A dynamic collaboration system and method for heterogeneous multi-agent

Abstract

The application relates to a dynamic cooperation system and method for heterogeneous multi-agent, the system comprising a perception evaluation module, an interruption triggering module and a mask recombination module. The perception evaluation module obtains observation data of each agent and extracts observation feature codes to evaluate individual contribution of each agent to the task of the group it is currently in; the interruption triggering module calculates the time sequence change rate of the individual contribution, and when the relative difference between the time sequence change rate of the individual contribution of the target agent and the average change rate of the group it is in exceeds a set threshold, it is determined that the asynchronous interruption condition is met and an asynchronous interruption signal is generated; the mask recombination module responds to the asynchronous interruption signal, calculates the similarity between the target agent and the existing group based on the observation feature codes to screen the candidate associated group, constructs a local recombination mask based on the candidate associated group, performs forward calculation of the dynamic grouping network based on the local recombination mask, and only performs state updating and recombination on the target agent and the candidate associated group.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent control technology, and in particular to a dynamic collaborative system and method for heterogeneous multi-agent systems. Background Technology

[0002] In the field of multi-agent systems, with the continuous growth of the number of tasks and the scale of agents, adopting task allocation strategies to rationally distribute heterogeneous agents (such as autonomous vehicles and drones) to different task groups for parallel execution according to needs is a mainstream approach to improve the overall collaborative efficiency of the system. Currently, the closest existing technology usually adopts a method based on team value decomposition (such as combined with the QMIX architecture) for dynamic task allocation of heterogeneous multi-agent systems. This method evaluates the Q-value contribution of heterogeneous individuals by constructing an independent value network, and when an individual's Q-value is lower than the weighted sum of the average value of its peers and the overall average value, a dynamic removal mechanism is initiated to remove the individual from the current group; subsequently, a fusion strategy based on the Q-value attention mechanism is used to aggregate the observation features of the remaining individuals to generate unified group state information to guide subsequent decisions.

[0003] However, when the aforementioned existing technologies are applied in highly dynamic and non-stationary environments, they reveal the following significant technical defects and system bottlenecks.

[0004] Existing technologies for adjusting agent groups typically rely on fixed, periodic grouping step sizes (e.g., adjusting every fixed step) or solely on absolute difference thresholds for individual removal. This static or rigid approach fails to capture, in real-time and adaptively, sudden, asynchronous, and close interactions between agents in the environment, leading to delayed regrouping responses. More seriously, this judgment metric is highly susceptible to interference from single-frame noise or regular output fluctuations. Due to the lack of effective multi-frame noise immunity verification and adaptive dynamic threshold adjustment mechanisms, the system is prone to frequent misjudgments and erroneous regrouping, severely impacting its robustness.

[0005] When an agent is removed or a group changes, existing techniques often require re-executing the complete feature encoding and attention network forward propagation for the reconstructed new group. This approach leads to exponentially increasing computational overhead as the multi-agent scale expands, lacking fine-grained branch masking and low-overhead prior selection mechanisms. Furthermore, in asynchronous dynamic environments, the lack of explicit temporal anchoring and double-buffered memory protection makes frequent dynamic reorganizations highly susceptible to disrupting the temporal continuity of Markov Decision Processes (MDPs), leading to decision turbulence.

[0006] In existing technologies, once a threshold condition is met, the agent is directly and hard-removed from the current group. Since group state information is obtained by attention-weighted aggregation of the representations of all individuals within the group, the sudden loss of an individual causes a drastic instantaneous discontinuity in the weight distribution of the attention mechanism (i.e., feature jump), leading to significant fluctuations in the network output. Simultaneously, during the backpropagation phase of training, due to the lack of a strong truncation mechanism for cross-branch gradient flow, spurious gradients caused by local reorganization can couple into other unaffected static groups, resulting in severe cross-branch gradient contamination.

[0007] For example, CN120494091A discloses a method, system, and medium for optimizing the performance of a multi-agent system. The method includes: performing round-based dialogue on the multi-agent system, and then using a large language model to obtain the contribution score of each agent and the communication value score of each communication path in the current round of dialogue; based on the contribution score and communication value score of a preset round, determining the agents and communication paths to be eliminated, and removing them from the multi-agent system; obtaining the current task's development stage within its project, and then recording the development stage, current task requirements, and the agents and communication paths to be eliminated as a preset historical optimal elimination mode. This technical solution uses a round-based semantic scoring mechanism based on a large language model, which has inference latency and cannot meet the real-time response requirements of this application for high-frequency burst interactions and single-frame noise from physical sensors. Furthermore, its logical elimination of the agent list does not address technical issues such as tensor dimension changes and cross-branch gradient pollution caused by dynamic recombination under a deep reinforcement learning architecture. Therefore, this text-layer scheduling scheme differs from the underlying computational architecture of this application and cannot solve the dynamic cooperative control problem.

[0008] In view of the above-mentioned deficiencies of the prior art, this application aims to provide a control method and system for dynamic reorganization of multiple agents, so as to solve the technical problems of the prior art being easily triggered by single-frame noise interference during dynamic grouping, having large computational overhead and easy timing disorder in local reorganization, and having policy divergence caused by feature jumps due to individual removal.

[0009] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0010] Existing technologies employ a fixed step-size mechanism in the dynamic grouping of heterogeneous multi-agent systems, which suffers from technical drawbacks such as delayed reorganization response and susceptibility to misjudgments caused by single-frame noise interference from the environment. Furthermore, frequent reorganization during individual removal and group changes leads to increased system computational overhead, impaired temporal continuity of the Markov decision process, and network output instability caused by feature jumps.

[0011] To address the shortcomings of existing technologies, this invention provides a dynamic collaborative system for heterogeneous multi-agent systems from a first aspect. The system includes a perception and evaluation module, an interruption triggering module, and a mask reassembly module.

[0012] The perception and evaluation module acquires observation data from each agent and extracts observation feature codes to evaluate the individual contribution of each agent to the task of its current group. The interruption triggering module calculates the temporal change rate of individual contributions. When the relative difference between the temporal change rate of the target agent's individual contribution and the average change rate of its group exceeds a set threshold, the asynchronous interruption condition is met, and an asynchronous interruption signal is generated. The mask reconstruction module responds to the asynchronous interruption signal by calculating the similarity between the target agent and existing groups based on the observation feature codes to filter candidate related groups. It then constructs a local reconstruction mask based on the candidate related groups and performs forward computation of the dynamic grouping network based on the local reconstruction mask to shield network computation branches for non-candidate related groups, performing state updates and reconstruction only on the target agent and candidate related groups. The system of this invention can capture sudden interactions through relative change rates, avoiding response lag caused by fixed step sizes; at the same time, it uses local masks to shield irrelevant branches, limiting the search to related groups and reducing system computational overhead.

