Multi-agent system cooperative behavior pattern recognition method, efficiency evaluation method and device and storage medium

Abstract

The invention discloses a multi-agent system cooperative behavior mode identification method, an efficiency evaluation method and device and a storage medium. The identification method comprises the following steps: determining a node model corresponding to the multi-agent system according to received agent characteristic parameters of the multi-agent system; according to the node model and a preset cooperative behavior identification condition, determining a hyperedge corresponding to the intelligent agent system; determining a hypergraph neural network model corresponding to the agent system according to the node model and the hyperedge; through a hypergraph attention mechanism, according to a hypergraph neural network model, determining a collaborative behavior representation vector of a hyperedge; and according to the cooperative behavior representation vector, identifying a cooperative behavior mode corresponding to the hyperedge.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-agent system, in particular to a multi-agent system cooperative behavior pattern recognition method and device, an efficiency evaluation method and device, and a storage medium. BACKGROUND

[0002] Traditional graph neural network structure mainly models binary relationships, i.e., a single edge can only connect two nodes. This structure is relatively effective in describing one-to-one interaction between agents, but it is difficult to express complex cooperative behaviors such as many-to-many or one-to-many. For example, multiple agents perform parallel operations around a sub-task at the same time, and several individuals cooperate to complete a perception-judgment-execution chain. In these "group cooperation" behaviors, there are multiple subjects acting on a single task, and traditional graph structures cannot fully capture such high-order interaction patterns.

[0003] To solve this problem, the hypergraph modeling method provides more flexible structure expression capability. Hypergraph is a structure that can associate multiple nodes through a hyperedge. It can not only depict complex multi-element relationships, but also capture potential group cooperation structures between agents while preserving individual information.

[0004] Based on the hypergraph structure, the hypergraph neural network (HGNN) model proposed in recent years can realize bidirectional propagation of information between nodes and hyperedges, thereby effectively extracting high-order relationships between nodes. This mechanism is naturally suitable for modeling multi-agent cooperative behaviors such as heterogeneous collaboration, role complementation, and task pairing. However, the current HGNN still faces the following challenges in the application of multi-agent system efficiency evaluation:

[0005] However, the inventors of the present application have found that the existing construction of hypergraph neural networks has the problems of low automation degree and limited behavior pattern recognition capability. For example, in a multi-agent system, how to identify group behavior from running data and convert it into a hyperedge structure still lacks an efficient and universal method. Existing behavior pattern recognition methods mostly take a single agent as the analysis unit, making it difficult to systematically identify and induce cooperative patterns among multiple individuals.

[0006] The content of the background section merely represents the knowledge of the discloser and does not necessarily represent the state of the art. SUMMARY

[0007] According to an aspect of the present application, the present application provides a method for identifying a cooperative behavior pattern of a multi-agent system, the method comprising: determining a node model corresponding to the multi-agent system according to received agent feature parameters of the multi-agent system; determining a hyperedge corresponding to the multi-agent system according to the node model and a preset cooperative behavior identification condition; determining a hypergraph neural network model corresponding to the multi-agent system according to the node model and the hyperedge; determining a cooperative behavior representation vector of the hyperedge according to the hypergraph neural network model through a hypergraph attention mechanism; and identifying a cooperative behavior pattern corresponding to the hyperedge according to the cooperative behavior representation vector.

[0008] According to some embodiments of the present application, determining a node model corresponding to the multi-agent system according to received agent feature parameters of the multi-agent system comprises: determining state sequence data of the multi-agent system according to received agent feature parameters at fixed time steps; determining node features of the agent according to the state sequence data; and determining the node model according to the node features.

[0009] According to some embodiments of the present application, determining a hyperedge corresponding to the multi-agent system according to the node model and the preset cooperative behavior identification condition comprises: determining nodes corresponding to the preset cooperative behavior identification condition according to the node model and the preset cooperative behavior identification condition; and aggregating the nodes corresponding to the preset cooperative behavior identification condition to determine the hyperedge.

[0010] According to some embodiments of the present application, determining a cooperative behavior representation vector according to the hypergraph neural network model through the hypergraph attention mechanism comprises: performing linear mapping on node features of the node model to obtain features in a unified representation space of the node features; determining a first cooperative behavior representation vector of the hyperedge according to the features in the unified representation space of the node features and a hyperedge feature vector; determining updated features of the node features according to the first cooperative behavior representation vector; and outputting the cooperative behavior representation vector of the hyperedge through preset hierarchical stacking according to the updated features.

[0011] According to some embodiments of the present application, identifying a cooperative behavior pattern corresponding to the hyperedge according to the cooperative behavior representation vector comprises: identifying a cooperative behavior pattern corresponding to the hyperedge according to the cooperative behavior representation vector through a clustering algorithm, wherein the clustering algorithm comprises a K-means clustering algorithm, a density clustering algorithm, and a spectral clustering algorithm.

