A multi-agent formation control method based on machine learning

Abstract

The present application relates to the technical field of multi-agent cooperative control, and discloses a multi-agent formation control method based on machine learning. The method comprises establishing a cooperative control framework defining information interaction rules and decision levels. Based on the framework, the position, heading and communication state of each agent are collected, and a machine learning model is used for pattern recognition to dynamically extract a feature sequence reflecting the overall stability of the formation. The control performance is evaluated in real time according to the change of the feature sequence, and a strategy optimization stage is started. In this stage, historical data and real-time features are combined to generate adaptive control instructions. The instructions are decomposed into individual tasks and distributed to each agent. The system integrates the state confirmation information fed back after the execution of the individual tasks, updates the overall control model parameters, and iteratively calculates to finally output the optimized control sequence. Through data driving and online learning, the method improves the adaptive ability and control robustness of the formation system in complex environments.

Description

Technical Field

[0001] This invention relates to the field of multi-agent cooperative control technology, specifically to a multi-agent formation control method based on machine learning. Background Technology

[0002] Multi-agent formation control is a core technology in fields such as robot collaborative operations and drone swarms. Its goal is to enable multiple agents to maintain a specific formation and move collaboratively through local interactions. Existing technologies primarily employ model-based control strategies, such as the leader-follower method, the virtual structure method, or behavior rule-based methods. These methods rely on pre-established, accurate dynamic models and ideal environment assumptions, calculating motion commands for each agent through fixed control laws. Under conditions of accurate models and controllable environments, these methods can achieve basic formation-keeping functionality.

[0003] Real-world application environments are often fraught with uncertainty, exposing the limitations of traditional methods. The fundamental problem lies in the fact that control strategies are static and pre-defined, failing to understand the overall operational state of the formation and lacking the ability to adjust based on real-time situations. When unexpected disturbances occur, the system can only react based on local and superficial sensor data, easily leading to decreased control performance, formation instability, or even mission failure. Existing technologies lack the ability to extract deep features from system operational data, and the decision-making process lacks learning and evolutionary mechanisms.

[0004] The core challenge facing current technology is how to overcome the limitations of model-driven approaches, enabling formation systems to dynamically perceive overall stability and achieve online self-optimization of control strategies. This requires new control methods that can automatically identify the essential characteristics of formation states and form a closed loop of perception, decision-making, execution, and learning, thereby maintaining robust and adaptive collaborative control in complex dynamic environments. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-agent formation control method based on machine learning to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a multi-agent formation control method based on machine learning, the method comprising:

[0007] A collaborative control framework for a multi-agent formation is established, wherein the collaborative control framework defines the information interaction rules and decision-making hierarchy among the agents;

[0008] Based on the cooperative control framework, the operational status data of each intelligent agent in the formation is collected, including position, heading and communication status.

[0009] Machine learning models are used to perform pattern recognition on the collected operational status data and extract feature sequences that reflect the overall stability of the formation.

[0010] The real-time performance of formation control is evaluated based on the dynamic changes of the feature sequence, and a performance evaluation report is generated.

[0011] The strategy optimization phase is initiated based on the performance evaluation report. The strategy optimization phase generates adaptive control instructions based on historical control data and real-time feature sequences.

[0012] The adaptive control instructions are decomposed into individual control tasks corresponding to each agent;

[0013] Distribute individual control tasks to each agent and receive status confirmation information from the agents after they have executed the tasks;

[0014] Integrate all status confirmation information and update the parameters of the overall formation control model;

[0015] The next control cycle is iteratively calculated based on the updated control model parameters;

[0016] After completing the iterative calculations, the final multi-agent formation control sequence is output.

[0017] Preferably, the establishment of the cooperative control framework for multi-agent formation includes:

[0018] Define the role types of agents in the formation, including leader agents and follower agents;

[0019] A data communication protocol is defined between the leading agent and the following agent, and the data communication protocol specifies the data transmission format and response time limit;

[0020] Construct a topology-based formation communication network, wherein the formation communication network determines the direct communication links and indirect communication paths between agents;

[0021] A decision center is established within the collaborative control framework. This decision center is responsible for aggregating global information and generating top-level control strategies.