[0013] According to a preferred embodiment, the system further includes a decoupled training module. The decoupled training module is configured to: during the network parameter training phase, perform gradient backpropagation on the network computation branches corresponding to the target agent and candidate associated groups that performed the state update and reconstruction based on the state update and reconstruction results of the mask reconstruction module, and perform gradient truncation on the network computation branches containing groups that did not undergo updates.

[0014] The decoupled training module cuts off the backpropagation link of the static group during the backpropagation stage, blocks the coupling of false gradients caused by local reorganization to the outside, solves the problem of cross-branch gradient pollution, and ensures the convergence stability of the joint strategy during the training process.

[0015] According to a preferred embodiment, the threshold is set as a dynamic adaptive threshold based on task load modulation; the interrupt triggering module is configured to: generate a baseline triggering threshold by statistically analyzing historical samples of relative differences through a sliding window; obtain the task load rate in the current local environment; and modulate the baseline triggering threshold using the task load rate to generate a dynamic adaptive threshold.

[0016] The interrupt triggering module dynamically updates the cross-group transfer conditions based on the task intensity of the local environment. When the task load is high, the interrupt triggering module increases the decision tolerance to ensure response speed, and decreases the tolerance when the task load is low to suppress invalid transfers and maintain system stability.

[0017] According to a preferred embodiment, when the temporal change rate of individual contribution is calculated and it is determined that the asynchronous interruption condition is met, the interrupt triggering module is configured to: perform smoothing preprocessing on individual contribution to suppress single-frame fluctuations; count the number of trigger frames in a set window where the relative difference exceeds a set threshold; generate an asynchronous interrupt signal when the proportion of the number of trigger frames in the set window meets a first preset condition; and after generating the asynchronous interrupt signal, release the abnormal state when the relative difference meets the release condition for multiple consecutive frames.

[0018] The interrupt triggering module uses smooth preprocessing and window proportion statistics to filter out false triggers caused by physical noise in a single frame of the sensor; the interrupt triggering module provides damping for state transitions by setting up a hysteresis exit logic with continuous verification of multiple frames, thus solving the problem of frequent network command jumps caused by routine exploration actions.

[0019] According to a preferred embodiment, when constructing a local reassembly mask and performing forward computation of the dynamic grouping network based on the local reassembly mask, the mask reassembly module is configured to: construct a binarized mask matrix; for candidate association groups, set the corresponding elements in the binarized mask matrix to a first value, and set the corresponding elements of non-candidate association groups to a second value; combine the binarized mask matrix with the original matching scores output by the dynamic grouping network to correct the matching scores corresponding to non-candidate association groups to a penalty value; and perform a probability normalization operation on the corrected matching scores to mask the network computation branches for non-candidate association groups.

[0020] The mask reconstruction module cuts off invalid cooperative links without changing the dimension of the input matrix by setting a penalty value and correcting the matching score matrix before probability normalization. The module uses matrix element-level algebraic operations to achieve branch masking, avoiding tensor dimension mismatch and execution anomalies caused by direct node removal.

[0021] According to a preferred embodiment, the system further includes a state timing alignment module, which is configured to: set a state buffer for each existing group; when performing state updates and reorganization, retain historical state snapshots and freeze their parameters for groups that have not performed updates, while marking the state of each existing group with a timestamp; when performing system global value calculation, ensure that groups that have performed state updates and reorganization use the new state at the current moment, and groups that have not performed updates use historical state snapshots with historical timestamps.

[0022] The state timing alignment module maintains historical feature alignment for static groups during asynchronous reassembly by allocating double buffers and timestamping states. This module ensures data consistency across different clock steps, maintaining the temporal continuity of the Markov decision process.

[0023] According to a preferred embodiment, if a target agent is removed from its existing group, the mask reconstruction module is configured to: employ virtual placeholder absorption nodes to take over the weights of the removed target agent, thereby reconstructing the group state information; wherein the individual contribution corresponding to the virtual placeholder absorption node decays over time. This invention, by introducing virtual placeholder absorption nodes to take over the weights of the removed agent, further solves the problem of feature jumps caused by individual removal leading to policy oscillations, and maintains the continuity of the numerical distribution.

[0024] The present invention provides a dynamic cooperation method for heterogeneous multi-agent systems from a second aspect. The method includes: acquiring observation data of each agent and extracting observation feature codes to evaluate the individual contribution of each agent to the task of its current group; calculating the temporal change rate of the individual contribution; when the relative difference between the temporal change rate of the individual contribution of the target agent and the average change rate of its group exceeds a set threshold, determining that the asynchronous interruption condition is met and generating an asynchronous interruption signal; responding to the asynchronous interruption signal, determining the candidate association groups of the target agent, constructing a local reorganization mask based on the candidate association groups, and performing forward computation of a dynamic grouping network based on the local reorganization mask to shield the network computation branches of non-candidate association groups, and performing state updates and reorganization only on the target agent and the candidate association groups.

[0025] This dynamic collaboration method generates an asynchronous interrupt signal when the relative difference in individual contributions exceeds a set threshold, shortening the system's topology reorganization response time for sudden interactions and avoiding control delays caused by fixed-time-step grouping. In the forward computation of this method, this step uses matrix element-level algebraic operations to mask non-candidate associated groups using a local reorganization mask without changing the dimension of the input matrix. This limits the reorganization computation to associated groups, reducing system data processing overhead and computing power costs when the scale of multi-agent operations expands. It also ensures the continuity of existing control commands for other static groups, reduces the task failure rate caused by local environmental changes, and improves the completion rate of multi-agent collaborative operations.

[0026] According to a preferred embodiment, the step of performing forward computation of a dynamic grouping network based on a local recombination mask to shield network computation branches for non-candidate association groups includes: constructing a binarized mask matrix; setting the corresponding elements in the binarized mask matrix to first values ​​for candidate association groups and setting the corresponding elements for non-candidate association groups to second values; combining the binarized mask matrix with the original matching scores output by the dynamic grouping network to correct the matching scores corresponding to non-candidate association groups to penalty values; and performing probability normalization on the corrected matching scores to shield network computation branches for non-candidate association groups.