[0012] According to another aspect of the present application, the present application also provides a method for evaluating performance of a multi-agent system, the method comprising: determining a node model corresponding to the multi-agent system according to received agent feature parameters of the multi-agent system; determining a hyperedge corresponding to the multi-agent system according to the node model and a preset cooperative behavior recognition condition; determining a hypergraph neural network model corresponding to the multi-agent system according to the node model and the hyperedge; determining a cooperative behavior representation vector of the hyperedge according to the hypergraph neural network model through a hypergraph attention mechanism; recognizing a cooperative behavior mode corresponding to the hyperedge according to the cooperative behavior representation vector; and determining a performance evaluation parameter according to the cooperative behavior mode.

[0013] According to another aspect of the present application, the present application also provides a device for recognizing a cooperative behavior mode of a multi-agent system, the device comprising a processing module. The processing module determines a node model corresponding to the multi-agent system according to received agent feature parameters of the multi-agent system; the processing module determines a hyperedge corresponding to the multi-agent system according to the node model and a preset cooperative behavior recognition condition; the processing module determines a hypergraph neural network model corresponding to the multi-agent system according to the node model and the hyperedge; the processing module determines a cooperative behavior representation vector of the hyperedge according to the hypergraph neural network model through a hypergraph attention mechanism; and the processing module recognizes a cooperative behavior mode corresponding to the hyperedge according to the cooperative behavior representation vector.

[0014] According to some embodiments of the present application, the processing module further determines a performance evaluation parameter according to the cooperative behavior mode.

[0015] According to another aspect of the present application, the present application also provides a non-volatile computer-readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement the recognition method as described above.

[0016] According to another aspect of the present application, the present application also provides an electronic device, comprising: one or more processors; and a storage device configured to store one or more programs, the one or more programs being executable by the one or more processors to enable the one or more processors to implement the recognition method as described above.

[0017] According to another aspect of the present application, the present application also provides a computer program product, comprising: a computer program stored on a computer-readable storage medium; and the computer program comprising program instructions executable by a computer to cause the computer to perform the recognition method as described above.

[0018] According to another aspect of the present application, the present application also provides a non-volatile computer-readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement the performance evaluation method as described above.

[0019] According to another aspect of the present application, the present application further provides an electronic device, comprising: one or more processors; a storage device configured to store one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement the performance evaluation method as described above.

[0020] According to another aspect of the present application, the present application further provides a computer program product, comprising: a computer program stored on a computer readable storage medium; the computer program comprising program instructions which, when executed by a computer, cause the computer to perform the performance evaluation method as described above. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort based on these drawings.

[0022] Figure 1 A flowchart of a recognition method 1000 according to an embodiment of the present application is shown;

[0023] Figure 2 A flowchart of step S110 according to an embodiment of the present application is shown;

[0024] Figure 3 A flowchart of step S120 according to an embodiment of the present application is shown;

[0025] Figure 4 A flowchart of step S150 according to an embodiment of the present application is shown;

[0026] Figure 5 A flowchart of a performance evaluation method 2000 according to an embodiment of the present application is shown;

[0027] Figure 6 A structural diagram of a recognition device according to an embodiment of the present application is shown;

[0028] Figure 7 A framework structural diagram of a hypergraph neural network model according to an embodiment of the present application is shown.

[0029] LEGEND OF DRAWINGS

[0030] A recognition device 30; a processing module 31. DETAILED DESCRIPTION

[0031] Example embodiments now will be described more fully hereinafter with reference to the accompanying drawings. Example embodiments, however, can be implemented in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of example embodiments to those skilled in the art. Like reference numerals refer to like elements throughout the several views.

[0032] The described features, structures, or characteristics can be combined in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the disclosure. One skilled in the relevant art will recognize, however, that the technology can be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In these instances, well-known structures, methods, devices, implementations, materials, and so forth can not have been described in detail in order to avoid obscuring aspects of the technology.

[0033] Furthermore, the term "comprising" and "including" and their variants are intended to cover both the case where only the stated elements are present and the case where additional elements are also present. For example, a process, method, system, product, or apparatus that comprises or includes one step or unit can optionally also include other steps or units not listed.

[0034] The terms "first", "second", and the like, in the description and in the claims, do not necessarily denote any ordinal, but are used to distinguish different objects.

[0035] The technical solutions in the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0036] The English and its English full name and corresponding Chinese interpretation involved in the present application are as follows:

[0037] GNN, Graph Neural Networks, graph neural network;

[0038] K-means, K-means clustering;

[0039] DBSCAN, Density-Based Spatial Clustering of Applications with Noise, density clustering;

[0040] Spectral Clustering

[0041] HGNN, Hypergraph Neural Networks

[0042] ID, Identification

[0043] GAT, Graph Attention Networks

[0044] GCN, Graph Convolutional Networks

[0045] According to an aspect of the present application, the present application provides a method for identifying a cooperative behavior pattern of a multi-agent system. The identification method 1000 can be executed by a computer system. Exemplarily, the computer system can be a host (or server) with data processing capability.

[0046] Referring to Figure 1 , the identification method 1000 can include steps S110-S150.

[0047] In step S110, the computer system determines a node model corresponding to the multi-agent system according to the received agent feature parameters of the multi-agent system.

