Intelligent weapon target distribution method based on graph neural network

Abstract

The invention relates to the technical field of intelligent simulation, and discloses a weapon target intelligent distribution method based on a graph neural network. Modeling a weapon target allocation problem into a learning task of a heterogeneous graph structure, constructing a bipartite graph model comprising weapon nodes, target nodes and associated edges thereof, and performing feature extraction and message transmission by using a multilayer graph attention network and a graph isomorphic network; feature coding is carried out on weapons and target nodes, and associated attributes such as interception rate are processed; multi-neighborhood information aggregation is realized through an attention mechanism and a multi-layer perceptron; and finally, calculating a distribution probability and generating a weapon target distribution scheme. And a batch graph processing mechanism and a sub-graph distinguishing method are adopted, multi-scene parallel computing is supported, and the method has the capability of dynamically adapting to different scale problems. The method can effectively capture the topological relation between weapons and targets, has the advantage of calculation efficiency in processing variable-scale distribution tasks, and is suitable for command decision scenes such as air defense and anti-missile, unmanned aerial vehicle cluster combat and the like.

Description

Technical Field

[0001] This invention relates to the field of intelligent simulation technology, specifically to a method for intelligent weapon target allocation based on graph neural networks. Background Technology

[0002] Weapon Target Assignment (WTA) is one of the core decision problems in modern command and control systems. Its objective is to achieve optimal or near-optimal firepower allocation against multiple enemy targets under limited weapon resources, maximizing overall combat effectiveness, such as minimizing enemy threat, maximizing the probability of destruction, or minimizing resource consumption. WTA is widely applied in scenarios such as air defense and missile defense, multi-platform coordinated strikes, and UAV swarm operations, exhibiting high complexity and real-time requirements. Essentially a large-scale combinatorial optimization problem, its solution space grows exponentially with the number of weapons and targets. It typically involves nonlinear constraints, multi-objective optimization, and real-time response requirements in dynamic environments, thus being recognized as an NP-hard (Non-deterministic Polynomial) problem.

[0003] Traditional WTA modeling typically relies on mathematical optimization methods, using integer programming or mixed-integer programming frameworks and employing methods such as branch and bound and Lagrange relaxation to find solutions. While these methods can yield relatively optimal solutions, their high computational complexity makes them unsuitable for real-time decision-making in high-dimensional, dynamic, and uncertain real-world battlefield environments. To address this issue, researchers have subsequently proposed various intelligent optimization methods based on metaheuristic algorithms, including genetic algorithms, particle swarm optimization, simulated annealing, and ant colony optimization. These methods demonstrate good global optimum-solving performance for small- to medium-scale problems, but their performance is highly dependent on parameter tuning and they are prone to getting trapped in local optima as the problem size increases, resulting in a significant decrease in convergence speed. In recent years, with the development of artificial intelligence technology, methods based on traditional feedforward neural networks and reinforcement learning have gradually become research hotspots. For example, Deep Q-Networks (DQNs) and policy gradient methods have been used to model the sequential decision-making process of WTA, learning to assign policies through interaction with the environment and exhibiting a certain degree of generalization ability. However, these methods typically treat weapons and targets as independent vector inputs, ignoring their inherent spatial and relational structures, resulting in insufficient ability to express the essential characteristics of the problem. Furthermore, traditional neural networks can only handle problems with fixed topologies. While padding techniques can be used to supplement the input size of neural networks when facing dynamic-scale planning tasks, these introduced empty neurons also affect the accuracy and efficiency of problem solving.

[0004] It is worth noting that the interaction between weapons and targets in the WTA problem naturally constitutes a dense graph structure: each weapon node has equal attributes such as damage probability, distance, and priority with multiple target nodes, and there may also be cooperative or competitive relationships between weapons and targets. This graph structure contains rich topological information and semantic relationships, which are key factors influencing allocation decisions. However, most existing studies have "flattened" this graph structure into vector or matrix inputs, failing to effectively utilize the geometric characteristics and relational reasoning capabilities of the graph structure. Although Graph Neural Networks (GNNs) have demonstrated excellent graph structure modeling capabilities in recent years in areas such as social network analysis, molecular structure prediction, and recommender systems, their application in command and decision problems such as WTA is still in its early stages. Existing neural network-based WTA methods mostly use fully connected or sequence models, lacking explicit modeling of the complex dependencies between weapons and targets, which limits the interpretability and generalization performance of the models.