[0027] This dynamic collaboration method sets a penalty value and modifies the matching score matrix before probability normalization. This masks network computation branches for non-candidate association groups without changing the dimension of the input matrix, ensuring control accuracy during the network topology transformation transition phase. This step achieves branch masking through matrix element-level algebraic operations, avoiding input tensor dimension mismatch caused by directly removing agent nodes. It eliminates the risk of system execution anomalies due to sudden changes in the underlying matrix dimension, simplifying operation and maintenance. Simultaneously, this method uses a binarized mask matrix and normalization operations to bring the output probability corresponding to non-candidate association groups close to zero, reducing numerical fluctuations in network output and providing smooth trajectory control for user-operated physical agents. Attached Figure Description

[0028] Figure 1 This is a simplified schematic diagram of the module connection relationship of a dynamic cooperative system for heterogeneous multi-agent systems provided by the present invention; Figure 2 This is a schematic diagram of the workflow and module interaction of a dynamic collaborative system for heterogeneous multi-agent systems provided by the present invention; Figure 3 This is a flowchart illustrating a dynamic collaboration method for heterogeneous multi-agent systems provided by the present invention.

[0029] List of reference numerals 100: Central Processing Unit; 101: Motherboard; 102: System Bus; 103: Memory Bus; 104: Peripheral Component Interconnect Bus; 200: Memory; 300: Graphics Processing Unit; 400: Communication Interface; 401: Radio Frequency Antenna; 500: Execution Terminal; 501: Local Processor; 502: Wireless Communication Module; 503: Sensor; 504: Unmanned Vehicle; 505: Unmanned Aerial Vehicle; 506: High-Speed ​​Data Bus; 600: Perception and Evaluation Module; 700: Interrupt Trigger Module; 800: Mask Reassembly Module; 900: State Timing Alignment Module; 1000: Decoupling Training Module. Detailed Implementation

[0030] The following is a detailed explanation with reference to the accompanying drawings.

[0031] This invention provides explanations and clarifications for some terms and concepts.

[0032] Heterogeneous multi-agent: refers to multiple independent execution units that differ in hardware configuration, sensor type, or observation dimension in a collaborative task environment.

[0033] Observation features include not only directly acquired raw physical signals (such as image pixels, radar point clouds), but more preferably, high-dimensional feature vectors or observation feature codes generated by deep neural network (such as CNN, Transformer or MLP) encoders.

[0034] Individual contribution: A weighted index representing the influence of a single agent in the local or global task achievement process. In specific embodiments, it can be implemented using the individual action value function (Q-value) in reinforcement learning, marginal contribution reward, or the attention weight distribution output by the attention mechanism network; this invention does not limit this to any particular method.

[0035] Penalty value: In mathematical computation, after subsequent probability normalization operations (such as the Softmax function), the output probability of the corresponding branch approaches zero (e.g., less than 10). -5 A negative real number (or equivalent to the lower limit of floating-point numbers in a computer). Ideally, it could be negative infinity (-inf); however, in practical code engineering, it could also be a negative number with a very large absolute value (such as 10). -9 All of these are equivalent embodiments of the present invention.

[0036] Virtual placeholder absorbing node: A virtual data volume used to maintain structural consistency in matrix dimensional operations. Physically, it does not correspond to any real intelligent agent; in the network data structure, it manifests as a placeholder tensor with zero values ​​or specific initialization feature encodings. Its core function is to absorb the parameter weights of removed entities, preventing parameter collapse caused by sudden dimensional changes during the forward propagation of group state information in the network.

[0037] Task load rate: Its quantitative calculation method can be configured heterogeneously according to the application scenario: it can be based on business indicators (such as the density of enemy targets to be attacked in the local field of view), or based on physical indicators (such as the current remaining power / battery endurance ratio of the intelligent agent), or computing power indicators (real-time memory or CPU usage percentage of the onboard computing unit).

[0038] State update and reorganization: In terms of data structure, this is reflected in updating the topological adjacency matrix or attention interaction mask of the multi-agent system, so that when the target agent propagates forward in the next time step, its information interaction channel is disconnected from the original group and a new information interaction connection is established with the candidate associated group.

[0039] Candidate association groups refer to groups selected from the existing group pool globally after the target agent triggers an asynchronous interruption, following a feature evaluation phase. These groups are prepared to establish new information interaction and collaborative topological relationships with the target agent. Physically or logically, they represent the new team the target agent is about to migrate to or join. In the computer program's code implementation, candidate association groups are represented as a set of activated group indices, cluster center labels, or a sequence of target nodes in a topological adjacency matrix. It is important to note that the number of candidate association groups in this invention is not limited to a single group. In mutually exclusive grouping (hard clustering) scenarios, it can be only the single target group with the highest similarity (i.e., Top-1); in overlapping grouping or collaborative attention allocation (soft clustering) scenarios, it can also be a set of multiple groups with similarity exceeding a certain threshold (i.e., Top-K). Whether the target agent ultimately joins a single group or establishes connections across multiple groups, it falls within the protection scope of this invention.

[0040] Example 1 In the dynamic collaborative system of this invention, heterogeneous multi-agent refers to multiple execution units with different sensor 503 configurations in the collaborative environment. For example, the system deploys an unmanned vehicle 504 equipped with a high-dimensional lidar and a drone 505 that relies on low-dimensional feature perception. The observation data acquired by the target agents differ in data modality and dimensionality.

[0041] When the local environment changes dynamically, the existing algorithm logic suffers from response latency. The original dynamic grouping algorithm, in order to reduce computational load and maintain system stability, adjusts the grouping every fixed step. The current application scenario involves high-frequency interactions and sudden adversarial situations. This fixed, periodic triggering mechanism leads to system response latency.

[0042] During the interval between two group computations, if the agent encounters a sudden resource contention, it maintains its original group identity and historical strategy to execute the task. Because the agent cannot leave its current group in real time based on the sudden event, and cannot perform state updates and reorganization in a timely manner, the failure rate of some tasks increases.

[0043] For example, taking a specific scenario of heterogeneous multi-agent collaborative target tracking task as an example, we will objectively explain the reasons for the above-mentioned technical problems.

