An unmanned vehicle cooperative control method based on a gated-hypergraph neural network

Abstract

The application discloses a kind of unmanned vehicle cooperative control method based on gate-supernetwork neural network, belong to vehicle networking and automatic driving technical field, including: obtaining the real-time state information of several vehicle clusters in traffic scene;Supernetwork representation model is constructed based on real-time state information, node feature matrix, hyperedge feature matrix and correlation matrix are obtained based on supernetwork representation model, input into the action value and weighted global feature obtained by pre-constructed gate-supernetwork neural network;Based on action value and weighted global feature, obtain the joint action of unmanned vehicle cluster.The application models high-order interaction relationship between vehicles through supernetwork structure, breaks through the binary limitation of traditional graph model, significantly improves the multi-element interaction modeling capability of complex mixed traffic scene.Combined with the dynamic screening of multi-head attention features by dimension, effectively filter noise and enhance key decision variable identification, realize more robust, stable cooperative control under observation limitation or environmental disturbance.

Description

Technical Field

[0001] This invention belongs to the fields of vehicle networking and autonomous driving technology, and particularly relates to a cooperative control method for unmanned vehicles based on a gated hypergraph neural network. Background Technology

[0002] In intelligent transportation systems, the autonomous decision-making capabilities of unmanned vehicles directly impact overall traffic safety and operational efficiency. With the increasing prevalence of mixed traffic environments, the differences in intelligence levels, communication capabilities, and behavioral patterns among different types of vehicles are intensifying, making traditional rule-based and static strategy-dependent methods inadequate for handling dynamic and uncertain scenarios. Therefore, collaborative decision-making methods for unmanned vehicle systems have gradually become a focus of innovation. These methods model the interaction relationships between vehicles, guiding unmanned vehicles to achieve global behavioral optimization under local observation conditions, thereby effectively improving the adaptability and collaborative efficiency of unmanned vehicles in complex environments.

[0003] In hybrid autonomous transportation systems, unmanned vehicles face challenges such as dense interactions among heterogeneous participants and information asymmetry. Their decision-making process heavily relies on understanding and modeling the relationships between multiple vehicles within the traffic environment. Current unmanned vehicle decision-making methods generally suffer from insufficient interaction modeling capabilities when dealing with complex traffic scenarios, making it difficult to capture high-order collaborative relationships among vehicle clusters. Furthermore, existing methods lack effective structure perception mechanisms to address information asymmetry and target differences arising from heterogeneous vehicle clusters, hindering the achievement of refined strategy adjustments. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a cooperative control method for unmanned vehicles based on a gated hypergraph neural network, thereby resolving the issues present in the prior art.

[0005] To achieve the above objectives, this invention provides a cooperative control method for unmanned vehicles based on a gated hypergraph neural network, comprising:

[0006] Obtain real-time status information of several vehicle clusters in a traffic scenario;

[0007] A hypergraph representation model is constructed based on the real-time state information, and node feature matrix, hyperedge feature matrix and association matrix are obtained based on the hypergraph representation model.

[0008] An initial neural network is constructed based on a multi-head attention mechanism. A gating mechanism is then introduced into the initial neural network to obtain a gated hypergraph neural network. The gated hypergraph neural network includes a local network, a global network, and a hybrid network.

[0009] The node feature matrix, the hyperedge feature matrix, and the correlation matrix are input into the gated-hypergraph neural network to obtain local features and weighted global features.

[0010] The joint actions of the unmanned vehicle swarm are obtained based on the local features and weighted global features.

[0011] Optionally, the set of node hyperedges in the hypergraph representation model includes three types of interactions: homogeneous interaction edges, heterogeneous interaction edges, and hybrid interaction edges.

[0012] Optionally, the local network includes a two-layer hypergraph convolutional structure and a gating mechanism; the two-layer hypergraph convolutional structure includes a non-attention hypergraph convolutional layer and a multi-head attention hypergraph convolutional layer.