[0048] According to an example embodiment, the multi-agent system can be a system composed of heterogeneous agents such as drones, unmanned vehicles, and carrying robots, which completes complex tasks through interaction and cooperation. The agent can be an intelligent device such as a drone, an unmanned vehicle, and a carrying robot.

[0049] The agent feature parameter can be a parameter describing the current state of the agent. The agent feature parameter can include kinematic state parameters, behavior state parameters, perception ability parameters, communication characteristic parameters, energy consumption characteristic parameters, and type label parameters.

[0050] The node model can be a relationship model between nodes in the hypergraph neural network model corresponding to the multi-agent system. Each agent corresponds to a node, and the computer system can regard each agent as a node v i .

[0051] For example, the computer system can determine the state sequence data of the agent system according to the agent feature parameters received at fixed time steps; the computer system can determine the node features of the agent according to the state sequence data; and the computer system can determine the node model according to the node features.

[0052] During the running of the multi-agent system, the computer system can continuously record time series data of the state of the agent, and collect task state information such as task assignment, task execution progress, task completion, and environmental change data such as obstacles, interference sources, and emergencies, to form a complete multi-agent system behavior trajectory data set. The computer system can also record background information such as task state information, environmental change data, and system configuration to form a complete running log.

[0053] In step S120, the computer system determines the super-edge corresponding to the agent system according to the node model and the preset cooperative behavior recognition condition.

[0054] According to an example embodiment, the super-edge can be a same cooperative unit at any time when the plurality of agents meet the preset cooperative behavior recognition condition. One super-edge can connect any number of nodes.

[0055] The preset cooperative behavior recognition condition can be a condition for judging whether different nodes are cooperative behaviors. For example, the system behavior recognition condition can include one of the following conditions: different nodes jointly participate in a same subtask (such as cooperatively carrying a same object, jointly detecting a same target area, etc.), different nodes are in a same subregion and maintain a specific topological structure (such as maintaining a circle or maintaining on a straight line), and different nodes have continuous information interaction or dependence with each other.

[0056] According to an example embodiment, the computer system can determine nodes that meet the preset cooperative behavior recognition condition according to the node model and the preset cooperative behavior recognition condition; and the computer system aggregates the nodes that meet the preset cooperative behavior recognition condition to determine the super-edge.

[0057] In step S130, the computer system determines the supergraph neural network model corresponding to the agent system according to the node model and the super-edge.

[0058] According to an example embodiment, the supergraph neural network model can be a multi-element neural network model corresponding to the multi-agent system.

[0059] The computer system can construct the supergraph neural network model through the node model and all super-edges.

[0060] In step S140, the computer system determines the cooperative behavior representation vector of the super-edge according to the supergraph neural network model through a supergraph attention mechanism.

[0061] According to an example embodiment, the supergraph attention mechanism can be a mechanism for processing complex high-order relationships of the supergraph neural network model.

[0062] The cooperative behavior representation vector can be a feature vector representing the cooperative behavior pattern of the super-edge. Each super-edge can correspond to a cooperative behavior representation vector.

[0063] For example, the computer system can perform linear mapping on the node features of the node model to obtain features of a uniform representation space of the node features; the computer system can determine a first collaborative behavior representation vector of the hyperedge according to the features of the uniform representation space of the node features and the hyperedge feature vector; the computer system can determine updated features of the node features according to the first collaborative behavior representation vector; and the computer system can output a collaborative behavior representation vector of the hyperedge through a preset hierarchical stacking according to the updated features.

[0064] In step S150, the computer system identifies a collaborative behavior mode corresponding to the hyperedge according to the collaborative behavior representation vector.

[0065] According to an example embodiment, the collaborative behavior mode can be a behavior mode with similar structures and consistent task attributes.

[0066] The collaborative behavior mode can include a local collaboration mode, a long-chain dependency mode, a hub coordination mode, and a distributed parallel mode. The local collaboration mode is characterized by short distance and few nodes. The long-chain dependency mode is characterized by asynchronous task relay. The hub coordination mode is characterized by a core node and multiple edge nodes. The distributed parallel mode is characterized by multiple sub-teams with no intersection for parallel operation.

[0067] For example, the computer system can classify multiple collaborative behavior units with similar structures and consistent task attributes into the same collaborative behavior mode according to the collaborative behavior representation vector through a clustering algorithm.

[0068] According to the above-described embodiments, the technical solution of the present application can determine a node model corresponding to a multi-agent system through an agent feature parameter. The technical solution of the present application can determine a hyperedge corresponding to the multi-agent system through the node model and a preset collaborative behavior recognition condition. The technical solution of the present application can determine a hypergraph neural network model corresponding to the multi-agent system through the node model and the hyperedge. The technical solution of the present application can determine a collaborative behavior representation vector of the hyperedge according to the hypergraph neural network model through a hypergraph attention mechanism. The technical solution of the present application can identify a collaborative behavior mode corresponding to the hyperedge through the collaborative behavior representation vector.

