Multi-agent collaborative decision-making system and method based on grouping decision-making and dynamic execution

Abstract

The invention discloses a multi-agent collaborative decision-making system and method based on grouping decision-making and dynamic execution, and relates to the technical field of multi-agent application. The method comprises the steps of 1, formulating a grouping strategy for intelligent agent execution tasks; step 2, making a dynamic execution strategy: step 21, constructing a strategy decision-making tree, and selecting indexes as input dimensions of the decision-making tree, the indexes including a task emergency degree, a resource competition index and a historical success rate sliding window; step 22, predicting the degree of parallelism of the tasks by adopting an XGBoost classifier, and executing a decision by utilizing a decision tree node according to prediction; and step 3, performing fault-tolerant coordination in the execution process.

Description

Technical Field

[0001] The present invention discloses a multi-agent collaborative decision-making system and method based on group decision-making and dynamic execution, and relates to the technical field of multi-agent applications. Background Art

[0002] In today's digital age, operators face increasingly complex network environments and massive business demands. Multi-agent systems, as a technology that simulates the collaborative work of multiple individuals in human society, have broad application prospects in operator scenarios, such as network optimization, troubleshooting, and resource scheduling. However, existing Reasoning-Action multi-agent collaboration mechanisms have numerous issues that severely restrict their effectiveness in operator scenarios, such as:

[0003] Efficiency bottleneck: Only one agent can be activated at a time to make decisions. For example, when processing network topology optimization involving over 200 subtasks, the system response time exceeds 15 minutes. This serial decision-making approach wastes a significant amount of time waiting for agent switching during task execution, underutilizing computing resources. This results in extremely low overall task processing efficiency, making it difficult to meet operators' requirements for real-time and high efficiency.

[0004] Idle resources: Because agents execute decision-making tasks serially, other agents remain idle during the intervals between decisions made by one agent, hindering the full utilization of computing resources. Statistics show that the average agent utilization rate is only 38%, resulting in significant waste of computing resources and increased operating costs.

[0005] Rigid strategies: A fixed serial execution model is used, making it impossible to dynamically adjust execution strategies based on the actual characteristics of tasks, such as task urgency, resource requirements, and inter-task dependencies. In operator scenarios, business requirements are complex and ever-changing, and different tasks may require different execution methods to achieve optimal results. The traditional fixed serial model is difficult to adapt to these diverse needs, resulting in poor task execution results. Furthermore, existing improvements to multi-agent systems in the industry primarily focus on communication optimization, task scheduling, and resource allocation, but all have significant limitations. Summary of the Invention

[0006] In response to the problems of the prior art, the present invention provides a multi-agent collaborative decision-making system and method based on group decision-making and dynamic execution, which conveniently expands templates and generates codes based on the templates.

[0007] The specific scheme proposed by the present invention is:

[0008] The present invention provides a multi-agent collaborative decision-making method based on group decision-making and dynamic execution, comprising:

[0009] Step 1: Develop a grouping strategy for agents to perform tasks:

[0010] Step 11: DAG division based on task dependency graph: Analyze the input tasks, construct a task dependency graph, use a directed acyclic graph (DAG) to represent the sequence and dependency between tasks, divide interrelated and non-conflicting tasks into the same group, and make the agents perform tasks in groups.

[0011] Step 12: Define the conflict detection matrix. The dimension of the conflict detection matrix is ​​the number of agents × resource type. Use the conflict detection matrix to record the occupation of different resources by each agent when performing tasks. By analyzing the conflict detection matrix, detect resource conflicts between different task groups.

[0012] Step 13: Construct a mixed integer programming model to minimize the resource consumption of each task group after grouping, and ensure that the resources used by the agents in the same task group do not conflict.

[0013] Step 2: Develop a dynamic execution strategy:

[0014] Step 21: Build a policy decision tree and select indicators as input dimensions of the decision tree. The indicators include task urgency, resource competition index, and historical success rate sliding window;

[0015] Step 22: Use the XGBoost classifier to predict the parallelism of the task and use the decision tree nodes to make decisions based on the prediction;

[0016] Step 3: Fault-tolerant coordination during execution:

[0017] Step 31: For tasks within the group: Use blockchain-style accounting to record every operation and status change during the task execution process. When an exception occurs in the execution of a task within the group, rollback is performed based on the accounting information.

