convolutional networks (GCNs). The general structure of neural networks is used as the basic processing unit of GCN, as shown in Figure 4. Its core structure realizes the mapping from input to features through the stacking of linear transformation (or convolution) and activation functions.

[0025] Taking a multi-agent interaction system in a game scenario as an example for further explanation, the input layer in Figure 4 is used to receive multi-source input features and can simultaneously receive various types of feature data, including but not limited to: Dialogue graph: semantic features extracted based on inter-agent communication records, such as the feature vector formed after encoding the communication dialogue between subject A and subject B as shown at the bottom of Figure 2; Behavior graph: behavioral coordination features extracted based on the action records of the subjects (such as movement, attack, support, etc.); Intent graph: intent-related features that can be used as prior information or contextual feature input to assist in inference at the current moment.

[0026] It should be noted that the dialogue graph, behavior graph, and intent graph shown in Figure 4 are only exemplary representations of input features. In actual applications, the types and combinations of input features can be flexibly adjusted according to specific tasks, and together they constitute the multi-source input feature layer for subsequent processing.

[0027] Convolutional / Linear Transformation Layer: This layer performs convolutional operations or linear transformations on the input features to extract a higher-level abstract representation. This layer corresponds to the various encoding and estimation functions mentioned earlier, such as node encoding functions. Relationship estimation function First intention inference function and second intention inference function, etc.

[0028] Activation function layers are used to introduce nonlinear transformations and enhance the expressive power of the model. Commonly used activation functions include ReLU, Sigmoid, and Softmax.

[0029] The output layer is used to output the encoded or estimated feature results, including but not limited to node representation, edge representation, relation score, relation strength weight, relation type probability distribution, intent probability distribution, etc.

[0030] (2) Perform edge encoding to generate edge representation: Based on node representation, the main body pair Generating edge representations includes constructing edge input features: ;in, It is the edge input feature, For element-wise multiplication, This indicates vector concatenation.

[0031] If interaction evidence exists, such as communication or collaborative behavior, its encoded vectors can be concatenated. end.

[0032] In one embodiment, to illustrate the extraction of richer semantic features from the original interaction data to construct edge representations, taking the extraction of information from two complementary channels as an example, as shown in Figure 5, the information in the communication records is encoded by a message encoder (Enc), and the "communication semantic similarity" is obtained by calculating the cosine similarity of the two message vectors; simultaneously, the collaborative action records of the subjects are used to calculate the behavioral synergy degree using a consistency scoring function. These features, together with the initial features of the graph nodes, constitute the input basis for subsequent estimation of cooperative relationships.

[0033] Figure 5 is a schematic diagram illustrating the edge representation construction and cooperative relationship estimation based on multi-source feature fusion in this embodiment. The main body (nodes) in the figure includes at least , representing the various entities in a multi-agent system; , , This represents a continuous time step divided by a time window; Entities , , The corresponding time step extracted in step S100 The system extracts the subject state features. In the upper communication information processing channel, the original communication record serves as the input of interactive evidence. After processing by the encoder (Enc), the historical (previous time step) communication information encoding and the current time step encoding are obtained. Then, the cosine similarity between the two is calculated to obtain the semantic similarity of communication between the subjects, reflecting the consistency of communication content. In the lower behavior information processing channel, the collaborative behavior record serves as another type of interactive evidence input. After behavior extraction, the behavior collaboration degree is calculated to reflect the degree of consistency of actions between the subjects. Then, the subject state features are fused with the communication semantic similarity and behavior collaboration degree input feature fusion module (e.g., MLP, multilayer perceptron) to output the edge input features. Based on the edge input features, combined with the historical collaboration relationship representation transmitted from the previous time step, the system performs temporal information fusion and update through a gated recurrent unit (GRU) to output the collaboration relationship representation of the current time step. This GRU uses its gating mechanism (update gate, reset gate) to adaptively determine the degree of retention of historical collaboration patterns and the fusion weight of current interactive evidence, thereby capturing the dynamic evolution of collaboration relationships over time. The final output of the collaboration relationship is the collaboration relationship estimation result (relationship strength weight or relationship type probability distribution) described in step S300.