[0069] The identification method provided by the present application can effectively capture complex collaboration relationships such as one-to-many and many-to-many between multiple agents by constructing a hypergraph neural network model containing multiple types of nodes and multiple granularity hyperedges, determine a hyperedge through a preset collaborative behavior recognition condition, and have good structure generalization ability through a hypergraph attention mechanism, which can adapt to multiple system types and task structures and support unified modeling in a heterogeneous node and heterogeneous task environment.

[0070] The identification method of the present application can break through the limitation of traditional graph structure supporting only binary relationship, and is suitable for team cooperation, resource scheduling and dynamic coordination behavior identification in complex tasks.

[0071] Optionally, referring to Figure 2 , step S110 can include steps S111-S113.

[0072] In step S111, the computer system determines state sequence data of the agent system according to the received agent feature parameters at fixed time steps.

[0073] According to an example embodiment, the fixed time step can be an interval of each step of the agent. For example, the fixed time step Δt can be set to 1s-10s.

[0074] The agent feature parameters can include kinematic state parameters, behavior state parameters, perception ability parameters, communication characteristic parameters, energy consumption characteristic parameters, and type label parameters, etc.

[0075] The kinematic state parameters can include the position, velocity and acceleration of the agent, etc. The behavior state parameters can include whether the agent performs a task, the stage of the agent, and the like. The perception ability parameters can include the sensor range of the agent and the accuracy of the agent, etc. The communication characteristic parameters can include the communication radius of the agent and the data throughput rate of the agent, etc. The energy consumption characteristic parameters can include the remaining energy of the agent and the execution consumption of the agent, etc. The type label parameters can include the type of the agent, such as transportation type, detection type, perception type, and decision type, etc.

[0076] The state sequence data can be a series of discrete or continuous state observation sets generated by the evolution of the agent of the multi-agent system over time.

[0077] The computer system can receive all the agent feature parameters in real time through a unified data transmission interface, or collect them in real time through sensors, and sample the state sequence data at fixed time steps Δt, as shown in the following formula:

[0078] S={s 0 ,s 1 ,s 2 ,...,s T},s t ={f(v1) t ,...,f(v N ) t};

[0079] Wherein f(v1) t is the node feature of node v1 at time t; f(v N ) tFor node v N The node features at time t; s t Let s be the set of node features of all nodes at time t; 0 S represents the set of node features for all nodes at the initial time; S represents the state sequence data.

[0080] In step S112, the computer system determines the node characteristics of the agent based on the state sequence data.

[0081] According to the example embodiment, node features can be vector representations used to represent the attributes or states carried by a node. Node features consist of multiple dimensions, including position vectors, velocity and acceleration vectors, energy consumption state vectors, communication feature vectors, task participation information vectors, and type label vectors.

[0082] For example, a computer system can determine node characteristics using the following formula:

[0083] f(v i )=[p i ,u i ,γ i ,β i ,t i ,θ i ];

[0084] Where, f(v) i ) is node v i Node characteristics; p i For node v i Two-dimensional or three-dimensional spatial position vector; u i For node v i velocity and acceleration vectors; γ i For node v i The energy consumption state vector, such as remaining power and energy consumption rate. β i For node v i The communication characteristic vectors, such as bandwidth and delay; t i For node v i The task participation information vector, such as the current task ID and stage; θ i This is a type label vector (such as transportation, perception, etc.).

[0085] In step S113, the computer system determines the node model based on the node characteristics.

[0086] According to an example embodiment, a computer system can construct a node model based on nodes and their corresponding node features.

[0087] Through the above embodiments, the technical scheme of the application can construct a dynamic node model through the agent feature parameters of the multi-agent system, thereby constructing a dynamic hypergraph neural network model.

[0088] Optionally, referring to Figure 3 , step S120 can include step S121 and step S122.

[0089] In step S121, the computer system determines the nodes satisfying the preset cooperative behavior recognition condition according to the node model and the preset cooperative behavior recognition condition.

[0090] For example, the computer system can determine whether the nodes satisfy the cooperative behavior recognition condition according to the following formula:

[0091]

[0092] wherein d(v i ,v j ) is the distance between node v i and node v j at a certain moment; δ t (v i ,v j ) is the continuous cooperation duration of node v i and node v j ; ∈ is a space threshold value; τ is a time threshold value.

[0093] In the case where the computer system determines that φ(v i ,v j ) = 1, node v i and node v j constitute a cooperative pair. The computer system can traverse all nodes to determine all cooperative pairs.

[0094] In step S122, the computer system aggregates the nodes satisfying the preset cooperative behavior recognition condition to determine a hyperedge.

[0095] According to example embodiments, all cooperative pairs can be aggregated to form a cooperative behavior unit, i.e., a hyperedge, through a heuristic method or density clustering (such as DBSCAN). The hyperedge can be multiple.

[0096] Each hyperedge e k is accompanied by a hyperedge feature vector

[0097] The hyperedge feature vector can be a feature vector representing the properties of the hyperedge. The hyperedge feature vector can include task type (such as carrying, detection, and response, etc.), duration (the length of time the behavior lasts), success rate (task completion rate), cost (resource consumption evaluation), and scale (number of participating nodes).