[0044] The system is initially configured to perform a global group topology update every 10 time steps. At time step 0, the unmanned vehicle 504 equipped with a high-dimensional lidar and the drone 505 equipped with a low-dimensional camera are grouped into the same existing group and jointly perform a local area tracking task. At time step 2, the drone 505 encounters sudden signal interference or physical obstruction, causing changes in its acquired observation data and a significant decrease in its individual contribution to the current group's tracking task. Under the fixed-step grouping mechanism, the system must wait until time step 10 to trigger the next forward computation for network reorganization. During the interval between time steps 3 and 9, the system maintains the original group state information. During this period, the unmanned vehicle 504 maintains its cooperative connection with the drone 505 and continues to execute historical strategies in the absence of effective aerial observation features. Simultaneously, the system does not reorganize the unmanned vehicle 504 with other normally observed target agents in the neighboring area. During this interval, the system fails to respond to the sudden interaction event that occurred at time step 2. The delay in recombining calculations caused the autonomous vehicle 504 to continuously receive invalid cooperative parameters across multiple time steps, without any state update or recombination. This process directly resulted in the failure of the tracking task in that local area.

[0045] The technical problem solved by this invention is: how to achieve real-time response to high-frequency burst interactions in a heterogeneous multi-agent system that adopts a fixed step-size grouping mechanism, so as to avoid reorganization lag and local task failure.

[0046] This invention provides a dynamic collaborative system and method for heterogeneous multi-agent systems. It also provides a decoupled training method for multi-agent joint strategies, used to truncate the gradients of the global and local computational graphs during the backpropagation update phase of network parameters to block gradient coupling across branches. Furthermore, it provides a method for fusing heterogeneous multimodal features, introducing a regularization penalty term when calculating joint attention weights to balance the weight distribution of sensor data from different dimensions. Finally, it provides a threshold adaptive adjustment method based on dynamic load, using task load characteristics to drive cross-group transition conditions for adaptive adjustment. Finally, it provides an electronic device and a computer-readable storage medium for storing corresponding program instructions and executing them by a processor to implement the above steps.

[0047] like Figure 1As shown, the dynamic collaborative system for heterogeneous multi-agent systems of the present invention has a physical hardware structure including a central processing unit 100, a graphics processing unit 300, a memory 200, a communication interface 400, and an execution terminal 500 composed of heterogeneous multi-agent systems. The heterogeneous multi-agent systems include an unmanned vehicle 504 equipped with lidar and a drone 505 equipped with a camera. The memory 200 includes random access memory and read-only memory. The memory is used to store program instructions, network parameters, state buffers, and intermediate computation data.

[0048] like Figure 1 As shown, the central processing unit 100, graphics processor 300, memory 200, and communication interface 400 are physically connected via the system bus 102 on the motherboard 101. The central processing unit 100 and memory 200 are connected via the memory bus 103. The central processing unit 100 and graphics processor 300 are physically connected via the peripheral component interconnect bus 104. The communication interface 400 is connected to the radio frequency antenna 401 via a physical feeder. The sensor 503 inside the execution terminal 500 establishes a physical connection with its internal local processor 501 and wireless communication module 502 via the high-speed data bus 506. In high-frequency, dense sensing scenarios in a local environment, the sensor 503, carrying high-dimensional observation data, generates massive amounts of raw signals such as 3D point clouds. These signals are directly transmitted to the local processor 501 for front-end feature dimensionality reduction and pooling operations. After feature encoding extraction, the local processor 501 generates low-frequency physical execution instructions and lightweight status control messages, which are then carried and transmitted via the high-speed data bus 506. This layered processing architecture effectively avoids CAN bus paralysis caused by the direct impact of high-dimensional data. The communication interface 400 establishes a bidirectional wireless transmission link with the wireless communication module 502 of the execution terminal 500 through the radio frequency antenna 401. The cloud communication interface 400 sends a broadcast signal containing a global system timestamp and an asynchronous trigger baseline downward through the radio frequency antenna 401, establishing full-duplex communication with the wireless communication module 502, enabling the cloud motherboard 101 to form a closed-loop trajectory control feedback with the local processor 501 of the execution terminal.

[0049] like Figure 1 As shown, the perception evaluation module 600, interrupt triggering module 700, mask reassembly module 800, state timing alignment module 900, and decoupling training module 1000 are deployed as a program instruction set in memory 200. The central processing unit 100 and the graphics processing unit 300 execute the program instruction set. The output of the perception evaluation module 600 is connected to the input of the interrupt triggering module 700. The signal output of the interrupt triggering module 700 is connected to the trigger control terminal of the mask reassembly module 800. The state timing alignment module 900 establishes a data connection with the dual-state buffer in memory 200. The decoupling training module 1000 is connected to the backpropagation computation graph branch in the graphics processing unit 300.

[0050] like Figure 1 As shown, the execution terminal 500 sends observation data to the communication interface 400. The communication interface 400 receives the observation data from the execution terminal 500 and sends it to the memory 200. The memory 200 sends the observation data to the perception evaluation module 600. The perception evaluation module 600 receives the observation data from the memory 200, generates observation feature data and individual contribution data, sends the individual contribution data to the interrupt triggering module 700, and sends the observation feature data to the memory 200. The interrupt triggering module 700 receives the individual contribution data from the perception evaluation module 600, generates time-series change rate data, relative difference data, and asynchronous interrupt signal, and sends the asynchronous interrupt signal to the mask reconstruction module 800. The mask reconstruction module 800 receives the asynchronous interrupt signal from the interrupt triggering module 700, receives the observation feature data from the memory 200, generates similarity data, candidate association group data, and local reconstruction mask data, and sends the local reconstruction mask data to the graphics processor 300. The graphics processor 300 receives local reconstructed mask data from the mask reconstruction module 800, generates state update and reconstruction result data, and sends the state update and reconstruction result data to the mask reconstruction module 800. The mask reconstruction module 800 receives the state update and reconstruction result data from the graphics processor 300, sends the state update and reconstruction result data to the state timing alignment module 900 and the decoupling training module 1000, and simultaneously sends the state update and reconstruction result data to the communication interface 400 via the system bus 102. The communication interface 400 sends the result data to the corresponding execution terminal 500 via the radio frequency antenna 401, so that the corresponding local processor 501 can output specific physical execution trajectory control instructions based on the reconstruction result. The state timing alignment module 900 receives the state update and reconstruction result data from the mask reconstruction module 800, generates historical state snapshot data and timestamp data, and sends the historical state snapshot data and timestamp data to the memory 200. The decoupled training module 1000 receives state update and recombination result data from the mask recombination module 800, generates gradient backpropagation data and gradient truncation data, and sends the gradient backpropagation data and gradient truncation data to the graphics processor 300.