[0034] In step S300, cooperation relationship estimation (edge ​​weight estimation / relationship type estimation) is performed to obtain the initial cooperation relationship estimation result. The cooperation relationship estimation includes at least one of the following two estimations: (1) Edge weight estimation (continuous value output) through the relationship estimation function. Edge input features The relationship score is obtained through processing. It outputs continuous initial relation strength weights. : ;in, For the Sigmoid function, The relationship estimation function It is a multilayer perceptron.

[0035] (2) Relation type estimation obtains the discrete relation type probability distribution by normalizing the exponential function: ;in, for Probability distribution of various relation types yes The Middle The probability of each relation type ; It is a weight matrix. It is the bias vector; It is the indicator of the relationship type with the highest probability; It is a normalized exponential function.

[0036] The initial cooperation relationship estimation result includes at least one of the following: initial relationship strength weights or relationship type probability distributions.

[0037] In step S400, behavioral intention inference (subject-level or subject-pair level) is performed to obtain the initial behavioral intention estimation result.

[0038] The behavioral intent inference includes at least one of the following: subject-level intent inference: aggregation and subject All relevant edge input features, and the node representation of the subject. After fusion, the inference function is derived from the first intent. Obtain the main body Intent probability distribution: Among them, Agg Indicates the subject Aggregation of relevant side input features (e.g., weighted average, attention-weighted, etc.); the first intent inference function Used for subject-level intent inference, its input is the current time step. Lower body Node representation With the aggregated To achieve hierarchical reasoning, which progresses from local interactions to layer-by-layer abstraction of the global structure, The multilayer perceptron (MLP) architecture is adopted, specifically including: an input layer, which... and The dimension obtained after splicing is The vector, where Let be the dimension represented by the node. The dimension of the input features is defined by the following layer: Hidden layers, each containing at least one fully connected layer followed by a non-linear activation function (such as ReLU), are used to extract high-level semantic features; the dimension of the hidden layers is set according to the task complexity; the output layer maps the output of the hidden layers to the number of intent categories. The main body is output through the Softmax function. Intention probability distribution .

[0039] Subject-level intent inference: Input edge features With the corresponding initial relation strength weight After fusion, the second intention inference function is used. Get the subject Intent probability distribution: The initial relation strength weights are used as auxiliary inputs for intent inference, ensuring consistency between intent and relation; the second intent inference function... Used for subject-level intent inference, its input is edge input features. Weights of initial relationship strength This function is also implemented based on a multilayer perceptron, specifically including: an input layer, which... and After concatenation, the resulting dimension is The vector, where As auxiliary input, it is used to establish consistency constraints between intent and relation in the early stages of inference; the hidden layer contains at least one fully connected layer, each followed by a non-linear activation function (such as ReLU) to fuse edge features and relation weight information; the output layer maps the output of the hidden layer to the number of intent categories. And output the main body through the Softmax function. Intention probability distribution .

[0040] The initial behavioral intent estimation results should include at least the subject-level intent probability distribution. Or the probability distribution of the subject's intention to the level One of them.

[0041] The intent inference of this invention is not a simple, planar process, but rather employs a hierarchical graph convolutional network. By combining multiple layers of GCN with multi-head attention, it abstracts layer by layer from local interactions to the global structure (cross-layer multi-dimensional information), ultimately achieving advanced inference goals such as deception detection and hidden intent recognition. The bottom-level GCN and attention heads are used to aggregate neighbor information and capture local interaction patterns; the higher-level modules then perform deeper inference on this basis to output the final intent probability distribution.

[0042] To achieve collaborative optimization of "relationship-intent", this invention designs a closed-loop feedback mechanism, which enables the intention inference result to correct the collaborative relationship estimation in reverse, and at the same time uses relationship changes to constrain the intention update, thereby reducing the instability caused by fragmented modeling.

[0043] Preferably, in the architecture of a hierarchical graph convolutional network, the first intent inference function High-level modules deployed in graph convolutional networks, their input features and Having undergone multiple rounds of aggregation and abstraction using the underlying GCN and multi-head attention, it is able to capture global interaction patterns across subjects, supporting advanced reasoning tasks such as deception detection and hidden intent recognition; the second intent inference function... It can be deployed after the underlying GCN to capture local interaction semantics, or it can be used as a component of a high-level module in a graph convolutional network to output fine-grained interaction intents. The output results will be used in subsequent closed-loop feedback correction steps, specifically manifested as: subject-to-subject hierarchical intent probability distribution. Used to calculate the intent consistency gating coefficient, thereby correcting the initial cooperation relationship estimation results.

