Multi-agent collaborative adaptation degree screening method based on hierarchical feature learning

Abstract

The invention discloses a multi-agent collaborative adaptation degree screening method based on hierarchical feature learning. The method comprises the steps of preprocessing a data set, extracting key features, constructing a hierarchical feature learning network, training the hierarchical feature learning network and testing the hierarchical feature learning network. According to the construction method of the hierarchical feature learning network, a multi-scale operation knowledge graph module, a multi-scale operation feature extraction module and a multi-source feature fusion and adaptation degree screening module are included, and screening of the adaptation degrees of task execution capabilities of multiple agents is completed. The defects that in an existing multi-agent task allocation and collaborative screening method, feature expression is single, collaborative relation modeling is insufficient, and adaptability evaluation is inaccurate are overcome. The method has the advantages of being comprehensive in feature expression, complete in collaborative relation modeling, accurate in adaptability evaluation, high in screening precision, high in system operation efficiency, high in task success rate and the like, and can be applied to task scheduling platforms of industrial automation, multi-robot collaboration, intelligent manufacturing and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent control technology, specifically to the degree of multi-agent collaborative adaptation. Background Technology

[0002] With the development of artificial intelligence and automated systems, multi-agent systems have been widely used in complex task collaboration, distributed control, and group decision-making. Multi-agent systems typically consist of multiple agents with heterogeneous functions and capabilities, working together through communication and interaction to complete complex tasks. In practical applications, significant differences exist among agents in terms of execution capabilities, response speed, skill levels, and resource consumption, leading to limitations in task collaboration efficiency and overall performance.

[0003] Existing methods for multi-agent task allocation and collaborative optimization mainly include rule-based allocation methods, reinforcement learning-based policy optimization methods, and graph neural network-based relationship modeling methods. Among them, rule-based methods require manual pre-setting of a large number of task allocation conditions and have poor generalization ability; reinforcement learning-based methods have slow convergence speed in high-dimensional task spaces and are prone to getting trapped in local optima; while graph neural network-based methods can capture the relationship structure between agents to a certain extent, most of them only focus on single-layer or single-scale features and cannot fully express the multi-level dependencies and multi-scale collaborative characteristics between multiple agents.

[0004] Traditional methods often only assess fit at the task or individual level, lacking joint modeling of the multi-layered relationships between "single agent capabilities, group collaborative capabilities, and task requirements," resulting in unstable fit selection results and difficulty in coping with dynamically changing task environments.

[0005] In the fields of artificial intelligence and intelligent control technology, there is an urgent need for a hierarchical feature learning method that can simultaneously extract and fuse features at multiple scales, including single agent features, group collaboration, and task relationships. This would enable the selection of the suitability of multiple agents under complex tasks, provide a scientific basis for agent collaborative decision-making, and improve the success rate of task completion and system operating efficiency. Summary of the Invention

[0006] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a multi-agent collaborative adaptation screening method based on hierarchical feature learning, which has comprehensive feature expression, complete collaborative relationship modeling, accurate adaptability evaluation, and high screening accuracy.

[0007] The technical solution adopted to solve the above technical problems is:

[0008] (1) Dataset preprocessing

[0009] The multi-scale knowledge graph was used as a dataset and divided into a training set and a test set in an 8:2 ratio.

[0010] (2) Extracting key features

[0011] Convolutional neural networks are used to extract single agent response data and its corresponding task mapping matrix, multi-agent collaborative operation data and its corresponding task relationship matrix, initial features of operation relationship, initial features of single agent, task node features, and operation skill node features.

[0012] (3) Constructing a hierarchical feature learning network

[0013] The hierarchical feature learning network consists of a multi-scale knowledge graph module, a multi-scale feature extraction module, a multi-source feature fusion module, and a fit degree screening module connected in series.

[0014] (4) Training the hierarchical feature learning network

[0015] 1) Constructing the loss function

[0016] Construct the loss function as follows :

[0017]

[0018] in, This represents the loss function for single-agent cooperative tasks. This represents the loss function for multi-agent cooperative tasks. It is the weight decay parameter, with a value range of 0.0001 to 0.1.

[0019] 2) Training a hierarchical feature learning network

[0020] The training set was input into the hierarchical feature learning network for training. The training parameters were as follows: the server graphics card used for training was an NVIDIA GeForce RTX 3090, the initial learning rate was 0.001, the number of training epochs was 100, the batch size for training was 32, the Adam optimizer was used, and training was continued until the loss function converged.

[0021] (5) Testing the hierarchical feature learning network

[0022] The test set is input into the trained hierarchical feature learning network for testing, and the output is the degree of multi-agent collaborative adaptation.

[0023] In step (3) of the present invention, the construction of the hierarchical feature learning network is composed of a single agent running state knowledge graph module and a multi-agent collaborative state knowledge graph module connected in series.

[0024] In step (3) of the present invention, the hierarchical feature learning network is constructed by connecting a single agent scale feature module and a multi-agent scale feature module in parallel.