[0098] Optionally, in step S130, the computer system can determine the hypergraph neural network model according to the following formula:

[0099]

[0100] wherein, is the hypergraph neural network model; V = {v1,...,v N} is a node set, corresponding to all agents; ε = {e1,...,e M} is a hyperedge set, e1 is a hyperedge, e M is a hyperedge, M is the number of hyperedges, each hyperedge connects two or more nodes, representing a collaborative behavior unit.

[0101] For example, as shown in Figure 7 , the framework structure of the hypergraph neural network model of Figure 7 is different from the classic GAT or GCN, one hyperedge of the hypergraph neural network model can connect multiple nodes, and the node and hyperedge types are explicitly distinguished.

[0102] The hyperedge can connect multiple nodes, for example, the hyperedge e1 can connect v2, v3 and v6, and the hyperedge e3 can connect v1, v2 and v5.

[0103] Optionally, referring to Figure 4 , step S1140 can include steps S141-S144.

[0104] In step S141, the computer system linearly maps the node features of the node model to obtain the features of the unified representation space of the node features.

[0105] According to an example embodiment, the computer system can project the initial feature vector f(v i ) of each node v i to the unified representation space via linear transformation:

[0106]

[0107] wherein, is a training parameter, d is the projection dimension, F is the original feature dimension, is a real number field. is the feature of the unified representation space of the node features of the current node.

[0108] In step S142, the computer system determines the first collaborative behavior representation vector of the hyperedge according to the features of the unified representation space of the node features and the hyperedge feature vector.

[0109] According to an example embodiment, the first collaborative behavior representation vector can be a feature vector representing a collaborative behavior pattern of a hyperedge in an intermediate layer of the hypergraph neural network model. Each hyperedge can correspond to a first collaborative behavior representation vector.

[0110] For example, for a hyperedge e k The set of connected nodes is where v k1 is a node connected to the hyperedge e k is a node connected to the hyperedge e km is a node connected to the hyperedge e k is a node connected to the hyperedge e

[0111] The computer system can determine the first collaborative behavior representation vector of the hyperedge according to the following formula:

[0112]

[0113] where a ik is the attention weight of node i on the hyperedge e k ; W1 is a projection parameter that can be optimized by gradient descent method; and σ is a nonlinear activation function. is the first collaborative behavior representation vector.

[0114] The computer system can calculate the attention weight a ik according to the following scoring function:

[0115]

[0116] a is a learnable weight vector; W2 is a trainable weight matrix. Softmax is a standardization function that converts multiple attention scores into a probability distribution. LeakyReLU is an activation function.

[0117] In step S143, the computer system determines the updated feature of the node feature according to the first collaborative behavior representation vector.

[0118] According to an example embodiment, the updated feature of the node feature can be a feature of the set of hyperedges connected to the node that is updated and propagated in reverse. The computer system can determine the updated feature according to the following formula:

[0119]

[0120] where u is the updated feature; β ki is an attention coefficient that can be obtained by a structure symmetry or independent learning mechanism; and W3 is a trainable weight matrix.

[0121] In step S144, the computer system outputs the collaborative behavior representation vector of the hyperedge according to the updated feature through the preset layer stack.

[0122] According to an example embodiment, the preset layer can be a hierarchical structure of a hypergraph neural network. The computer system can obtain the collaborative behavior representation vector of each hyperedge through the stack of L layers, which can also be represented as

[0123] Optionally, step S150 can be specifically: the computer system identifies the collaborative behavior pattern corresponding to the hyperedge according to the collaborative behavior representation vector through a clustering algorithm.

[0124] According to an example embodiment, the clustering algorithm can include a K-means clustering algorithm, a density clustering algorithm, and a spectral clustering algorithm.

[0125] For example, clustering all hyperedge embedding vectors Extract typical collaborative behavior patterns, which can be performed through a DBSCAN clustering algorithm. The DBSCAN clustering algorithm can automatically identify behavior clustering structures of arbitrary shapes, does not need to preset the number of clusters, and is robust to noisy hyperedges.

[0126] The clustering algorithm can output a set of collaborative behavior patterns Each class (C K ) represents a common collaborative behavior structure, i.e., a collaborative behavior pattern. The collaborative behavior pattern can also include sparse parallel collaboration (nodes work independently at a long distance), tightly coupled core-edge collaboration (there is a central agent), relay link collaboration (task chain transmission), and distributed complementary collaboration (heterogeneous nodes have clear division of labor).

[0127] The categories of collaborative behavior patterns can be associated with task labels to form an adaptive behavior dictionary.

[0128] Through the above embodiments, the technical scheme of the present application can improve the structural expression and adaptability of the hypergraph neural network model through the hypergraph attention mechanism.

[0129] The present inventors have also found that current HGNNs still have the problem of single performance index structure in multi-agent system performance evaluation. For example, existing evaluation frameworks focus on quantitative indicators such as task completion rate and resource consumption, and lack in-depth modeling of behavior organization structure and collaboration strength.

[0130] According to another aspect of the present application, the present application provides a multi-agent system performance evaluation method 2000. The performance evaluation method 2000 can be executed by a computer system.