[0044] In step S500, the step of correcting the initial cooperation relationship estimation result using the initial behavioral intention estimation result includes: (1) calculating the intention consistency gating coefficient based on the initial behavioral intention estimation result (for example, three types of intentions: cooperation / neutrality / competition). .

[0045] First, utilize the subject pair Collaboration example intent probability distribution Probability distribution of competing example intents We obtain the initial values ​​of the intent consistency gating coefficient: ; and will Mapped to The final intent consistency gating coefficient is obtained as follows: (2) The initial cooperative relationship estimation result is weighted and fused with the gating coefficient to obtain the corrected cooperative relationship.

[0046] If the cooperation relationship estimation uses edge weight estimation, then the modified cooperation relationship includes the following final modified edge weights: ;in, The feedback strength coefficient (recommended range: 0.1 to 0.7).

[0047] If the cooperation relationship estimation chooses relationship type estimation, the revised cooperation relationship includes the following pairs. The reweighted probability distribution obtained after reweighting: ;in, A reweighted vector generated from the intent probability. It is a normalization function.

[0048] Furthermore, in step S500, the smoothing constraint applied to the initial behavioral intent estimation result using the modified cooperation relationship includes: considering that changes in the relationship between adjacent time steps will reflect changes in the cooperation state, and constructing the change amount of the modified cooperation relationship: ;like If the value is too small, for example, less than a preset lower threshold, the intention should not fluctuate drastically; if If the value is too large, for example, greater than a preset upper threshold, then the intention to adjust quickly is allowed.

[0049] This leads to the construction of a constraint loss term intended to smooth out the subject, used to calculate the obtained subject. At time step Corrected probability distribution of behavioral intentions The intention to smooth the constraint loss term is given by the following formula: ;in, Represents the square of the L2 norm, main body The intention is to smooth out weights It can be determined by the change in the relevant edges of the subject: In the above formula, It is the main body The neighborhood set of subjects, yes The number of subjects contained therein.

[0050] The intent smoothing constraint loss term is used to penalize drastic fluctuations in the initial behavioral intent estimation results between adjacent iterations, and the penalty strength is proportional to the intent smoothing weight. Positive correlation. Thus, when the relationship structure changes drastically, The larger the value, the faster the intent updates; when the structure is stable, it suppresses intent jitter.

[0051] In step S500, to enhance the model's robustness against strategic deception, an adversarial training process is introduced into the closed-loop feedback mechanism of this invention. A typical process of this adversarial training is as follows: First, the communication text is encoded by an encoder such as Bidirectional Encoder Representations from Transformers (BERT), while the behavioral sequence is encoded by a Long Short-Term Memory (LSTM) network; then, the multimodal features are fused; finally, an adversarial game is played between a generator (i.e., the intent inference model of this invention) and a discriminator. This adversarial process forces the model to learn a more robust and consistent intent representation, thereby effectively dealing with "inconsistent words and actions" strategic deception and avoiding any sensitive information.

[0052] In step S600, the iteration termination condition can be any of the following termination conditions: Condition 1: Maximum number of iterations: The number of iterations reaches... (Recommended 1-5 iterations); Condition 2: Convergence condition: If the change in the corrected cooperation relationship between two consecutive iterations satisfies: ;in, and They represent the first The second iteration and the first The final corrected edge weights describe the modified cooperative relationships in the next iteration.

[0053] Then convergence is determined and the process stops, with a preset threshold. exist ~ Values ​​can be taken within a range.

[0054] The method of updating the graph representation based on the modified collaboration relationship and behavioral intention includes: using the modified collaboration relationship as a new edge weight for message aggregation in step S200; and / or encoding the modified behavioral intention and adding it to the subject state features to update the initial features of the graph nodes in the next iteration.

[0055] In one embodiment, to more accurately capture the spatiotemporal evolution patterns of alliance relationships and behavioral intentions during iterative optimization, this invention can draw on the idea of ​​multi-scale feature extraction. It uses LSTM to capture short-term instantaneous changes, attention mechanisms to model medium-term temporal dependencies, and pooling operations such as max pooling to summarize long-term macro trends. These multi-scale features can be fused and used to update graph representations or correct inference results, thereby improving the accuracy of long-horizon prediction.