[0022] Preferably, the collection of operational status data for each agent within the formation includes:

[0023] The navigation data reported by the navigator is periodically acquired through the formation communication network;

[0024] Simultaneously, it receives state update messages from the following agent, the state update messages containing relative position and motion vector;

[0025] The navigation data and status update messages are timestamped to form a synchronized set of running status data.

[0026] The operational status data set is standardized and transformed to eliminate the dimensional differences between data sources from different agents.

[0027] Preferably, the step of using a machine learning model to perform pattern recognition on the collected operational status data includes:

[0028] The standardized set of operational status data is input into a pre-trained time series analysis model;

[0029] Use a time series analysis model to extract the fluctuation characteristics of operational status data in the time dimension;

[0030] The fluctuation characteristics are matched with the preset formation mode template to identify the current formation operation mode;

[0031] Output the feature sequence corresponding to the formation operation mode, and the feature sequence is in the form of a multi-dimensional vector.

[0032] Preferably, the evaluation of the real-time performance of formation control includes:

[0033] Calculate the deviation between the feature sequence and the features of the ideal formation state;

[0034] Calculate the comprehensive performance index by combining the real-time connectivity quality of the formation communication network;

[0035] The stability level of formation control performance is determined based on the changing trend of comprehensive performance indicators;

[0036] Generate a performance evaluation report that includes deviation, comprehensive performance indicators, and stability level.

[0037] Preferably, the startup strategy optimization phase includes:

[0038] When the stability level in the performance evaluation report is lower than the set threshold, the control strategy optimization process is triggered.

[0039] Retrieve historical scenarios and corresponding successful control strategies that are similar to the current formation operation mode from the historical database;

[0040] By using reinforcement learning algorithms in machine learning, candidate control instruction sets are generated by combining real-time feature sequences with historical control strategies.

[0041] The candidate control instruction set is simulated and verified to select the optimal adaptive control instruction.

[0042] Preferably, the step of decomposing the adaptive control instructions into individual control tasks corresponding to each agent includes:

[0043] Analyze the overall control objective of adaptive control instructions and determine the contribution weight of each agent in the objective;

[0044] Based on the contribution weight, the overall control objective is quantified into the expected state change of each agent.

[0045] By combining the dynamic constraints of each agent, the expected state change is converted into an executable individual control task;

[0046] Each individual control task is assigned a priority and execution time limit.

[0047] Preferably, the distribution of individual control tasks to each intelligent agent includes:

[0048] Individual control tasks are sent to the corresponding intelligent agents via a formation communication network;

[0049] Start the task execution timer and monitor whether the intelligent agent sends back status confirmation information within the execution time limit;

[0050] If no status confirmation is received within the timeout period, the individual control task will be resent via the backup communication path.

[0051] Record log information for all agent task distributions.

[0052] Preferably, the step of integrating all state confirmation information and updating the overall formation control model parameters includes:

[0053] Verify the completeness and consistency of the received status confirmation information;

[0054] The status confirmation information is compared with the expected goals of the individual control task to calculate the task execution consistency.

[0055] Adjust the agent reliability coefficients in the overall formation control model based on the task execution consistency, and update the cooperative parameters of the control model using the latest reliability coefficients.

[0056] Preferably, the iterative calculation for the next control cycle includes:

[0057] Re-collect formation operation status data based on the updated control model parameters;

[0058] The newly collected operational status data is input into the time series analysis model to generate new feature sequences;

[0059] The new feature sequences are used for performance evaluation, and a new round of performance evaluation reports are generated.

[0060] The decision on whether further strategy optimization is needed will be made based on the new performance evaluation report, thus forming a closed-loop control cycle.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] Machine learning models are used to perform pattern recognition on multi-agent operational status data, automatically extracting feature sequences reflecting the overall stability of the formation. This elevates traditional control decisions based on raw data to a deeper understanding of the system's overall behavioral patterns. Machine learning models can extract abstract features characterizing formation coordination quality, formation maintenance, and potential risks from multi-dimensional, non-linear data such as position, heading, and communication status. These feature sequences dynamically reflect the evolutionary trends of the system's macroscopic state, achieving a deep perception and early warning of the formation's health status. This allows the system to perceive subtle state deviations and their cumulative effects, basing control decisions on a quantitative understanding of its inherent operational laws, thus improving the accuracy and foresight of state assessments.