[0013] Optionally, after the node feature matrix, the hyperedge feature matrix, and the association matrix are input into the local network, initial adjacency features are extracted through an attentionless hypergraph convolutional layer. The initial adjacency features are then processed through a multi-head attention hypergraph convolutional layer to obtain multi-head spliced ​​features. After weighting the multi-head spliced ​​features through a gating mechanism, they are mapped to the dimension of the action space through two fully connected layers and the local features are output. The gating mechanism includes a linear mapping layer and a Sigmoid activation function.

[0014] Optionally, the global network includes a two-layer hypergraph convolutional structure, a gating mechanism, a global average pooling layer, and a global attention layer; the two-layer hypergraph convolutional structure includes a non-attention hypergraph convolutional layer and a multi-head attention hypergraph convolutional layer.

[0015] Optionally, after the node feature matrix, the hyperedge feature matrix, and the association matrix are input into the global network, initial adjacency features are extracted through an attentionless hypergraph convolutional layer. The initial adjacency features are then processed through a multi-head attention hypergraph convolutional layer to obtain multi-head spliced ​​features. After the multi-head spliced ​​features are weighted through a gating mechanism, they are processed through a global average pooling layer. The output of the global average pooling layer is then input into the global attention layer to obtain weighted global features.

[0016] Optionally, the hybrid network maps the input local features and weighted global features to a high-dimensional latent space through a feature mapping layer, and then processes them sequentially through a hypergraph convolutional layer, a multi-head hypergraph attention layer, and a gated multi-head attention mechanism. The processed local features and weighted global features are aggregated and dimensionality reduced through a global average pooling layer, and then processed through a global attention layer to obtain a global feature representation. The global feature representation is then mapped to the joint action value function output space through two fully connected layers to obtain a decision signal. Based on the decision signal, the joint action of the unmanned vehicle cluster is obtained.

[0017] Optionally, residual connections are used before and after the non-attention hypergraph convolutional layer, the multi-head attention hypergraph convolutional layer, and the gating mechanism.

[0018] The present invention provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0020] Compared with the prior art, the present invention has the following advantages and technical effects:

[0021] 1) This invention proposes a hypergraph structure for modeling the relationships between traffic participants, which can depict high-order interaction relationships between multiple vehicles. It breaks through the limitation of traditional graph models that can only express pairwise relationships, and significantly improves the ability to model multiple interaction patterns in complex traffic scenarios. It is especially suitable for mixed traffic environments with clustered collaborative behavior and heterogeneous communication structures.

[0022] 2) This invention combines a gating mechanism to dynamically filter the features of multi-head attention output dimension by dimension, effectively filtering redundant and noisy information, improving the model's ability to identify key decision variables, and exhibiting stronger robustness and stability under conditions of limited vehicle observation or large environmental disturbances. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0024] Figure 1 This is a schematic diagram of a gated hypergraph neural network according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the hypergraph and association matrix according to an embodiment of the present invention;

[0026] Figure 3 This is a schematic diagram illustrating the construction method of the inter-cluster association matrix according to an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram illustrating the construction method of the inter-cluster hyperedge feature matrix in an embodiment of the present invention. Detailed Implementation

[0028] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0030] Example 1

[0031] Graph Neural Networks (GNNs), as a deep learning method capable of depicting structured relationships, are suitable for expressing the dynamic connections between various vehicles in a transportation system. By introducing graph structure representations, the model can fully integrate the state information of neighboring vehicles, achieving joint perception of the local and global environment. This provides a more accurate state representation for the decision-making process, thereby enhancing the responsiveness and collaborative level of autonomous vehicles in complex traffic situations.

[0032] In real-world traffic environments, the interactions between vehicles exhibit both structural and dynamic characteristics. Graph neural networks can capture the coupling features between individual behaviors and group patterns by modeling nodes and their connections. In autonomous vehicle decision-making systems, the introduction of graph structures not only enhances the model's ability to perceive local traffic conditions but also provides structured prior information for policy generation. Through dynamic updates of node representations and adaptive adjustments of edge weights, the system can more accurately reflect the degree of influence between different vehicles, thereby guiding autonomous vehicles to make more rational action choices and execute strategies in highly interactive scenarios.

