Multi-agent collaborative decision-making method and device thereof
By employing a multi-agent collaborative decision-making method, utilizing data perceptrons, temporal incremental consistency integration, and multi-level weighted feature pyramid algorithms, the problems of high cost of heterogeneous adaptation, communication latency, and representation differences among multi-agents are solved. This achieves efficient and stable collaborative perception and decision-making, improving the collaborative performance of multi-agents in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YUNNAN NORMAL UNIV
- Filing Date
- 2026-04-14
- Publication Date
- 2026-06-30
AI Technical Summary
Existing multi-agent cooperative methods suffer from high adaptation costs under heterogeneous hardware and model conditions, spatial and semantic misalignment caused by communication latency, sensitivity of unified representation to differences, and insufficient utilization of the complementary information of multiple agents, resulting in poor cooperative decision-making performance.
A multi-agent collaborative decision-making method is adopted, which generates multimodal representation vectors through a data perceptron, and uses a time-series incremental consistency ensemble algorithm and a multi-level weighted feature pyramid algorithm for representation fusion and compression, thereby achieving efficient information sharing and collaborative decision-making among agents.
It improves the adaptability and universality of multi-agent collaborative decision-making under heterogeneous hardware and model conditions, reduces communication overhead, enhances the ability to deeply understand complex environments, and improves the accuracy, stability and robustness of collaborative perception and decision-making.
Smart Images
Figure CN122021785B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a multi-agent collaborative decision-making method and apparatus, belonging to the field of multi-agent collaborative decision-making technology. Background Technology
[0002] Currently, multi-agent collaborative systems, due to their distributed, autonomous, and robust characteristics, have become an important technological form supporting cloud-edge-device collaborative applications and have been widely used in complex scenarios such as intelligent transportation, smart cities, industrial internet, intelligent manufacturing, and smart healthcare. In these applications, the system typically operates collaboratively by multiple distributed agents. Each agent needs to complete information perception, data interaction, and joint decision-making in a dynamic and open environment. For example, in typical scenarios such as autonomous driving, multi-agent collaborative decision-making can achieve joint perception of blind spot targets and complex environments, improving system safety and reliability.
[0003] In related technologies, current multi-agent cooperative methods mainly rely on adapting the features or representations of each agent to achieve the unification of perceptual information through single-step or multi-step transformation. Although these methods have made progress in feature transformation and representation alignment, as the system scales up, the heterogeneity of agents in terms of sensing devices, model structures, and data distribution becomes increasingly prominent, resulting in the following shortcomings in practical deployment:
[0004] (1) The training cost of heterogeneous adaptation for multi-agent agents is high. Existing multi-agent collaborative methods usually require retraining adaptation modules for different agents or sharing part of the network structure. In safety-critical applications such as autonomous driving, the perception model is highly coupled with downstream tasks and is difficult to directly replace or retrain. When each pair of heterogeneous agents needs to be adapted and trained separately, the training complexity and deployment cost will be significantly increased.
[0005] (2) Communication delays among multiple agents lead to spatial and semantic misalignment. Existing methods mostly rely on cloud or centralized nodes for information exchange, which requires data uploading, centralized processing and result feedback. Communication and computation delays are unavoidable. Different agents may have asynchronous observations of the same target, which in turn leads to spatial position deviations and semantic inconsistencies.
[0006] (3) Differences in the unified representation of multiple agents limit the stability of collaborative decision-making. Existing methods reduce training costs by aligning the representations of different agents to a unified space, but this unified representation is usually based on a specific agent. When other agents differ significantly from it in modality or feature distribution, the representation alignment effect is difficult to guarantee, which can easily lead to semantic information loss and affect the collaborative decision-making effect.
[0007] (4) Insufficient utilization of information complementarity among multiple agents. Existing methods focus on single mapping or transformation at the feature level, failing to fully explore the complementary advantages of multiple modalities and agents in terms of perspective, accuracy, and temporal information, thus limiting further improvement in collaborative perception and decision-making performance.