[0025] In step (3) of the present invention, the construction of the hierarchical feature learning network is composed of a single agent scale feature extraction layer, a multi-agent community division and contribution calculation layer, and feature aggregation layer 1 connected in series; the multi-agent scale feature module is composed of an interaction subgraph feature construction layer, a multi-agent node feature extraction layer, and feature aggregation layer 2 connected in series.

[0026] The multi-source feature fusion and adaptation degree screening module consists of a multi-agent community feature fusion module, a multi-agent collaborative feature fusion module, and a joint objective function loss calculation layer connected together; the outputs of the multi-agent community feature fusion module and the multi-agent collaborative feature fusion module are connected to the joint objective function loss calculation layer.

[0027] The multi-agent community feature fusion module is composed of a running fact probability prediction layer and a single agent skill loss calculation layer connected in series; the multi-agent collaborative feature fusion module is composed of a multi-agent collaborative capability feature fusion layer, a task execution result prediction layer, and a multi-agent collaborative loss calculation layer connected in series.

[0028] In step (3) of this invention, which involves constructing a hierarchical feature learning network, the method for constructing the multi-scale knowledge graph module is as follows:

[0029] 1) Construct a knowledge graph module for the operational state of a single agent.

[0030] The single-agent operational state knowledge graph includes a normal operation state skill hypergraph. and abnormal operation status skill supermap , ∈ , ∈ ,in yes OK A real matrix of columns, = + , The number of response types to run. The value range is 2 to 10; =| |For multi-agent The number of entities in the middle, The value range is 2 to 10.

[0031] Determine the normal operating status of the skill supermap using the following formula and abnormal operation status skill supermap

[0032]

[0033]

[0034] in, It is a matrix of all zeros. ∈ , for OK A matrix of real numbers in columns; This represents the association matrix between the running state type and each entity. ∈ , yes OK A matrix of real numbers in columns; yes The transpose of the matrix, ∈ , yes OK A matrix of real numbers in columns; and This represents the runtime log matrix within a multi-agent system. ∈ , ∈ , yes OK A column of real numbers.

[0035] 2) Construct a multi-agent collaborative state knowledge graph module

[0036] Multi-agent cooperative state association graphs include positive cooperative state graphs. Negative Cooperative State Diagram Its construction method is the same as that of the single agent runtime state knowledge graph.

[0037] In step (3) of this invention, which involves constructing a hierarchical feature learning network, the method for constructing the multi-scale running feature extraction module is as follows:

[0038] 1) Extracting single-agent scale features

[0039] ① Initialization

[0040] Initialize the characteristics of entity relations and operational relations of multi-agent systems using the following formula. :

[0041]

[0042] in, This represents the initialization characteristics of two types of operational relationships. ∈ , yes A real matrix with d rows and d columns, where d is the feature dimension of the initial embedding, and its value ranges from 1 to 2. ; Representing multi-agent Initialization features of a single agent ∈ , yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within the range of 2 to 50; Representing multi-agent Initialization features of task nodes ∈ yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within the range of 2 to 30; ∈ Representing multi-agent Initialization features of running skill nodes in the middle, ∈ yes A real matrix with d rows and d columns, Representing multi-agent The number of corresponding nodes within the range of 1 to 4; (·) represents the feature splicing operation.

[0043] ② Learning characteristics

[0044] Learn normal operating characteristics using the following formula and abnormal operation characteristics :

[0045]

[0046]

[0047] in, This is the single-agent operating state under normal response conditions. This refers to the single-agent operating state under abnormal response conditions. and They are and The normalization degree matrix, and These are the two-view HGCN at the 1st The weight matrix of the layer, This is the activation function.

[0048] ③ Fusion characteristics

[0049] Use the following formula to capture dual-view interaction information of normal and abnormal operating states:

[0050]

[0051] =( )

[0052] =( )

[0053] in, For the operational relationships of multi-agent systems, This is a characteristic of a single agent. Features of task nodes Features of running skill nodes, and The first one learned on the dual-view HGCN Two characteristics of the layer.

[0054] By following three steps—initialization, feature learning, and feature fusion—features at the single-agent scale are obtained.

[0055] 2) Divide multi-agent communities and determine their computational contributions

[0056] Calculate the contribution of the multi-agent community using the following formula. :

[0057]

[0058]

[0059]

[0060] in, ∈ , It is a 1-row, d-column real matrix. This is a characteristic of a single agent; The number of single agents, The value range is 1 to 50; As for contribution level, , This is the weight matrix. ∈ , Let be a real matrix with d rows and d columns. ∈ , It is a real matrix with d rows and 1 column;

[0061] 3) Construct feature aggregation layer 1

[0062] Aggregate all multi-agent community features using the following formula :

[0063] in, For any multi-agent community's capability features, the mapping function F(·) is used to enhance the correlation between the features of the multi-agent community.