[0051] The dynamic cooperation system for heterogeneous multi-agent systems of the present invention is used to execute a dynamic cooperation method for heterogeneous multi-agent systems. The method includes steps S100 to S400, such as... Figure 3 As shown.

[0052] S100: Acquire observation data from each agent and extract observation feature codes to evaluate the individual contribution of each agent to the task of its current group, such as... Figure 2 As shown.

[0053] Heterogeneous multi-agent systems perform collaborative tasks in a local environment and collect multimodal environmental information through their own configured heterogeneous sensor networks to generate observation data. The execution terminal 500 sends the observation data to the communication interface 400. The communication interface 400 receives the observation data from the execution terminal 500 and sends it to the memory 200. The memory 200 sends the observation data to the perception and evaluation module 600. The perception and evaluation module 600 receives the observation data from the memory 200 and, after inputting the observation data, independently extracts observation features of a unified dimension through the dedicated network computing branches of each agent. Based on the encoded observation features, the perception and evaluation module 600 calculates the individual contribution of each agent in the current task to generate individual contribution data.

[0054] In one specific implementation, the heterogeneous multi-agent system includes an unmanned vehicle 504 equipped with LiDAR and a drone 505 equipped with a camera. The unmanned vehicle 504, positioned on a ground exploration path, collects 3D environmental information to generate high-dimensional point cloud observation data, while the drone 505, hovering above the local environment, collects 2D environmental information to generate low-dimensional image observation data. The perception and evaluation module 600 internally includes a point cloud processing branch and an image processing branch, which respectively receive high-dimensional and low-dimensional observation data to generate local environmental feature vectors in parallel.

[0055] In another optional implementation, the perception evaluation module 600 includes a feature alignment sub-network that performs dimensionality reduction pooling operations on high-dimensional data and zero-value dimensionality increase filling operations on low-dimensional data, uniformly mapping data from different modalities to a preset constant-dimensional space to generate observation feature data of the same size. Through the aforementioned network computation branches and the feature alignment sub-network, the underlying tensor conflicts caused by direct fusion computation are avoided, enabling the system to independently quantify responses to local environmental changes after eliminating data heterogeneity, regardless of the global state.

[0056] S200: Calculate the temporal change rate of individual contribution. When the relative difference between the temporal change rate of the target agent's individual contribution and the average change rate of its group exceeds a set threshold, it is determined that the asynchronous interruption condition is met and an asynchronous interruption signal is generated. Figure 2 As shown.

[0057] To prevent false triggering caused by single-frame fluctuations, the interrupt triggering module 700 receives individual contribution data from the perception evaluation module 600 and then executes a three-level noise reduction mechanism that includes smoothing preprocessing, sliding window counting, and hysteresis determination.

[0058] The interrupt trigger module 700 performs smoothing preprocessing on the raw data within consecutive time steps to suppress single-frame fluctuations. The calculation formula for the preprocessing is as follows: .

[0059] In the above formula, The individual contribution level after smoothing at the current moment. The original individual contribution at the current moment. The contribution of smoothed individuals in the previous time step. This is the smoothing coefficient. .

[0060] The interrupt trigger module 700 performs counting verification through a sliding window. The interrupt trigger module 700 sets the length of the setting window to... And calculate the relative difference within this setting window. Less than the dynamic adaptive threshold (i.e., satisfy) < The number of trigger frames) The triggering frame rate within the set window satisfies the first preset condition if and only if the percentage of the triggering frame rate within the set window meets the first preset condition (i.e. ,in The interrupt triggering module 700 only determines that the asynchronous interrupt condition is met and generates an asynchronous interrupt signal when the relative difference reaches a threshold (the threshold percentage is not met). To avoid repeated oscillations at the set threshold boundary, the interrupt triggering module 700 introduces a delayed exit mechanism. After generating the asynchronous interrupt signal, when the relative difference is continuous... The frame satisfies the release condition (i.e.) ,in When the hysteresis occurs, the interrupt trigger module 700 will then remove the abnormal state of the target agent. The specific principle behind this is explained below.

[0061] The interrupt trigger module 700 generates relative difference data for individual contributions based on the individual contribution level and the group average level. The formula for calculating the relative difference is: .

[0062] In the above formula, The relative difference in individual contributions. The temporal change rate of the individual contribution of the target intelligent agent. This represents the average rate of change for agents of the same type within the group. The average rate of change for the group as a whole. They are weighted coefficients of the same type, and .

[0063] .

[0064] Indicates the group to which agent i belongs at time t. Rate of change of Q within, This represents the system's microstep size.

[0065] .

[0066] .

[0067] In the above formula, Let k be the set of agents of class k within the group. For the corresponding quantity, G is the set of all agents in the group.

[0068] The relative attenuation index is constructed using the above formula. To deduct the fluctuation components of the group as a whole and the average of the same category. The larger the absolute negative value, the more the agent's contribution declines significantly beyond the group average, thus accurately preserving the abnormal contribution fluctuations of the target agent itself (i.e., individual sudden anomalies, rather than a decline in overall group returns). Its core function is: if and only if the relative difference... When the threshold is less than the dynamic threshold, the interrupt triggering module 700 determines that a sudden interaction event of the target agent has occurred, thus eliminating the high-frequency false triggers caused by other teammates in the same group performing suboptimal exploration actions that lower the expected benefits of the whole team.

[0069] To balance the interference of environmental noise and the characteristics of different task stages, the interrupt triggering module 700 adopts a two-layer judgment logic that combines historical sample statistics within a sliding window with task load rate modulation to generate a dynamic adaptive threshold.

[0070] The interrupt triggering module 700 extracts historical samples of relative differences and calculates the baseline triggering threshold based on the mean and standard deviation. Specifically, the interrupt triggering module 700 extracts historical samples of relative differences to construct a sliding statistical baseline: it takes the relative differences within the most recent W micro-time steps. Sample, calculate the historical mean within the sliding window. Compared with historical standard deviation The calculation formula is as follows: .

[0071] .

[0072] Based on the above sliding statistics, we introduce The baseline trigger threshold is calculated in principle. The formula for calculating the baseline trigger threshold is: .

[0073] In the above formula, The baseline trigger threshold, This represents the historical mean of the relative differences within the sliding window. For historical standard deviation, For safety factors. Preferably, The value is 3.

[0074] The interrupt triggering module 700 modulates the baseline triggering threshold based on the current local environment's task load rate, generating a dynamically adaptive threshold. The formula for calculating the dynamically adaptive threshold is: .