[0005] Therefore, there is an urgent need for an intelligent allocation method that can fully exploit the inherent graph structure features in the WTA problem. Summary of the Invention

[0006] This invention relates to a weapon target intelligent allocation method based on graph neural networks, and more particularly to a method for modeling and optimizing the allocation problem between weapons and targets using graph neural networks. Addressing the problem of weapon target allocation in a virtual simulation game environment through target threat assessment and interception rate assessment of the simulation system, this invention provides an intelligent weapon target allocation scheme from a macroscopic perspective, thereby improving the effectiveness and reliability of the virtual simulation game system in achieving simulation tasks.

[0007] Step 1: Construct a bipartite graph model consisting of weapon nodes and target nodes, where weapon nodes and target nodes carry attribute features, and edges represent the attack relationship between weapons and targets and carry interception rate features;

[0008] Step 2: Perform high-dimensional mapping on weapon node features and target node features using the node embedding module to generate initial node embedding vectors;

[0009] Step 3: Map edge attributes to high-dimensional edge embedding vectors using the edge feature embedding module;

[0010] Step 4: Message passing is performed using a multi-layer graph attention network and a graph isomorphic network layer. First, neighborhood information is aggregated and node features are updated through an attention mechanism. Then, global topology information is further fused through the graph isomorphic network layer.

[0011] Step 5: The output module calculates the allocation scores of weapons and targets and generates the optimal allocation scheme.

[0012] The beneficial effects of this invention are:

[0013] First, a mathematically formalized batch processing mechanism is adopted, which simplifies the storage complexity.

[0014] Secondly, the system has dynamic adaptability and can handle... Diverse battlefield scenarios support real-time decision-making at a rate of thousands per second.

[0015] Finally, through an end-to-end deep learning architecture, the optimal allocation scheme is directly output, avoiding the design limitations of traditional heuristic rules.

[0016] In summary, this invention demonstrates outstanding practical performance in command and control, UAV scheduling, and other fields, especially in complex scenarios such as multi-theater collaborative operations, where it can achieve intelligent resource optimization and allocation. Detailed Implementation

[0017] The present invention will be further described in detail below with reference to specific embodiments.

[0018] This invention proposes a weapon-target intelligent allocation method based on graph neural networks (GNNs). For the first time, it models the WTA problem as a graph structure learning task, utilizing GNNs for end-to-end feature learning and relational reasoning on the heterogeneous graph composed of weapons and targets, effectively capturing the topological dependencies and semantic associations within the system. Compared to traditional methods, this method not only more closely approximates the essential structure of the WTA problem but also possesses good scalability, adapting to weapon and target sets of different sizes, overcoming the limitations of existing mathematical programming and metaheuristic algorithms that are sensitive to problem size. Furthermore, the proposed method supports end-to-end training and online inference, exhibiting strong real-time performance and robustness, providing a new technical path for intelligent firepower allocation in complex battlefield environments. This invention constructs a bipartite graph model, utilizing multi-layer graph attention networks (GATs) and graph isomorphic networks (GINs) for feature extraction and message passing, ultimately outputting the optimal allocation scheme between weapons and targets.

[0019] This invention is achieved through the following technical solution:

[0020] 1. System Architecture Overview:

[0021] The core of this invention is a weapon target allocation model based on graph neural networks, whose architecture includes the following key modules:

[0022] (1) Node embedding module: performs initial feature embedding on weapon nodes and target nodes.

[0023] (2) Edge feature embedding module: embeds the edge features between the weapon and the target.

[0024] (3) Message passing module: Node features are updated through multi-layer graph attention network (GAT) and graph isomorphism network (GIN).

[0025] (4) Output module: Calculate the probability of the weapon and the target being assigned.

[0026] 2. Node Embedding Module:

[0027] Weapons and target nodes typically possess their own attributes upon input (such as the weapon's attack capability, the target's threat level, etc.). This invention first embeds features of the weapon and target nodes using a fully connected neural network, with the specific steps as follows:

[0028] Weapon node embedding:

[0029] Let the initial characteristic of the weapon node be a scalar. It is mapped to a high-dimensional vector through a two-layer fully connected network. :

[0030] (1);

[0031] in, , For learnable parameters, For embedded dimensions, It represents Activation function.

[0032] Target node embedding:

[0033] Similarly, the initial features of the target node Embedded using the following formula:

[0034] (2);

[0035] in, , These are parameter matrices that are different from those embedded in the weapon.

[0036] 3. Edge Feature Embedding Module:

[0037] The edges between weapons and targets typically possess associated attributes (such as attack success rate, distance, etc.). This invention uses an edge feature embedding module to embed edge features... Mapped to a high-dimensional vector:

[0038] (3);

[0039] in, , These are the edge embedding parameters.