[0063] Based on real-time performance evaluation reports and historical control data, adaptive control commands are generated, and the control model parameters are iteratively updated using status confirmation information from the intelligent agent. This process constructs a complete "perception-decision-execution-learning" closed-loop optimization mechanism. The strategy optimization phase no longer relies on fixed control laws but dynamically adjusts based on performance feedback. The introduction of historical data enables control commands to incorporate past experience and possess contextual adaptability. Distributing commands and receiving status confirmation information ensures the effective execution and verification of control actions and provides real-time, accurate closed-loop data for model parameter updates. This endows the formation control system with online self-learning and continuous evolution capabilities. The control strategy can autonomously optimize based on actual operating results, thus exhibiting stronger robustness and adaptability in the face of uncertainties such as internal parameter fluctuations, external environmental disturbances, or unexpected communication interruptions, maintaining the long-term stability and efficiency of the formation's operation. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the machine learning-based multi-agent formation control method described in this invention.

[0065] Figure 2 A flowchart for machine learning pattern recognition;

[0066] Figure 3 A scatter plot showing the correlation between communication quality and overall performance indicators of multi-agent formations;

[0067] Figure 4 A flowchart for decomposing control instructions;

[0068] Figure 5 This is a graph showing the dynamic changes in agent reliability coefficients and cooperative parameters during a multi-agent formation control cycle. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides a multi-agent formation control method based on machine learning. The method includes: establishing a cooperative control framework for the multi-agent formation, which defines the information interaction rules and decision-making hierarchy among the agents; collecting operational status data of each agent within the formation based on the cooperative control framework, including position, heading, and communication status; using a machine learning model to perform pattern recognition on the collected operational status data and extract feature sequences reflecting the overall stability of the formation; evaluating the real-time performance of the formation control based on the dynamic changes of the feature sequences and generating a performance evaluation report; initiating a strategy optimization phase based on the performance evaluation report, whereby adaptive control commands are generated based on historical control data and real-time feature sequences; decomposing the adaptive control commands into individual control tasks corresponding to each agent; distributing individual control tasks to each agent and receiving status confirmation information fed back by the agents after executing the tasks; integrating all status confirmation information and updating the parameters of the overall formation control model; performing iterative calculations for the next control cycle based on the updated control model parameters; and outputting the final multi-agent formation control sequence after completing the iterative calculations.

[0071] Example 1: In specific implementation, the role types of agents in the formation are defined, including a lead agent and follower agents. The lead agent is responsible for guiding the overall movement direction of the formation, while the follower agents adjust their own state according to the instructions of the lead agent. A data communication protocol is set between the lead agent and the follower agents, specifying the data transmission format and response time limit. The data transmission format adopts a structured message form, including field identifiers and checksums, and the response time limit is dynamically adjusted according to the communication distance. A formation communication network based on a topology structure is constructed, which determines the direct communication links and indirect communication paths between agents. The topology structure uses a graph theory model to represent the node connection relationships. A decision center is established within the collaborative control framework. The decision center is responsible for summarizing global information and generating top-level control strategies. The decision center runs on a central server or distributed nodes. In some embodiments, the definition of role types can be extended to a multi-level leader-follower structure, with the lead agent further divided into a main lead agent and an auxiliary lead agent to adapt to complex formation configurations.