[0064] 4) Constructing interactive subgraph features

[0065] Construct interactive subgraph features using the following formula:

[0066]

[0067]

[0068] in, Interaction diagram indicating normal operation Interactive subgraph features, Interaction diagram representing abnormal operation Interactive subgraph features, Represents the identity matrix. and These are positive cooperative state diagrams. Negative Cooperative State Diagram The corresponding degree matrix.

[0069] 5) Extract features from multi-agent nodes

[0070] Extract multi-agent node features using the following formula:

[0071]

[0072]

[0073] in, and They are the first Characteristics of nodes that are operating normally and abnormally in the layer. and These are the learnable coefficients on the dual-view CSGCN. (x) is a Chebyshev polynomial. and These are the normal operation interaction diagrams. Interaction diagram with abnormal operation The normalized Laplace matrix.

[0074] 6) Construct feature aggregation layer 2

[0075] Aggregate the feature information of each entity within a multi-agent system using the following formula. :

[0076]

[0077]

[0078]

[0079] in, This represents the capability characteristics of multiple agents under normal operating conditions. This represents the capability characteristics of multiple agents under abnormal operating conditions. Representing multi-agent The total number of internally running entities, For finite positive integers, Represents an entity index. Represents the feature dimension index and =1,2,..., , (·) represents the feature fusion function. This represents the global interaction state of the merged multi-agent system.

[0080] In step (3) of this invention, the construction method of the multi-source feature fusion and adaptation degree screening module is as follows:

[0081] 1) Predicting the probability of running facts

[0082] Predict the probability of running facts using the following formula :

[0083]

[0084]

[0085]

[0086] MLP stands for Feedforward Artificial Neural Network. There are two types of relational features; Characteristics of a single agent running a task. Features of the task unit For a single agent running a task Features As a characteristic of operational skills, For task units Its characteristics.

[0087] 2) Calculate the skill loss of a single agent

[0088] Calculate single agent skill loss using the following formula :

[0089]

[0090] in, For any multi-agent The number of tasks executed in the process This represents the actual score of a single agent performing a task. This indicates the predicted score.

[0091] 3) Integrating multi-agent collaborative capabilities

[0092] The following formula integrates the characteristics of multi-agent collaborative capabilities. :

[0093] in, Features at the internal structural scale of multi-agent systems. Features of the source interaction scale, (·) is a linear transformation function.

[0094] 4) Predict the results of task execution

[0095] Predict the task execution result using the following formula :

[0096]

[0097] in, For task units Difficulty / sensitivity to change characteristics For task units Distinguishing features For task units The characteristics of the operational skills involved.

[0098] 5) Calculate the multi-agent cooperative loss

[0099] Calculate the multi-agent cooperative loss using the following formula :

[0100]

[0101] in, This represents the actual score of multiple agents performing the task.

[0102] This invention proposes a multi-agent collaborative adaptation screening method based on hierarchical feature learning. It constructs a knowledge graph of single-agent operational states and a knowledge graph of multi-agent collaborative states to model agent behavioral features at two levels: operational state and collaborative relationship. It employs dual-view convolution and high-order community partitioning to fuse single-agent and multi-agent capability features. Through multi-source feature fusion and task execution prediction, it achieves quantitative screening of agent adaptation, improving the accuracy of task allocation and collaborative efficiency. This addresses the technical problem in practical applications where significant differences in execution capabilities, response speeds, skill levels, and resource consumption among different agents limit task collaboration efficiency and overall performance. This invention overcomes the shortcomings of existing multi-agent task allocation and collaborative screening methods, such as limited feature representation, insufficient collaborative relationship modeling, and inaccurate adaptation assessment. It can be applied to task scheduling platforms in industrial automation, multi-robot collaboration, and intelligent manufacturing to support task-agent matching, bottleneck identification, and dynamic task allocation, thereby improving system operating efficiency and task success rate. Attached Figure Description

[0103] Figure 1 This is a flowchart of Example 1 of the present invention.

[0104] Figure 2 This is a schematic diagram of the structure of a hierarchical feature learning network.

[0105] Figure 3 yes Figure 2 A schematic diagram of the structure of the knowledge graph module operating at multiple scales.

[0106] Figure 4 yes Figure 2 A schematic diagram of the multi-scale feature extraction module.

[0107] Figure 5 yes Figure 2 A schematic diagram of the multi-source feature fusion and adaptation degree screening module. Detailed Implementation

[0108] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0109] Example 1

[0110] The multi-agent cooperative adaptation screening method based on hierarchical feature learning in this embodiment consists of the following steps (see...). Figure 1 ):

[0111] (1) Dataset preprocessing

[0112] The multi-scale knowledge graph was used as a dataset and divided into a training set and a test set in an 8:2 ratio.

[0113] (2) Extracting key features

[0114] Convolutional neural networks are used to extract single agent response data and its corresponding task mapping matrix, multi-agent collaborative operation data and its corresponding task relationship matrix, initial features of operation relationship, initial features of single agent, task node features, and operation skill node features.