[0056] Overall, taking the multi-subject interactive system in a game scenario as an example, Figure 2 fully illustrates the entire process framework of the method of the present invention from raw data input to final result output.

[0057] First, extract interaction evidence from the historical game states and actions at the top (example of communication records at the bottom), and construct the dynamic graph structure in the middle to complete the data preprocessing and time alignment in step S100.

[0058] Subsequently, based on the constructed dynamic graph structure, node representations and edge representations are generated through "graph representation encoding" (corresponding to step S200). Building upon this, two core tasks are executed in parallel: first, "cooperation relationship representation" (corresponding to step S300), outputting an initial estimate of the cooperation relationship between the subjects; and second, "behavioral intention inference" (corresponding to step S400), outputting an initial estimate of the subject's behavioral intention. The "individual layer / local layer / global layer" on the right side of Figure 2 illustrates the multi-layered architecture of behavioral intention inference—from individual feature extraction to local interaction semantic fusion, and then to global policy intention reasoning.

[0059] Next, the closed-loop feedback correction mechanism in the method of this invention is intuitively demonstrated through the "intention probability distribution feedback" process (corresponding to step S500): the result of behavioral intention inference is fed back to the cooperation relationship estimation module for gating correction, and the corrected cooperation relationship is used to smooth the intention inference result, thereby achieving the synergistic optimization of relationship and intention.

[0060] Finally, after the iterative optimization judgment in step S600, if the termination condition is not met, the graph representation is updated based on the correction result and the iteration continues; if the termination condition is met, the final collaborative relationship and behavioral intention result is output. This game scenario example fully presents the practical application process of the method of this invention in a multi-agent interactive system.

[0061] The above method involves designing the loss function and training method (supervised or weakly supervised) for parameter training, including: if collaborative relationship labels exist. or behavioral intention label Supervised training can be used; if the labels are incomplete, weakly supervised or semi-supervised training can be used.

[0062] (1) Relationship prediction loss (marginal weight regression or classification) regression loss ;in, Indicating the subject's relationship to the regression task At time step The true relationship strength value, superscript " "The final corrected edge weights used to distinguish the model output" .

[0063] Similarly, cross-entropy (classification) loss can be obtained (for...) (After taking the logarithm, sum the results using the collaboration relationship labels as weighting coefficients).

[0064] (2) Intending to predict losses ;in, It is a label for genuine behavioral intent. Representing the subject The true intention of the behavior belongs to the first Class (e.g., behavioral intention categories include cooperation / neutrality / competition, etc.), otherwise, if the subject The true intent of the action does not belong to the first kind, ,but ; Representing the subject At time step Belongs to the (3) Closed-loop consistency and stability constraints ;in, This is the weighting coefficient (recommended value: 0.1 to 1.0).

[0065] Finally, to verify the effectiveness and feasibility of the method of the present invention, the performance differences between "independent modeling" and "collaborative closed-loop modeling" were compared under the dataset and unified evaluation criteria used in the present invention, and the contributions of key mechanisms such as closed-loop optimization strategy (alternating optimization and end-to-end fine-tuning), shared representation space alignment, and structural embedding to the results were further verified.

[0066] (I) Experimental parameters (1) Data set and task settings: This experiment uses the human game dataset, which contains the interaction records of multiple agents at discrete time steps. It can be used to construct a dynamic graph of time evolution and supports two types of tasks: Collaboration relationship prediction task: predict the collaboration relationship between agents at a given time step and output the relationship strength or the risk of relationship change; the evaluation index is the F1 value (the higher the F1, the more accurate the prediction).

[0067] Behavioral Intent Inference Task: Classify and infer the behavioral intent of an entity or entity pair (three types of intent in this invention), and use the F1 score as the evaluation index (and statistical indicators such as macro average accuracy in some analysis tables).

[0068] (2) Comparison settings (most similar to existing solutions) To reflect the improvement effect of the present invention compared with the existing technology, the following comparison is set: Independent modeling solution A: Independent collaboration relationship prediction model, that is, only the collaboration relationship prediction module is trained, and the intention inference module is not introduced to participate in training and correction.