[0018] Step 32: For cross-group tasks: tasks involving collaboration among multiple task groups, split a large distributed task into multiple local tasks based on the Saga transaction model, and define compensation operations for each local task. When a local task fails, execute the corresponding compensation operations in sequence.

[0019] Furthermore, the mixed integer programming model constructed in step 13 of the multi-agent collaborative decision-making method based on group decision-making and dynamic execution is formulated as follows:

[0020]

[0021] Where G represents the set of all task groups, g represents a specific task group, and α = 0.7 and β = 0.3 are preset as resource weight coefficients, representing the importance of CPU resources and memory resources respectively; C g CPU and C g mem Respectively represent the CPU resource consumption and memory resource consumption of task group g; R a Represents the resource set required by agent a. By solving the model, the optimal agent grouping scheme is obtained to achieve reasonable allocation and efficient utilization of resources.

[0022] Furthermore, in step 21 of the multi-agent collaborative decision-making method based on group decision-making and dynamic execution, the preset task urgency value range is 0-10, and the priority of the task is measured by the task urgency; the preset resource competition index value range is 0-1, and the resource competition index reflects the degree of resource tension during the task execution process; the preset historical success rate sliding window size is 50, and the historical success rate sliding window is used to predict the execution difficulty and success probability of the current task based on the execution status of a certain number of tasks in the past, and the indicators are analyzed and decided through the strategy decision tree to select a suitable execution strategy for the task.

[0023] Furthermore, in step 21 of the multi-agent collaborative decision-making method based on group decision-making and dynamic execution, the decision tree is pruned and optimized:

[0024] The cost-complexity pruning method is adopted to determine the optimal subtree through cross-validation to avoid overfitting. During pruning, the key paths that are strongly related to resource conflict detection and grouping results are retained to ensure a balance between decision-making efficiency and accuracy.

[0025] Furthermore, step 3 of the multi-agent collaborative decision-making method based on group decision-making and dynamic execution also includes step 33: establishing an exception propagation blocker to monitor and analyze abnormal information in real time during task execution. When an anomaly is detected, the impact range of the anomaly is determined, and the anomaly is isolated within a radius of ≤3 hops to prevent the anomaly from spreading in the multi-agent system.

[0026] The present invention also provides a multi-agent collaborative decision-making system based on group decision-making and dynamic execution, comprising: a grouping strategy module, an execution strategy module and a fault-tolerant coordination module.

[0027] The grouping strategy module formulates the grouping strategy for agents to perform tasks:

[0028] Step 11: DAG division based on task dependency graph: Analyze the input tasks, construct a task dependency graph, use a directed acyclic graph (DAG) to represent the sequence and dependency between tasks, divide interrelated and non-conflicting tasks into the same group, and make the agents perform tasks in groups.

[0029] Step 12: Define the conflict detection matrix. The dimension of the conflict detection matrix is ​​the number of agents × resource type. Use the conflict detection matrix to record the occupation of different resources by each agent when performing tasks. By analyzing the conflict detection matrix, detect resource conflicts between different task groups.

[0030] Step 13: Construct a mixed integer programming model to minimize the resource consumption of each task group after grouping, and ensure that the resources used by the agents in the same task group do not conflict.

[0031] The execution strategy module formulates dynamic execution strategies:

[0032] Step 21: Build a policy decision tree and select indicators as input dimensions of the decision tree. The indicators include task urgency, resource competition index, and historical success rate sliding window;

[0033] Step 22: Use the XGBoost classifier to predict the parallelism of the task and use the decision tree nodes to make decisions based on the prediction;

[0034] Fault-tolerant coordination is performed during the execution of the fault-tolerant coordination module:

[0035] Step 31: For tasks within the group: Use blockchain-style accounting to record every operation and status change during the task execution process. When an exception occurs in the execution of a task within the group, rollback is performed based on the accounting information.

[0036] Step 32: For cross-group tasks: tasks involving collaboration among multiple task groups, split a large distributed task into multiple local tasks based on the Saga transaction model, and define compensation operations for each local task. When a local task fails, execute the corresponding compensation operations in sequence.