[0115] (3) Constructing a hierarchical feature learning network

[0116] Figure 2 A schematic diagram of the hierarchical feature learning network in this embodiment is provided. Figure 2 In this embodiment, the hierarchical feature learning network is composed of a multi-scale knowledge graph module, a multi-scale feature extraction module, a multi-source feature fusion and adaptation degree screening module connected in series.

[0117] Figure 3 Given Figure 2 A schematic diagram of the structure of the multi-scale knowledge graph module. Figure 3 In this embodiment, the multi-scale operational knowledge graph module is composed of a single agent operational state knowledge graph module and a multi-agent collaborative state knowledge graph module connected in series.

[0118] The method for constructing the multi-scale operational knowledge graph module in this embodiment is as follows:

[0119] 1) Construct a knowledge graph module for the operational state of a single agent.

[0120] The single-agent operational state knowledge graph includes a normal operation state skill hypergraph. and abnormal operation status skill supermap , ∈ , ∈ ,in yes OK A real matrix of columns, = + , The number of response types to run. The value range is 2 to 10, in this embodiment The value range is 6; =| |For multi-agent The number of entities in the middle, The value range is 2 to 10, in this embodiment The value range is 6.

[0121] Determine the normal operating status of the skill supermap using the following formula and abnormal operation status skill supermap :

[0122]

[0123]

[0124] in, It is a matrix of all zeros. ∈ , for OK A matrix of real numbers in columns; This represents the association matrix between the running state type and each entity. ∈ , yes OK A matrix of real numbers in columns; yes The transpose of the matrix, ∈ , yes OK A matrix of real numbers in columns; and This represents the runtime log matrix within a multi-agent system. ∈ , ∈ , yes OK A column of real numbers.

[0125] 2) Construct a multi-agent collaborative state knowledge graph module

[0126] Multi-agent cooperative state association graphs include positive cooperative state graphs. Negative Cooperative State Diagram Its construction method is the same as that of the single agent runtime state knowledge graph.

[0127] A method for constructing a multi-scale operational knowledge graph module.

[0128] Figure 4 Given Figure 2 A schematic diagram of the multi-scale feature extraction module. Figure 4 In this embodiment, the multi-scale operation feature extraction module is composed of a single-agent scale feature module and a multi-agent scale feature module connected in parallel.

[0129] The single-agent scale feature module of this embodiment consists of a single-agent scale feature extraction layer, a multi-agent community partitioning and contribution calculation layer, and feature aggregation layer 1 connected in series. The multi-agent scale feature module of this embodiment consists of an interaction subgraph feature construction layer, a multi-agent node feature extraction layer, and feature aggregation layer 2 connected in series.

[0130] The construction method of the multi-scale running feature extraction module in this embodiment is as follows:

[0131] 1) Extracting single-agent scale features

[0132] ① Initialization

[0133] Initialize the characteristics of the entity relations and operational relations of the multi-agent system according to formula (1). :

[0134] (1)

[0135] in, This represents the initialization characteristics of two types of operational relationships. ∈ , yes A real matrix with d rows and d columns, where d is the dimension of the initial embedding features, and the value of d ranges from 1 to 2. In this embodiment, the value of d is 26. Representing multi-agent Initialization features of a single agent ∈ , yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within, The value range is 2 to 50, in this embodiment The value range is 30; Representing multi-agent Initialization features of task nodes ∈ yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within, The value range is 2 to 30, in this embodiment The value range is 20; ∈ Representing multi-agent Initialization features of running skill nodes in the middle, ∈ yes A real matrix with d rows and d columns, Representing multi-agent The number of corresponding nodes within, The value range is 1 to 4, in this embodiment The value is 3; (·) represents the feature splicing operation.

[0136] ② Learning characteristics

[0137] Learn normal operating characteristics using the following formula and abnormal operation characteristics :

[0138]

[0139]

[0140] in, This is the single-agent operating state under normal response conditions. This refers to the single-agent operating state under abnormal response conditions. and They are and The normalization degree matrix, and These are the two-view HGCN at the 1st The weight matrix of the layer, For activation functions;

[0141] ③ Fusion characteristics

[0142] Use the following formula to capture dual-view interaction information of normal and abnormal operating states:

[0143]

[0144] =( )

[0145] =( )

[0146] in, For the complex operational relationships of multi-agent systems, This is a characteristic of a single agent. Features of task nodes Features of running skill nodes, and The first one learned on the dual-view HGCN Two characteristics of the layer.

[0147] The features at the single agent scale are obtained by following three steps: initialization, feature learning, and feature fusion.

[0148] 2) Divide multi-agent communities and determine their computational contributions

[0149] Calculate the contribution of the multi-agent community according to formula (2). :

[0150]

[0151]

[0152]

[0153] in, ∈ , It is a 1-row, d-column real matrix. is the characteristic of a single agent; u is the number of single agents, and the value of u ranges from 1 to 50. In this embodiment, the value of u is 30. As for contribution level, , This is the weight matrix. ∈ , Let be a real matrix with d rows and d columns. ∈ , It is a real matrix with d rows and 1 column; .