[0131] According to another aspect of the present application, the present application provides a performance evaluation method 2000 of a multi-agent system. The performance evaluation method 2000 can be executed by a computer system.

[0132] Referring to Figure 5 The performance evaluation method 2000 can include steps S210-S260.

[0133] Steps S210-S250 are the same as steps S110-S150 in the identification method 1000, and thus will not be described again here.

[0134] In step S260, the computer system determines a performance evaluation parameter according to the cooperative behavior pattern.

[0135] According to an example embodiment, the performance evaluation parameter can be a parameter for evaluating the performance of the multi-agent system. The performance evaluation parameter can include a behavior pattern-based evaluation parameter, a node participation level indicator parameter, and a system overall evaluation indicator.

[0136] The behavior pattern-based evaluation parameter can include cooperative efficiency, cooperative robustness, node contribution, cooperative concentration, and behavior diversity.

[0137] The cooperative efficiency can be the average completion time per behavior, and the computer system can determine the cooperative efficiency according to the following formula:

[0138]

[0139] wherein, is the cooperative efficiency; C k is the cooperative behavior pattern in the kth round; L z is the completion time of the cooperative behavior z; and z is the cooperative behavior.

[0140] The cooperative robustness can be the behavior persistence within the cooperative behavior pattern, and the computer system can determine the cooperative robustness according to the following formula:

[0141]

[0142] wherein, is the cooperative robustness; is the average duration of the cooperative behavior.

[0143] The node contribution can be the number of nodes participating in the cooperative behavior pattern, and the computer system can determine the node contribution according to the following formula:

[0144] G i = | {e k ∈ E | v i ∈ e k} | ;

[0145] wherein G i is the node contribution degree.

[0146] The coordination concentration can be the degree of dependence of the central node, and the computer system can determine the coordination concentration according to the following formula:

[0147]

[0148] wherein E cent is the coordination concentration.

[0149] The behavior diversity can be the pattern structure difference, and the computer system can determine the behavior diversity according to the following formula:

[0150]

[0151] wherein D div is the behavior diversity; p k is the probability of occurrence of the cooperative behavior pattern in the kth round.

[0152] The behavior pattern evaluation parameter can evaluate the organization and coordination capability of the multi-agent system in the face of different task environments and resource distributions, and can be used to identify typical behavior structures, evaluate pattern advantages and disadvantages, and provide structured reference basis for task allocation strategies.

[0153] The node participation level index parameter can include node contribution degree, node centrality and node coordination role bias.

[0154] The node contribution degree can also be the number of cooperative behavior units participated by a single node, which is used to identify key nodes, core resources or overloaded nodes.

[0155] The node centrality can be the degree of a node being in a "hub" position in the cooperative network (such as hypercentrality), which is used to locate indispensable "connection bridge" type agents in cooperative transmission.

[0156] The node coordination role bias can be which type of task (transportation type / monitoring type, etc.) the node is more inclined to participate in, which is used to evaluate whether the roles in the system are reasonably allocated, and to avoid structural bias.

[0157] The node participation level index parameter can evaluate the role performance and collaboration characteristics of individual agents in the multi-agent system. By statistically analyzing the participation frequency, structural position and behavior bias of a single node, the node participation level index parameter can find the possible "key nodes", "edge nodes" or "role imbalance" phenomena in the multi-agent system, and then assist in system-level decisions such as agent scheduling, load balancing and communication optimization.

[0158] The system overall evaluation index can include a synergy coverage rate, a synergy concentration rate, and a synergy imbalance rate.

[0159] The synergy coverage rate can be a proportion of super edges identified as a synergy behavior by clustering to all behaviors, used to measure the synergy utilization efficiency in the system and identify inefficient independent behaviors.

[0160] The synergy concentration rate can be whether there is an excessive dependence on part of the high-frequency nodes, preventing the "overloaded center" phenomenon of the synergy structure, used to check the balance of task distribution and the fairness of the scheduling mechanism.

[0161] The synergy imbalance rate can be a variance or a range of node cooperation frequency, used to determine whether there is a risk of "part of the multi-agent system being overloaded and part of the multi-agent system being idle for a long time".

[0162] The system overall evaluation index can evaluate important synergy indicators of the multi-agent system at the overall level and the macro level, mainly used to reflect whether the synergy structure distribution is balanced, whether there is a problem of insufficient resource utilization or core node overload, and other systemic problems. The system overall evaluation index is also applicable to high-level decision analysis, system health monitoring, and scheduling strategy optimization scenarios.

[0163] Through the above embodiments, the technical scheme of the present application can determine the performance evaluation parameters through the synergy behavior mode. The technical scheme of the present application outputs a series of multi-dimensional and interpretable multi-agent system performance evaluation parameters through behavior clustering and index quantification calculation, providing direct support for multi-agent system optimization and task scheduling.

[0164] The technical scheme of the present application can be integrated with a real-time running platform, a system simulator, or a visual analysis platform to realize dynamic monitoring, behavior tracking, and abnormal identification of the multi-agent system running state, having strong engineering adaptability and deployment value.

