Multi-agent collaborative communication method and system based on graph neural network and variational encoder

Abstract

The application provides a kind of multi-agent collaborative communication method and system based on graph neural network and variational encoder, comprising: obtaining local state feature and local communication interaction graph topology information;Input it into the graph structure value evaluation network to obtain the expected information gain;When the expected information gain is greater than the single communication cost threshold, trigger the communication request, and input the high-dimensional belief state of itself into the variational bottleneck encoder for compression to obtain the low-dimensional hidden variable feature, and broadcast the low-dimensional hidden variable feature to the neighbor agent;Decoding is carried out using the decoder of the variational bottleneck encoder to obtain the neighbor approximate belief state;Fuse and update the local maintained belief state to obtain the fused belief state;Generate and execute the motion control instruction of the current time step, so that the application jointly optimizes the communication opportunity and the communication content, greatly reduces the communication frequency and the transmission bandwidth, saves the core energy consumption and resources, and ensures the high collaborative control and formation accuracy.

Description

Technical Field

[0001] This invention relates to the field of multi-agent cooperative control technology, specifically to a multi-agent cooperative communication method and system based on graph neural networks and variational encoders. Background Technology

[0002] In practical applications such as warehousing and logistics fleets, disaster relief teams, and air / satellite constellations, communication constraints are a core bottleneck in multi-agent cooperative control. In these systems, communication costs include bandwidth, energy consumption, latency, and packet reliability under unstable links. Therefore, the core requirement of the system is not "how to conduct more communication," but rather "how to maintain the formation quality of multiple agents within a tight communication budget." Existing technologies for addressing communication constraints in multi-agent systems mainly include: while classical formation and consensus control can provide stable trajectory tracking, they typically assume frequent or fixed-period communication, resulting in significant waste of communication resources.

[0003] Event-triggered and self-triggered strategies reduce communication by triggering communication when local errors exceed a set threshold. However, this relies heavily on human experience to adjust the threshold and is difficult to adapt to dynamically changing scenarios.

[0004] Learning-based communication scheduling incorporates methods such as graph neural networks to improve communication scheduling. However, these methods typically optimize "communication timing" and "communication content" as two separate problems. The free energy principle states that the decision-making objective of an agent is to minimize expected uncertainty (i.e., the expected information gain should be greater than the operational cost). However, directly applying this principle to distributed systems faces two major challenges: first, the magnitude of information gain cannot be predicted before communication; second, the transmission of the entire belief state is extremely wasteful of bandwidth. This invention, inspired by the free energy principle, introduces graph neural networks to predict information gain and combines this with a variational encoder to compress information, thereby completely solving the aforementioned problems. Summary of the Invention

[0005] This invention is made to solve the above-mentioned problems, and aims to provide a multi-agent cooperative communication method and system based on graph neural networks and variational encoders.

[0006] This invention provides a multi-agent cooperative communication method based on graph neural networks and variational encoders, characterized by the following steps: Step S1, acquiring local state features and local communication interaction graph topology information at the current time step; Step S2, inputting the local state features and local communication interaction graph topology information into a pre-trained graph structure value evaluation network to obtain the expected information gain; Step S3, comparing the expected information gain with a preset single communication cost threshold. If the expected information gain is greater than the single communication cost threshold, a communication request is triggered, and the high-dimensional belief state is input into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features, which are then broadcast to neighboring agents within the communication range; Step S4, receiving the low-dimensional latent variable features broadcast from neighboring agents and decoding them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state; Step S5, fusing and updating the neighbor's approximate belief state with the locally maintained belief state to obtain a fused belief state; Step S6, generating and executing motion control commands for the current time step based on the fused belief state.

[0007] The multi-agent cooperative communication method based on graph neural network and variational encoder provided by the present invention may also have the following features: wherein the local state features include: normalized position tracking error, velocity, self-belief uncertainty, number of neighbors, average neighbor distance and normalized timestamp.

[0008] The multi-agent cooperative communication method based on graph neural networks and variational encoders provided in this invention may also have the following features: the graph structure value evaluation network adopts a multi-layer graph convolutional network structure containing residual connections and has dual outputs, one output head outputting the probability of triggering communication and the other output head outputting a scalar estimate of the expected information gain.

[0009] The multi-agent cooperative communication method based on graph neural network and variational encoder provided by this invention may also have the following features: the dimension of the high-dimensional belief state is 18-dimensional, and the dimension of the low-dimensional latent variable features is 8-dimensional.