[0033] like Figure 1 As shown, this embodiment provides a cooperative control method for unmanned vehicles based on a gated hypergraph neural network, including:

[0034] The system acquires real-time state information of several vehicle clusters in a traffic scenario; constructs a hypergraph representation model based on the real-time state information, and obtains node feature matrices, hyperedge feature matrices, and correlation matrices based on the hypergraph representation model; constructs an initial neural network based on a multi-head attention mechanism, and introduces a gating mechanism into the initial neural network to obtain a gated-hypergraph neural network; the gated-hypergraph neural network includes local networks, global networks, and hybrid networks; inputs the node feature matrices, hyperedge feature matrices, and correlation matrices into the gated-hypergraph neural network to obtain local features and weighted global features; and obtains the joint actions of the unmanned vehicle cluster based on the local features and weighted global features.

[0035] Furthermore, by constructing a cluster-level hypergraph representation, the multi-dimensional relationships between different vehicles are modeled using hyperedges, and a gating mechanism is introduced to filter key feature information. Local networks capture behavioral dependencies within the cluster, while the global network integrates cross-cluster interaction information and achieves feature aggregation under the guidance of a multi-head attention mechanism. Finally, the hybrid network outputs a joint action value function to assign optimal control commands to various types of unmanned vehicles, achieving efficient and stable multi-vehicle collaborative strategy generation.

[0036] ;

[0037] in, Let be the joint action value function, representing the cumulative discounted reward that the system expects to obtain when performing the joint action u in global state s. It is the set of states consisting of the states of all agents. It is the set of joint actions of all agents. For parameters of the hybrid network, This represents the set of all parameters. The joint action value function takes the local action values ​​(Q-values) of each agent obtained from the local network and the global features provided by the global network as input, and achieves nonlinear aggregation and mapping through a hybrid network.

[0038] The Mixing Network architecture is designed to fuse feature information from the outputs of the Local Network and the Global Network to generate a Joint Action-Value Function suitable for collaborative control of multiple CAV clusters. Specifically, the Mixing Network is implemented using a Hypergraph Neural Network architecture, which includes several core modules such as feature mapping layers, hypergraph convolutional layers, multi-head hypergraph attention layers, gated multi-head attention mechanism layers, global average pooling layers, and global attention layers. The specific design process is as follows: First, the input local features and weighted global features are mapped to a high-dimensional latent space through a feature mapping layer; then, the high-order interaction relationship between nodes and hyperedges is captured through a hypergraph convolutional layer; next, the weight information of multi-scale interaction between different clusters is further learned through a multi-head hypergraph attention layer; then, the gated multi-head attention mechanism is used to dynamically filter and enhance the features; next, the above features are aggregated and dimensionality reduced through a global average pooling layer to obtain the overall feature representation of the scene; finally, a fine global feature representation is obtained through a global attention layer, and this global feature is further mapped to the joint action value function output space through two fully connected layers to generate the final decision signal.

[0039] The local network is responsible for processing and outputting the feature information and local action value function (local Q-value) within each CAV cluster, capturing the local control strategy of a single cluster. The global network is used to capture the global state features of the entire traffic environment to provide global contextual information support. The hybrid network plays a crucial role in bridging the gap between the local and global networks, effectively fusing the local Q-values ​​of each cluster output by the local network with the global feature information provided by the global network, ultimately outputting a joint action value function applicable to the entire unmanned vehicle system. Therefore, the hybrid network plays a pivotal role in the network structure and information flow path, and is a key network structure for achieving efficient fusion of local and global decisions. The hybrid network structure is shown in Table 1.

[0040] Table 1

[0041] in, For the first The number of CAVs within a cluster, here is .

[0042] The detailed parameters of the gated attention layer network set in this embodiment are shown in Table 2.