[0165] Optionally, the computer system can also determine a comprehensive synergy evaluation parameter according to the performance evaluation parameter.

[0166] According to an example embodiment, the comprehensive synergy evaluation parameter can be an evaluation parameter for evaluating the overall performance of the multi-agent system.

[0167] The computer system can determine the comprehensive synergy evaluation parameter according to the behavior mode-based evaluation parameter and the corresponding weight, the node participation level index parameter and the corresponding weight, and the system overall evaluation index and the corresponding weight.

[0168] For example, the computer system can determine the comprehensive synergy evaluation parameter according to the following formula:

[0169]

[0170] wherein, for Standardized indicators; λ1 is Weighting coefficients; for Standardized index; λ2 is Weighting coefficients; For G i Standardized indicators; λ3 is The weighting coefficients; λ4 is the weighting coefficient of D. div The weighting coefficients.

[0171] Comprehensive collaborative evaluation parameters can be used to support operations such as scoring and ranking, model selection, and task decision-making in engineering deployment.

[0172] Through the above embodiments, the comprehensive collaborative evaluation parameters of this application provide a unified framework for multi-index fusion, enabling multi-agent systems to possess tunability, adaptability, and engineering deployment value. The comprehensive collaborative evaluation parameters can serve as an objective function for task scheduling optimization, a reward function for reinforcement learning during training, or a health scoring indicator for the operating system, supporting real-time feedback and closed-loop optimization.

[0173] According to another aspect of this application, this application also provides a recognition device 30 for the cooperative behavior pattern of a multi-agent system, which can execute the recognition method 1000 described above. See also Figure 6 The identification device 30 includes a processing module 31.

[0174] The processing module 31 determines the node model corresponding to the multi-agent system based on the received agent feature parameters of the multi-agent system.

[0175] The processing module 31 determines the hyperedge corresponding to the intelligent agent system based on the node model and preset collaborative behavior recognition conditions.

[0176] The processing module 31 determines the hypergraph neural network model corresponding to the agent system based on the node model and hyperedges.

[0177] Processing module 31 determines the cooperative behavior representation vector of the hyperedges based on the hypergraph neural network model through the hypergraph attention mechanism.

[0178] The processing module 31 identifies the cooperative behavior pattern corresponding to the hyperedge based on the cooperative behavior representation vector.

[0179] The multi-agent system, agent feature parameters, node model, preset cooperative behavior recognition conditions, hypergraph attention mechanism, cooperative behavior representation vector and cooperative behavior pattern have been described in the above recognition method 1000, so they will not be repeated here.

[0180] The application can effectively capture the complex collaboration relationship between multiple agents, such as many-to-one and many-to-many, by constructing a hypergraph neural network model containing multiple types of nodes and multiple granularity hyperedges, determine the hyperedge by the preset collaborative behavior recognition condition, and has good structure generalization ability through the hypergraph attention mechanism, which can adapt to various system types and task structures, and support unified modeling in heterogeneous node and heterogeneous task environment.

[0181] The application can break through the limitation of traditional graph structure supporting only binary relationship, and is suitable for team collaboration, resource scheduling and dynamic collaborative behavior recognition in complex tasks.

[0182] Optionally, the processing module 31 determines the state sequence data of the agent system according to the agent feature parameters received according to fixed time steps.

[0183] The processing module 31 determines the node feature of the agent according to the state sequence data.

[0184] The processing module 31 determines the node model according to the node feature.

[0185] The state sequence data and the node feature have been described in the above-mentioned identification method 1000, and will not be repeated here.

[0186] Optionally, the processing module 31 determines the node corresponding to the preset collaborative behavior recognition condition according to the node model and the preset collaborative behavior recognition condition.

[0187] The processing module 31 aggregates the nodes corresponding to the preset collaborative behavior recognition condition to determine the hyperedge.

[0188] The preset collaborative behavior recognition condition has been described in the above-mentioned identification method 1000, and will not be repeated here.

[0189] Optionally, the processing module 31 linearly maps the node feature of the node model to obtain the feature of the unified representation space of the node feature.

[0190] The processing module 31 determines the first collaborative behavior representation vector of the hyperedge according to the feature of the unified representation space of the node feature and the hyperedge feature vector.

[0191] The processing module 31 determines the updated feature of the node feature according to the first collaborative behavior representation vector.

[0192] The processing module 31 outputs the collaborative behavior representation vector of the hyperedge through the preset hierarchical stacking according to the updated feature.

[0193] The first collaborative behavior representation vector and the updated feature have been described in the above-mentioned identification method 1000, and will not be repeated here.

[0194] Optionally, the processing module 31 identifies the cooperative behavior pattern corresponding to the hyperedge according to the cooperative behavior representation vector by a clustering algorithm, where the clustering algorithm includes a K-means clustering algorithm, a density clustering algorithm, and a spectral clustering algorithm.

[0195] The clustering algorithm has been described in the above-mentioned identification method 1000, and thus will not be described again.

[0196] Optionally, the identification device 30 can perform the above-mentioned performance evaluation method 2000. The processing module 31 further determines the performance evaluation parameter according to the cooperative behavior pattern.