[0008] To address the aforementioned issues, this invention proposes a multi-agent collaborative decision-making method and apparatus, which enables efficient collaboration under heterogeneous hardware and model conditions, alleviates semantic loss caused by communication latency and unified representation, and fully utilizes the complementarity of multi-agent and multi-modal information, thereby improving the real-time performance, reliability, and stability of collaborative decision-making. It is suitable for multi-agent collaborative perception and decision-making applications in complex dynamic environments. Summary of the Invention
[0009] The technical problem to be solved by the present invention is to provide a multi-agent collaborative decision-making method and apparatus to overcome the problems of high adaptation cost under heterogeneous hardware and model conditions, spatial and semantic misalignment caused by communication delay, sensitivity of unified representation to differences, and insufficient utilization of the complementarity of multi-agent information in existing multi-agent collaborative methods. In this way, efficient, stable and real-time collaborative decision-making can be achieved in complex dynamic environments, and the reliability and robustness of multi-agent collaborative perception and decision-making can be improved.
[0010] The technical solution of this invention is: a multi-agent cooperative decision-making method, the specific steps of which are as follows:
[0011] S1: Each agent performs joint perception and feature encoding on the currently acquired multimodal input data through the data perceptron to obtain a multimodal representation vector;
[0012] S2: Each agent introduces a time-series incremental consistency integration algorithm through a data updater to fuse and update the multimodal representation vector and the historical representation set, thereby obtaining a fused representation with enhanced consistency.
[0013] S3: Each agent compresses the consistency-enhanced fusion representation through a data transmitter to obtain a compressed message, and sends the compressed message to neighboring agents;
[0014] S4: Each agent receives compressed messages from neighboring agents through a data receiver, and performs decoding and feature reconstruction processing on the compressed messages to obtain the representation information of neighboring agents. Each agent aligns and fuses the representation information of neighboring agents with its own fused representation with enhanced consistency to generate a joint representation for collaborative decision-making.
[0015] S5: Each agent's data updater uses a multi-level weighted feature pyramid algorithm to update the joint representation in layers, generating new historical representation data to support the next round of multi-agent collaborative decision-making.
[0016] Optionally, S1 specifically includes:
[0017] S1.1: Each agent participating in collaborative decision-making receives multimodal input data from multiple data sources, including but not limited to visual modality data. Voice modal data and text modal data This forms the current multimodal input set for the corresponding intelligent agent. Multiple agents participating in collaborative decision-making constitute an agent ensemble. ,in, Indicates the first One intelligent agent;
[0018] S1.2: Each agent inputs the received multimodal input data into its corresponding data perceptron for joint perception and feature encoding. The data perceptron is based on a multimodal coding model. Feature extraction and fusion processing are performed on data from different modalities to generate a unified multimodal representation vector. ,in, Represents the dimension of the feature space.
[0019] Optionally, S2 specifically includes:
[0020] S2.1: Each intelligent agent By introducing a time-series incremental consistency ensemble algorithm through a data updater, the current multimodal representation vector is updated. Its historical representation set The fusion process is performed, in which... , Indicating collaborative decision-making in history The multimodal representation vectors obtained in each round are used by the temporal incremental consistency ensemble algorithm to evaluate the consistency relationship between different temporal representations, and to perform weighted fusion of the current representation and historical representations to generate a fused representation with enhanced consistency. ,in, This represents a time-incremental consistency integration function;
[0021] S2.2: Each agent uses a data updater to update the fused representation. Perform state update processing to fuse representations As new historical representations, they are written into the set of historical representations to form an updated set of historical representations. .
[0022] Optionally, S3 specifically includes:
[0023] S3.1: Each intelligent agent The fusion representation obtained by fusing data updaters has enhanced consistency. The data is input to the corresponding data transmitter, which compresses the fused representation using an encoding compression function to generate a compressed message for inter-agent communication. ,in, Indicates the first The compression mapping function corresponding to the data transmitter of each intelligent agent. This represents the feature dimension of the compressed message, used to reduce communication bandwidth consumption and improve information transmission efficiency. Represents the dimension of the feature space;
[0024] S3.2: Each intelligent agent The generated compressed message Send to the set of neighboring agents with whom it has a communication connection Neighboring agents receive the compressed message for alignment, fusion, and collaborative decision-making.
[0025] Optionally, S4 specifically includes:
[0026] S4.1: Each intelligent agent Through the corresponding data receiver, from neighboring agent set The middle receives compressed messages sent from neighboring agents. ,in, Indicates the first The compression mapping function corresponding to the data transmitter of each intelligent agent, and the data receiver for compressed messages. Decoding and feature reconstruction are performed to obtain the representation information of neighboring agents. ,in, Indicates the first Decompression mapping functions corresponding to the data receivers of each intelligent agent;
[0027] S4.2: Each agent will reconstruct the representation information of its neighboring agents. Enhanced consistency of fusion representation after fusion with itself Alignment and fusion processes are performed to generate joint representations for collaborative decision-making. ,in, This represents the cross-agent representation alignment and fusion function, whereby each agent aligns and fuses the representations based on the joint representations. Generate collaborative decision-making results, thereby enabling information sharing and collaborative decision-making among multiple agents.