[0069] Independent modeling scheme B: Independent behavioral intent inference model, which trains only the intent inference module and does not introduce a collaborative relationship prediction module to provide structured relationship constraints.

[0070] Collaborative Closed-Loop Scheme (Invention): The collaborative framework (MA-SRHG) is a unified framework in which collaborative relationship prediction and behavioral intention inference are jointly trained and mutually corrected and optimized through a closed-loop feedback mechanism.

[0071] (3) Training strategy and closed-loop optimization parameters To verify the contribution of the closed-loop optimization strategy to the final effect, the experiment further compared two training strategies: Alternating Optimization Only: During the training process, the cooperative relationship prediction and intention inference modules are updated alternately, so that the two gradually enter a stable cooperative state; Alternating + End-to-End Fine-Tuning: After the alternating optimization converges, end-to-end joint fine-tuning is performed to further eliminate residual mismatches between modules and improve the final performance.

[0072] Furthermore, this invention performs stage statistics on the alignment process of the shared representation space, dividing it into iteration intervals: Initial (0–50), Early (51–100), Middle (101–200), Late (201–300), and Converged (300+). The alignment similarity (mean / variance / minimum / maximum) and alignment loss are also statistically analyzed to reflect the convergence characteristics of closed-loop collaborative training.

[0073] (II) Experimental Results For ease of explanation, the core results corresponding to the tables in this invention are summarized in Tables 1 to 4. Table 1 presents a quantitative comparison of the collaborative effects of existing independent modeling and the collaborative closed-loop optimization method in this invention, involving evaluation metrics including collaborative relationship F1, intent F1, training time, and number of parameters. Table 2 illustrates the contribution of end-to-end fine-tuning to the final performance of the closed-loop framework, involving evaluation metrics including collaborative relationship F1, intent F1, stability error, and total loss. Table 3 presents the training dynamic statistics (convergence evidence of closed-loop collaboration) results of shared representation space alignment, with relevant statistical items including mean similarity, variance, minimum / maximum similarity, and alignment loss. Table 4 presents a comparison of the impact of structural embedding on intent inference performance (verification of key components in the closed-loop framework), involving evaluation metrics including accuracy of three types of intent, macro-average accuracy, and total F1.

[0074] Table 1. Quantitative comparison results of module synergy effects (independent modeling vs. collaborative closed loop)

[0075] Table 2. Schematic diagram of the contribution of end-to-end fine-tuning to the final performance of the closed-loop frame.

[0076] Table 3. Training dynamic statistics of shared representation space alignment

[0077] Table 4 Comparison of the impact of structural embedding on intent inference performance

[0078] (III) Results Analysis (1) Collaborative closed-loop can simultaneously improve the performance of two types of tasks and reduce training costs. As shown in Table 1, the collaborative closed-loop framework of this invention has achieved significant improvements in both core tasks: the F1 of collaborative relationship has increased from 0.78 in independent modeling to 0.86 (+10.3%), and the F1 of intention has increased from 0.74 in independent modeling to 0.83 (+12.2%). This indicates that separating the prediction of collaborative relationship and the inference of behavioral intention will result in the loss of the gains brought by information complementarity. However, this invention enables the two to be collaboratively optimized in a unified representation space through closed-loop feedback, thereby obtaining higher accuracy.

[0079] Meanwhile, from an engineering cost perspective, the collaborative framework reduces redundant feature extraction through parameter sharing, resulting in a total number of parameters (12.1M) less than the sum of two independent models (15.1M), and a training time (4.5h) less than the sum of training them separately (6.0h). Therefore, this invention not only improves performance but also efficiency, making it more suitable for engineering deployment.

[0080] (2) End-to-end fine-tuning can further eliminate residual mismatches between modules and improve stability. Table 2 shows that after adding end-to-end fine-tuning on the basis of alternating optimization, both the cooperative relationship F1 and the intention F1 are further improved (0.84→0.86, 0.81→0.83), while the stability error (MAE) decreases (0.067→0.061) and the total loss decreases (0.73→0.69). This indicates that the closed-loop framework still has "residual errors" that can be jointly adjusted end-to-end after completing alternating training. End-to-end fine-tuning helps to further align the relationship module and the intention module under the same goal, thereby obtaining a more stable final output.