[0072] In practice, navigation data reported by the lead agent is periodically acquired through the formation communication network. This navigation data includes GPS coordinates and inertial measurement unit (INS) readings. Simultaneously, state update messages from following agents are received. These messages contain relative positions and motion vectors, with the relative positions expressed in polar or Cartesian coordinates. The navigation data and state update messages are timestamped to form a synchronized operational state data set. This timestamping is synchronized using a network time protocol. The operational state data set is then standardized to eliminate dimensional differences between data sources from different agents. This standardization is achieved through linear scaling, with the specific formula as follows:

[0073]

[0074] in: Represents the original data value. and These are the minimum and maximum values ​​in the dataset, respectively. This is the standardized result. Optionally, the standardization transformation can also use the Z-score method, calculated based on the mean and standard deviation. In some embodiments, timestamp alignment processing can introduce interpolation algorithms to compensate for transmission jitter and improve data synchronization accuracy.

[0075] Example 2: See Figure 2 In specific implementation, the standardized operational status data set is input into a pre-trained time-series analysis model. This model can employ recurrent neural network structures such as Long Short-Term Memory (LSTM) networks or gated recurrent units (GRUs). The time-series analysis model extracts the fluctuation characteristics of the operational status data over time. These fluctuation characteristics include periodic changes in the agent's position sequence, trend drifts in heading angles, and instantaneous abrupt changes in communication signal strength. The fluctuation characteristics are matched against a pre-defined formation pattern template, which stores a library of standard feature patterns corresponding to typical formations such as linear queues, triangular arrays, or circular encirclements. After identifying the current formation operation mode, the corresponding feature sequence is output. The feature sequence is in multi-dimensional vector form, with the vector dimension consistent with the feature dimension. In some embodiments, the pre-training process of the time-series analysis model uses historical formation operation data as the training set and employs a time-series prediction task for supervised learning.

[0076] In practice, the deviation between the feature sequence and the ideal formation state features is calculated. The ideal formation state features are predefined in the feature space according to task requirements. The deviation is calculated using a distance metric in the feature space, with the specific formula as follows:

[0077]

[0078] in: Indicates the degree of deviation. For feature dimension, For the i-th dimension component of the current feature sequence, Let i be the i-th dimension component of the ideal formation state feature. The weight coefficients for the i-th feature are denoted as . A comprehensive performance index is calculated by combining the real-time connectivity quality of the formation communication network. Real-time connectivity quality is quantified by packet success rate and latency. The comprehensive performance index calculation weights and fuses deviation and connectivity quality parameters to generate a scalar evaluation value. The stability level of formation control performance is determined based on the trend of the comprehensive performance index, and the stability level is divided into multiple discrete levels. A performance evaluation report containing deviation, comprehensive performance index, and stability level is generated and stored in a structured data format. Optionally, the deviation can also be calculated using the Mahalanobis distance algorithm. When using the Mahalanobis distance algorithm in deviation calculation, it is necessary to estimate the statistical characteristics of the feature sequence in the multi-dimensional feature space, including the covariance relationship between each dimension, based on historical formation operation data or a preset ideal formation state feature distribution. The implementation process of the Mahalanobis distance algorithm first calculates the mean vector and covariance matrix of the ideal formation state features, and then compares the current real-time feature sequence vector with the ideal state, obtaining the deviation by measuring the distance between the two in the Mahalanobis space. In some embodiments, a sliding window mechanism can be introduced to determine the stability level, and the statistical characteristics of the comprehensive performance index within the time window can be analyzed.

[0079] See Figure 3 This graph uses communication quality as the horizontal axis and overall performance index as the vertical axis, illustrating the relationship between the two through scatter plots and presenting the overall pattern with a trend line. Communication quality and overall performance show a positive correlation: as communication quality improves, the overall performance index gradually increases. This aligns with the design logic of integrating deviation and connectivity quality into the overall performance index. Communication quality is the foundation of formation coordination, and its improvement directly enhances the overall performance of formation control. The fluctuations in the scatter plots reflect the complexity of real-world scenarios: even with similar communication quality, overall performance can differ due to factors such as agent position deviation and heading stability, demonstrating that the overall performance index is a weighted fusion of multi-dimensional parameters. The upward slope of the trend line visually reflects the contribution weight of communication quality to overall performance, serving as a key visual basis for evaluating the stability of formation control performance.