[0043] Table 2

[0044] For mixed traffic scenarios, the hyperedge feature matrix, node feature matrix, and correlation matrix are first obtained through cluster-based hypergraph representation, including both local and global (local / global hyperedge feature matrix, local / global node feature matrix, and local / global correlation matrix). These are then input into the local and global networks, respectively. For the local network, node features are first mapped to the hidden layer feature space; subsequently, the processed features are processed through hypergraph convolution; the processed data is then guided by a multi-head hypergraph attention mechanism, utilizing hyperedge features to guide information propagation; next, a gated multi-head attention mechanism is used to process the features; finally, two fully connected layers map the features to the action space dimension and output local features (joint action value function (Q-value)). The global network differs from the local network in that after weighting the features using a gated multi-head attention mechanism, a global average pooling layer is used for further processing, followed by a global attention layer to obtain weighted global features. Furthermore, residual connections are used before and after hypergraph convolution, hypergraph attention, and the gating mechanism throughout the entire process. The outputs of the two networks can eventually be fed into algorithms such as MAPPO (Multi-Agent Proximal Policy Optimization), QMIX, and COMA (Counterfactual Multi-Agent Policy Gradients) to generate joint actions for each CAV cluster.

[0045] The basic principle of hypergraphs:

[0046] Hypergraph can be used To indicate, among which, Indicates a hypergraph A set of nodes, and The diagram shows... A hypergraph is a set of hyperedges. Unlike traditional graph representations, edges in a hypergraph can connect any number of nodes. In other words, each hyperedge... Both are sets of nodes A subset, namely By using hyperedges to represent the multivariate relationships between multiple nodes, hypergraphs can more naturally represent complex relationships in applications such as mixed transportation, social groups, and recommendation systems.

[0047] The incidence matrix and the hyperedge feature matrix can be used to mathematically represent a hypergraph.

[0048] Correlation Matrix It can be used to represent the connection relationship between a node and a hyperedge, where, The number of nodes in the hypergraph. Let be the number of hyperedges in the hypergraph. The construction rules for the incidence matrix are as follows: Figure 2 As shown, for node and super edge ,have:

[0049] .

[0050] The hyperedge feature matrix can be used to represent the features of the hyperedge. For a given hyperedge... There are 10 superedges, and the feature dimension of each superedge is 0. The hypergraph whose hyperedge feature matrix It can be represented as:

[0051] ;

[0052] in, For super-edge eigenvectors.

[0053] Cluster-based hypergraph representation:

[0054] The cluster-based hypergraph representation method in this embodiment is an extension of the cluster-based graph representation method. Assuming that in... In the moment-to-moment scene A cluster of CAVs (Connected and Automated Vehicles), for the cluster It contains Vehicles of CAVs, and these CAVs observed a total of The remaining vehicles. Then the entire transportation system is... Hypergraph representation of time It can be jointly characterized by the observation hypergraph of all clusters at that moment, that is:

[0055] ;

[0056] in, Indicates at time No. The observation hypergraph structure of a CAV cluster The features are stored in the node feature matrix and hyperedge feature matrix In the cluster, the topological connection relationship between vehicle nodes and hyperedges is specifically manifested as the attribution and connection structure between nodes and hyperedges, which is determined by the correlation matrix. express, Let be the correlation matrix of the i-th cluster at time t.

[0057] These three concepts will be introduced one by one.

[0058] First, for the node feature matrix (In the hypergraph structure at time t, the first...) The system constructs a feature representation matrix for each node. Each row of this matrix typically corresponds to a node (e.g., an autonomous vehicle or a human-driven vehicle), and each column represents the specific characteristics of the node at that moment (e.g., vehicle speed, position, acceleration, safety factor, etc.). Secondly, for the association matrix, based on the vehicle types and interaction properties in the TMA (Traffic Mixed Autonomy System), this embodiment defines three hyperedge categories to accurately characterize the conflicts and potential cooperation relationships in different vehicle clusters.

[0059] Similar interaction edge Within the same CAV cluster, all vehicles possess the same perception capabilities and control strategies, thus exhibiting a close collaborative mode. Therefore, Connecting CAV nodes of the same type to the same superedge highlights information sharing and control interaction during collaborative driving within the cluster, thereby capturing the cooperative intent of resource allocation within the cluster.

[0060] Heterogeneous interaction edge Different CAV clusters often have different task objectives and perception accuracy, and can influence each other in the same environment. To address cross-cluster interactions between CAVs, this embodiment defines combinations containing different types of CAV nodes as heterogeneous interaction edges in the association matrix. It connects two or more CAV nodes from different clusters, thus representing the conflicts between them caused by path planning and communication protocols.