[0040] 4. Message passing module:

[0041] The message passing module is the core of this invention. It updates node features step by step through a multi-layer graph neural network, specifically including the following steps: (1) Graph Attention Network (GAT) layer:

[0042] GAT aggregates information from neighboring nodes through an attention mechanism. Let node... The neighbor is Its update formula is:

[0043] (4);

[0044] Among them, attention coefficient Calculated using the following formula:

[0045] (5);

[0046] in, For learnable vectors, This indicates a splicing operation. Represents artificial neural networks Activation function Representing neural networks Normalized function.

[0047] In this invention, a three-layer Graph Attention Network (GAT) structure is employed. The feature dimension mapping and the number of attention heads are explained below:

[0048] 1) The first layer will - dimensional feature mapping is -Dimensional, multi-head attention (4 heads).

[0049] 2) The second layer will - dimensional feature mapping is -Dimensional, single-head attention.

[0050] 3) The third layer will - dimensional feature mapping is -Dimensional, single-head attention.

[0051] (2) Graph Isomorphic Network (GIN) Layer:

[0052] GIN further extracts global features through a multilayer perceptron (MLP). :

[0053] (6);

[0054] in, It is a two-layer fully connected network:

[0055] (7);

[0056] The second-layer GIN directly outputs the final node features:

[0057] (8);

[0058] 5. Output module:

[0059] Weapon nodes are calculated based on node features. With the target node Score distribution:

[0060] (9);

[0061] in, Represents element-wise product. , These are the parameters for the output layer. The final allocation scheme is generated using either softmax or the Hungarian algorithm.

[0062] 6. Batch image processing mechanism:

[0063] The specially designed batch graph processing mechanism includes: merging bipartite graphs from multiple independent scenes into a batch graph. (Equation 10) Parallel computation is achieved through node cascading and edge index offset; batch pointers are maintained. (Equation 13) and batch vector (Equation 15) is a molecular diagram structure.

[0064] In weapon target allocation problems, it is often necessary to process multiple independent scenarios simultaneously (such as task allocation across multiple theaters). This invention merges multiple bipartite graphs into a single batch graph, and the processing procedure is described below:

[0065] Let the input be a list of K independent bipartite graphs. Each of the graphs Include:

[0066] Node set (Weapon node ∪ Target node)

[0067] Edge set

[0068] Batch Image The construction process is as follows:

[0069] (10);

[0070] In practice:

[0071] (1) Node concatenation: The feature matrices of all nodes are concatenated along the 0th dimension.

[0072] (11);

[0073] in, , For feature dimensions.

[0074] (2) Edge index offset: The edge index of the k-th subgraph needs to be increased by the offset of the number of nodes in the previous k-1 subgraphs.

[0075] (12);

[0076] in, It is a matrix of all 1s.

[0077] 7. Subgraph differentiation mechanism

[0078] The subgraph differentiation mechanism specifically includes: accumulating node count vectors. The range of nodes in the subgraph is located (Equation 14); based on the nodes at both ends of the edge. Value-based edge attribution determination supports parallel loss calculation at the subgraph level.

[0079] To track the original subgraph structure in batch graphs, the system maintains two core attributes:

[0080] (1) Setting the batch pointer (ptr)

[0081] Define the cumulative node count vector :

[0082] (13);

[0083] This property supports:

[0084] Subgraph node location: The range of nodes in the k-th subgraph is

[0085] Conversion between global and local indexes: For global node indexes in a batch graph Its subgraph satisfies:

[0086] (14);

[0087] Among them, the local index is .

[0088] (2) Batch processing vector

[0089] Define the node affiliation vector :

[0090] (15);

[0091] This property achieves the following:

[0092] Edge ownership determination: For edges its subgraph ;

[0093] Parallel computing grouping: via The operation enables parallel processing at the subgraph level.

[0094] Example 1:

[0095] Example 1 provides a case study of weapon target assignment reasoning. This example details the entire process of weapon target assignment reasoning based on graph neural networks. Assume a battlefield scenario, which includes... weapons and The objectives are as follows. The specific implementation steps for this scenario are as follows:

[0096] (1) Input data preparation:

[0097] Target threat vector (Threat level range: 5-100):

[0098] (16);

[0099] Interception Rate Matrix :

[0100] (17);

[0101] (2) Data transformation and graph construction:

[0102] 1) Generation of node features in graph models:

[0103] The weapon node's characteristic is the row average of its interception rate:

[0104] (18);

[0105] The features of the target node are normalized:

[0106] (19);

[0107] in, These are the target node features after normalization.