[0080] Example 3: In specific implementation, when the stability level in the performance evaluation report is lower than a set threshold, the control strategy optimization process is triggered. The set threshold is a configurable numerical parameter; for example, optimization is triggered when the stability level is lower than the "good" level. Historical scenarios similar to the current formation operation mode and corresponding successful control strategies are retrieved from the historical database. The historical database stores feature sequences, control commands, and their execution effects recorded in past tasks. The retrieval process is based on the similarity measure of the feature sequences, identifying several records in the historical scenarios that are closest to the current formation operation mode. Using reinforcement learning algorithms in machine learning, a candidate control command set is generated by combining real-time feature sequences and historical control strategies. The reinforcement learning algorithm uses the current formation state as input and historical successful strategies as prior knowledge for exploration. The candidate control command set is simulated and verified to select the optimal adaptive control command. The simulation verifies the command execution effect in a digital twin environment. In some embodiments, content-based retrieval technology is used when retrieving similar historical scenarios, calculating the Euclidean distance between the current feature sequence and the historical feature sequence as a similarity measure.

[0081] In its implementation, the reinforcement learning algorithm employs an Actor-Critic framework. The Actor network is responsible for generating candidate control instructions, while the Critic network is responsible for evaluating the expected value of the instructions. The generation process of the candidate control instruction set follows the policy gradient method, exploring the neighborhood space surrounding historically successful control policies. The simulation verification stage evaluates the execution effect of each candidate control instruction in a multi-agent simulation environment. The effect evaluation is based on a predefined cost function, calculated as follows:

[0082]

[0083] in: Represents the total cost. This indicates a deviation in formation maintenance. Indicates overall energy consumption control. Indicates the task response time. These are the weighting coefficients for each component. The principle for selecting the optimal adaptive control command is to choose one that minimizes the total cost in simulation verification. The smallest candidate instruction. Optionally, random noise interference can be introduced into the simulation verification environment to test the robustness of the control instructions. In some embodiments, the reward function design of the reinforcement learning algorithm can comprehensively consider formation stability, communication efficiency, and energy consumption.

[0084] Example 4: See Figure 4In practical implementation, the overall control objective of the adaptive control command is analyzed, and the contribution weight of each agent in the objective is determined. For example, the overall control objective is to transform the current sector formation into a linear formation. The contribution weight is calculated based on the agent's role in the formation and its relative position to the target formation. Based on the contribution weight, the overall control objective is quantified into the expected state change for each agent. The expected state change includes the displacement vector, velocity change, and heading angle correction that need to be adjusted. Combined with the dynamic constraints of each agent, the expected state change is converted into executable individual control tasks. The dynamic constraints include the agent's maximum speed, minimum turning radius, and acceleration limit. Each individual control task is assigned a priority and execution time limit. High-priority individual control tasks involve key agents maintaining the overall stability of the formation, and the execution time limit is determined based on the task's urgency and communication latency. Refer to Table 1 for the specific attributes of individual control tasks.

[0085] Table 1: Individual Control Task Attribute Table

[0086]

[0087] In practical implementation, individual control tasks are sent to the corresponding agents via a formation communication network, employing a reliable transmission protocol to ensure data packet arrival. A task execution timer is started to monitor whether the agents send back status confirmation information within the execution time limit. This status confirmation information includes task reception confirmation and preliminary feasibility verification results. If no status confirmation information is received within the time limit, the individual control task is resent via a backup communication path, which may be implemented through other relay agents or multi-hop routing. Log information for all agent task distribution is recorded, including task sending timestamps, reception status, and retransmission counts. The pre-evaluation formula for task execution consistency can be expressed as:

[0088]

[0089] in: Indicates the estimated degree of agreement. This represents the agent's self-assessment value, which is the feedback from the agent. This indicates the capability threshold required for the task. It is understood that the reliability of task distribution directly impacts the overall effectiveness of formation coordination control. In some embodiments, priority marking can be further subdivided into multiple levels and associated with task scheduling policies. Optionally, the execution time limit setting can be dynamically adjusted considering the current network load.