[0028] Optionally, S5 specifically includes:
[0029] S5.1: Each intelligent agent The data updater for the joint representation generated by collaborative decision-making A multi-level weighted feature pyramid algorithm is introduced to perform hierarchical modeling and updating of the joint representation. The multi-level weighted feature pyramid algorithm maps the joint representation to multiple feature levels. Different feature levels correspond to feature representations of different scales or different degrees of semantic abstraction, among which, Representing joint representation The Each feature layer is divided into several feature layers, and corresponding weight coefficients are assigned to each feature layer. Weighted updates are performed on features at each level, where, Representing joint representation The The weight coefficients corresponding to each feature layer;
[0030] S5.2: Each agent aggregates the hierarchical features updated by the multi-level weighted feature pyramid algorithm to generate new historical representation data. The new historical representation data is then written into the corresponding historical representation set. This serves as historical input data for the next round of multi-agent collaborative decision-making, thereby enabling the continuous evolution and updating of historical knowledge during the multi-agent collaborative decision-making process. A fusion characterization for enhanced consistency.
[0031] Furthermore, this application also proposes a multi-agent cooperative decision-making device, which includes:
[0032] The data sensor is responsible for acquiring and modeling multimodal data from various agents, and for realizing joint multimodal perception and feature encoding.
[0033] The data updater is responsible for updating the historical representation data in two rounds: the first round uses a time-series incremental consistency fusion algorithm to fuse the current multimodal encoding with the historical multimodal representation, so as to achieve consistency maintenance across time series representations and update the historical representation data; the second round uses a multi-level weighted feature pyramid algorithm to perform hierarchical modeling and weighted fusion of the updated historical representations, generate new historical representation data, and use it for subsequent collaborative decision-making cyclic updates.
[0034] The data transmitter is responsible for compressing the fused representation and generating messages, and for enabling the transmission of compressed messages between agents and neighborhood communication.
[0035] The data receiver is responsible for receiving and reconstructing compressed messages from neighboring agents, and for achieving cross-agent representation alignment and fusion, as well as collaborative decision generation.
[0036] Furthermore, this application also proposes a multi-agent cooperative decision-making device, including a memory, a processor, and a multi-agent cooperative decision-making method program stored in the memory and executable on the processor. The processor executes the multi-agent cooperative decision-making method program to implement the steps of the multi-agent cooperative decision-making method as described above.
[0037] Furthermore, this application also proposes a computer-readable storage medium storing a multi-agent cooperative decision-making method program, which, when executed by a processor, implements the steps of the multi-agent cooperative decision-making method as described above.
[0038] The beneficial effects of this invention are:
[0039] (1) The present invention can support multiple heterogeneous agents to achieve efficient collaborative decision-making under the condition that there are differences in hardware configuration, perception modality and model structure. Compared with the existing methods, it has stronger adaptability and universality and is suitable for complex and diverse multi-agent application scenarios.
[0040] (2) This invention reduces the dependence on centralized cloud models and computing resources by directly performing feature interaction and difference information transmission between intelligent agents, effectively reducing the scale of communication data and interaction overhead, thereby improving the communication efficiency and real-time performance in the multi-agent collaboration process.
[0041] (3) By introducing the temporal incremental consistent integration algorithm and the multi-level weighted feature pyramid algorithm update mechanism, this invention fully explores the key semantic information of different intelligent agents in multi-view and multi-modal perception, effectively alleviates the semantic loss caused by unified representation and temporal asynchrony, and improves the ability to deeply understand targets and scenes in complex environments.