[0081] (3) The gradual alignment of the shared representation space is direct evidence of the effectiveness of closed-loop collaborative training. Table 3 shows the similarity statistics of "collaborative relationship representation - intention representation" during the training process: the mean similarity steadily increased from 0.42 to 0.81, the standard deviation decreased from 0.31 to 0.12, and the alignment loss decreased from 1.87 to 0.28. This result shows that closed-loop collaborative training is not "simple superposition of two tasks", but rather promotes the two types of representations to enter a consistent and stable shared space, so that intention inference can utilize the structured relational context, and relation prediction can also obtain feedback correction from high-level intention signals.

[0082] (4) Structural embedding significantly improves intent inference and is a key component of “relationship-intent coupling modeling”. Table 4 shows that the F1 of the complete configuration is 0.85, while the F1 drops to 0.72 after removing structural embedding, indicating that it is difficult to fully infer behavioral intent based on local features alone. Introducing structural embedding can incorporate the global context of the collaborative relationship between subjects into the inference process, thereby significantly improving the accuracy of intent recognition, especially for the “competitive intent” category (0.87 compared to 0.69).

[0083] In summary, this invention addresses the inherently time-varying and strongly coupled joint inference problem of predicting collaborative relationships and inferring behavioral intentions in multi-agent collaborative systems. To resolve issues in existing technologies such as the disconnect between collaborative relationship modeling and behavioral intention inference, unstable results due to unidirectional inference, and insufficient adaptability to noise and abnormal interactions, this invention proposes a closed-loop optimization method for multi-agent relationships and intentions based on dynamic graphs.

[0084] First, this invention constructs a unified modeling framework based on dynamic graphs, representing multi-agent interaction processes as time-evolving interaction relationship graphs. It encodes the state features of agents and the interaction features between agent pairs (across agents), forming a shared representation that can be used for relationship prediction and intent inference. Second, this invention simultaneously outputs collaborative relationship estimation results and behavioral intent inference results within the same framework, and designs a closed-loop feedback correction mechanism: it uses the intent inference results to perform gating correction on the collaborative relationship estimation, while introducing relationship change constraints to suppress meaningless jitter in intent output, thereby achieving collaborative optimization and stable convergence of relationships and intents. Furthermore, this invention provides implementable training strategies, including alternating optimization and end-to-end joint fine-tuning, and ensures the algorithm's controllability and feasibility through explicit iterative termination conditions.

[0085] Finally, experiments were conducted on the real multi-agent interaction dataset used in this invention to verify the findings, and the collaborative closed-loop scheme was compared with the independent modeling scheme. Experimental results show that compared with training the collaborative relationship prediction model and the intent inference model separately, the collaborative closed-loop framework of this invention can simultaneously improve the prediction performance of both types of tasks, and exhibits better engineering efficiency in terms of parameter sharing and training cost. Furthermore, end-to-end fine-tuning can further reduce stability errors and improve the final performance indicators, demonstrating that the closed-loop collaborative mechanism has clear technical effects and practical application value.

[0086] In one embodiment, the present invention provides a multi-agent relationship and intent collaborative closed-loop optimization device based on dynamic graphs, comprising: a first module for data preprocessing and time alignment, including dividing the original multi-agent interaction data into continuous time steps according to time windows, and extracting the subject state features, interaction evidence, and dynamic graph structure of each time step; a second module for graph representation encoding: encoding the subject state features and interaction features between subjects based on the dynamic graph structure of the current time step, generating node representations and edge representations; a third module for collaborative relationship estimation based on the edge representations, and outputting the initial collaborative relationship estimation result between subjects in the current time step; and a fourth module for behavioral intent estimation based on the node representations. The first module infers and outputs the initial behavioral intent estimation results for each subject in the current time step; the second module performs closed-loop feedback correction: it uses the initial behavioral intent estimation results to perform gating correction on the initial cooperation relationship estimation results to obtain the corrected cooperation relationship; it uses the corrected cooperation relationship to perform smoothing constraints on the initial behavioral intent estimation results to obtain the corrected behavioral intent; the third module performs iterative optimization: it determines whether the iteration termination condition is met; if the iteration termination condition is not met, it updates the graph representation based on the corrected cooperation relationship and behavioral intent, and returns to the steps of the second module or the third module to continue iterating; if the iteration termination condition is met, it outputs the final cooperation relationship and behavioral intent results.