[0010] The multi-agent cooperative communication method based on graph neural network and variational encoder provided by the present invention may also have the following feature: wherein, in step S5, the neighbor approximate belief state and the locally maintained belief state are fused and updated by an exponential moving average algorithm.

[0011] The multi-agent cooperative communication method based on graph neural network and variational encoder provided by the present invention may also have the following features: wherein step S6 includes the following sub-steps: step S61, generating planar velocity reference values ​​through the underlying planner according to the fused belief state; step S62, mapping the planar velocity reference values ​​to the linear velocity and angular velocity control commands of the nonholonomic constraint robot.

[0012] The multi-agent cooperative communication method based on graph neural network and variational encoder provided by the present invention may also have the following feature: wherein the loss function of the variational bottleneck encoder takes into account both reconstruction error and KL divergence constraint.

[0013] This invention provides a multi-agent cooperative communication system based on graph neural networks and variational encoders, characterized by the following features: a state acquisition module for acquiring local state features and local communication interaction graph topology information at the current time step; a value evaluation module for inputting the local state features and local communication interaction graph topology information into a pre-trained graph structure value evaluation network to obtain the expected information gain; a communication decision module for comparing the expected information gain with a preset single communication cost threshold, triggering a communication request when the expected information gain is greater than the single communication cost threshold, and inputting its own high-dimensional belief state into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features, and broadcasting the low-dimensional latent variable features to neighboring agents within the communication range; a communication decoding module for receiving the low-dimensional latent variable features broadcast from neighboring agents and decoding them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state; a fusion module for fusing and updating the neighbor's approximate belief state with the locally maintained belief state to obtain a fused belief state; and an instruction generation module for generating and executing motion control instructions for the current time step based on the fused belief state.

[0014] The role and effect of invention

[0015] The multi-agent cooperative communication method and system based on graph neural networks and variational encoders according to the present invention includes: step S1, acquiring local state features and local communication interaction graph topology information at the current time step; step S2, inputting the local state features and local communication interaction graph topology information into a pre-trained graph structure value evaluation network to obtain the expected information gain; step S3, comparing the expected information gain with a preset single communication cost threshold, and if the expected information gain is greater than the single communication cost threshold, triggering a communication request, and inputting its own high-dimensional belief state into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features, and then disseminating them to neighboring agents within the communication range. Step S4: Receive the low-dimensional latent variable features broadcast from neighboring agents and decode them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state; Step S5: Fuse and update the neighbor's approximate belief state with the locally maintained belief state to obtain the fused belief state; Step S6: Generate and execute the motion control command for the current time step based on the fused belief state. Therefore, the multi-agent cooperative communication method and system based on graph neural networks and variational encoders of the present invention, by jointly optimizing the communication timing and communication content under the free energy principle, extremely reduces the communication frequency and transmission bandwidth, saves core energy consumption and resources, and ensures extremely high cooperative control and formation accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a multi-agent collaborative communication method based on graph neural networks and variational encoders in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the communication framework of a multi-agent collaborative communication method based on graph neural networks and variational encoders in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of a multi-agent cooperative communication system based on graph neural networks and variational encoders in an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram comparing nominal formation control in an embodiment of the present invention.

[0020] Figure 5 This is a schematic diagram comparing the anti-interference robustness test in an embodiment of the present invention.

[0021] Figure 6 This is a diagram showing the ablation experiment results in an embodiment of the present invention.

[0022] Figure 7 This is a schematic diagram comparing the standardized open collaboration (MPE) benchmark tests in an embodiment of the present invention.

[0023] Figure 8 This is a schematic diagram comparing network congestion stress tests in an embodiment of the present invention.

[0024] Figure 9 This is a schematic diagram of the playback test comparison of the real physical radio log (COSMO-Bench) in an embodiment of the present invention. Detailed Implementation

[0025] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0026] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the multi-agent cooperative communication method and system based on graph neural networks and variational encoders of the present invention.

[0027] Example

[0028] Figure 1 This is a flowchart illustrating a multi-agent collaborative communication method based on graph neural networks and variational encoders in an embodiment of the present invention.

[0029] like Figure 1 As shown, this embodiment provides a multi-agent cooperative communication method based on graph neural networks and variational encoders, including:

[0030] Step S1: Obtain the local state features and local communication interaction graph topology information of the current time step.