[0042] (4) This invention makes full use of the complementary advantages of multiple agents in terms of perception perspective, modal information and time sequence features, and improves the accuracy, stability and robustness of collaborative perception and decision-making results through collaborative fusion and knowledge transfer mechanisms. Attached Figure Description
[0043] Figure 1 This is a flowchart of the steps of the present invention;
[0044] Figure 2 This is a schematic diagram of a multi-agent collaborative decision-making system according to Embodiment 1 of the present invention;
[0045] Figure 3 This is a schematic diagram of the specific process of one-to-many agent collaborative decision-making in Embodiment 1 of the present invention;
[0046] Figure 4 This is a schematic diagram of the time-series incremental consistency integration algorithm model in the data updater of Embodiment 1 of the present invention;
[0047] Figure 5 This is a schematic diagram of the compression process of the data transmitter and the decompression process of the data receiver in a single intelligent agent according to Embodiment 1 of the present invention. Figure 5 In the middle (a), it represents the data transmitter compression process. Figure 5 (b) represents the data receiver decompression process;
[0048] Figure 6 This is a schematic diagram of the multi-level weighted feature pyramid algorithm model in the data updater of Embodiment 1 of the present invention;
[0049] Figure 7 This is a schematic diagram of the hardware structure involved in Embodiment 1 of the present invention. Detailed Implementation
[0050] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0051] Example 1: As Figure 1 The diagram shows an overall schematic of a multi-agent collaborative decision-making method. The implementation process mainly includes four core components: S1: data sensor; S2 / S5: data updater; S3: data transmitter; S4: data receiver. Figure 3 The diagram illustrates the specific process of collaborative decision-making between a pair of intelligent agents. Each step is described in detail below:
[0052] S1: Specific implementation steps of the data sensor:
[0053] Each agent performs joint sensing and feature encoding on the currently acquired multimodal input data through a data perceptron to obtain a multimodal representation vector, specifically:
[0054] S1.1: Multimodal Data Acquisition and Modeling. Each agent participating in collaborative decision-making receives multimodal input data from multiple data sources. This multimodal input data includes, but is not limited to, visual modal data, speech modal data, and text modal data, forming the current multimodal input set for the corresponding agent. Different agents can simultaneously receive complete or partial modal data to support subsequent collaborative processing and dynamic division of labor. Specifically:
[0055] Define multiple agents participating in collaborative decision-making as an agent set. , among which, the A smart agent Receive multimodal input data from multiple data sources, wherein the multimodal input data is represented as ,in, , and These represent visual modal data, speech modal data, and text modal data, respectively.
[0056] S1.2: Multimodal Joint Perception and Feature Encoding. Each agent inputs the received multimodal input data into its corresponding data perceptron for joint perception and feature encoding. The data perceptron, based on a multimodal encoding model, extracts and fuses features from different modalities to generate a unified multimodal representation vector. This vector represents the semantic information of the current input data and provides a basic encoded representation for subsequent information interaction and collaborative decision-making among agents. Specifically:
[0057] The multimodal input data received by each agent The data is input to the corresponding data sensor, which includes a multimodal encoder. The encoding process is represented as ,in, The multimodal encoder, representing the feature space dimension, is used to extract and fuse features from data from different modalities to obtain a multimodal representation vector. The multimodal encoder includes multiple modal feature extraction sub-networks and a feature fusion module. Each modal feature extraction sub-network learns feature representations for its corresponding modal data to obtain feature vectors for each modality. The feature fusion module aligns, fuses, and maps features from different modalities to generate a unified multimodal representation vector.
[0058] S2: Specific implementation steps of the data updater:
[0059] Each agent introduces a time-incremental consistency ensemble algorithm through a data updater to fuse and update the multimodal representation vector with the historical representation set, resulting in a fused representation with enhanced consistency, specifically:
[0060] S2.1: Current-historical representation fusion based on temporal increment. (Each agent...) By introducing a time-series incremental consistency ensemble algorithm through a data updater, the current multimodal representation vector is updated. Its historical representation set The fusion process is performed, in which... , Indicating collaborative decision-making in history The multimodal representation vectors obtained in each round; the temporal incremental consistency ensemble algorithm evaluates the consistency relationship between different temporal representations, performs weighted fusion of the current representation and historical representations, and generates a fused representation with enhanced consistency. ,in, This represents a time-incremental consistent integration function. Specifically:
[0061] like Figure 4 The diagram shows a schematic of the time-series incremental consistency ensemble algorithm model in the data updater. For the first... An intelligent agent, at any time The system has stored the previous moment. Historical integration characteristics And obtain the representation of the current moment through the current perception and encoding process. The above and The inputs are shared with the temporal fusion module, which generates and updates the temporal fusion representation. The temporal fusion module's implementation includes: calculating the fusion weights at the current time step using a temporal weight generation network. The fusion weights are used to characterize the relative importance of current and historical information in the temporal dimension. Under the constraints of these fusion weights, a weighted integration method is used to fuse the current and historical representations, resulting in an updated, more consistent fused representation. The time-incremental consistency integration function is represented as follows:
[0062] (1)
[0063] in, This indicates element-weighted operations. For the updated consistency-enhanced fusion representation, On the one hand, it serves as the output representation of the current moment for subsequent task decisions; on the other hand, it serves as the output representation for the next moment. Historical representation input.