[0090] Example 5: In specific implementation, the completeness and consistency of the received state confirmation information are verified. The state confirmation information includes agent identifiers, task execution result codes, and timestamp data. Integrity verification checks whether all data fields are complete, and consistency verification compares the descriptions of the same spatiotemporal state in feedback information from different agents to check for logical conflicts. The state confirmation information is compared with the expected goals of the individual control task to calculate the task execution consistency. The expected goals are derived from the expected state change amount defined in the individual control task. Based on the task execution consistency, the agent reliability coefficients in the overall formation control model are adjusted. The agent reliability coefficient is a dynamically updated parameter used to reflect the historical performance of the agent in executing instructions. The latest reliability coefficients are used to update the coordination parameters of the control model. These coordination parameters are used to adjust the logic regarding the allocation of coordination weights among agents in the formation control algorithm. Task execution consistency. The calculation formula is:

[0091]

[0092] in: Indicates the degree of agreement between task execution and performance. The number of state dimensions. This represents the j-th dimension state value actually reached by the agent. Let represent the j-th dimension of the expected state value for the individual's control task. It can be understood that calculating the task execution fit provides a quantitative basis for evaluating the agent's actual performance.

[0093] In practical implementation, based on the updated control model parameters, the formation's operational status data is re-collected, and a new round of data requests and collection processes are initiated through the formation communication network. The re-collected operational status data is input into the time series analysis model to generate a new feature sequence. The time series analysis model and feature sequence generation method remain consistent with the previous implementation phase. Performance evaluation is performed using the new feature sequence, and a new performance evaluation report is generated. The performance evaluation report includes updated deviation, comprehensive performance indicators, and stability level information. Based on the new performance evaluation report, a decision is made as to whether further strategy optimization is needed, thus forming a closed-loop control cycle. The decision logic is based on the comparison result between the stability level and a preset threshold. In some embodiments, the agent reliability coefficient can be adjusted using a sliding window averaging method, focusing on the most recent task execution records. Optionally, the cycle length of the closed-loop control cycle can be dynamically configured according to the system's real-time requirements.

[0094] See Figure 5This graph, with the control cycle as the horizontal axis, simultaneously illustrates the changes in two core indicators: the left vertical axis represents the overall reliability coefficient of the agents, and the right vertical axis represents the formation coordination parameters. The coordination parameters show a stable, linear upward trend, indicating that the coordination mechanism of the agents within the formation is continuously optimized as the control cycle progresses. The reliability coefficient, on the other hand, shows a fluctuating upward trend, reflecting both the gradual improvement in the agents' performance with control iterations and the dynamic fluctuations in the agents' states in real-world scenarios. When the reliability coefficient fluctuates upward, the coordination parameters increase synchronously, confirming the logic of using the agent reliability coefficient to update the control model's coordination parameters. The higher the agent's execution reliability, the more reasonable the formation's coordination weight allocation, ultimately supporting the closed-loop optimization of formation control. This graph visually presents the dynamic iteration process of the model parameters and is a key visualization of the multi-agent formation's perception-decision-execution-learning closed loop.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-agent formation control method based on machine learning, characterized in that, The method includes: A collaborative control framework for a multi-agent formation is established, wherein the collaborative control framework defines the information interaction rules and decision-making hierarchy among the agents; Based on the cooperative control framework, the operational status data of each intelligent agent in the formation is collected, including position, heading and communication status. Machine learning models are used to perform pattern recognition on the collected operational status data and extract feature sequences that reflect the overall stability of the formation. The real-time performance of formation control is evaluated based on the dynamic changes of the feature sequence, and a performance evaluation report is generated. The strategy optimization phase is initiated based on the performance evaluation report. The strategy optimization phase generates adaptive control instructions based on historical control data and real-time feature sequences. The adaptive control instructions are decomposed into individual control tasks corresponding to each agent; Distribute individual control tasks to each agent and receive status confirmation information from the agents after they have executed the tasks; Integrate all status confirmation information and update the parameters of the overall formation control model; The next control cycle is iteratively calculated based on the updated control model parameters; After completing the iterative calculations, the final multi-agent formation control sequence is output.