[0154] 3) Construct feature aggregation layer 1

[0155] Aggregate all multi-agent community features using the following formula :

[0156] in, For any multi-agent community's capability features, the mapping function F(·) is used to enhance the correlation between the features of the multi-agent community.

[0157] 4) Constructing interactive subgraph features

[0158] Construct interactive subgraph features using the following formula:

[0159]

[0160]

[0161] in, Interaction diagram indicating normal operation Interactive subgraph features, Interaction diagram representing abnormal operation Interactive subgraph features, Represents the identity matrix. and These are positive cooperative state diagrams. Negative Cooperative State Diagram The corresponding degree matrix.

[0162] 5) Extract features from multi-agent nodes

[0163] Extract multi-agent node features using the following formula:

[0164]

[0165]

[0166] in, and They are the first Characteristics of nodes that are operating normally and abnormally in the layer. and These are the learnable coefficients on the dual-view CSGCN. (x) is a Chebyshev polynomial. and These are the normal operation interaction diagrams. Interaction diagram with abnormal operation The normalized Laplace matrix.

[0167] 6) Construct feature aggregation layer 2

[0168] Aggregate the feature information of each entity within a multi-agent system using the following formula. :

[0169]

[0170]

[0171]

[0172] in, This represents the capability characteristics of multiple agents under normal operating conditions. This represents the capability characteristics of multiple agents under abnormal operating conditions. Representing multi-agent The total number of internally running entities, For finite positive integers, Represents an entity index. Represents the feature dimension index and =1,2,..., , (·) represents the feature fusion function. This represents the global interaction state of the merged multi-agent system.

[0173] A method for constructing a multi-scale feature extraction module.

[0174] Figure 5 Given Figure 2 A schematic diagram of the multi-source feature fusion and adaptation degree screening module. Figure 5 In this embodiment, the multi-source feature fusion and adaptation degree screening module is composed of a multi-agent community feature fusion module, a multi-agent collaborative feature fusion module, and a joint objective function loss calculation layer. The outputs of the multi-agent community feature fusion module and the multi-agent collaborative feature fusion module are connected to the joint objective function loss calculation layer.

[0175] The multi-agent community feature fusion module in this embodiment consists of a running fact probability prediction layer and a single-agent skill loss calculation layer connected in series. The multi-agent cooperative feature fusion module in this embodiment consists of a multi-agent cooperative capability feature fusion layer, a task execution result prediction layer, and a multi-agent cooperative loss calculation layer connected in series.

[0176] The construction method of the multi-source feature fusion and adaptation degree screening module in this embodiment is as follows:

[0177] 1) Predicting the probability of running facts

[0178] Predict the probability of running facts using the following formula :

[0179]

[0180]

[0181]

[0182] MLP stands for Feedforward Artificial Neural Network. There are two types of relational features; Characteristics of a single agent running a task. Features of the task unit For a single agent running a task Features As a characteristic of operational skills, For task units Its characteristics.

[0183] 2) Calculate the skill loss of a single agent

[0184] Calculate single agent skill loss using the following formula :

[0185]

[0186] in, For any multi-agent The number of tasks executed in the process This represents the actual score of a single agent performing a task. This indicates the predicted score.

[0187] 3) Integrating multi-agent collaborative capabilities

[0188] The following formula integrates the characteristics of multi-agent collaborative capabilities. :

[0189] in, Features at the internal structural scale of multi-agent systems. Features of the source interaction scale, (·) is a linear transformation function.

[0190] 4) Predict the results of task execution

[0191] Predict the task execution result using the following formula :

[0192]

[0193] in, For task units Difficulty / sensitivity to change characteristics For task units Distinguishing features For task units The characteristics of the operational skills involved.

[0194] 5) Calculate the multi-agent cooperative loss

[0195] Calculate the multi-agent cooperative loss using the following formula :

[0196]

[0197] in, This represents the actual score of multiple agents performing the task.

[0198] A method for constructing a module for multi-source feature fusion and adaptation degree screening.

[0199] (4) Training the hierarchical feature learning network

[0200] 1) Constructing the loss function

[0201] Construct the loss function as follows :

[0202]

[0203] in, This represents the loss function for single-agent cooperative tasks. This represents the loss function for multi-agent cooperative tasks. It is the weight decay parameter. The value range is 0.0001 to 0.1, in this embodiment. The value can be 0.005, or it can be arbitrarily selected within the range of 0.0001 to 0.1.

[0204] 2) Training a hierarchical feature learning network

[0205] The training set was fed into the hierarchical feature learning network for training. The training parameters were as follows: the server graphics card used for training was an NVIDIA GeForce RTX 3090, the initial learning rate was 0.001, the number of training epochs was 100, the batch size was 32, the Adam optimizer was used, and training was performed until the loss function was reached. convergence.

[0206] (5) Testing the hierarchical feature learning network

[0207] The test set is input into the trained hierarchical feature learning network for testing, and the output is the degree of multi-agent collaborative adaptation.

[0208] A method for screening the degree of multi-agent collaborative adaptation based on hierarchical feature learning was developed.