[0037] Furthermore, the mixed integer programming model constructed when the grouping strategy module of the multi-agent collaborative decision-making system based on grouping decision-making and dynamic execution executes step 13 is as follows:

[0038]

[0039] Where G represents the set of all task groups, g represents a specific task group, and α = 0.7 and β = 0.3 are preset as resource weight coefficients, representing the importance of CPU resources and memory resources respectively; C g CPU and C g mem Respectively represent the CPU resource consumption and memory resource consumption of task group g; R a Represents the resource set required by agent a. By solving the model, the optimal agent grouping scheme is obtained to achieve reasonable allocation and efficient utilization of resources.

[0040] Furthermore, when the execution strategy module of the multi-agent collaborative decision-making system based on group decision-making and dynamic execution executes step 21, the preset task urgency value range is 0-10, and the priority of the task is measured by the task urgency; the preset resource competition index value range is 0-1, and the resource competition index reflects the degree of resource tension during the task execution process; the preset historical success rate sliding window size is 50, and the historical success rate sliding window is used to predict the execution difficulty and success probability of the current task based on the execution status of a certain number of tasks in the past, and the indicators are analyzed and decided through the strategy decision tree to select a suitable execution strategy for the task.

[0041] Furthermore, the execution strategy module of the multi-agent collaborative decision-making system based on group decision-making and dynamic execution performs pruning optimization on the decision tree when executing step 21:

[0042] The cost-complexity pruning method is adopted to determine the optimal subtree through cross-validation to avoid overfitting. During pruning, the key paths that are strongly related to resource conflict detection and grouping results are retained to ensure a balance between decision-making efficiency and accuracy.

[0043] Furthermore, the fault-tolerant coordination module of the multi-agent collaborative decision-making system based on group decision-making and dynamic execution also executes step 33: establishing an exception propagation blocker, real-time monitoring and analysis of abnormal information during task execution, and when an anomaly is detected, determining the impact range of the anomaly and isolating the anomaly within a radius of ≤3 hops to prevent the anomaly from spreading in the multi-agent system.

[0044] The benefits of the present invention are:

[0045] Improved task processing efficiency: In 5G core network optimization scenarios, task processing throughput increased from 82 tasks / min to 217 tasks / min, a 164% increase; average response latency decreased from 8.7s to 3.2s, a 63% reduction. In network expansion scenarios, the time required to optimize the parameters of 1,000 base stations was reduced from 53 minutes to 19 minutes, meeting real-time requirements.

[0046] Improved resource utilization: CPU utilization increased from 41% to 79%, a 93% improvement, reducing idle resources and saving operating costs.

[0047] Improved task execution accuracy: The task success rate increased from 89.3% to 91.5%, the error rate of complex tasks decreased by 0.8 percentage points, and reliability was enhanced.

[0048] Increased application benefits: Applied to operator networks, it has processed over 12 million work orders cumulatively, saving approximately 27 million yuan in annual operation and maintenance costs, with significant economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 It is a schematic flow chart of the method of the present invention.

[0051] Figure 2 It is a schematic diagram of the system application deployment of the present invention. DETAILED DESCRIPTION

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0053] The present invention provides a multi-agent collaborative decision-making method based on group decision-making and dynamic execution, comprising:

[0054] Step 1: Develop a grouping strategy for agents to perform tasks:

[0055] Step 11: Perform DAG partitioning based on the task dependency graph: Analyze the input tasks, construct a task dependency graph, and use a directed acyclic graph (DAG) to represent the sequence and dependency between tasks. Divide interrelated and conflict-free tasks into the same group, so that the intelligent agent can perform tasks in groups. For example, in network optimization tasks, optimization tasks for the same area or the same type of equipment can be divided into one group to reduce the switching overhead between tasks.

[0056] Step 12: Define a conflict detection matrix. The dimensions of the conflict detection matrix are agent number x resource type. This matrix records the resource usage of each agent while executing a task. By analyzing the conflict detection matrix, we can detect resource conflicts between different task groups and avoid task failures or reduced efficiency due to resource competition. For example, if two task groups need to use the computing resources of the same server at the same time, the conflict detection matrix can detect this and adjust the situation in a timely manner.