[0108] 2) Edge feature construction of the graph model: flatten the interception rate matrix into edge attributes :

[0109] (20);

[0110] Establish the edge connection relationship from weapon to target, that is, the set of node index tuples:

[0111] (twenty one);

[0112] (1) Model reasoning process:

[0113] Loading a pre-trained GNN model (parameter configuration) ):

[0114] (twenty two);

[0115] Forward propagation computation: Its unfolded form is as follows:

[0116] (twenty three);

[0117] Perform Softmax normalization on each row of the prediction result matrix to obtain the prediction probability of each weapon for each target: (twenty four);

[0118] (2) The model inference result is decoded into an allocation vector:

[0119] The allocation vector is achieved by assigning the target with the highest allocation probability to each weapon. Generation:

[0120] (25);

[0121] Using the allocation vector Threat level and interception rate matrix Combat effectiveness assessment can be calculated based on the formula for the expected value of escape threat assigned to weapon targets:

[0122] (26).

[0123] Example 2

[0124] Example 2 provides a computational instance of batch graph processing and subgraph differentiation. In this example, we consider the need to process two subgraphs in parallel:

[0125] 3 weapons + 2 targets );

[0126] 2 weapons + 3 targets );

[0127] Based on these two subgraphs, a batch graph with merged subgraphs can be generated using the batch graph processing method described above. The resulting batch graph has the following ptr and batch attributes:

[0128] ;

[0129] Based on this batch graph, the edge indices are transformed using the subgraph differentiation mechanism described above. An example result of the edge index transformation is as follows:

[0130] original edge In the batch diagram, it becomes ;

[0131] Reverse positioning: edge correspond Local edges .

[0132] Example 3

[0133] Example 3 describes the model training implementation process of the present invention. This example details the technical implementation of model training, focusing on three parts: generation of model training data, data preprocessing, and optimization training process.

[0134] (1) Generation of training and testing datasets for model training:

[0135] To utilize supervised learning mechanisms, training and testing datasets for model training need to be prepared in advance. This invention uses the CPlex solver to generate 20,000 optimization solutions and saves them to disk in JSON format, wherein:

[0136] The dataset contains problems whose sizes follow a random uniform distribution, i.e.: , ;

[0137] Solve the following objective function using the Cplex solver:

[0138] (27);

[0139] Simultaneously, the following constraints are required:

[0140] (18);

[0141] Because the Cplex solver is slow, it does not meet the requirements for online training. However, it can be pre-solved and saved to disk to achieve the purpose of fast training.

[0142] (2) Data preprocessing process:

[0143] Since the values ​​of threat level and interception rate do not conform to the data standardization specifications of machine learning, the following linear transformation is used to standardize the features: (29);

[0144] (30);

[0145] Based on this, the data is transformed from vectors into a graph structure:

[0146] (31);

[0147] The node features of the graph model are a combination of weapon features and target features:

[0148] (32);

[0149] Generate edge labels for the graph model, i.e., optimal assignment indicators, for supervised learning of the graph model of this invention:

[0150] (33);

[0151] The optimization process of model training:

[0152] Because of the imbalance between positive and negative samples in the dataset, this invention considers using a loss function with class balance as follows:

[0153] (34);

[0154] (35);

[0155] The batch training parameters for the model training process are set to batch_size=32, and the learning rate is set to gradually transition from 0.0001 to 0.00001.

[0156] (36);

[0157] During training, the Hamming accuracy metric was used to verify the performance changes of the model training process:

[0158] (37);

[0159] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above invention, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A weapon target intelligent allocation method based on graph neural networks, characterized in that, Includes the following steps: Step 1: Construct a bipartite graph model consisting of weapon nodes and target nodes, where weapon nodes and target nodes carry attribute features, and edges represent the attack relationship between weapons and targets and carry interception rate features; Step 2: Perform high-dimensional mapping on weapon node features and target node features using the node embedding module to generate initial node embedding vectors; Step 3: Map edge attributes to high-dimensional edge embedding vectors using the edge feature embedding module; Step 4: Message passing is performed using a multi-layer graph attention network and a graph isomorphic network layer. First, neighborhood information is aggregated and node features are updated through an attention mechanism. Then, global topology information is further fused through the graph isomorphic network layer. Step 5: The output module calculates the allocation scores of weapons and targets and generates the optimal allocation scheme.

2. The intelligent weapon target allocation method based on graph neural networks according to claim 1, characterized in that, Step 2 specifically includes: Weapon node embedding: Let the initial characteristic of the weapon node be a scalar. It is mapped to a high-dimensional vector of weapon nodes through a two-layer fully connected network. : （1）； in, , These are the learnable parameters. For embedded dimensions, represent Activation function; and These are the embedding layer biases for weapon node features; Target node embedding: Let the initial characteristics of the target node be scalars. It is mapped to a high-dimensional vector of the target node through a two-layer fully connected network. : （2）； in, , These are parameter matrices that are different from those embedded in the weapon; and These are the embedding layer biases of the target node features.