[0209] This invention constructs a knowledge graph of the single-agent operational state and a knowledge graph of the multi-agent collaborative state, modeling agent behavioral features from two layers: operational state and collaborative relationships. It employs dual-view convolution and high-order community partitioning to fuse single-agent and multi-agent capability features. Through multi-source feature fusion and task execution prediction, it achieves quantitative screening of agent suitability, improving the accuracy of task allocation and collaborative efficiency. This addresses the technical problem in practical applications where significant differences in execution capabilities, response speeds, skill levels, and resource consumption among different agents limit task collaborative efficiency and overall performance. This invention overcomes the shortcomings of existing multi-agent task allocation and collaborative screening methods, such as limited feature representation, insufficient collaborative relationship modeling, and inaccurate suitability assessment. It can be applied to task scheduling platforms in industrial automation, multi-robot collaboration, and intelligent manufacturing to support task-agent matching, bottleneck identification, and dynamic task allocation, thereby improving system operating efficiency and task success rate.

[0210] Example 2

[0211] The multi-agent cooperative adaptation screening method based on hierarchical feature learning in this embodiment consists of the following steps:

[0212] (1) Dataset preprocessing

[0213] The steps are the same as in Example 1.

[0214] (2) Extracting key features

[0215] The steps are the same as in Example 1.

[0216] (3) Constructing a hierarchical feature learning network

[0217] The hierarchical feature learning network structure is the same as that in Example 1.

[0218] The method for constructing the multi-scale operational knowledge graph module in this embodiment is as follows:

[0219] 1) Extracting single-agent scale features

[0220] ① Initialization

[0221] Initialize the characteristics of the entity relations and operational relations of the multi-agent system according to formula (1). :

[0222] (1)

[0223] in, This represents the initialization characteristics of two types of operational relationships. ∈ , yes A real matrix with d rows and d columns, where d is the dimension of the initial embedding features, and the value of d ranges from 1 to 2. In this embodiment, d is set to 25. Representing multi-agent Initialization features of a single agent ∈ , yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within, The value range is 2 to 50, in this embodiment The value is 2; Representing multi-agent Initialization features of task nodes ∈ yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within, The value range is 2 to 30, in this embodiment The value range is 2; ∈ Representing multi-agent Initialization features of running skill nodes in the middle, ∈ yes A real matrix with d rows and d columns, Representing multi-agent The number of corresponding nodes within, The value range is 1 to 4, in this embodiment The value is 1; (·) represents the feature splicing operation.

[0224] ② Learning characteristics

[0225] The steps are the same as in Example 1.

[0226] ③ Fusion characteristics

[0227] The steps are the same as in Example 1.

[0228] 2) Divide multi-agent communities and determine their computational contributions

[0229] Calculate the contribution of the multi-agent community according to formula (2). :

[0230] The expression of equation (2) is the same as that in Example 1.

[0231] In equation (2), u is the number of single agents, and the value of u ranges from 1 to 50. In this embodiment, the value of u ranges from 1. Other parameters and variables, as well as their value ranges, are the same as in embodiment 1.

[0232] The other steps in this step are the same as in Example 1. This completes the method for constructing the multi-scale operational knowledge graph module.

[0233] The other steps are the same as in Example 1. This completes the method for screening the degree of multi-agent collaborative adaptation based on hierarchical feature learning.

[0234] Example 3

[0235] The multi-agent cooperative adaptation screening method based on hierarchical feature learning in this embodiment consists of the following steps:

[0236] (1) Dataset preprocessing

[0237] The steps are the same as in Example 1.

[0238] (2) Extracting key features

[0239] The steps are the same as in Example 1.

[0240] (3) Constructing a hierarchical feature learning network

[0241] The hierarchical feature learning network structure is the same as that in Example 1.

[0242] The method for constructing the multi-scale operational knowledge graph module in this embodiment is as follows:

[0243] 1) Extracting single-agent scale features

[0244] ① Initialization

[0245] Initialize the characteristics of the entity relations and operational relations of the multi-agent system according to formula (1). :

[0246] (1)

[0247] In equation (1), This represents the initialization characteristics of two types of operational relationships. ∈ , yes A real matrix with d rows and d columns, where d is the dimension of the initial embedding features, and the value of d ranges from 1 to 2. In this embodiment, the value of d is 28. Representing multi-agent Initialization features of a single agent ∈ , yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within, The value range is 2 to 50, in this embodiment The value range is 50; Representing multi-agent Initialization features of task nodes ∈ yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within, The value range is 2 to 30, in this embodiment The value range is 30; ∈ Representing multi-agent Initialization features of running skill nodes in the middle, ∈ yes A real matrix with d rows and d columns, Representing multi-agent The number of corresponding nodes within, The value range is 1 to 4, in this embodiment The value range is 4; (·) represents the feature splicing operation.

[0248] ② Learning characteristics

[0249] The steps are the same as in Example 1.

[0250] ③ Fusion characteristics

[0251] The steps are the same as in Example 1.