[0057] Step 13: Construct a mixed integer programming model. Use the mixed integer programming model to minimize the resource consumption of each task group after grouping, and ensure that the resources used by the agents in the same task group do not conflict. The formula of the constructed mixed integer programming model is as follows:

[0058]

[0059] Where G represents the set of all task groups, g represents a specific task group, and α = 0.7 and β = 0.3 are preset as resource weight coefficients, representing the importance of CPU resources and memory resources respectively; C g CPU and C g mem Respectively represent the CPU resource consumption and memory resource consumption of task group g; R a Represents the resource set required by agent a. By solving the model, the optimal agent grouping scheme is obtained to achieve reasonable allocation and efficient utilization of resources.

[0060] Step 2: Develop a dynamic execution strategy:

[0061] Step 21: Build a strategy decision tree and select indicators as the input dimensions of the decision tree. The indicators include task urgency, resource competition index, and historical success rate sliding window.

[0062] The preset task urgency value range is 0-10, and the task priority is measured by task urgency; the preset resource competition index value range is 0-1, and the resource competition index reflects the resource intensity during task execution; the preset historical success rate sliding window size is 50, and the historical success rate sliding window is used to predict the execution difficulty and success probability of the current task based on the execution status of a certain number of tasks in the past. The indicators are analyzed and decided through the strategy decision tree to select the appropriate execution strategy for the task.

[0063] Prune and optimize the decision tree: Use the cost-complexity pruning method to determine the optimal subtree through cross-validation to avoid overfitting. During pruning, retain the key paths that are strongly related to resource conflict detection and grouping results to ensure a balance between decision efficiency and accuracy.

[0064] Step 22: Use the XGBoost classifier to predict the task's degree of parallelism and use the decision tree nodes to make decisions based on the predictions. The XGBoost classifier can learn the relationship between task characteristics and the optimal degree of parallelism, achieving an AUC of 0.89. It can accurately select the appropriate parallel execution method for each task, achieving efficient task execution. For example, for urgent tasks with sufficient resources, a higher degree of parallelism is predicted to speed up task processing. For tasks with intense resource competition, the degree of parallelism is appropriately adjusted to avoid task failure due to excessive resource competition.

[0065] The decision tree nodes are used to perform tasks, as shown below:

[0066] Root node: Input task feature vector [urgency, resource competition index, historical success rate],

[0067] First-level split node: With the task urgency (threshold set to 6) as the first judgment condition, tasks with urgency ≥ 6 enter the "high priority branch", otherwise they enter the "regular processing branch".

[0068] Secondary split node: In the high-priority branch, resource availability is determined by the resource competition index (threshold set to 0.4). When the competition index is < 0.4, the entire group is allowed to execute in parallel. Otherwise, the "resource allocation sub-strategy" (such as dynamically applying for a temporary resource pool) is triggered.

[0069] Leaf node: Outputs the execution policy combination {execution mode, resource quota, timeout threshold}, where the execution mode includes:

[0070] Parallel priority mode is suitable for scenarios with high urgency and sufficient resources, allowing agents in the same group to execute concurrently.

[0071] The pipeline mode is suitable for scenarios where there is weak dependency between tasks, and is executed in parallel in stages according to the DAG sequence.

[0072] The serial calibration mode is suitable for complex tasks with a historical success rate of <70%, forcing the single agent to execute step by step to reduce the error rate.

[0073] For example, when the task urgency is 8, which is high priority, the resource contention index is 0.3, which is low contention, and the historical success rate is 85%, the decision tree output is "parallel priority mode + CPU resource quota of 12 cores + timeout threshold of 90s";

[0074] When the task urgency is 4, it is normal, the resource competition index is 0.7, it is high competition, and the historical success rate is 65%, the output is "serial calibration mode + resource quota halved + step-by-step status verification" and so on.

[0075] Step 3: Fault-tolerant coordination during execution:

[0076] Step 31: For tasks within the group: Use blockchain-style accounting to record every operation and status change during the task execution process. When an exception occurs in the execution of a task within the group, rollback is performed based on the accounting information.