[0031] The local state features include: normalized location tracking error, velocity, self-belief uncertainty, number of neighbors, average neighbor distance, and normalized timestamp.

[0032] Step S2: Input the local state features and local communication interaction graph topology information into the pre-trained graph structure value evaluation network to obtain the desired information gain.

[0033] In this embodiment, the desired information gain is the desired entropy reduction. It is used to characterize the information gain that communication may bring.

[0034] The graph structure value assessment network adopts a multi-layer graph convolutional network structure with residual connections and has dual outputs. One output head outputs the probability of triggering communication, and the other output head outputs a scalar estimate of the expected information gain.

[0035] Step S3: Compare the expected information gain with the preset single communication cost threshold. If the expected information gain is greater than the single communication cost threshold, a communication request is triggered. The high-dimensional belief state is input into the pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features. The low-dimensional latent variable features are then broadcast to neighboring agents within the communication range.

[0036] Preset a fixed threshold for the cost of a single communication. When the expected entropy reduction When this occurs, a communication request is triggered, and the desired entropy reduction is achieved. At this time step, remain silent and directly execute step S4. In actual deployment, this depends on the bandwidth limitations and power budget of the communication hardware. In benchmark tests, the fixed threshold for event triggering is dynamically adjusted within the range of (1.5 ~ 2.0). By setting a unified fixed byte packet (8-dimensional latent variables + fixed frame header), the communication cost is modeled as a constant. Setting an appropriate trigger threshold ensures that the system does not send out random information when there is no sudden uncertainty.

[0037] When a communication request is triggered, the agent To one's own higher-dimensional belief state Input to a pre-trained variational bottleneck encoder ( The encoding end of the high-dimensional belief state. This high-dimensional belief state contains the agent's... The latent variable features include positional deviations, velocity deviations, and uncertainty variances of itself and its nearest neighbors. The variational bottleneck encoder maps and compresses these into a low-dimensional latent variable feature. Low-dimensional latent variable characteristics It retains the most critical information for collaborative tasks. (Agent) The latent variable feature is broadcast to neighbors within the communication radius via a wireless communication module. .

[0038] In this design, the high-dimensional belief state has 18 dimensions, the low-dimensional latent variable features have 8 dimensions, and the weight coefficient of the variational bottleneck encoder is 0.8. This parameter controls the balance between information compression and reconstruction error. Experiments demonstrate that when the weight coefficient is 0.8, the 18-dimensional high-dimensional belief features can be successfully compressed into 8-dimensional latent variable features (achieving 56% data dimensionality reduction) while maintaining extremely low reconstruction error, ensuring that the key information required for collaborative tasks is not lost.

[0039] The loss function of the variational bottleneck encoder takes into account both reconstruction error and KL divergence constraint.

[0040] Step S4: Receive the low-dimensional latent variable features broadcast from the neighboring agent and decode them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state.

[0041] In this embodiment, the intelligent agent Monitor and receive latent variable features sent from neighboring agents. The data is then input to the decoding end of the variational bottleneck encoder to reconstruct the neighbor approximate belief state. .

[0042] Step S5: Merge and update the neighbor's approximate belief state with the locally maintained belief state to obtain the merged belief state.

[0043] In step S5, the neighbor approximate belief state and the locally maintained belief state are merged and updated using the exponential moving average (EMA) algorithm.

[0044] Step S6: Generate and execute motion control instructions for the current time step based on the fused belief state.

[0045] Step S6 includes the following sub-steps:

[0046] Step S61: Based on the fused belief state, generate a planar velocity reference value through the underlying planner. ).

[0047] Step S62: Map the planar velocity reference value to the linear and angular velocity control commands of the nonholonomically constrained robot.

[0048] Calculate linear velocity The velocity reference value is projected based on the current heading angle and combined with the maximum speed limit. The angular velocity is then calculated. The deviation between the desired heading angle and the current heading angle is calculated, multiplied by the control gain, and then combined with the maximum angular velocity limit.

[0049] intelligent agent Perform physical movement based on the calculated linear and angular velocities to complete formation coordination at the current time step.

[0050] The control execution frequency for the multi-agent system is set to 10Hz. 10Hz is the standard frequency for closed-loop kinematic control of mobile robots (such as unmanned logistics vehicles and drones), which can ensure smooth tracking of formation trajectories and allow the system to complete communication timing judgment, information encoding and decoding, and state fusion calculation within each control cycle (100ms).