[0252] 2) Divide multi-agent communities and determine their computational contributions

[0253] Calculate the contribution of the multi-agent community according to formula (2). :

[0254] The expression of equation (2) is the same as that in Example 1.

[0255] In equation (2), u is the number of single agents, and the value of u ranges from 1 to 50. In this embodiment, the value of u is 50. Other parameters and variables, as well as their value ranges, are the same as in embodiment 1.

[0256] The other steps in this step are the same as in Example 1. This completes the method for constructing the multi-scale operational knowledge graph module.

[0257] The other steps are the same as in Example 1. This completes the method for screening the degree of multi-agent collaborative adaptation based on hierarchical feature learning.

Claims

1. A method for screening the cooperative adaptation degree of multi-agent agents based on hierarchical feature learning, characterized in that... It consists of the following steps: (1) Dataset preprocessing The multi-scale knowledge graph was used as the dataset and divided into a training set and a test set in an 8:2 ratio. (2) Extracting key features Convolutional neural networks are used to extract single agent response data and its corresponding task mapping matrix, multi-agent collaborative operation data and its corresponding task relationship matrix, initial features of operation relationship, initial features of single agent, task node features and operation skill node features. (3) Constructing a hierarchical feature learning network The hierarchical feature learning network consists of a multi-scale knowledge graph module, a multi-scale feature extraction module, a multi-source feature fusion and adaptation degree screening module connected in series. (4) Training the hierarchical feature learning network 1) Constructing the loss function Construct the loss function as follows : in, This represents the loss function for single-agent cooperative tasks. This represents the loss function for multi-agent cooperative tasks. This is the weight decay parameter, with a value range of 0.0001 to 0.1; 2) Training a hierarchical feature learning network The training set was input into the hierarchical feature learning network for training. The training parameters were as follows: the server graphics card used for training was an NVIDIA GeForce RTX 3090, the initial learning rate was 0.001, the number of training epochs was 100, the batch size for training was 32, the Adam optimizer was used, and training was carried out until the loss function converged. (5) Testing the hierarchical feature learning network The test set is input into the trained hierarchical feature learning network for testing, and the output is the degree of multi-agent collaborative adaptation.

2. The multi-agent cooperative adaptation degree screening method based on hierarchical feature learning according to claim 1, characterized in that: In step (3), the hierarchical feature learning network is constructed by connecting a single agent running state knowledge graph module and a multi-agent collaborative state knowledge graph module.

3. The multi-agent cooperative adaptation degree screening method based on hierarchical feature learning according to claim 1, characterized in that: In step (3), the hierarchical feature learning network is constructed by connecting a single agent scale feature module and a multi-agent scale feature module in parallel.

4. The multi-agent cooperative adaptation degree screening method based on hierarchical feature learning according to claim 3, characterized in that: In step (3) of constructing the hierarchical feature learning network, the single agent scale feature module is composed of a single agent scale feature extraction layer, a multi-agent community division and contribution calculation layer, and feature aggregation layer 1 connected in series; the multi-agent scale feature module is composed of an interaction subgraph feature construction layer, a multi-agent node feature extraction layer, and feature aggregation layer 2 connected in series.

5. The method for screening the degree of multi-agent cooperative adaptation based on hierarchical feature learning according to claim 1, characterized in that: The multi-source feature fusion and adaptation degree screening module consists of a multi-agent community feature fusion module, a multi-agent collaborative feature fusion module, and a joint objective function loss calculation layer connected together; the outputs of the multi-agent community feature fusion module and the multi-agent collaborative feature fusion module are connected to the joint objective function loss calculation layer. The multi-agent community feature fusion module is composed of a running fact probability prediction layer and a single agent skill loss calculation layer connected in series; the multi-agent collaborative feature fusion module is composed of a multi-agent collaborative capability feature fusion layer, a task execution result prediction layer, and a multi-agent collaborative loss calculation layer connected in series.

6. The multi-agent cooperative adaptation degree screening method based on hierarchical feature learning according to claim 1, characterized in that: In step (3), the construction method of the multi-scale knowledge graph module in building the hierarchical feature learning network is as follows: 1) Construct a knowledge graph module for the operational state of a single agent. The single-agent operational state knowledge graph includes a normal operation state skill hypergraph. and abnormal operation status skill supermap , ∈ , ∈ ,in yes OK A real matrix of columns, = + , The number of response types to run. The value range is 2 to 10; =| |For multi-agent The number of entities in the middle, The value range is 2 to 10; Determine the normal operating status of the skill supermap using the following formula and abnormal operation status skill supermap in, It is a matrix of all zeros. ∈ , for OK A matrix of real numbers in columns; This represents the association matrix between the running state type and each entity. ∈ , yes OK A matrix of real numbers in columns; yes The transpose of the matrix, ∈ , yes OK A matrix of real numbers in columns; and This represents the runtime log matrix within a multi-agent system. ∈ , ∈ , yes OK A matrix of real numbers in columns; 2) Construct a multi-agent collaborative state knowledge graph module Multi-agent cooperative state association graphs include positive cooperative state graphs. Negative Cooperative State Diagram Its construction method is the same as that of the single agent runtime state knowledge graph.