2. The multi-agent formation control method based on machine learning according to claim 1, characterized in that, The aforementioned collaborative control framework for establishing a multi-agent formation includes: Define the role types of agents in the formation, including leader agents and follower agents; A data communication protocol is defined between the leading agent and the following agent, and the data communication protocol specifies the data transmission format and response time limit; Construct a topology-based formation communication network, wherein the formation communication network determines the direct communication links and indirect communication paths between agents; A decision center is established within the collaborative control framework. This decision center is responsible for aggregating global information and generating top-level control strategies.

3. The multi-agent formation control method based on machine learning according to claim 2, characterized in that, The collected operational status data of each agent within the formation includes: The navigation data reported by the navigator is periodically acquired through the formation communication network; Simultaneously, it receives state update messages from the following agent, the state update messages containing relative position and motion vector; The navigation data and status update messages are timestamped to form a synchronized set of running status data. The operational status data set is standardized and transformed to eliminate the dimensional differences between data sources from different agents.

4. The multi-agent formation control method based on machine learning according to claim 3, characterized in that, The process of using machine learning models to perform pattern recognition on the collected operational status data includes: The standardized set of operational status data is input into a pre-trained time series analysis model; Use a time series analysis model to extract the fluctuation characteristics of operational status data in the time dimension; The fluctuation characteristics are matched with the preset formation mode template to identify the current formation operation mode; Output the feature sequence corresponding to the formation operation mode, and the feature sequence is in the form of a multi-dimensional vector.

5. The multi-agent formation control method based on machine learning according to claim 4, characterized in that, The evaluation of the real-time performance of formation control includes: Calculate the deviation between the feature sequence and the features of the ideal formation state; Calculate the comprehensive performance index by combining the real-time connectivity quality of the formation communication network; The stability level of formation control performance is determined based on the changing trend of comprehensive performance indicators; Generate a performance evaluation report that includes deviation, comprehensive performance indicators, and stability level.

6. The multi-agent formation control method based on machine learning according to claim 5, characterized in that, The startup strategy optimization phase includes: When the stability level in the performance evaluation report is lower than the set threshold, the control strategy optimization process is triggered. Retrieve historical scenarios and corresponding successful control strategies that are similar to the current formation operation mode from the historical database; By using reinforcement learning algorithms in machine learning, candidate control instruction sets are generated by combining real-time feature sequences with historical control strategies. The candidate control instruction set is simulated and verified to select the optimal adaptive control instruction.

7. The multi-agent formation control method based on machine learning according to claim 6, characterized in that, The process of decomposing adaptive control instructions into individual control tasks corresponding to each agent includes: Analyze the overall control objective of adaptive control instructions and determine the contribution weight of each agent in the objective; Based on the contribution weight, the overall control objective is quantified into the expected state change of each agent. By combining the dynamic constraints of each agent, the expected state change is converted into an executable individual control task; Each individual control task is assigned a priority and execution time limit.

8. The multi-agent formation control method based on machine learning according to claim 7, characterized in that, The distribution of individual control tasks to each agent includes: Individual control tasks are sent to the corresponding intelligent agents via a formation communication network; Start the task execution timer and monitor whether the intelligent agent sends back status confirmation information within the execution time limit; If no status confirmation is received within the timeout period, the individual control task will be resent via the backup communication path. Record log information for all agent task distributions.

9. The multi-agent formation control method based on machine learning according to claim 8, characterized in that, The process of integrating all status confirmation information and updating the overall formation control model parameters includes: Verify the completeness and consistency of the received status confirmation information; The status confirmation information is compared with the expected goals of the individual control task to calculate the task execution consistency. Adjust the agent reliability coefficients in the overall formation control model based on the task execution consistency, and update the cooperative parameters of the control model using the latest reliability coefficients.

10. A multi-agent formation control method based on machine learning according to claim 9, characterized in that, The iterative calculation for the next control cycle includes: Re-collect formation operation status data based on the updated control model parameters; The newly collected operational status data is input into the time series analysis model to generate new feature sequences; The new feature sequences are used for performance evaluation, and a new round of performance evaluation reports are generated. The decision on whether further strategy optimization is needed will be made based on the new performance evaluation report, thus forming a closed-loop control cycle.

Images