[0087] Specifically, the present invention can be implemented as the following modules, which are executed in sequence: (1) First module (data preprocessing module): aligning time steps and generating main features; (2) Second module (graph coding module): generating node representations. edge representation (3) Third module (initial estimation module of cooperation relationship): output or (4) Fourth Module (Initial Inference Module for Behavioral Intent): Output (5) Fifth module (closed-loop feedback correction module): intention → relation gating correction, relation → intention smoothing constraint; (6) Sixth module (iteration control module): judge the termination condition and control the iteration; final output results: final relation, intention and optional consistency score.

[0088] The data processing relationship between modules is as follows: preprocessing → graph coding → relation estimation → intent inference → feedback correction → (if not terminated, return to graph coding / relation estimation) → output.

[0089] On the other hand, in one embodiment, the present invention provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the dynamic graph-based multi-agent relationship and intent collaborative closed-loop optimization method provided in any of the above embodiments. The computer device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores sample data. The network interface of the computer device is used for communication with external terminals via a network connection.

[0090] On the other hand, in one embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs provided in any of the above embodiments are implemented.

[0091] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0092] Matters not covered in this invention are common knowledge.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs, characterized in that, include: Step S100: Perform data preprocessing and time alignment, including dividing the original multi-subject interaction data into continuous time steps according to time windows, and extracting the subject state features, interaction evidence, and dynamic graph structure of each time step; Step S200: Perform graph representation encoding: based on the dynamic graph structure of the current time step, encode the subject state features and interaction features between subjects to generate node representations and edge representations; Step S300: Estimate the cooperation relationship based on the edge representations and output the initial cooperation relationship estimation results between subjects in the current time step; Step S400: Infer the behavioral intent based on the node representations and output the initial behavioral intent estimation results of each subject in the current time step. Step S500: Perform closed-loop feedback correction: Use the initial behavioral intention estimation result to perform gating correction on the initial cooperation relationship estimation result to obtain the corrected cooperation relationship; The modified behavioral intent is obtained by applying a smoothing constraint to the initial behavioral intent estimation result using the modified cooperative relationship. Step S600: Perform iterative optimization: Determine whether the iteration termination condition is met; If the iteration termination condition is not met, the graph representation is updated based on the corrected collaboration relationship and behavioral intention, and the process returns to step S200 or step S300 to continue the next iteration; if the iteration termination condition is met, the final collaboration relationship and behavioral intention results are output.

2. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 1, characterized in that, In step S100, the dynamic graph structure uses an adjacency matrix. or candidate edge set The candidate edge set is provided in the form of; wherein, the candidate edge set is provided in the form of Generate based on interaction evidence or prior rules in the original multi-agent interaction data.

3. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 2, characterized in that, In step S200, the graph representation encoding includes: node encoding: through a node encoding function. Nodes / Main Subjective state characteristics Mapped to node representation: ;in, It is a time step. ， It represents the total number of time steps in the partition; node encoding function. Linear transformation or multilayer perceptron; edge encoding: for the subject pairs in the dynamic graph structure Based on the corresponding node representation and Generating edge representations includes constructing the following edge input features: ;in, It is the edge input feature, For element-wise multiplication, The vector concatenation is represented; if there is interactive evidence, the encoded vector of the interactive evidence is concatenated to the end of the edge representation.

4. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 3, characterized in that, In step S300, the cooperative relationship estimation includes at least one of the following two estimation methods: edge weight estimation: through a relationship estimation function. For the edge input features The relationship score is obtained through processing. It outputs continuous initial relation strength weights. ;in, For the Sigmoid function, The relationship estimation function For multilayer perceptrons; relation type estimation: through normalized exponential functions. Obtain the discrete relation type probability distribution: ； ;in, for Probability distribution of various relation types yes The Middle The probability of each relation type ； It is a weight matrix. It is the bias vector; It is the relationship type indicator with the highest probability; the initial cooperation relationship estimation result includes at least one of the initial relationship strength weight or relationship type probability distribution.

5. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 4, characterized in that, In step S400, the behavioral intent inference includes at least one of the following: subject-level intent inference: aggregation and subject All relevant edge input features, and the main body Node representation After fusion, the inference function is derived from the first intent. Obtain the main body Intent probability distribution: Among them, Agg Indicates the subject Aggregation of relevant edge input features; First intent inference function Implemented using a multilayer perceptron, deployed in a high-level module of a hierarchical graph convolutional network; subject-level intent inference: edge input features. With the corresponding initial relation strength weight After fusion, the second intention inference function is used. Get the subject Intent probability distribution: ; where the second intention inference function Implemented using a multilayer perceptron, it is deployed after a lower-level graph convolutional network to capture local interaction semantics, or as a component of a higher-level module of the graph convolutional network; the initial behavioral intent estimation result includes at least a subject-level intent probability distribution. Or the probability distribution of the subject's intention to the level One of them.

6. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 5, characterized in that, In step S500, the step of correcting the initial cooperation relationship estimation result using the initial behavioral intent estimation result includes: calculating the intent consistency gating coefficient based on the initial behavioral intent estimation result. The initial cooperation relationship estimation result is weighted and fused with the gating coefficients to obtain the corrected cooperation relationship: if the cooperation relationship estimation uses edge weight estimation, the corrected cooperation relationship includes the following final corrected edge weights: ;in, For feedback strength coefficients; if the cooperation relationship estimation chooses relationship type estimation, then the corrected cooperation relationship includes a reweighted probability distribution: ;in, A reweighted vector generated from the intent probability. It is a normalization function.

7. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 6, characterized in that, In step S500, the smoothing constraint applied to the initial behavioral intent estimation result using the modified cooperative relationship includes: calculating the change in the modified cooperative relationship between adjacent time steps. Based on the aforementioned changes, intentional smoothing weights are constructed for each subject: ;in, It is the main body The neighborhood set of subjects, yes The number of subjects included; the intention is to smooth the constraint loss term, which is used to calculate the number of subjects. At time step Corrected probability distribution of behavioral intentions The intention to smooth the constraint loss term is given by the following formula: ;in, The square of the L2 norm is represented; the intention smoothing constraint loss term is used to penalize drastic fluctuations in the initial behavioral intention estimation results between adjacent iterations, and the penalty strength is proportional to the intention smoothing weight. Positive correlation.

8. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to any one of claims 1-7, characterized in that, In step S600, the iteration termination condition includes at least one of the following: the number of iterations reaches a preset upper limit for the number of iterations; the change in the corrected cooperative relationship output between two adjacent iterations is less than a preset threshold.

9. The multi-agent relationship and intent collaborative closed-loop optimization method based on dynamic graphs according to claim 1, characterized in that, In step S600, updating the graph representation based on the corrected collaboration relationship and behavioral intent includes: using the corrected collaboration relationship as a new edge weight for message aggregation in step S200; and / or encoding the corrected behavioral intent and adding it to the subject state features to update the initial features of the graph nodes in the next iteration.

10. A multi-agent relationship and intent collaborative closed-loop optimization device based on dynamic graphs, characterized in that, include: The first module is used for data preprocessing and time alignment, including dividing the original multi-subject interaction data into continuous time steps according to time windows, and extracting the subject state features, interaction evidence and dynamic graph structure of each time step; The second module is used for graph representation encoding: based on the dynamic graph structure of the current time step, the main state features and the interaction features between the main entities are encoded to generate node representations and edge representations; The third module is used to estimate the cooperation relationship based on the edge representation and output the initial cooperation relationship estimation result between the subjects in the current time step; the fourth module is used to infer the behavioral intention based on the node representation and output the initial behavioral intention estimation result of each subject in the current time step. The fifth module is used for closed-loop feedback correction: the initial cooperation relationship estimation result is gating and corrected using the initial behavioral intention estimation result to obtain the corrected cooperation relationship; The modified behavioral intent is obtained by applying a smoothing constraint to the initial behavioral intent estimation result using the modified cooperative relationship. The sixth module is used for iterative optimization: determining whether the iteration termination condition is met; If the iteration termination condition is not met, the graph representation is updated based on the corrected collaboration relationship and behavioral intention, and the iteration continues from the steps of the second module or the third module; if the iteration termination condition is met, the final collaboration relationship and behavioral intention results are output.

Images