[0064] S2.2: Maintaining consistency of multimodal representations and updating historical representation data. Each agent uses a data updater to update the fused representation. Perform state update processing, As new historical representations, they are written into the set of historical representations to form an updated set of historical representations. By continuously updating the historical representation set, the multimodal representations maintain semantic consistency and stability during multiple rounds of collaborative decision-making, thereby effectively avoiding representation drift and knowledge forgetting caused by changes in data distribution or modal differences, and providing a reliable feature foundation for subsequent inter-agent communication and collaborative decision-making.
[0065] S3: Specific implementation steps of the data transmitter:
[0066] Each agent compresses the consistency-enhanced fused representation through a data transmitter to obtain a compressed message, and then sends the compressed message to its neighboring agents, specifically as follows:
[0067] S3.1: Compression of fused representations and message generation. (Each agent...) The fusion representation obtained by fusing data updaters has enhanced consistency. The data is input to the corresponding data transmitter, which compresses the fused representation using an encoding compression function to generate a compressed message for inter-agent communication. ,in, Indicates the first The compression mapping function corresponding to the data transmitter of each intelligent agent. The feature dimensions representing the compressed message are used to reduce communication bandwidth usage and improve information transmission efficiency. Specifically:
[0068] like Figure 5 The diagram shown illustrates the compression process of a data transmitter and the decompression process of a data receiver in a single intelligent agent. Figure 5 In section (a), the compression process of the data transmitter is represented, and the compression mapping function is described. Implemented by a data transmitter, whose input is the consistency-enhanced fused representation obtained by the data updater. The output is a compressed message used for inter-agent communication. .
[0069] Specifically, the compression mapping function first applies the consistency-enhanced fusion representation. The input is fed into a recombination module for recombination processing. This module sequentially includes interpolation, channel consistency, and convolutional mapping operations. The interpolation operation resamples and aligns the feature resolution. Subsequently, the channel consistency operation normalizes the number of feature channels to ensure structural consistency in subsequent feature mapping processes. Finally, the convolutional mapping operation performs local aggregation and dimensionality compression on the features, resulting in a compact intermediate feature representation.
[0070] Further, the intermediate feature representation is input to an aligner module, which sequentially includes a fusion directional attention operation, a layer normalization operation, and a multilayer perceptron operation. The fusion directional attention operation is based on the query vector. Key vector AND value vector We employ weighted modeling of features to highlight semantic information that is more important for the current collaborative decision-making of the agent; layer normalization is used to stabilize the feature distribution; and multilayer perceptron operations are used to enhance the nonlinear expressive power of features.
[0071] After the above processing, the compressed message is obtained. Its feature dimension is a preset low-dimensional space, which is used to reduce the communication bandwidth occupation and improve the information transmission efficiency between intelligent agents while ensuring the ability to express key information.
[0072] S3.2: Compressed message transmission between agents and neighborhood communication. Each agent... The generated compressed message Send to the set of neighboring agents with whom it has a communication connection Neighboring agents receive the compressed messages for subsequent alignment, fusion, and collaborative decision-making. By employing a neighborhood communication mechanism using compressed messages in a multi-agent system, efficient information sharing is achieved, communication overhead is reduced, and key representational information required for collaborative decision-making is maintained.
[0073] S4: Specific implementation steps for the data receiver:
[0074] Each agent receives compressed messages from neighboring agents via a data receiver, and decodes and reconstructs the compressed messages to obtain the representation information of the neighboring agents. Each agent then aligns and fuses the representation information of the neighboring agents with its own fused representation, which enhances consistency, to generate a joint representation for collaborative decision-making. Specifically:
[0075] S4.1: Each intelligent agent Through the corresponding data receiver, from neighboring agent set The middle receives compressed messages sent from neighboring agents. ,in, Indicates the first The compression mapping function corresponding to the data transmitter of each intelligent agent, and the data receiver for compressed messages. Decoding and feature reconstruction are performed to obtain the representation information of neighboring agents. ,in, Indicates the first The decompression mapping function corresponding to the data receiver of each agent. Specifically:
[0076] like Figure 5 As shown in (b), the data receiver decompression process is illustrated. The decompression mapping function is implemented by the data receiver, and its input is the compressed message received from the neighboring agent. The output is the reconstructed representation information of neighboring agents. .
