CN-122021785-A - Multi-agent collaborative decision-making method and device thereof

CN122021785ACN 122021785 ACN122021785 ACN 122021785ACN-122021785-A

Abstract

The invention relates to a multi-agent collaborative decision-making method and a device thereof, belonging to the technical field of multi-agent collaborative decision-making. The method comprises the steps of enabling each intelligent agent to conduct joint perception and feature coding on acquired multi-mode input data through a data sensor, enabling a data updater to conduct fusion updating on a current multi-mode representation vector and a history representation set through a time sequence increment consistency integration algorithm, enabling the data sender to compress and send the fused codes to adjacent intelligent agents, enabling each intelligent agent to receive coding information from the adjacent intelligent agents through a data receiver and conduct alignment fusion with the fusion codes of the adjacent intelligent agents, enabling the data updater to conduct updating on fusion representation through a multi-level weighting feature pyramid algorithm, and generating new history representation data to support next round collaborative decision. The invention improves the reliability and the robustness of multi-agent collaborative decisions through the four core components of the data sensor, the data updater, the data transmitter and the data receiver.

Inventors

Xia yuelong
LI JICHAO
LIANG YAN
XIE BAOLING
CHEN YANING

Assignees

云南师范大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (9)

1. A multi-agent collaborative decision-making method, characterized in that the method comprises the steps of: S1, each intelligent agent performs joint sensing and feature coding on currently acquired multi-modal input data through a data sensor to obtain a multi-modal characterization vector; S2, introducing a time sequence increment consistency integration algorithm by each agent through a data updater, and carrying out fusion updating on the multi-mode characterization vector and the history characterization set to obtain a consistency enhanced fusion characterization; S3, each agent compresses the fusion characterization with enhanced consistency through a data transmitter to obtain a compressed message, and sends the compressed message to the adjacent agents; S4, each intelligent agent receives the compressed information from the adjacent intelligent agent through the data receiver, decodes and performs characteristic reconstruction processing on the compressed information to obtain the characteristic information of the adjacent intelligent agent, and each intelligent agent performs alignment and fusion processing on the characteristic information of the adjacent intelligent agent and the fusion characteristic with enhanced consistency after self fusion to generate a joint characteristic for collaborative decision; and S5, each agent data updater adopts a multi-stage weighted feature pyramid algorithm to update the joint characterization in a layering way, and generates new historical characterization data for supporting the next round of multi-agent collaborative decision process.
2. The multi-agent collaborative decision-making method according to claim 1, wherein S1 is specifically: S1.1 Each agent involved in collaborative decisions receives multimodal input data from a plurality of data sources, including but not limited to visual modality data Speech modality data And text modality data Forming a current multi-mode input set corresponding to the intelligent agent Multiple agents participating in collaborative decision-making to form an agent set , wherein, Represent the first An agent; s1.2, each intelligent agent inputs the received multi-mode input data to a corresponding data sensor to perform joint sensing and feature coding, wherein the data sensor is based on a multi-mode coding model Feature extraction and fusion processing are carried out on different modal data to generate a unified multi-modal characterization vector , wherein, Representing the feature space dimension.
3. The multi-agent collaborative decision-making method according to claim 1, wherein S2 is specifically: S2.1. each agent Introducing a sequential increment consistent integration algorithm through a data updater to represent vectors in current multiple modes With historical characterization sets A fusion process is performed, wherein, , Represented in historical collaborative decisions The multi-mode characterization vector obtained in the turn, the time sequence increment consistency integration algorithm evaluates consistency relations among different time sequence characterizations, performs weighted fusion on the current characterization and the historical characterization, and generates a consistency enhanced fusion characterization , wherein, Representing a sequential increment consistent integration function; S2.2, each agent utilizes a data updater to characterize the fusion Status updating processing is carried out, and fusion characterization is carried out Writing the history token set as a new history token to form an updated history token set 。
4. The multi-agent collaborative decision-making method according to claim 1, wherein S3 is specifically: S3.1. each agent Fusion characterization of consistency enhancement by fusion of data updaters Input to a corresponding data transmitter that compresses the fusion token by means of a coded compression function to generate a compressed message for inter-agent communication , wherein, Represent the first The compression mapping function corresponding to each agent data sender, Representing the characteristic dimension of the compressed message, for reducing the occupation of communication bandwidth and improving the information transmission efficiency, Representing a feature space dimension; s3.2. each agent Compressed message to be generated To a collection of adjacent agents with which a communication connection exists And the adjacent agent receives the compressed message for alignment fusion and collaborative decision processing.
5. The multi-agent collaborative decision-making method according to claim 1, wherein S4 is specifically: s4.1. each agent By means of corresponding data receivers, from Adjacent agent sets of (a) In receiving compressed message representations sent from neighboring agents , wherein, Represent the first Compression mapping function corresponding to intelligent agent data transmitter, data receiver pair compressed message Decoding and characteristic reconstruction processing are carried out to obtain the characterization information of the adjacent agent , wherein, Represent the first A decompression mapping function corresponding to the data receiver of each intelligent agent; s4.2, representing information of adjacent agents obtained by reconstructing each agent Fusion characterization with consistency enhancement after self-fusion Alignment and fusion processing is carried out to generate joint characterization for collaborative decision , wherein, Representing cross-agent token alignment and fusion functions, each agent based on the joint token And generating a collaborative decision result, thereby realizing information sharing and collaborative decision among multiple intelligent agents.
6. The multi-agent collaborative decision-making method according to claim 1, wherein S5 is specifically: s5.1. each agent Data updater of (1) for collaborative decision-making generated joint characterization Introducing a multi-stage weighted feature pyramid algorithm to perform hierarchical modeling and updating processing on the joint characterization, wherein the multi-stage weighted feature pyramid algorithm maps the joint characterization to a plurality of feature levels Different feature levels correspond to feature representations of different scales or different degrees of semantic abstraction, wherein, Representing joint characterization Is the first of (2) Each feature layer is allocated with corresponding weight coefficient The weighting updates are performed for each level of features, wherein, Representing joint characterization Is the first of (2) The weight coefficients corresponding to the feature layers; S5.2, each agent aggregates the layered features updated by the multi-level weighted feature pyramid algorithm to generate new historical representation data And writing the new history characterization data into a corresponding history characterization set The historical input data is used as the historical input data in the next round of multi-agent collaborative decision-making process, thereby realizing the continuous evolution and updating of the historical knowledge in the multi-agent collaborative decision-making process, wherein, Is fusion characterization with enhanced consistency.
7. An apparatus for implementing a multi-agent collaborative decision-making method according to claim 1, wherein the apparatus comprises: the data perceptron is responsible for acquiring and modeling multi-modal data of each agent and realizing multi-modal joint perception and feature coding; The data updater is in charge of two-round updating of the history characterization data, wherein the first round adopts a time sequence increment consistency integration algorithm to fuse the current multi-mode code with the history multi-mode characterization to realize consistency maintenance of cross-time sequence characterization and the history characterization data updating; the data transmitter is responsible for integrating the compression of the characterization and the generation of the message and realizing the transmission and neighborhood communication between the agents compressing the message; The data receiver is responsible for receiving and reconstructing the compression messages of the adjacent agents and realizing cross-agent characterization alignment fusion and collaborative decision generation.
8. A multi-agent collaborative decision-making apparatus comprising a memory, a processor and a multi-agent collaborative decision-making method program stored on the memory and operable on the processor, the processor executing the steps of the multi-agent collaborative decision-making method program to implement the multi-agent collaborative decision-making method of any of claims 1-6.
9. A computer readable storage medium, characterized in that it has stored thereon a multi-agent co-decision method program, which when executed by a processor, implements the steps of the multi-agent co-decision method according to any of claims 1 to 6.