[0061] Hybrid Interactive Edge In the TMA system, there is a high degree of interaction between connected autonomous vehicles (CAVs) and human vehicles (HVs). Therefore, this invention will combine nodes containing at least one CAV and one or more HVs, and use a hybrid interaction edge. They are concatenated to map the interaction between CAV and HV in the association matrix.

[0062] Based on the above analysis, the construction method of the inter-cluster association matrix is ​​as follows: Figure 3 As shown.

[0063] The inter-cluster association matrix can be constructed by concatenating all intra-cluster association matrices, i.e.:

[0064] ;

[0065] No. The shape of the intra-cluster association matrix within each cluster is: , This represents the total number of vehicle clusters in the scene.

[0066] For the hyperedge eigenma matrix The local hyperedge feature matrix at time t is , can be represented as:

[0067] ;

[0068] in, for Always Beyond the Edge The eigenvectors of can be represented as:

[0069] .

[0070] like Figure 4 As shown, the hyperedge feature matrix between clusters The construction of the global correlation matrix is ​​similar to that of the global correlation matrix; it can be formed by concatenating all the local hyperedge feature matrices, i.e.:

[0071] ;

[0072] in, This represents the feature vector dimension of the hyperedge in the hypergraph model.

[0073] The basic principle of gating mechanisms:

[0074] Gating is a technique used to dynamically control the flow of information in neural networks. Its core lies in controlling the flow of information through one or more gates. Each gate is a weight between 0 and 1, representing the degree to which information is passed through or blocked. A gating mechanism mainly consists of three parts: 1. An input gate, which controls the influence of the current input; 2. A forget gate, which controls the influence of previous states, determining how much past memory needs to be forgotten; and 3. An output gate, which determines which information from the current state needs to be output.

[0075] Assume the input of the model is Its previous state was Each gate is calculated based on a linear transformation of the input and the state at the previous time step, plus a bias, and then obtained through a nonlinear transformation using an activation function.

[0076] Input gate The calculation method is as follows:

[0077] ;

[0078] in, For activation function, Here is the weight matrix of the input gate. This is the weight matrix from the previous state to the input gate. For bias terms, Represents batch size. and These represent the feature dimensions of the input and the dimensions of the hidden state, respectively.

[0079] Forgotten Gate The calculation method is as follows:

[0080] ;

[0081] in, Here is the weight matrix for the forget gate. This is the weight matrix from the previous state to the forget gate. This is a bias term.

[0082] Output gate The calculation method is as follows:

[0083] ;

[0084] in, This is the weight matrix of the output gate. This is the weight matrix from the previous state to the output gate. This is a bias term.

[0085] Next, by calculating the candidate states This represents the combination of the current input and the state at the previous time step:

[0086] ;

[0087] Then, update the current state:

[0088] ;

[0089] in, This represents element-wise multiplication. Finally, a linear transformation is performed to obtain the final output. The calculation is as follows:

[0090] .

[0091] in, This refers to the weight matrix of the final fully connected layer in the hybrid network described in this embodiment. Its specific function is to transform the high-dimensional hidden state feature vectors obtained after the network undergoes multiple nonlinear processing. Linear mapping is applied to the output space to generate action values ​​or control signals that can be used for decision-making. The bias vector of the final fully connected layer serves to provide an appropriate offset during the linear mapping process described above, thereby improving the flexibility and expressiveness of the output and ensuring that the final output value can accurately fit the actual required action space.

[0092] By introducing gating mechanisms at each moment, the network can dynamically adjust the flow of information, thereby selectively remembering or forgetting past information.

[0093] The basic principle of hypergraph neural networks:

[0094] Hypergraph Neural Network (HGNN) is an extension of GNN based on hypergraph structures, used to handle complex high-order structures. Commonly used models include Hypergraph Convolutional Neural Network (HGCN) and Hypergraph Attention (HGAT).