3. The intelligent weapon target allocation method based on graph neural networks according to claim 2, characterized in that, Step 3 specifically includes: The edge features are embedded through the edge feature embedding module. Mapped to high-dimensional vectors of edge features : （3）； in, , These are the edge embedding parameters; and These are the embedding layer biases for the weapon-target relationship features.

4. The intelligent weapon target allocation method based on graph neural networks according to claim 3, characterized in that, Step 4 specifically includes: Step 4.1: The graph attention network aggregates information from neighboring nodes through an attention mechanism; Set weapon and target nodes The neighbor is Its update formula is: （4）； in, For updated weapon and target node characteristics; Attention coefficient Features of neighbors; For weapon and target nodes The neighboring nodes; The attention coefficient is calculated using the following formula: （5）； in, For learning vectors, This indicates a splicing operation. Representing artificial neural networks Activation function Representing neural networks Normalization function; For the nodal characteristics of weapons and targets; It is the transpose symbol; Step 4.2: The graph isomorphic network further extracts global features through a multilayer perceptron. ; （6）； in, Updated node features for weapon nodes and target nodes; It is a two-layer fully connected network: （7）； in, represent Input; These are the weights of the second layer network; These are the weights of the first layer of the network; This is the bias of the first layer network; This is the bias for the second layer network; The second-layer graph isomorphic network directly outputs the final node features. : （8）； in, It is a Layer 2 fully connected network.

5. The intelligent weapon target allocation method based on graph neural networks according to claim 4, characterized in that, In step 4.1, the implementation of the graph attention network layer includes a three-layer attention network structure. The first layer transfers the feature dimension from... Compress to And using 4-head attention, the second layer is restored to Single-head attention, third layer extended to Single-headed attention.

6. The intelligent weapon target allocation method based on graph neural networks according to claim 4, characterized in that, In step 5, the score allocation is calculated as follows: Weapon nodes are calculated based on node features. With the target node Distribution of scores : （9）； in, Represents element-wise product. , These are the output layer parameters; The final target node features output by the second-layer graph isomorphic network; The final weapon node features output by the second-layer graph isomorphic network; To allocate the internal bias of the computation network; To allocate external biases for the computational network.

7. The intelligent weapon target allocation method based on graph neural networks according to claim 1, characterized in that, A batch graph processing mechanism is used for allocation: multiple independent bipartite graphs are merged into a single batch graph, and parallel computation is achieved through node concatenation and edge index offset; a batch pointer is maintained. and batch vector Based on molecular diagram structure; When dividing a subgraph into subgraphs, a subgraph differentiation mechanism is used: the range of subgraph nodes is located by accumulating node count vector pointers; and the batch processing vectors of the nodes at both ends of an edge are used. Determine edge ownership and support parallel loss calculation at the subgraph level.

8. The intelligent weapon target allocation method based on graph neural networks according to claim 7, characterized in that, The bipartite graphs are merged into a single batch graph, as follows: Let the input be a list of K independent bipartite graphs. Each bipartite graph Includes: Node set and edge set ; For weapon node set, For the target node set; Batch Image The construction process is as follows: （10）； in, For feature splicing operations; In practice: (1) Node concatenation: All node feature matrices are concatenated along the 0th dimension, as shown below: （11）； in, The number of nodes. For feature dimensions; For network input during batch processing, For the first One sample; (2) Edge index offset: list The edge index of the k-th subgraph needs to be added with the offset of the number of nodes in the first k-1 subgraphs, as shown below: （12）； in, The edge number for batch processing; The index of the k-th edge; For the first The number of nodes in each sample; It is a matrix of all 1s.

9. The intelligent weapon target allocation method based on graph neural networks according to claim 8, characterized in that, The specific subgraph differentiation mechanism is as follows: 1) Define the cumulative node count vector : （13）； in, for The space of the natural logarithm of dimensionality; This is the k-th element of the cumulative node counting vector; Used to implement: Subgraph node location: The range of nodes in the k-th subgraph is ; Conversion between global and local indexes: For global node indexes in a batch graph Its subgraph satisfies: （14）； Among them, the local index is ; 2) Define the node affiliation vector : ；（15）； in, Represents the node affiliation vector The One element; It is a shorthand for "if and only if"; Used to implement: Edge ownership determination: For edges Its subgraph ; For the edge Endpoints; Parallel computing grouping: via The operation enables parallel processing at the subgraph level.

Images