[0075] In the above formula, For dynamic adaptive threshold, The baseline trigger threshold, The lower limit coefficient of load modulation and , This refers to the task load rate. The task load rate is specifically calculated as the ratio of the number of remaining unfinished tasks in the current local environment to the initial total number of tasks.

[0076] Through the above two-layer judgment logic, the dynamic collaboration system realizes the dynamic update of cross-group transfer conditions: when the task load rate is high due to intensive tasks and intense resource competition, the absolute value of the modulated dynamic adaptive threshold is reduced accordingly, making it easier for the target agent to meet the asynchronous interruption condition, thereby ensuring the response speed under high-frequency burst interaction; while when the task enters the final sparse stage and the task load rate is low, the absolute value of the modulated dynamic adaptive threshold is increased accordingly, thereby effectively suppressing the frequent invalid transfer behavior of heterogeneous multi-agents between residual tasks and avoiding state oscillation.

[0077] The interrupt trigger module 700 performs statistical verification to determine the proportional condition for triggering abnormal states. The formula for calculating the proportional condition is: .

[0078] In the above formula, To meet the trigger frame count for determining the difference, To set the total number of frames in the window, This is the threshold for triggering the percentage.

[0079] In one specific implementation, the interrupt trigger module 700 is set with a fixed hysteresis constant, and generates an exception release signal when the relative difference data meets the release condition. The formula for calculating the release condition is: .

[0080] In the above formula, The relative difference in individual contributions. For dynamic adaptive threshold, This is the hysteresis constant.

[0081] In another optional implementation, the interrupt trigger module 700 reads the total number of unassigned task targets and the maximum rated capacity at full load, calculates the floating-point ratio between the two, and uses it as the input formula to the aforementioned task load rate parameter Load(t) for nonlinear annealing adjustment. The aforementioned decision logic isolates the normal exploration interference from teammates in the same group, and introduces a hysteresis constant to form physical damping to eliminate command jumps caused by single-frame noise from sensor 503.

[0082] Existing technologies (such as CN112396187A) employ heuristic static grouping, which is prone to failure due to rule rigidity in highly dynamic environments. Conventional event-triggered mechanisms (such as CN116794987A) rely on absolute error thresholds, which are susceptible to high-frequency command jumps when encountering single-frame physical noise from sensors. The interrupt triggering module 700 in this application utilizes a baseline triggering threshold based on sliding window statistics and an additional hysteresis constant to accurately isolate sudden interactions and achieve noise immunity against global routine exploration actions, significantly reducing the computational overhead of reorganization determination.

[0083] S300: In response to an asynchronous interrupt signal, it calculates the similarity between the target agent and existing groups based on observed feature encoding to filter candidate associated groups, such as... Figure 3 As shown.

[0084] The mask reconstruction module 800 receives an asynchronous interrupt signal from the interrupt triggering module 700 and reads the currently cached individual feature codes of the target agent and the group state codes of each existing group from the memory 200. The mask reconstruction module 800 calculates the cosine similarity between the target agent and each existing group based on vector inner product operations. The formula for calculating the cosine similarity is: .

[0085] In the above formula, To calculate the obtained cosine similarity, Encode vectors for the individual features of the target intelligent agent. This is the group state encoding vector for the existing groups, where T is the matrix transpose symbol. Let L2 be the norm of the vector.

[0086] The mask reconstruction module 800 sorts the calculated cosine similarity in descending order and extracts specific groups whose similarity values ​​meet the preset truncation conditions to generate candidate associated group data.

[0087] In one specific implementation, the mask reassembly module 800 extracts the top 3 existing groups with the highest cosine similarity values ​​in descending order, and performs a list concatenation operation between the top 3 existing groups and the original historical groups to generate a set of data containing 4 related objects, which is then sent to the forward computing node.

[0088] In another alternative implementation, the mask reconstruction module 800 acquires the absolute spatial Euclidean distance data between the target agent and each existing group, and compares it numerically with the rated communication physical radius parameter. The mask reconstruction module 800 truncates the search paths for groups whose Euclidean distance data is greater than the rated parameter, and generates a filtered list of available groups to perform the aforementioned similarity calculation.

[0089] The mask reconstruction module 800 only performs a complete forward computation of the reconstruction discriminant network on the selected candidate association group set, skipping all branches corresponding to other groups. This prior screening operation reduces the time complexity of the forward propagation reconstruction search from a full traversal... Significantly reduced to Where N is the total number of agents, G is the total number of groups, and K is a constant much smaller than G. This mechanism implements precise on-demand computing power allocation, effectively avoiding the exponential increase in network forward inference computational complexity caused by the scaling up of heterogeneous multi-agent dynamic cooperative systems.

[0090] S400: Constructs a local recombination mask based on candidate association groups, and performs forward computation of the dynamic grouped network based on the local recombination mask to mask the network computation branches for non-candidate association groups, such as... Figure 3 As shown, state updates and reorganization are performed only on the target agent and candidate associated groups, such as... Figure 3 As shown.

[0091] The mask reassembly module 800 specifically includes local reassembly mask matrix construction logic, penalty value algebraic superposition operator, and probability normalization execution unit. The penalty value algebraic superposition operator corrects the original matching score matrix of non-candidate association groups, and then the probability normalization execution unit masks the corresponding branches, thereby eliminating the risk of system tensor dimension mismatch in the computation flow.

[0092] Specifically, the mask reassembly module 800 performs forward computation of the dynamic grouping network based on a local reassembly mask to shield the network computation branches of non-candidate association groups. The steps include: constructing a binarized mask matrix; setting the corresponding elements in the binarized mask matrix as the first value for candidate association groups and setting the corresponding elements of non-candidate association groups as the second value; combining the binarized mask matrix with the original matching scores output by the dynamic grouping network to correct the matching scores corresponding to non-candidate association groups to a penalty value; and performing probability normalization on the corrected matching scores to shield the network computation branches of non-candidate association groups.

[0093] Specifically, the mask reconstruction module 800 constructs a binary mask matrix of dimension N×G, setting the matrix elements corresponding to candidate association groups to constant 1 and non-candidate groups to constant 0. The mask reconstruction module 800 obtains the original matching score output by the network and superimposes it with the mask matrix to generate a corrected score. The formula for calculating the corrected score is: .

[0094] In the above formula, To correct the score matrix, L is the original matching score matrix, and M is the binarized mask matrix. The symbol for the Hadama product is used. It is a negative infinity constant.