[0077] Step 32: For cross-group tasks: tasks involving collaboration among multiple task groups, split a large distributed task into multiple local tasks based on the Saga transaction model, and define compensation operations for each local task. When a local task fails, execute the corresponding compensation operations in sequence.

[0078] Step 33: Establish an anomaly propagation blocker to monitor and analyze abnormal information in real time during task execution. When an anomaly is detected, determine the impact range of the anomaly and isolate the anomaly within a radius of ≤3 hops to prevent the anomaly from spreading in the multi-agent system.

[0079] After the implementation of the present invention, multiple scenarios were tested and verified:

[0080] Fault handling scenarios: In a backbone network fault simulation platform tested on over 1,000 network elements and over 50 fault types, fault location time was reduced from 420s to 150s, a 64% improvement. The cross-domain fault recovery SLA compliance rate increased from 78% to 94%, and the success rate for complex fault handling increased from 81% to 92.3%.

[0081] Resource scheduling scenario: For the dynamic allocation of 5G slice resources, simulating 2,000+ virtual nodes and real-time traffic loads, the resource allocation throughput increased from 500 times / minute to 1,300 times / minute, a 160% improvement; the resource fragmentation rate decreased from 28% to 12%; and the response latency in hot spots decreased from 2.1s to 0.8s.

[0082] Experimental design and verification methods

[0083] Control group setting: Compared with the existing serial decision system and the existing optimal distributed multi-agent system, the system of the present invention adopts dynamic grouping + strategy decision tree + fault tolerance mechanism.

[0084] Statistical test: Continuous variables, including throughput and latency, were tested using a two-tailed t-test (α = 0.05), and categorical variables, including task success rate and resource utilization, were tested using a chi-square test (P < 0.01). Significant differences were found.

[0085] Sample and repeatability: The test sample size for each scenario is ≥1000 times, covering multiple working conditions; key indicators take the average value of 10 independent experiments, and the standard deviation is controlled within 5% of the mean.

[0086] Example 2

[0087] The present invention also provides a multi-agent collaborative decision-making system based on group decision-making and dynamic execution, comprising: a grouping strategy module, an execution strategy module and a fault-tolerant coordination module.

[0088] The grouping strategy module formulates the grouping strategy for agents to perform tasks:

[0089] Step 11: DAG division based on task dependency graph: Analyze the input tasks, construct a task dependency graph, use a directed acyclic graph (DAG) to represent the sequence and dependency between tasks, divide interrelated and non-conflicting tasks into the same group, and make the agents perform tasks in groups.

[0090] Step 12: Define the conflict detection matrix. The dimension of the conflict detection matrix is ​​the number of agents × resource type. Use the conflict detection matrix to record the occupation of different resources by each agent when performing tasks. By analyzing the conflict detection matrix, detect resource conflicts between different task groups.

[0091] Step 13: Construct a mixed integer programming model to minimize the resource consumption of each task group after grouping, and ensure that the resources used by the agents in the same task group do not conflict.

[0092] The execution strategy module formulates dynamic execution strategies:

[0093] Step 21: Build a policy decision tree and select indicators as input dimensions of the decision tree. The indicators include task urgency, resource competition index, and historical success rate sliding window;

[0094] Step 22: Use the XGBoost classifier to predict the parallelism of the task and use the decision tree nodes to make decisions based on the prediction;

[0095] Fault-tolerant coordination is performed during the execution of the fault-tolerant coordination module:

[0096] Step 31: For tasks within the group: Use blockchain-style accounting to record every operation and status change during the task execution process. When an exception occurs in the execution of a task within the group, rollback is performed based on the accounting information.

[0097] Step 32: For cross-group tasks: tasks involving collaboration among multiple task groups, split a large distributed task into multiple local tasks based on the Saga transaction model, and define compensation operations for each local task. When a local task fails, execute the corresponding compensation operations in sequence.

[0098] The information interaction, execution process and other contents between the modules in the above system are based on the same concept as the embodiment of the method of the present invention. For specific contents, please refer to the description in the embodiment of the method of the present invention and will not be repeated here.