[0077] Specifically, the data receiver first receives compressed messages sent from neighboring agents. The compressed message is then input into a converter module for decoding. This converter module sequentially includes a fusion directional attention operation, a layer normalization operation, and a multilayer perceptron operation. Specifically, the fusion directional attention operation generates a corresponding key vector from the compressed message. AND value vector and the query vector generated by the receiving agent's own representation. Interactive modeling is performed to achieve cross-agent semantic alignment and information selection; layer normalization is used to stabilize the feature distribution during the decoding process; and multilayer perceptron operations are used to enhance the nonlinear expressive power of features.
[0078] The features processed by the converter are input to the reconstructor module, which sequentially includes a convolutional mapping operation, an interpolation operation, and a channel consistency operation. Specifically, the convolutional mapping operation is used to reconstruct local features from the decoded features; the interpolation operation is used to restore the spatial resolution of the features; and the channel consistency operation is used to adjust the number of feature channels to match the current feature space of the receiving agent.
[0079] Through the above decoding and feature reconstruction processes, the representation information of neighboring agents is obtained. This is used for subsequent cross-agent feature fusion and collaborative decision-making.
[0080] S4.2: Cross-agent representation alignment and fusion, and collaborative decision generation. Each agent reconstructs the representation information of its neighboring agents. Enhanced consistency of fusion representation after fusion with itself Alignment and fusion processes are performed to generate joint representations for collaborative decision-making. ,in, This represents the cross-agent representation alignment and fusion function, whereby each agent aligns and fuses the representations based on the joint representations. Generate collaborative decision-making results, thereby achieving information sharing and collaborative decision-making among multiple agents. Specifically:
[0081] In the implementation process, the cross-agent representation alignment and fusion function first performs scale and resolution alignment on the representation information of neighboring agents through interpolation operations, so that it is consistent with the representation of the current agent in the feature dimension; then, the representation information of neighboring agents after interpolation alignment is spliced with the fused representation of the current agent to form a joint feature representation in the feature dimension, which is used as a joint representation output for subsequent collaborative decision-making.
[0082] S5: Specific implementation steps of the data updater:
[0083] Each agent's data updater employs a multi-level weighted feature pyramid algorithm to hierarchically update the joint representation, generating new historical representation data to support the next round of multi-agent collaborative decision-making. Specifically:
[0084] S5.1: Hierarchical update of fused features based on a multi-level weighted feature pyramid. (Each agent...) The data updater for the joint representation generated by collaborative decision-making A multi-level weighted feature pyramid algorithm is introduced to perform hierarchical modeling and updating of the joint representation. This algorithm maps the joint representation to multiple feature levels. Different feature levels correspond to feature representations of different scales or different degrees of semantic abstraction, among which, Representing joint representation The Each feature layer This indicates the number of layers in the joint representation; and is achieved by assigning corresponding weight coefficients to each feature layer. Weighted updates are applied to features at each level to enhance the expressive power of multi-source features at different scales. Specifically,
[0085] like Figure 6 The diagram shows a schematic of the multi-level weighted feature pyramid algorithm model in the data updater. Let the first level be... The self-fusion representation of an agent at the current moment is as follows: From neighboring agents The received reconstructed representation information is The data updater first updates... and Spatial or dimensional alignment is performed, and resolution or feature dimensions are unified through interpolation. Then, a concatenation operation is used to fuse the results, yielding the initial joint representation for collaborative decision-making. ,Right now This represents the joint representation of collaborative decision-making at layer 1, corresponding to "Joint Representation Layer 1 of Collaborative Decision-Making" in the diagram. Next, the data updater... Using this as input, a multi-level collaborative decision-making joint representation is constructed step by step, forming a multi-level feature pyramid structure. , among which, the Layer joint representation It is obtained from the features of the previous layer through feature transformation or scale mapping, and is used to characterize collaborative information under different semantic abstraction levels or different perceptual scales.
[0086] For any i Hierarchical collaborative decision joint representation The data updater generates its own and neighboring agents' intermediate feature representations at that level. and Subsequently, according to the first The semantic importance of layer features is assigned a corresponding weight coefficient to each feature layer. This is used to adjust the contribution of features at different levels to historical updates. Under weight modulation, hierarchical difference compensation features are generated separately. and The difference compensation feature is used to characterize the complementary information between itself and neighboring agents in collaborative perception and decision-making at the current level.
[0087] Difference compensation features generated at each level and By splicing and aggregating, we can obtain the aggregated features at each level. Then aggregate features at each level Cross-level aggregation processing is performed. Low-level features are upsampled to higher-level spatial scales and then concatenated with corresponding level features to generate new historical representation data. .