[0095] The mask reconstruction module 800 performs a standard probability normalization operation on the corrected score. The formula for probability normalization is: .

[0096] In the above formula, The output probability of assigning the target agent to a specific group is given by exp, which is an exponential function. To correct the corresponding elements in the score matrix, G represents the total number of groups.

[0097] When the target agent is removed, the mask reassembly module 800 generates a code containing virtual placeholder features. The virtual placeholder absorbing node has a regularized Q-value that decays exponentially over time. The formula for calculating the exponential decay is: .

[0098] In the above formula, The regularized Q-value of the virtual placeholder node at the current moment. Let t be the initial regularized Q-value of the agent being removed, and t be the current time. Let exp be the shift-out time, and let exp be the exponential function. This is the smoothing time constant, used to control the transition duration.

[0099] In one specific implementation, the mask reassembly module 800 extracts the feature vectors of all resident agents in the original group within the most recent historical time step, performs an arithmetic average pooling operation to generate a historical average vector, and assigns it to the virtual placeholder feature encoding. .

[0100] In another optional implementation, the mask reassembly module 800 extracts the attention probability distribution of the observation features of each modality in the joint attention aggregation stage and calculates the attention distribution information entropy. The formula for calculating the attention distribution information entropy is: .

[0101] In the formula, For the information entropy of the attention distribution, Attention probability distribution assigned to each modality observation feature within the group. It is a logarithmic function.

[0102] The mask reconstruction module 800 multiplies the information entropy by an adjustment coefficient to generate regularization penalty data, which is then added to the model's forward loss function. The mask reconstruction module 800 utilizes a masking mechanism to perform local updates algebraically within the framework, and introduces absorbing nodes to maintain the continuity of the underlying numerical distribution when the agent detaches.

[0103] In one specific implementation, when the target intelligent agent meets the asynchronous interrupt condition and performs state update and reorganization, the execution terminal 500 exhibits synchronous response characteristics in the physical entity's motion state parameters and hardware signals. For example, when performing a target tracking task, when the UAV 505 leaves its original formation and establishes a local reorganization mask with the candidate associated group, the onboard local processor 501 of the UAV 505 outputs trajectory correction instructions based on the reorganized forward calculation results. Specifically, the network decision actions output by the mask reorganization and state update are transformed into specific physical execution instructions of the execution terminal (such as adjusting the motor PWM duty cycle, changing the rotor speed or servo yaw angle, etc.), thereby completing the actual trajectory control of the physical entity. Driven by this physical instruction, the UAV 505 exhibits non-smooth step changes in flight heading angle and spatial acceleration. Observing the high-speed data bus 506 through an external message analysis device reveals that the transmission frequency of the cooperative control message for the original group decreases within the preset clock cycle after the asynchronous interrupt signal is generated, while the bandwidth occupancy of the feature interaction message for the new associated group increases significantly. By observing sudden changes in the trajectory of terminal devices or detecting changes in the message frequency of the bus, it is possible to externally verify whether the system has triggered the dynamic mask reassembly logic.

[0104] S500: Sets up state buffers for each existing group and performs state timing alignment.

[0105] The state timing alignment module 900 independently allocates dual-state buffers (current and previous frame) for each existing group within the local environment in the memory 200. The state timing alignment module 900 only overwrites the data of the active state group in the current buffer and generates historical state snapshot data containing group characteristics and physical timestamps. During global value aggregation calculation, the state timing alignment module 900 synchronously outputs new state data for the active group and historical replay state data for groups not involved in the reorganization.

[0106] In one specific implementation, the state timing alignment module 900 generates a system asynchronous reassembly step identifier bit and binds it to the data frame. It then calculates the physical time difference between the historical timestamp corresponding to the stationary group and the current system time and appends it to the identifier bit. The global hybrid network performs a decay calculation on the initial discount factor based on the physical time difference data to generate the corresponding reward discount coefficient.

[0107] In another alternative implementation, the state timing alignment module 900 sets a maximum tolerance staleness constant parameter. When the time span between the historical timestamp and the current clock exceeds this constant parameter, a refresh command signal is generated. Upon receiving this signal, the mask reassembly module 800 initiates a full state update by activating the network branch corresponding to the static group. Data isolation is achieved through a dual-state buffer and timestamp marking architecture to ensure the timing integrity of the Markov decision process.

[0108] S600: During the training phase of network parameters, decoupled training is performed based on the state update and reorganization results.

[0109] like Figure 2 As shown, the decoupled training module 1000 performs gradient backpropagation on the network computation branches corresponding to the target agent and candidate association groups that perform state updates and reorganizations, and performs gradient truncation on the network computation branches of groups that have not undergone updates. The decoupled training module 1000 reads the state update and reorganization results and generates gradient truncation data for the corresponding computation branches for non-candidate association groups.

[0110] The formula for calculating gradient cutoff is: ; .

[0111] In the above formula, The partial derivative of the overall error gradient for the static branch computation graph of non-candidate association groups. To maintain the frozen set of network parameters from the previous time step, The observation encoding network parameters for the i-th agent are... Encode network parameters for the individual information of the i-th agent. A set of agents that have not triggered any exceptions. For the group information hyperparameters of the g-th group, the network parameters are... For the group Q-value hyperparameters of the g-th group, the network parameters are... This is the set of static groups that are not covered.

[0112] While performing masking blocking, the decoupled training module 1000 performs regular parameter updates on the activated network branches of the target agent and candidate association groups involved in the reconstruction. The calculation formula for the regular parameter update is: , .

[0113] In the above formula, To activate the set of parameter tensors for the computational branch of the network, The model learning rate parameter, To activate the overall error gradient tensor of the branch backpropagation, For the set of agents that trigger the reorganization, For the set of candidate association groups involved, The assignment update symbol.

[0114] In one specific implementation, the decoupled training module 1000 constructs a sequence of binary mask vectors and performs a Hadamard product operation on these vectors and the gradient partial derivative tensors corresponding to the global hybrid network weight matrix. The formula for calculating the Hadamard product is as follows: .

[0115] In the above formula, Let the gradient partial derivative tensor of the weight matrix be... The weight matrix of the global hybrid network (corresponding to the parameters of the global hybrid hyperparameter network). ), It is a sequence of binary mask vectors. This is the symbol for the Hadama product.