[0099] Likewise, the benefits of the system of the present invention are:

[0100] Improved task processing efficiency: In 5G core network optimization scenarios, task processing throughput increased from 82 tasks / min to 217 tasks / min, a 164% increase; average response latency decreased from 8.7s to 3.2s, a 63% reduction. In network expansion scenarios, the time required to optimize the parameters of 1,000 base stations was reduced from 53 minutes to 19 minutes, meeting real-time requirements.

[0101] Improved resource utilization: CPU utilization increased from 41% to 79%, a 93% improvement, reducing idle resources and saving operating costs.

[0102] Improved task execution accuracy: The task success rate increased from 89.3% to 91.5%, the error rate of complex tasks decreased by 0.8 percentage points, and reliability was enhanced.

[0103] Increased application benefits: Applied to operator networks, it has processed over 12 million work orders cumulatively, saving approximately 27 million yuan in annual operation and maintenance costs, with significant economic value.

[0104] It should be noted that not all steps and modules in the above-mentioned processes and system structures are required, and certain steps or modules can be omitted according to actual needs. The execution order of each step is not fixed and can be adjusted as needed. The system structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or may be implemented by certain components in multiple independent devices.

[0105] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A multi-agent collaborative decision-making method based on group decision-making and dynamic execution, characterized by include: Step 1: Develop a grouping strategy for agents to perform tasks: Step 11: DAG division based on task dependency graph: Analyze the input tasks, construct a task dependency graph, use a directed acyclic graph (DAG) to represent the sequence and dependency between tasks, divide interrelated and non-conflicting tasks into the same group, and make the agents perform tasks in groups. Step 12: Define the conflict detection matrix. The dimension of the conflict detection matrix is ​​the number of agents × resource type. Use the conflict detection matrix to record the occupation of different resources by each agent when performing tasks. By analyzing the conflict detection matrix, detect resource conflicts between different task groups. Step 13: Construct a mixed integer programming model to minimize the resource consumption of each task group after grouping, and ensure that the resources used by the agents in the same task group do not conflict. Step 2: Develop a dynamic execution strategy: Step 21: Build a policy decision tree and select indicators as input dimensions of the decision tree. The indicators include task urgency, resource competition index, and historical success rate sliding window; Step 22: Use the XGBoost classifier to predict the parallelism of the task and use the decision tree nodes to make decisions based on the prediction; Step 3: Fault-tolerant coordination during execution: Step 31: For tasks within the group: Use blockchain-style accounting to record every operation and status change during the task execution process. When an exception occurs in the execution of a task within the group, rollback is performed based on the accounting information. Step 32: For cross-group tasks: tasks involving collaboration among multiple task groups, split a large distributed task into multiple local tasks based on the Saga transaction model, and define compensation operations for each local task. When a local task fails, execute the corresponding compensation operations in sequence.

2. A multi-agent collaborative decision-making method based on group decision-making and dynamic execution according to claim 1, characterized in that The mixed integer programming model constructed in step 13 is as follows: Where G represents the set of all task groups, g represents a specific task group, and α = 0.7 and β = 0.3 are preset as resource weight coefficients, representing the importance of CPU resources and memory resources respectively; C g CPU and C g mem Respectively represent the CPU resource consumption and memory resource consumption of task group g; R a Represents the resource set required by agent a. By solving the model, the optimal agent grouping scheme is obtained to achieve reasonable allocation and efficient utilization of resources.

3. The multi-agent collaborative decision-making method based on group decision-making and dynamic execution according to claim 1 is characterized by: In step 21, the value range of the task urgency is preset to 0-10, and the priority of the task is measured by the task urgency; The preset resource competition index has a value range of 0-1, which reflects the degree of resource tension during task execution. The preset historical success rate sliding window size is 50. The historical success rate sliding window is used to predict the execution difficulty and success probability of the current task based on the execution status of a certain number of tasks in the past. The indicators are analyzed and decided through the strategy decision tree to select the appropriate execution strategy for the task.

4. A multi-agent collaborative decision-making method based on group decision-making and dynamic execution according to claim 1 or 3, characterized in that In step 21, the decision tree is pruned and optimized: The cost-complexity pruning method is adopted to determine the optimal subtree through cross-validation to avoid overfitting. During pruning, the key paths that are strongly related to resource conflict detection and grouping results are retained to ensure a balance between decision-making efficiency and accuracy.

