CN-121980215-A - Multi-agent result consistency verification method and system based on dynamic weighting and context awareness

CN121980215ACN 121980215 ACN121980215 ACN 121980215ACN-121980215-A

Abstract

The invention discloses a multi-agent result consistency verification method and system based on dynamic weighting and context awareness, which are used for acquiring context information of a cooperation task, dividing the key level of the cooperation task, independently executing the cooperation task by all agents and generating respective local results, broadcasting the local results to other agents by all agents, changing the current reputation score of each agent into the voting weight of the verification, selecting a corresponding consistency verification strategy from a plurality of preset verification strategies according to the key level of the cooperation task, executing consistency verification on the local results based on the voting weight of each agent to achieve a consensus result, confirming and distributing the consensus result to all agents, and dynamically updating the reputation scores of all agents according to the verification process. The invention can dynamically, efficiently and safely realize the result consistency verification among multiple agents according to the task context and the self state of the agents.

Inventors

DENG XUTAO
LI XING
HAN XIN
CHEN BINGYUAN
QIN ZHANPENG
LV JIANHAO
HUANG YAN

Assignees

东方电气集团数字科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (9)

1. The multi-agent result consistency verification method based on dynamic weighting and context awareness is characterized by comprising the following steps of: step S1, task initialization and context analysis, wherein one or one lightweight coordination module in the intelligent agent acquires context information of a collaboration task and classifies the criticality level of the collaboration task according to a preset rule or a machine learning model; s2, all the agents participating in the cooperation independently execute the cooperation task according to the received task instruction and generate respective local results; Step S3, broadcasting local results with identity IDs to other related agents or verification modules by each agent through a communication network, inquiring the current reputation score of each agent according to the ID of each agent, and converting the reputation score into voting weight for verification; Step S4, selecting a corresponding consistency verification strategy from a plurality of preset verification strategies according to the key grade of the cooperation task in the step 1, and executing consistency verification on the local result based on the voting weight of each agent so as to achieve a consensus result; and S5, after the final consensus is achieved through the step S4, the consensus result is confirmed and distributed to all the agents to guide follow-up actions, and the reputation scores of all the participating agents are dynamically updated according to the verification process.
2. The multi-agent result consistency verification method based on dynamic weighting and context awareness according to claim 1, wherein in step S1, the criticality classes of the collaborative task are classified into three classes of high criticality, medium criticality, and low criticality, wherein: Highly critical tasks refer to tasks that, once made wrong, would lead to serious consequences; Medium critical tasks refer to tasks that affect system performance and efficiency but do not immediately raise security concerns; the low-criticality task refers to an auxiliary task with low requirements on real-time and accuracy.
3. The method for verifying the consistency of multi-agent results based on dynamic weighting and context awareness as in claim 1, wherein in step S3, the reputation score is obtained based on a comprehensive evaluation of the historical performance of the agent, and the determination of the voting weight converts the reputation score into the voting weight for the verification through a linear or nonlinear conversion function.
4. The multi-agent result consistency verification method based on dynamic weighting and context awareness as claimed in claim 1, wherein in step S4, the selecting a verification policy according to the task' S criticality class comprises: For a high-criticality task, calculating the total weight of each possible result by adopting a weighted Bayesian-busy-tolerant protocol, wherein when the total weight of a certain result exceeds a preset high threshold value, the result is accepted as a final consensus; For a medium critical task, adopting a weighted majority voting protocol, and selecting a result with highest total weight as a final consensus; For low criticality tasks, a lightweight authentication protocol is employed.
5. The method for dynamically weighted and context aware-based multi-agent result consistency verification of claim 4, wherein the lightweight verification protocol includes directly taking the result of the agent with the highest reputation score or taking a round of simple weighted voting.
6. The multi-agent result consistency verification method based on dynamic weighting and context awareness as claimed in claim 1, wherein in step S5, the dynamically updating reputation score comprises: Increasing the reputation score of the agent when its local outcome is consistent with the final consensus; when the local result of the agent is inconsistent with the final consensus, reducing the reputation score of the agent; The update amplitude of the dynamically updated reputation score is a fixed value or a value that is dynamically adjusted according to the degree of deviation.
7. A multi-agent result consistency verification system based on dynamic weighting and context awareness implementing the multi-agent result consistency verification method based on dynamic weighting and context awareness of any one of claims 1-6, comprising: The task execution module is deployed on each agent and is used for independently executing tasks and generating local results; The communication module is deployed on each intelligent agent and is used for broadcasting local results and receiving consensus results; A verification module, the verification module comprising: a context analyzer for evaluating a criticality class of a collaborative task; A reputation manager for storing and updating reputation scores of the respective agents; A consistency verification engine for executing a specific verification protocol according to the context and weight.
8. The multi-agent result consistency verification system based on dynamic weighting and context awareness of claim 7, wherein the verification module employs a distributed deployment.
9. The multi-agent result consistency verification system based on dynamic weighting and context awareness of claim 7, wherein the reputation manager maintains a database of reputation scores for each agent in real time and is dynamically updated based on each verification result.