[0051] Figure 2This is a schematic diagram of the communication framework of a multi-agent collaborative communication method based on graph neural networks and variational encoders in an embodiment of the present invention.

[0052] like Figure 2 As shown, the principle of this embodiment is as follows:

[0053] An operation is only beneficial when the expected reduction in uncertainty (information gain) exceeds the operational cost. Mapped to communication algorithms, the expected information gain is defined as the desired reduction in entropy. That is, the difference between the entropy of the agent's current belief state and the posterior entropy of the belief state after receiving the potential communication message.

[0054] Define the decrease in net expected free energy as: Only when The algorithm only allows communication to be triggered when the entropy decrease is greater than 0. Since an agent cannot directly calculate the actual entropy reduction before receiving information from neighboring agents, this problem is transformed into a deep learning supervised regression problem. Specifically, during offline training, data generated by classic high-frequency communication algorithms (such as EC-DAIF) is used to calculate the actual entropy reduction after actual communication as a label. A graph-structured value evaluation network (GNN) is then used to approximate this expected value. The complex free energy formula is simplified into a trigger calculation formula for ultra-fast online inference:

[0055]

[0056] in, For indicator functions, Output the activation function for the network. This is the final operation threshold. For graph neural network evaluators, For intelligent agents At time step The "final communication trigger decision variable".

[0057] Each agent in this embodiment At each time step The online communication and control loop process is as follows:

[0058] Input: Local state feature matrix Local communication interaction graph topology Cost per communication Output: Control commands and communication actions.

[0059] 1. Obtain current state features: Extract features from the current state itself and its neighbors to form... = [Normalized location error, velocity, belief uncertainty, number of neighbors, average neighbor distance, timestamp].

[0060] 2. Expected Free Energy Gain Assessment (Timing Judgment): Invoking the Graph Neural Network Estimator :

[0061] Predicted expected information gain: ( , )

[0062] 3. Triggering determination mechanism:

[0063] IF ( > THEN

[0064] / / Trigger communication to perform content compression

[0065] The 18-dimensional high-dimensional belief containing information about the three nearest neighbors. Input encoder Generate 8-dimensional latent variables:

[0066] Broadcast hidden variables via radio to within the communication radius

[0067] END IF

[0068] 4. Message reception and decoding integration:

[0069] FOR each received from a neighbor Hidden variable messages D0

[0070] Call the decoder Restoring approximate beliefs:

[0071] )

[0072] Based on the Exponential Moving Average (EMA) algorithm, Merge and update with the locally maintained global belief state

[0073] END FOR

[0074] 5. Perform motion control:

[0075] Input the merged and updated belief state into the underlying control planner;

[0076] Calculate and output the linear velocity for a nonholonomically constrained robot (such as a unicycle model). and angular velocity instruction.

[0077] The core of this invention relies on a specific neural network structure and training method, specifically configured as follows:

[0078] 1. The structure of the Graph Neural Network (GNN) includes:

[0079] Network architecture: A 2-layer graph convolutional network (GCN) is used.

[0080] Hidden layer settings: The hidden layer dimension is set to 128, the ReLU activation function is used, and residual connections are added between layers to prevent gradient vanishing.

[0081] Output layer (dual-head design):

[0082] First: Output the probability of Sigmoid activation (classification task).

[0083] First two: Output scalar value estimate (regression task, predict expected entropy reduction).

[0084] Loss function: Combining binary cross-entropy loss (BCE) and mean squared error loss (MSE), i.e.:

[0085]

[0086] in, This represents the total loss function of the graph neural network evaluator. (Binary Cross-Entropy) represents the binary cross-entropy loss. Describing the intelligent agent of network prediction The probability of communication triggering, Describe the agent in offline training data Communication trigger tag, This represents the weighting adjustment coefficient. Describing the intelligent agent of network prediction The expected value of communication This represents the actual ex-post information gain calculated based on historical exercise data. This represents the mean square error.

[0087] 2. Variational bottleneck encoder ( The structural settings include:

[0088] Input feature definition: The input is an 18-dimensional vector, consisting of the summation of the states of the three nearest neighbors. Each neighbor contains 6 features: ( If there are fewer than 3 neighboring units, fill in the remaining units with 0.