[0197] The performance evaluation parameter has been described in the above-mentioned performance evaluation method 2000, and thus will not be described again.

[0198] The technical solution of the present application outputs a series of multi-dimensional and interpretable performance evaluation parameters of the multi-agent system by behavior clustering and index quantization calculation, thereby providing direct support for multi-agent system optimization and task scheduling.

[0199] According to another aspect of the present application, the present application further provides a non-volatile computer readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement the identification method as described above.

[0200] According to another aspect of the present application, the present application further provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the identification method as described above.

[0201] According to another aspect of the present application, the present application further provides a computer program product, comprising: a computer program stored on a computer readable storage medium; the computer program comprising program instructions, when the program instructions are executed by a computer, the computer executes the identification method as described above.

[0202] According to another aspect of the present application, the present application further provides a non-volatile computer readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement the performance evaluation method as described above.

[0203] According to another aspect of the present application, the present application further provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the performance evaluation method as described above.

[0204] According to another aspect of the present application, the present application also provides a computer program product, comprising: a computer program stored on a computer readable storage medium; the computer program comprising program instructions which, when executed by a computer, cause the computer to perform the performance evaluation method as described above.

[0205] Finally, it should be noted that the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application is described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions of the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying a coordinated behavior pattern of a multi-agent system, characterized in that, The identification method comprises: According to the received multi-agent system agent feature parameters, the node model corresponding to the multi-agent system is determined; According to the node model and the preset cooperative behavior recognition condition, the super edge corresponding to the agent system is determined; According to the node model and the super edge, the supergraph neural network model corresponding to the agent system is determined; According to the supergraph neural network model, the cooperative behavior representation vector of the super edge is determined through the supergraph attention mechanism; According to the cooperative behavior representation vector, the cooperative behavior mode corresponding to the super edge is identified.

2. The identification method according to claim 1, characterized in that, According to the received multi-agent system agent feature parameters, the node model corresponding to the multi-agent system is determined, comprising: According to the received multi-agent system agent feature parameters, the node model corresponding to the multi-agent system is determined, comprising: According to the state sequence data, the node feature of the agent is determined; According to the node feature, the node model is determined.

3. The identification method according to claim 1, characterized in that, According to the node model and the preset cooperative behavior recognition condition, the node corresponding to the preset cooperative behavior recognition condition is determined; The nodes corresponding to the preset cooperative behavior recognition condition are aggregated to determine the super edge. According to the node model and the super edge, the supergraph neural network model corresponding to the agent system is determined, comprising:

4. The identification method according to claim 1, characterized in that, The node feature of the node model is linearly mapped to obtain the feature of the unified representation space of the node feature; According to the feature of the unified representation space of the node feature and the super edge feature vector, the first cooperative behavior representation vector of the super edge is determined; According to the first cooperative behavior representation vector, the updated feature of the node feature is determined; According to the updated feature, the cooperative behavior representation vector of the super edge is output through the preset hierarchical stacking. According to the cooperative behavior representation vector, the cooperative behavior mode corresponding to the super edge is identified, comprising:

5. The identification method according to claim 1, characterized in that, According to the cooperative behavior representation vector, the cooperative behavior mode corresponding to the super edge is identified through clustering algorithm, wherein the clustering algorithm comprises K-means clustering algorithm, density clustering algorithm and spectral clustering algorithm. The efficiency evaluation method comprises:

6. A method for performance evaluation of a multi-agent system, characterized in that, According to the received multi-agent system agent feature parameters, the node model corresponding to the multi-agent system is determined; According to the node model and the preset cooperative behavior recognition condition, the super edge corresponding to the agent system is determined; According to the node model and the super edge, the supergraph neural network model corresponding to the agent system is determined; According to the supergraph neural network model, the cooperative behavior representation vector of the super edge is determined through the supergraph attention mechanism; According to the cooperative behavior representation vector, the cooperative behavior mode corresponding to the super edge is identified; According to the cooperative behavior mode, the efficiency evaluation parameter is determined. The identification device comprises:

7. An apparatus for recognizing a cooperative behavior pattern of a multi-agent system, characterized by The processing module determines the node model corresponding to the multi-agent system according to the received multi-agent system agent feature parameters; The processing module determines the node model corresponding to the multi-agent system according to the received multi-agent system agent feature parameters; The processing module determines a hyper-edge corresponding to the intelligent agent system according to the node model and a preset cooperative behavior identification condition; The processing module determines a hypergraph neural network model corresponding to the intelligent agent system according to the node model and the hyper-edge; The processing module determines a cooperative behavior representation vector of the hyper-edge according to the hypergraph neural network model through a hypergraph attention mechanism; The processing module identifies a cooperative behavior mode corresponding to the hyper-edge according to the cooperative behavior representation vector.

8. The identification device of claim 7, wherein, The processing module further determines an efficiency evaluation parameter according to the cooperative behavior mode.

9. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the identification method of any one of claims 1-5.

10. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the efficiency evaluation method of claim 6.

Images