[0116] In another optional implementation, the decoupled training module 1000 extracts the gradient matrix of the activated network computation branches and calculates and generates L2 norm data. When the L2 norm data is greater than a preset network stability boundary constant, the decoupled training module 1000 performs a scaling operation to proportionally reduce the gradient matrix and generate corrected gradient backpropagation data.

[0117] The decoupled training module 1000 adds a special asynchronous identifier to the samples of the asynchronous reassembly step in the experience replay pool, and automatically enables the hierarchical gradient truncation rule based on this identifier during training sampling. By cutting off the backhaul link of the invariant region through the above hierarchical truncation, it effectively avoids the unaffected static groups and remaining agents from receiving unreasonable false credit assignments due to the asynchronous departure of teammates; at the same time, it ensures that the global hybrid network only updates the weight branches corresponding to the activated groups, and sets the gradients of the weight branches corresponding to the frozen groups to zero, thereby ensuring that the global joint value decomposition always conforms to the Individual-Global Maximization (IGM) principle constraint in the asynchronous environment, and consolidates the convergence stability of the multi-agent joint strategy.

[0118] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A dynamic cooperative system for heterogeneous multi-agent systems, characterized in that, The system includes: The perception evaluation module (600) is used to acquire the observation data of each agent and extract the observation feature encoding to evaluate the individual contribution of each agent to the task of its current group. The interrupt triggering module (700) is used to calculate the temporal change rate of the individual contribution. When the relative difference between the temporal change rate of the individual contribution of the target agent and the average change rate of its group exceeds a set threshold, it is determined that the asynchronous interruption condition is met and an asynchronous interruption signal is generated. The mask reassembly module (800) is used to respond to the asynchronous interrupt signal, calculate the similarity between the target agent and the existing groups based on the observed feature encoding to filter candidate associated groups, construct a local reassembly mask based on the candidate associated groups, and perform forward computation of the dynamic grouping network based on the local reassembly mask to shield the network computation branches of non-candidate associated groups, and perform state update and reassembly only on the target agent and candidate associated groups.

2. The system according to claim 1, characterized in that, It also includes a decoupled training module (1000), which is configured as follows: During the training phase of the network parameters, gradient backpropagation is performed on the network computation branches corresponding to the target agent and candidate associated groups that have performed state updates and recombinations based on the state update and recombination results of the mask recombination module (800), and gradient truncation is performed on the network computation branches of groups that have not performed updates.

3. The system according to claim 1 or 2, characterized in that, The set threshold is a dynamic adaptive threshold based on task load modulation; the interrupt triggering module (700) is configured to: Historical samples of the relative differences are statistically analyzed using a sliding window to generate a baseline trigger threshold; Get the task load rate in the current local environment; The baseline trigger threshold is modulated using the task load rate to generate the dynamic adaptive threshold.

4. The system according to any one of claims 1 to 3, characterized in that, When the time-series change rate of the individual contribution is calculated and the asynchronous interrupt condition is met, the interrupt triggering module (700) is configured as follows: Perform smoothing preprocessing on the individual contribution values ​​to suppress single-frame fluctuations; The number of trigger frames whose relative difference exceeds the set threshold within the specified window is counted; When the proportion of the number of trigger frames within the set window meets the first preset condition, the asynchronous interrupt signal is generated; After the asynchronous interrupt signal is generated, the abnormal state is lifted when the relative difference satisfies the release condition for multiple consecutive frames.

5. The system according to any one of claims 1 to 4, characterized in that, When constructing a local reassembly mask and performing forward computation of a dynamic packet network based on the local reassembly mask, the mask reassembly module (800) is configured as follows: Construct a binary mask matrix. For the candidate association groups, set the corresponding elements in the binary mask matrix as the first value, and set the corresponding elements of the non-candidate association groups as the second value. The binarized mask matrix is ​​combined with the original matching score output by the dynamic grouping network to correct the matching score corresponding to the non-candidate association group to a penalty value. A probability normalization operation is performed on the corrected matching score to mask the network computation branch for the non-candidate association groups.

6. The system according to any one of claims 1 to 5, characterized in that, When the dynamic grouping network aggregates the observed feature codes of different modalities through joint attention, the mask reassembly module (800) is further configured to: Extract the attention probability distribution of each modality's observation features during joint attention aggregation, and calculate the information entropy of the attention distribution; The attention distribution information entropy is multiplied by an adjustment coefficient to generate regularization penalty data, and the regularization penalty data is superimposed on the forward loss function of the dynamic grouping network.

7. The system according to any one of claims 1 to 6, characterized in that, The system further includes a state timing alignment module (900), which is configured to: Set up a status buffer for each existing group; When performing the state update and reorganization, a snapshot of the historical state of the groups that have not been updated is retained and their parameters are frozen, while a timestamp is added to the state of each existing group. When performing system-wide value calculation, groups that perform the state update and reorganization use the new state at the current moment, while groups that do not perform the update use the historical state snapshot with historical timestamps.

8. The system according to any one of claims 1 to 7, characterized in that, If the target agent is removed from its existing group, the mask reconstruction module (800) is configured to: Virtual placeholder absorption nodes are used to take over the weights of the removed target agents in order to reconstruct the group state information; The individual contribution of the virtual occupant absorption node decreases over time.

9. A dynamic cooperation method for heterogeneous multi-agent systems, characterized in that, The method includes: Acquire observation data from each agent and extract observation feature codes to evaluate the individual contribution of each agent to the task of its current group; Calculate the temporal change rate of the individual contribution. When the relative difference between the temporal change rate of the individual contribution of the target agent and the average change rate of its group exceeds a set threshold, it is determined that the asynchronous interruption condition is met and an asynchronous interruption signal is generated. In response to the asynchronous interrupt signal, candidate association groups of the target agent are determined, and a local reorganization mask is constructed based on the candidate association groups. Forward computation of the dynamic group network is performed based on the local reorganization mask to shield the network computation branches of non-candidate association groups, and state updates and reorganization are performed only on the target agent and candidate association groups.

10. The method according to claim 9, characterized in that, The step of performing forward computation of the dynamic grouping network based on the local reassembly mask to mask the network computation branches of non-candidate association groups includes: Construct a binary mask matrix, and set the corresponding element in the binary mask matrix to the first value for the candidate association group, and set the corresponding element of the non-candidate association group to the second value; The binarized mask matrix is ​​combined with the original matching score output by the dynamic grouping network to correct the matching score corresponding to the non-candidate association group to a penalty value. A probability normalization operation is performed on the corrected matching score to mask the network computation branch for the non-candidate association groups.

Images