[0089] Encoder topology: Fully connected layer architecture, with dimensions decreasing from 18 to 64 to 32 to 8, ultimately outputting an 8-dimensional mean and log-variance.

[0090] Decoder topology: A structure symmetrical to the encoder, with dimensions increasing from 8 to 32 to 64 to 18 for feature reconstruction.

[0091] The loss function formula is:

[0092] in, This represents the total loss function of the variational bottleneck encoder. This represents the original high-dimensional belief state input to the encoder. This represents the low-dimensional, compact latent variables generated after encoding. Indicates the use of latent variables The reconstructed approximate belief state To reconstruct the error term, This is the weighting adjustment coefficient. The constraint term for the KL divergence balances reconstruction error and information compression.

[0093] 3. Training environment and strategies include:

[0094] The optimizer used is the Adam optimizer, with a learning rate set to 10⁻³.

[0095] Generalization Training Distribution: To ensure the robustness of the algorithm across various scenarios, a total of 10,000 episodes were used for pre-training. The scenario distribution was strictly set as follows: Static (no disturbance): 40%, Obstacles: 30%, Perturbation (location change disturbance): 20%, Dynamic (object dynamic change): 10%.

[0096] This training configuration directly supports the beneficial effect of the present invention, which has stronger environmental adaptability compared to the prior art.

[0097] Figure 3 This is a schematic diagram of a multi-agent cooperative communication system based on graph neural networks and variational encoders in an embodiment of the present invention.

[0098] like Figure 3 As shown, this embodiment provides a multi-agent cooperative communication system 100 based on graph neural networks and variational encoders, including: a state acquisition module 10, a value assessment module 20, a communication decision module 30, a communication decoding module 40, a fusion module 50, and an instruction generation module 60.

[0099] The state acquisition module 10 is used to acquire the local state characteristics and local communication interaction graph topology information of the current time step.

[0100] The value assessment module 20 is used to input local state features and local communication interaction graph topology information into a pre-trained graph structure value assessment network to obtain the desired information gain.

[0101] The communication decision module 30 compares the expected information gain with a preset single communication cost threshold. When the expected information gain is greater than the single communication cost threshold, a communication request is triggered, and its own high-dimensional belief state is input into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features. The low-dimensional latent variable features are then broadcast to neighboring agents within the communication range.

[0102] The communication decoding module 40 is used to receive low-dimensional latent variable features broadcast from neighboring agents and decode them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state.

[0103] The fusion module 50 is used to merge and update the neighbor's approximate belief state with the locally maintained belief state to obtain the fused belief state.

[0104] The instruction generation module 60 is used to generate and execute motion control instructions for the current time step based on the fused belief state.

[0105] According to the multi-agent cooperative communication method and system based on graph neural networks and variational encoders involved in this embodiment, it includes: Step S1, acquiring local state features and local communication interaction graph topology information at the current time step; Step S2, inputting the local state features and local communication interaction graph topology information into a pre-trained graph structure value evaluation network to obtain the expected information gain; Step S3, comparing the expected information gain with a preset single communication cost threshold, and if the expected information gain is greater than the single communication cost threshold, triggering a communication request, and inputting its own high-dimensional belief state into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features, and then communicating with neighboring agents within the communication range. Step S4: Receive the low-dimensional latent variable features broadcast from neighboring agents and decode them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state; Step S5: Fuse and update the neighbor's approximate belief state with the locally maintained belief state to obtain the fused belief state; Step S6: Generate and execute the motion control command for the current time step based on the fused belief state. Therefore, the multi-agent cooperative communication method and system based on graph neural networks and variational encoders of the present invention optimizes the communication timing and communication content under the free energy principle, thereby drastically reducing the communication frequency and transmission bandwidth, saving core energy consumption and resources, and ensuring extremely high cooperative control and formation accuracy.

[0106] Figure 4 This is a schematic diagram comparing nominal formation control in an embodiment of the present invention.

[0107] like Figure 4 As shown, in a nominal formation control experiment involving five agents, the comparison was made with an existing communication control algorithm based on a classic event-triggered mechanism (EC-DAIF algorithm). Experimental results show that within the same task cycle, the average number of communication triggers in this embodiment is 42.9, a reduction of 36.7% compared to the 67.8 triggers of the comparison algorithm. Regarding communication bandwidth, because this embodiment compresses the 18-dimensional original state features into 8-dimensional latent variable features through a variational bottleneck encoder, the total bandwidth consumption is 1372.8 KB, a significant reduction of 71.9% compared to the 4881.6 KB consumed by the comparison algorithm when transmitting the full state. Simultaneously, the average formation tracking error recorded in this embodiment is 0.0446, essentially the same as the 0.0443 of the comparison algorithm. These experimental data objectively demonstrate that this embodiment significantly reduces communication resource consumption without causing a substantial decrease in system control accuracy.