[0088] S5.2: Historical Representation Data Generation and Collaborative Decision-Making Iterative Update. Each agent aggregates the hierarchical features updated by the multi-level weighted feature pyramid algorithm to generate new historical representation data. The new historical representation data is then written into the corresponding historical representation set. This serves as historical input data for the next round of multi-agent collaborative decision-making, thereby enabling the continuous evolution and stable updating of historical knowledge during the multi-agent collaborative decision-making process.
[0089] This application also proposes a multi-agent cooperative decision-making device, such as... Figure 7 The diagram shown is a schematic diagram of the hardware operating environment of a multi-agent collaborative decision-making method program running device involved in the embodiments of this application.
[0090] like Figure 7 As shown, a multi-agent collaborative decision-making device includes: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between various component modules. The user interface 1003 may include a display screen, an input unit such as a keyboard, etc. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.
[0091] Those skilled in the art will understand that Figure 7 The structure shown does not constitute a limitation on a multi-agent cooperative decision-making process and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.
[0092] Optionally, the memory 1005 is connected to the processor 1001. The processor 1001 can be used to control the operation of the memory 1005 and can also read data in the memory 1005 to implement a multi-agent cooperative decision-making method.
[0093] Optionally, such as Figure 7 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and a multi-agent collaborative decision-making method program.
[0094] Optionally, in Figure 7 In the multi-agent collaborative decision-making device shown, the network interface 1004 is mainly used for data communication with other devices; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the multi-agent collaborative decision-making device of this application can be set in a multi-agent collaborative decision-making device device.
[0095] like Figure 7 As shown, the multi-agent cooperative decision-making device calls a multi-agent cooperative decision-making method program stored in the memory 1005 through the processor 1001, and executes it as follows: Figure 2 As shown, the four core components of a multi-agent collaborative decision-making method provided in Embodiment 1 of this application are: S1: data sensor; S2 / S5: data updater; S3: data transmitter; S4: data receiver.
[0096] Furthermore, this application also proposes a computer-readable storage medium storing a multi-agent cooperative decision-making method program, which, when executed by a processor, implements the steps of the multi-agent cooperative decision-making method as described above.
[0097] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0098] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. This application can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.
[0099] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0100] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A multi-agent collaborative decision-making method, characterized in that, The method includes the following steps: S1: Each agent performs joint perception and feature encoding on the currently acquired multimodal input data through the data perceptron to obtain a multimodal representation vector; S2: Each agent introduces a time-series incremental consistency integration algorithm through a data updater to fuse and update the multimodal representation vector and the historical representation set, thereby obtaining a fused representation with enhanced consistency. S3: Each agent compresses the consistency-enhanced fusion representation through a data transmitter to obtain a compressed message, and sends the compressed message to neighboring agents; S4: Each agent receives compressed messages from neighboring agents through a data receiver, and performs decoding and feature reconstruction processing on the compressed messages to obtain the representation information of neighboring agents. Each agent aligns and fuses the representation information of neighboring agents with its own fused representation with enhanced consistency to generate a joint representation for collaborative decision-making. S5: Each agent's data updater uses a multi-level weighted feature pyramid algorithm to update the joint representation in layers, generating new historical representation data to support the next round of multi-agent collaborative decision-making process; Specifically, S1 is: S1.1: Each agent participating in collaborative decision-making receives multimodal input data from multiple data sources, including visual modal data. Voice modal data and text modal data This forms the current multimodal input set for the corresponding intelligent agent. X i Multiple agents participating in collaborative decision-making constitute an agent ensemble. ,in, Indicates the first One intelligent agent; S1.2: Each agent inputs the received multimodal input data into its corresponding data perceptron for joint perception and feature encoding. The data perceptron is based on a multimodal coding model. Feature extraction and fusion processing are performed on data from different modalities to generate a unified multimodal representation vector. ,in, Represents the dimension of the feature space; Specifically, S2 is: S2.1: Each intelligent agent By introducing a time-series incremental consistency ensemble algorithm through a data updater, the current multimodal representation vector is updated. Its historical representation set The fusion process is performed, in which... , Indicating collaborative decision-making in history The multimodal representation vectors obtained in each round are used by the temporal incremental consistency ensemble algorithm to evaluate the consistency relationship between different temporal representations, and to perform weighted fusion of the current representation and historical representations to generate a fused representation with enhanced consistency. ,in, This represents a time-incremental consistency integration function; S2.2: Each agent uses a data updater to update the fused representation. Perform state update processing to fuse representations As new historical representations, they are written into the set of historical representations to form an updated set of historical representations. .