5. The multi-agent collaborative decision-making method based on group decision-making and dynamic execution according to claim 1 is characterized by: Step 3 also includes step 33: establishing an anomaly propagation blocker to monitor and analyze abnormal information in real time during task execution. When an anomaly is detected, the impact range of the anomaly is determined and the anomaly is isolated within a radius of ≤3 hops to prevent the anomaly from spreading in the multi-agent system.

6. A multi-agent collaborative decision-making system based on group decision-making and dynamic execution, characterized by include: Grouping strategy module, execution strategy module and fault-tolerant coordination module, The grouping strategy module formulates the grouping strategy for agents to perform tasks: Step 11: DAG division based on task dependency graph: Analyze the input tasks, construct a task dependency graph, use a directed acyclic graph (DAG) to represent the sequence and dependency between tasks, divide interrelated and non-conflicting tasks into the same group, and make the agents perform tasks in groups. Step 12: Define the conflict detection matrix. The dimension of the conflict detection matrix is ​​the number of agents × resource type. Use the conflict detection matrix to record the occupation of different resources by each agent when performing tasks. By analyzing the conflict detection matrix, detect resource conflicts between different task groups. Step 13: Construct a mixed integer programming model to minimize the resource consumption of each task group after grouping, and ensure that the resources used by the agents in the same task group do not conflict. The execution strategy module formulates dynamic execution strategies: Step 21: Build a policy decision tree and select indicators as input dimensions of the decision tree. The indicators include task urgency, resource competition index, and historical success rate sliding window; Step 22: Use the XGBoost classifier to predict the parallelism of the task and use the decision tree nodes to make decisions based on the prediction; Fault-tolerant coordination is performed during the execution of the fault-tolerant coordination module: Step 31: For tasks within the group: Use blockchain-style accounting to record every operation and status change during the task execution process. When an exception occurs in the execution of a task within the group, rollback is performed based on the accounting information. Step 32: For cross-group tasks: tasks involving collaboration among multiple task groups, split a large distributed task into multiple local tasks based on the Saga transaction model, and define compensation operations for each local task. When a local task fails, execute the corresponding compensation operations in sequence.

7. A multi-agent collaborative decision-making system based on group decision-making and dynamic execution according to claim 6, characterized in that The mixed integer programming model constructed by the grouping strategy module in step 13 is as follows: Where G represents the set of all task groups, g represents a specific task group, and α = 0.7 and β = 0.3 are preset as resource weight coefficients, representing the importance of CPU resources and memory resources respectively; C g CPU and C g mem Respectively represent the CPU resource consumption and memory resource consumption of task group g; R a Represents the resource set required by agent a. By solving the model, the optimal agent grouping scheme is obtained to achieve reasonable allocation and efficient utilization of resources.

8. The multi-agent collaborative decision-making system based on group decision-making and dynamic execution according to claim 6 is characterized by: In step 21, the execution strategy module presets the task urgency value range from 0 to 10, and measures the priority of the task by the task urgency; The preset resource competition index has a value range of 0-1, which reflects the degree of resource tension during task execution. The preset historical success rate sliding window size is 50. The historical success rate sliding window is used to predict the execution difficulty and success probability of the current task based on the execution status of a certain number of tasks in the past. The indicators are analyzed and decided through the strategy decision tree to select the appropriate execution strategy for the task.

9. A multi-agent collaborative decision-making system based on group decision-making and dynamic execution according to claim 6 or 8, characterized in that In step 21, the strategy module is executed to prune the decision tree: The cost-complexity pruning method is adopted to determine the optimal subtree through cross-validation to avoid overfitting. During pruning, the key paths that are strongly related to resource conflict detection and grouping results are retained to ensure a balance between decision-making efficiency and accuracy.

10. The multi-agent collaborative decision-making system based on group decision-making and dynamic execution according to claim 6 is characterized by: The fault-tolerant coordination module also executes step 33: establishing an exception propagation blocker to monitor and analyze the exception information in real time during the task execution process. When an exception is detected, the impact range of the exception is determined and the exception is isolated within a radius of ≤3 hops to prevent the exception from spreading in the multi-agent system.

Images