[0108] Figure 5 This is a schematic diagram comparing the anti-interference robustness test in an embodiment of the present invention.

[0109] like Figure 5 As shown, this embodiment can adaptively allocate the communication budget when facing high uncertainties, such as interference from external physical collisions or dynamic changes in the target, exhibiting excellent anti-interference robustness and rapid recovery capability. In the robustness test for high uncertainty environments, interference conditions exceeding the normal training intensity were set, with the existing self-triggered communication algorithm (Minimal-CAVI algorithm) as the comparison object.

[0110] In the positional change interference experiment (applying a set instantaneous position offset to the agent), the average communication volume of this embodiment was 349.5 times, which was 63.9% less than the 967.9 times of the comparison algorithm. The average time required for the agent formation to recover to a steady state was 147.3 control steps, which was faster than the 162.3 control steps of the comparison algorithm.

[0111] In a dynamic target tracking experiment (where the formation target undergoes an expected rotational change), this embodiment, while maintaining the same level of tracking error (2.2304 compared to the comparative algorithm's 2.2188), reduces the total communication requirements from 905.4 times in the comparative algorithm to 359.1 times (a reduction of 60.3%). This demonstrates that this embodiment can effectively avoid ineffective concurrent communication broadcasts and improve the system's adaptive recovery efficiency when environmental uncertainties surge.

[0112] Figure 6This is a diagram showing the ablation experiment results in an embodiment of the present invention.

[0113] like Figure 6 As shown, this embodiment demonstrates that by using GNN to optimize timing and VAE to compress content, the combined use of these two methods yields greater benefits compared to using them individually. Ablation experiments with controlled variables further validate the synergistic effect of each module. Under the same formation task conditions, based on the basic communication strategy without the optimization of this invention (communication volume 403.4 times): if only GNN is applied for communication timing optimization, the communication volume is reduced to 301.6 times (a reduction of 25.2%); if only VAE is applied for communication content compression, the communication volume is reduced to 182.6 times (a reduction of 54.7%); while in this embodiment, when both are applied together, the communication volume is further reduced to 137.8 times, with a total reduction rate of 65.8%. The above data proves that the "joint optimization of timing decision and content compression" scheme adopted in this embodiment produces synergistic gains, overcoming the limitations of optimizing both in isolation in existing technologies.

[0114] In four distinct public benchmark tests—MPE, LBF, RWARE, and Overcooked—this embodiment reduced communication by an average of 87.2% compared to Full-Comm, and statistical tests showed no significant decrease in task rewards. It has a strong generalization advantage in complex real-world network fluctuations and various standardized collaborative tasks.

[0115] Experimental statistics show that, in the four types of tasks mentioned above, this embodiment reduces the amount of communication messages sent by an average of 87.2% compared to the benchmark.

[0116] Figure 7 This is a schematic diagram comparing standardized public collaboration (MPE) benchmark tests in an embodiment of the present invention. Figure 7 As shown, taking the MPE benchmark test as an example, the average number of communications decreased from 150.0 times in the comparison benchmark to 17.6 times. Furthermore, statistical tests showed no significant difference in task execution rewards and coverage errors between the two algorithms (0.850 vs. 0.839 for the comparison algorithm). This objectively verifies that the communication strategy of this embodiment has reliable generalization performance in different types of collaborative tasks.

[0117] Figure 8 This is a schematic diagram comparing network congestion stress tests in an embodiment of the present invention.

[0118] like Figure 8As shown, this embodiment simulates a real network under "packet loss + latency" stress tests, comparing it with the fixed-periodic-comm method in the face of moderate to severe network degradation. Under moderate network degradation conditions, the adaptive event-triggered communication volume of this embodiment is 19.4 times (lower than the 21.0 times of fixed-periodic communication), and the task coverage error is reduced to 0.873 under this budget (better than the 0.891 of fixed-periodic communication). The event-triggered communication volume of this embodiment is not only lower than that of fixed-periodic communication across the board, but also achieves lower task coverage error in 3 out of 4 stress levels. This indicates that this embodiment can prioritize the transmission of high-value feature information when the network is congested.