2. The multi-agent cooperative decision-making method according to claim 1, characterized in that, Specifically, S3 is: S3.1: Each intelligent agent The fusion representation obtained by fusing data updaters has enhanced consistency. The data is input to the corresponding data transmitter, which compresses the fused representation using an encoding compression function to generate a compressed message for inter-agent communication. ,in, Indicates the first The compression mapping function corresponding to the data transmitter of each intelligent agent. This represents the feature dimension of the compressed message, used to reduce communication bandwidth consumption and improve information transmission efficiency. Represents the dimension of the feature space; S3.2: Each intelligent agent The generated compressed message Send to the set of neighboring agents with whom it has a communication connection Neighboring agents receive the compressed message for alignment, fusion, and collaborative decision-making.
3. The multi-agent cooperative decision-making method according to claim 1, characterized in that, Specifically, S4 is: S4.1: Each intelligent agent Through the corresponding data receiver, from neighboring agent set The middle receives compressed messages sent from neighboring agents. ,in, Indicates the first The compression mapping function corresponding to the data transmitter of each intelligent agent, and the data receiver for compressed messages. Decoding and feature reconstruction are performed to obtain the representation information of neighboring agents. ,in, Indicates the first Decompression mapping functions corresponding to the data receivers of each intelligent agent; S4.2: Each agent will reconstruct the representation information of its neighboring agents. Enhanced consistency of fusion representation after fusion with itself Alignment and fusion processes are performed to generate joint representations for collaborative decision-making. ,in, This represents the cross-agent representation alignment and fusion function, whereby each agent aligns and fuses the representations based on the joint representations. Generate collaborative decision-making results, thereby enabling information sharing and collaborative decision-making among multiple agents.
4. The multi-agent cooperative decision-making method according to claim 1, characterized in that, Specifically, S5 is: S5.1: Each intelligent agent The data updater for the joint representation generated by collaborative decision-making A multi-level weighted feature pyramid algorithm is introduced to perform hierarchical modeling and updating of the joint representation. The multi-level weighted feature pyramid algorithm maps the joint representation to multiple feature levels. Different feature levels correspond to feature representations of different scales or different degrees of semantic abstraction, among which, Representing joint representation The Each feature layer is divided into several feature layers, and corresponding weight coefficients are assigned to each feature layer. Weighted updates are performed on features at each level, where, Representing joint representation The The weight coefficients corresponding to each feature layer; S5.2: Each agent aggregates the hierarchical features updated by the multi-level weighted feature pyramid algorithm to generate new historical representation data. The new historical representation data is then written into the corresponding historical representation set. This serves as historical input data for the next round of multi-agent collaborative decision-making, thereby enabling the continuous evolution and updating of historical knowledge during the multi-agent collaborative decision-making process. A fusion characterization for enhanced consistency.
5. An apparatus for implementing a multi-agent cooperative decision-making method as described in claim 1, characterized in that, The device includes: The data sensor is responsible for acquiring and modeling multimodal data from various agents, and for realizing joint multimodal perception and feature encoding. The data updater is responsible for updating the historical representation data in two rounds: the first round uses a time-series incremental consistency integration algorithm to fuse the current multimodal encoding with the historical multimodal representation, so as to achieve consistency maintenance across time series representations and update the historical representation data; the second round uses a multi-level weighted feature pyramid algorithm to perform hierarchical modeling and weighted fusion of the updated historical representations, generate new historical representation data, and use it for subsequent collaborative decision-making and iterative updates. The data transmitter is responsible for compressing the fused representation and generating messages, and for enabling the transmission of compressed messages between agents and neighborhood communication. The data receiver is responsible for receiving and reconstructing compressed messages from neighboring agents, and for achieving cross-agent representation alignment and fusion, as well as collaborative decision generation.
6. A multi-agent collaborative decision-making device, characterized in that, The system includes a memory, a processor, and a multi-agent cooperative decision-making method program stored in the memory and executable on the processor. The processor executes the multi-agent cooperative decision-making method program to implement the steps of the multi-agent cooperative decision-making method as described in any one of claims 1 to 4.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a multi-agent cooperative decision-making method program, which, when executed by a processor, implements the steps of the multi-agent cooperative decision-making method as described in any one of claims 1 to 4.