Description

Multi-agent result consistency verification method and system based on dynamic weighting and context awareness Technical Field The invention relates to the technical field of artificial intelligence and multi-agent systems, in particular to a multi-agent result consistency verification method and system based on dynamic weighting and context awareness. Background With the development of artificial intelligence and Internet of things technology, multi-agent systems have demonstrated great application potential in numerous fields, such as autopilot fleet coordination, unmanned aerial vehicle cluster formation, distributed sensor network monitoring, smart grid scheduling, flexible manufacturing robot coordination, and the like. In these systems, multiple agents communicate and interact together to accomplish complex tasks that a single agent cannot accomplish independently. One core and critical challenge is result consistency. That is, in a distributed, dynamic, and potentially interfering environment, how all agents participating in the collaboration are guaranteed to agree on a certain key state, decision, or final output. The lack of an effective consistency verification mechanism may lead to confusion of system behaviors, task failure, and even security incidents. At present, the following technical schemes are mainly adopted for solving the problem of multi-agent consistency, but the following obvious defects and defects exist: (1) Center verification method: The technical description is that a central node or leader is set, and all agents send respective local results to the central node. The central node is responsible for comparing, arbitrating or fusing all results and finally distributing the unified decision result to all agents. The defects and shortcomings are that once a central node fails or is attacked maliciously, the verification mechanism of the whole system is paralyzed, and the robustness is poor. And the communication bottleneck is that all data need to flow through the central node, and when the number of the intelligent agents is increased or the data volume is increased, the central node becomes the bottleneck of communication and calculation, so that the expandability of the system is limited. The real-time performance is poor, communication delay is increased when data comes to and comes from a central node, and the requirements on application scenes (such as automatic driving obstacle avoidance) needing quick response are difficult to meet. (2) Traditional distributed consensus algorithm: description of the art classical distributed consensus algorithms such as Paxos, raft and bayer fault-tolerant algorithms (e.g. PBFT) are used. These algorithms ensure that non-faulty nodes eventually agree upon the presence of partial node failures or malicious behavior through multiple rounds of message passing. Drawbacks and disadvantages of this type of algorithm, which typically requires O (n 2) -level message complexity (n being the number of nodes), is cost prohibitive to implement on resource-constrained agents (e.g., drones, sensors). High latency algorithms typically require multiple stages (e.g., prepare, pre-commit, commit) to reach consensus, resulting in longer decision delays. The lack of context-aware flexibility and adaptability is that these algorithms are usually "cut-through" and do not dynamically adjust the verification policy according to the specific nature of the task (e.g., importance, urgency) and current environmental conditions (e.g., network quality), which is too "heavy" for non-critical tasks, resulting in resource waste, and for highly critical tasks, which may also affect real-time performance due to flow stiffness. (3) Simple voting or majority voting methods: The technical description is that the intelligent agents directly exchange results, each result calculates a ticket, and the result with the largest number of tickets is adopted as the final consensus. Defects and deficiencies: The Bayesian fault cannot be handled, and the method assumes that the agent only "goes down" and does not "do nothing. When malicious agents for sending fake or contradictory results exist, simple voting is very easy to operate, and the correctness of the results cannot be guaranteed. Ignoring agent variability, the method considers all agents as being completely equivalent, ignoring their differences in computing power, sensor accuracy, historical reliability, etc. It is obviously unreasonable that a high-precision, high-reliability "ticket" of an agent has the same weight as a low-precision, frequently errant "ticket" of an agent. It may occur that the ticket number is flat or cannot be halved, which may result in an inability to make an effective decision when the number of agents is even or the opinion is scattered. In view of the above, it is difficult in the prior art to achieve a good balance among robustness, efficiency, scalability and adaptability. Therefore, a new method f