Description

Multi-agent collaborative decision-making method and device thereof Technical Field The invention relates to a multi-agent collaborative decision-making method and a device thereof, belonging to the technical field of multi-agent collaborative decision-making. Background The current multi-agent cooperative system has become an important technical form for supporting cloud edge cooperative application due to the characteristics of distributed, autonomous cooperation and high robustness, and is widely applied to complex scenes such as intelligent transportation, smart cities, industrial Internet, intelligent manufacturing, intelligent medical treatment and the like. In such applications, the system is usually operated cooperatively by a plurality of distributed agents, each agent needs to complete information sensing, data interaction and joint decision in a dynamic and open environment, for example, in typical scenes such as automatic driving, the multi-agent cooperative decision can realize joint sensing of a blind area target and a complex environment, and the safety and reliability of the system are improved. In the related art, the current multi-agent collaboration method mainly relies on adapting features or characteristics among agents, and unification of perception information is achieved through single-step or multi-step conversion. Although these methods make progress in feature transformation and alignment of characterization, as the scale of the system expands, the heterogeneity of the intelligent agents in terms of sensing devices, model structures, data distribution, etc. is increasingly prominent, so that these methods still have the following drawbacks in actual deployment: (1) The multi-agent heterogeneous adaptation training cost is high. Existing multi-agent collaboration methods typically require retraining adaptation modules or sharing portions of the network structure for different agents. In safety critical applications such as automatic driving, the perception model is highly coupled with a downstream task, so that direct replacement or retraining is difficult, and when each pair of heterogeneous intelligent bodies are required to be respectively subjected to adaptive training, the training complexity and the deployment cost are obviously increased. (2) The multi-agent communication latency results in spatial and semantic misalignment. In the existing method, information interaction is carried out through a cloud or a centralized node, the processes of data uploading, centralized processing, result feedback and the like are needed to be carried out, communication and calculation time delay are unavoidable, and the problem of time asynchronism exists when different intelligent agents observe the same target, so that spatial position deviation and semantic inconsistency are caused. (3) The multi-agent unified characterization differences make synergistic stability limited. Existing methods reduce training costs by aligning representations of different agents into a unified space, but the unified representations are typically based on a particular agent. When the difference between other intelligent agents and the intelligent agents in the mode or the feature distribution is large, the characteristic alignment effect is difficult to ensure, semantic information loss is easy to cause, and the collaborative decision effect is influenced. (4) The complementarity of multi-agent information is underutilized. The existing method focuses on single mapping or conversion of the characteristic layer in multiple modes, and complementary advantages of multiple modes and multiple agents in view angle, precision and time sequence information are not fully mined, so that further improvement of collaborative perception and decision performance is limited. Aiming at the problems, the invention provides a multi-agent collaborative decision-making method and a device thereof, which can realize efficient collaboration under heterogeneous conditions of hardware and models, relieve semantic loss caused by communication time delay and unified characterization, fully utilize complementarity of multi-agent and multi-mode information, thereby improving instantaneity, reliability and stability of collaborative decision-making, and are suitable for multi-agent collaborative sensing and decision-making application in complex dynamic environments. Disclosure of Invention The technical problem to be solved by the invention is to provide a multi-agent collaborative decision-making method and a device thereof, so as to solve the problems of high adapting cost, space and semantic dislocation caused by communication delay, sensitivity of unified characterization to difference, insufficient utilization of multi-agent information complementarity and the like of the traditional multi-agent collaborative decision-making method under heterogeneous conditions of hardware and models, thereby realizing efficient, stable and real-t