[0095] HGCN is an extension of traditional GCN, specifically designed for processing hypergraph structured data. Unlike traditional GCN, hypergraph convolution aggregates information from multiple nodes through hyperedges. Its process is as follows:

[0096] ;

[0097] in, The vertex metric matrix describes the nodes. The number of connected hyperedges, i.e.:

[0098] ;

[0099] Here is the hyperedge metric matrix, which describes the hyperedge. The number of connected nodes, i.e.:

[0100] ;

[0101] This is the hyperedge weight matrix. For the first The node feature matrix of the layer, for and The weight matrix between, and It is a non-linear activation function.

[0102] HGAT enhances the representation of hypergraph convolution by introducing an attention mechanism. This method learns the relationship between different nodes and hyperedges, thereby dynamically adjusting the weights for information propagation. For nodes... and the hyperedge connected to it Its attention coefficient The calculation method is as follows:

[0103] ;

[0104] in, For nodes The neighboring nodes, For nodes eigenvectors, For super-edge eigenvectors, These are learnable attention weights. The calculated attention coefficients are used. This allows us to perform a weighted summation of the information from neighboring nodes, thereby updating the node. The information, namely:

[0105] ;

[0106] in, The weight matrix is ​​a learnable matrix. According to The node is updated based on its attention coefficient and information about surrounding nodes. Information.

[0107] Multi-head HGAT can further enhance the model's expressive power. It allows the model to simultaneously decelerate the attention coefficient from multiple angles and integrate the results, thereby capturing richer relational information. Its calculation formula is as follows:

[0108] ;

[0109] in, For the number of attention heads, and The first Attention coefficients and learnable weight matrices for each attention head. For nodes The set of neighboring nodes in a hypergraph structure.

[0110] The following are embodiments of this application in highways and roundabouts. First, in traffic scenarios, hypergraphs can be used... To indicate, among which, Indicates a hypergraph A set of nodes, and The diagram shows... A set of edges. Adaptation matrix. It can be used to represent the connection relationship between a node and a hyperedge, where, The number of nodes in the hypergraph. Let be the number of hyperedges in the hypergraph. There are 10 superedges, and the feature dimension of each superedge is 0. The hypergraph whose hyperedge feature matrix It can be represented as:

[0111] ;

[0112] in, For super-edge eigenvectors.

[0113] The shape of the intra-cluster association matrix is:

[0114] ;

[0115] in, This represents the total number of vehicles (total number of nodes) in the scene. Represents cluster The total number of internal CAVs. In the above formula, 1 describes the cluster. Internal interaction, the first Cluster of How do each CAV interact with the rest of the cluster, and the second... Describes cluster of The interaction between CAVs and HVs.

[0116] The global association matrix can be constructed by concatenating all the association matrices within the cluster, that is:

[0117] ;

[0118] use Figure 3 The correlation matrix obtained by the construction method shown is:

[0119] .

[0120] For the edge feature matrix Local hyperedge feature matrix at time step It can be represented as:

[0121] ;

[0122] in, for Always Beyond the Edge The eigenvectors of can be represented as:

[0123] ;

[0124] in, Indicates the superedge The average velocity of all nodes. Indicates the superedge The average acceleration of all nodes. It can be represented as:

[0125] ;

[0126] Inter-cluster hyperedge feature matrix The construction of the global correlation matrix is ​​similar to that of the global correlation matrix; it can be formed by concatenating all the local hyperedge feature matrices, i.e.:

[0127] ;

[0128] The combination of the current input and the state at the previous time step:

[0129] ;

[0130] Then, update the current state:

[0131] ;

[0132] Get the final output :

[0133] ;

[0134] The state information of vehicle nodes is aggregated based on a gated hypergraph neural network. The network includes a two-layer hypergraph convolutional structure: the first layer is an attentionless hypergraph convolution used to extract initial adjacency features, and the second layer is a multi-head attention hypergraph convolution used to capture high-order interaction relationships between nodes and obtain a multi-head concatenated feature vector output. After the output of the second convolutional layer, a gating mechanism is introduced to dynamically weight the concatenated features. The gating mechanism includes: a gating network containing a linear mapping layer and a sigmoid activation function, used to compress and non-linearly map the concatenated feature vector, outputting gating weights corresponding to each channel; the gating weights are copied to match the dimension of the original features and multiplied element-wise with the original multi-head features, thereby weighting the importance of each channel, filtering out redundant information, and retaining key features.