7. The method for screening the degree of multi-agent cooperative adaptation based on hierarchical feature learning according to claim 1, characterized in that: In step (3), the construction method of the multi-scale running feature extraction module in building the hierarchical feature learning network is as follows: 1) Extracting single-agent scale features ① Initialization Initialize the characteristics of entity relations and operational relations of multi-agent systems using the following formula. : in, This represents the initialization characteristics of two types of operational relationships. ∈ , yes A real matrix with d rows and d columns, where d is the feature dimension of the initial embedding, and its value ranges from 1 to 2. ; Representing multi-agent Initialization features of a single agent ∈ , yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within the range of 2 to 50; Representing multi-agent Initialization features of task nodes ∈ yes A real matrix with d rows and d columns, It is a multi-agent The number of corresponding nodes within the range of 2 to 30; ∈ Representing multi-agent Initialization features of running skill nodes in the middle, ∈ yes A real matrix with d rows and d columns, Representing multi-agent The number of corresponding nodes within the range of 1 to 4; (·) represents the feature concatenation operation; ② Learning characteristics Learn normal operating characteristics using the following formula and abnormal operation characteristics : in, This is the single-agent operating state under normal response conditions. This refers to the single-agent operating state under abnormal response conditions. and They are and The normalization degree matrix, and These are the two-view HGCN at the 1st The weight matrix of the layer, For activation functions; ③ Fusion characteristics Use the following formula to capture dual-view interaction information of normal and abnormal operating states: =( ) =( ) in, For the operational relationships of multi-agent systems, This is a characteristic of a single agent. Features of task nodes Features of running skill nodes, and The first one learned on the dual-view HGCN Two characteristics of layers; The features at the single agent scale are obtained by following three steps: initialization, feature learning, and feature fusion. 2) Divide multi-agent communities and determine their computational contributions Calculate the contribution of the multi-agent community using the following formula. : in, ∈ , It is a 1-row, d-column real matrix. This is a characteristic of a single agent; The number of single agents, The value range is 1 to 50; As for contribution level, , This is the weight matrix. ∈ , Let be a real matrix with d rows and d columns. ∈ , It is a real matrix with d rows and 1 column; ; 3) Construct feature aggregation layer 1 Aggregate all multi-agent community features using the following formula : in, For any multi-agent community's capability characteristics, the mapping function F(·) is used to enhance the correlation between the characteristics of the multi-agent community; 4) Constructing interactive subgraph features Construct interactive subgraph features using the following formula: in, Interaction diagram indicating normal operation Interactive subgraph features, Interaction diagram representing abnormal operation Interactive subgraph features, Represents the identity matrix. and These are positive cooperative state diagrams. Negative Cooperative State Diagram The corresponding degree matrix; 5) Extract features from multi-agent nodes Extract multi-agent node features using the following formula: in, and They are the first Characteristics of nodes that are operating normally and abnormally in a layer. and These are the learnable coefficients on the dual-view CSGCN. (x) is a Chebyshev polynomial. and These are the normal operation interaction diagrams. Interaction diagram with abnormal operation The normalized Laplace matrix; 6) Construct feature aggregation layer 2 Aggregate the feature information of each entity within a multi-agent system using the following formula. : in, This represents the capability characteristics of multiple agents under normal operating conditions. This represents the capability characteristics of multiple agents under abnormal operating conditions. Representing multi-agent The total number of internally running entities, For finite positive integers, Represents an entity index. Represents the feature dimension index and =1,2,..., , (·) represents the feature fusion function. This represents the global interaction state of the merged multi-agent system.

8. The method for screening the degree of multi-agent cooperative adaptation based on hierarchical feature learning according to claim 1, characterized in that: In step (3), the construction method of the multi-source feature fusion and adaptation degree screening module in building the hierarchical feature learning network is as follows: 1) Predicting the probability of running facts Predict the probability of running facts using the following formula : MLP stands for Feedforward Artificial Neural Network. There are two types of relational features; Characteristics of a single agent running a task. Features of the task unit For a single agent running a task Features As a characteristic of operational skills, For task units Features; 2) Calculate the skill loss of a single agent Calculate single agent skill loss using the following formula : in, For any multi-agent The number of tasks executed in the process This represents the actual score of a single agent performing a task. Indicates the predicted score; 3) Integrating multi-agent collaborative capabilities The following formula integrates the characteristics of multi-agent collaborative capabilities. : in, Features at the internal structural scale of multi-agent systems. Features of the source interaction scale, (·) is a linear transformation function; 4) Predict the results of task execution Predict the task execution result using the following formula : in, For task units Difficulty / sensitivity to change characteristics For task units Distinguishing features For task units The characteristics of the operational skills involved; 5) Calculate the multi-agent cooperative loss Calculate the multi-agent cooperative loss using the following formula : in, This represents the actual score of multiple agents performing the task.

Images