[0119] Figure 9 This is a schematic diagram of the playback test comparison of the real physical radio log (COSMO-Bench) in an embodiment of the present invention.

[0120] like Figure 9 As shown, in the playback test using real physical WiFi and radio logs, the benchmark in this embodiment is the full communication method. Experimental data shows that the communication trigger rate of this embodiment is only 28.6% (i.e., saving 71.4% of communication resource usage), but it still achieves 100% coverage of critical information events and 100% capture rate of outlier events. Under the same limited communication rate, the critical information coverage rate of fixed-period communication drops significantly to 83.9%. In real-world, complex physical network deployment environments, this embodiment can balance communication bandwidth constraints with system communication reliability.

[0121] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-agent cooperative communication method based on graph neural networks and variational encoders, characterized in that, include: Step S1: Obtain the local state features and local communication interaction graph topology information of the current time step; Step S2: Input the local state features and the local communication interaction graph topology information into a pre-trained graph structure value evaluation network to obtain the desired information gain; Step S3: Compare the expected information gain with a preset single communication cost threshold. If the expected information gain is greater than the single communication cost threshold, trigger a communication request and input its own high-dimensional belief state into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features. Then, broadcast the low-dimensional latent variable features to neighboring agents within the communication range. Step S4: Receive the low-dimensional latent variable features broadcast from the neighboring agent and decode them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state. Step S5: The neighbor approximate belief state and the locally maintained belief state are fused and updated to obtain the fused belief state; Step S6: Generate and execute motion control commands for the current time step based on the fused belief state.

2. The multi-agent cooperative communication method based on graph neural networks and variational encoders according to claim 1, characterized in that: in, The local state features include: normalized location tracking error, velocity, self-belief uncertainty, number of neighbors, average neighbor distance, and normalized timestamp.

3. The multi-agent cooperative communication method based on graph neural networks and variational encoders according to claim 1, characterized in that: in, The graph structure value assessment network adopts a multi-layer graph convolutional network structure with residual connections and has dual outputs. One output head outputs the probability of triggering communication, and the other output head outputs the scalar estimate of the expected information gain.

4. The multi-agent cooperative communication method based on graph neural networks and variational encoders according to claim 1, characterized in that: in, The high-dimensional belief state has 18 dimensions, and the low-dimensional latent variable features have 8 dimensions.

5. The multi-agent cooperative communication method based on graph neural networks and variational encoders according to claim 1, characterized in that: in, In step S5, the neighbor approximate belief state and the locally maintained belief state are merged and updated using an exponential moving average algorithm.

6. The multi-agent cooperative communication method based on graph neural networks and variational encoders according to claim 1, characterized in that: in, Step S6 includes the following sub-steps: Step S61: Based on the fused belief state, generate a planar velocity reference value through the underlying planner; Step S62: Map the planar velocity reference value to the linear and angular velocity control commands of the nonholonomically constrained robot.

7. The multi-agent cooperative communication method based on graph neural networks and variational encoders according to claim 1, characterized in that: in, The loss function of the variational bottleneck encoder takes into account both reconstruction error and KL divergence constraint.

8. A multi-agent cooperative communication system based on graph neural networks and variational encoders, characterized in that, include: The state acquisition module is used to acquire the local state characteristics and local communication interaction graph topology information of the current time step; The value assessment module is used to input the local state features and the local communication interaction graph topology information into a pre-trained graph structure value assessment network to obtain the expected information gain. The communication decision module is used to compare the expected information gain with a preset single communication cost threshold. When the expected information gain is greater than the single communication cost threshold, a communication request is triggered, and its own high-dimensional belief state is input into a pre-trained variational bottleneck encoder for compression to obtain low-dimensional latent variable features. The low-dimensional latent variable features are then broadcast to neighboring agents within the communication range. The communication decoding module is used to receive the low-dimensional latent variable features broadcast from the neighboring agent and decode them using the decoder of the variational bottleneck encoder to obtain the neighbor's approximate belief state. The fusion module is used to fuse and update the neighbor's approximate belief state with the locally maintained belief state to obtain the fused belief state; The instruction generation module is used to generate and execute motion control instructions for the current time step based on the fused belief state.

Images