[0135] The state of each autonomous vehicle is modeled through a local policy network, which outputs an action value function. A global fusion network integrates local Q-values ​​and structural features to output a joint action value. The robust Q-Nexus engine (RQNE) algorithm is used to train the above network. At each time step, the system determines the optimal action based on the joint Q-value. Control commands are distributed to each vehicle controller by the centralized control module to drive the actual behavior. The experience quadruples generated during the interaction are stored in the experience replay buffer for offline updates of the policy network and the fusion network. Training continues until the number of rounds or performance convergence conditions are met.

[0136] This embodiment provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0137] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0138] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cooperative control method for unmanned vehicles based on gated hypergraph neural networks, characterized in that, Includes the following steps: Obtain real-time status information of several vehicle clusters in a traffic scenario; Based on the real-time status information, a hypergraph representation model is constructed with vehicle clusters as nodes and three types of interaction edges as hyperedges. Based on the hypergraph representation model, node feature matrix, hyperedge feature matrix and association matrix are obtained. An initial neural network is constructed based on a multi-head attention mechanism. A gating mechanism is then introduced into the initial neural network to obtain a gated hypergraph neural network. The gated hypergraph neural network includes a local network, a global network, and a hybrid network. The node feature matrix, the hyperedge feature matrix, and the correlation matrix are input into the gated-hypergraph neural network to obtain local features and weighted global features. The joint actions of the unmanned vehicle swarm are obtained based on the local features and weighted global features. The hypergraph representation model includes three types of interactions: edges with similar interactions, edges with dissimilar interactions, and edges with mixed interactions. The local network includes a two-layer hypergraph convolutional structure and a gating mechanism; the two-layer hypergraph convolutional structure includes a non-attention hypergraph convolutional layer and a multi-head attention hypergraph convolutional layer. After the node feature matrix, the hyperedge feature matrix, and the association matrix are input into the local network, initial adjacency features are extracted through an attentionless hypergraph convolutional layer. The initial adjacency features are then processed through a multi-head attention hypergraph convolutional layer to obtain multi-head concatenated features. After weighting the multi-head concatenated features through a gating mechanism, they are mapped to the dimension of the action space through two fully connected layers, and local features are output. The gating mechanism includes a linear mapping layer and a sigmoid activation function. The global network includes a two-layer hypergraph convolutional structure, a gating mechanism, a global average pooling layer, and a global attention layer; the two-layer hypergraph convolutional structure includes a non-attention hypergraph convolutional layer and a multi-head attention hypergraph convolutional layer. After the node feature matrix, the hyperedge feature matrix, and the correlation matrix are input into the global network, initial adjacency features are extracted through an attentionless hypergraph convolutional layer. The initial adjacency features are then processed through a multi-head attention hypergraph convolutional layer to obtain multi-head spliced ​​features. The multi-head spliced ​​features are then weighted through a gating mechanism and processed through a global average pooling layer. The output of the global average pooling layer is then input into the global attention layer to obtain weighted global features. The hybrid network maps the input local features and weighted global features to a high-dimensional latent space through a feature mapping layer, and then processes them sequentially through a hypergraph convolutional layer, a multi-head hypergraph attention layer, and a gated multi-head attention mechanism. The processed local features and weighted global features are aggregated and dimensionality reduced through a global average pooling layer, and then processed by a global attention layer to obtain a global feature representation. The global feature representation is then mapped to the joint action value function output space through two fully connected layers to obtain a decision signal. Based on the decision signal, the joint action of the unmanned vehicle cluster is obtained.

2. The unmanned vehicle cooperative control method based on gated hypergraph neural network according to claim 1, characterized in that, Residual connections are used before and after the non-attentional hypergraph convolutional layer, the multi-head attentional hypergraph convolutional layer, and the gating mechanism.

3. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method according to any one of claims 1-2.

4. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-2.

Images