Cluster algorithm training evolution system and method based on architecture model definition

Abstract

The cluster algorithm training and evolution system and method defined based on the architecture model include: a data acquisition unit that collects task objectives and task resources; an architecture unit that analyzes task resources based on the task objectives and outputs an architecture model; an algorithm training unit that performs algorithm training based on the architecture model and outputs a cluster decision-making strategy and task allocation scheme; a simulation evaluation unit that simulates and deduces the cluster decision-making strategy, evaluates the effectiveness of the task allocation scheme, and outputs a corresponding evaluation report; and a result output unit that outputs the evaluation report. The present invention first uniformly defines the functions and interfaces of multiple task resources using the architecture model to ensure compatibility, scalability, and interoperability among different task resources. Secondly, the architecture model is combined with the cluster decision-making algorithm to optimize decision-making efficiency and improve traceability and transparency.

Description

Technical Field

[0001] The present invention relates to the technical field of attack and defense confrontation simulation and algorithm training, and more specifically, to a cluster algorithm training evolution system and method based on system architecture model definition. Background Art

[0002] As modern warfare continues to evolve toward information technology and intelligentization, coordinated operations involving multiple arms and multiple weapon systems are becoming increasingly mainstream. The rapid deployment and joint operations of swarm drones, autonomous ground platforms, and long-range strike weapons in actual combat scenarios require the rapid generation or iterative optimization of cluster decision-making and control algorithms to cope with the dynamically changing battlefield environment. Traditional machine learning methods based on pure rules or limited to a single scenario often lack unified system standards, making it difficult to organically integrate the functional interfaces, operational logic, and operational processes of different weapon systems and combat units. This results in insufficient interpretability and credibility in simulation results when large-scale, multi-system collaboration is involved.

[0003] Currently, the use of system architecture models (such as DoDAF and UPDM) to define the relationships and logic between weapon systems and combat units is gradually emerging, serving as a unified data source for training and simulation. This allows for the completion of adversarial simulations under standardized and traceable conditions. However, traditionally, research or engineering teams have independently defined command rules, communication protocols, and functional interfaces for drones (or other weapon equipment), resulting in a lack of universal, standardized descriptions across different projects and platforms. Changes in weapon systems or operational requirements require manual reconfiguration or rewriting of numerous decision rules and control parameters, significantly limiting system scalability. Furthermore, using simple rules or black-box machine learning-based algorithms often makes it difficult to clearly and comprehensively understand the specific operating modes of each weapon system at different operational phases and functional interfaces. This is particularly true in large-scale, multi-weapon system coordinated combat scenarios. This lacks a systematic "system perspective" to understand why algorithms make certain decisions, making it difficult to pinpoint and correct errors in specific processes. Furthermore, the lack of a unified model for managing operational logic definition, platform modeling, and simulation scenario setup can easily lead to data fragmentation and disconnected interfaces. The final simulation results are difficult to accurately reflect the actual command process or troop action patterns, and their reference value is limited.

[0004] Therefore, the problems existing in the prior art need to be further improved and developed. Summary of the Invention

[0005] (1) Purpose of the Invention: To address the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a cluster algorithm training and evolution system and method based on the definition of an architecture model, aiming to address the deficiencies of existing cluster control, cluster decision-making and weapon system behavior logic algorithms in practical applications.

[0006] (2) Technical Solution: To solve the above technical problems, this technical solution provides a cluster algorithm training and evolution system based on the definition of the system architecture model, including a data acquisition unit, a system architecture unit, an algorithm training unit, a simulation evaluation unit, and a result output unit; the data acquisition unit collects task objectives and task resources;

[0007] The system architecture unit analyzes task resources according to task objectives and outputs a system architecture model;

[0008] The algorithm training unit performs algorithm training according to the system architecture model and outputs cluster decision-making strategies and task allocation plans;

[0009] The simulation evaluation unit simulates and deduces the cluster decision-making strategy, evaluates the effect of the task allocation scheme, and outputs a corresponding evaluation report;

[0010] The result output unit outputs an evaluation report.

[0011] Preferably, the data acquisition unit includes a resource acquisition module and a target input module; the resource acquisition module acquires relevant information data of the task resources, including the quantity, type, status, current location, energy, load, etc. of the task resources, initializes the relevant information data of the task resources, and obtains the corresponding task resource vector;

[0012] The target input module collects relevant information data of the task target, including task requirement priority, task requirement location, etc., and performs initialization processing on the relevant information data of the task target to obtain a corresponding task target vector.

[0013] Preferably, the task resource vector is [quantity, type, status, longitude normalization, latitude normalization, energy, load]; the task target vector is expressed as [priority, longitude normalization, latitude normalization].

[0014] Preferably, the system architecture unit includes a model building module and a model optimization module; the model building module establishes the framework of the system architecture model based on the mission objectives and mission resources: the system architecture pre-model. The system architecture pre-model includes functional components, interface protocols, communication relationships, tactics and strategies, state transition rules, etc., covering the interface compatibility and collaborative communication relationship between mission resources;

[0015] The model optimization module dynamically verifies the system architecture pre-model, including the correctness of task resources, interface compatibility, and optimization of tactical logic, to obtain an optimized system architecture model: a system architecture optimization model.

[0016] Preferably, the model building module analyzes the mission objective scenarios and specific tasks, and then constructs a pre-architecture model based on the mission resources. This includes key steps such as mission resource capability analysis, architecture performance design, and task allocation. By conducting a detailed analysis of the activity sequence and state transition rules within different mission objectives, a pre-architecture model encompassing multiple views is designed.

[0017] Preferably, the model optimization module adopts a self-verifying logic verification method, performs algorithm training according to the system architecture pre-model, obtains the cluster decision strategy and task allocation plan, the simulation evaluation unit evaluates the effect of the allocation plan, and the evaluation result is fed back to the model optimization module to perform adaptive training and adjustment on the system architecture pre-model; finally, the correctness decision of the system architecture pre-model is continuously adjusted according to the real-time feedback of task resources and task objectives, forming a closed-loop feedback mechanism to obtain the system architecture optimization model.

[0018] Preferably, the algorithm training unit regards different types of task resources as different intelligent agents, and learns to make optimal decisions in complex environments by repeatedly interacting with different environments of task resources, specifically including:

[0019] Step 301: Conduct agent modeling for different task resources;

[0020] Step 302: Train the agent's behavior through a reward mechanism;

[0021] Step 303: Optimize the behavior of each agent;

[0022] Step 304: Optimize and adjust the behavior of each agent in the cluster to obtain a cluster decision strategy.

[0023] Preferably, each agent receives corresponding rewards or penalties based on the actions it takes and the feedback from the environment; when the agent takes the right action, it is given positive rewards, and when it makes an incorrect decision, it is punished.

[0024] R t =λ1·Task completion reward -λ2·Resource consumption penalty -λ3·Task failure penalty,

[0025] Among them, λ1 is the reward coefficient, which is the feedback obtained by the agent from the environment after taking a certain action; λ2 and λ3 are penalty coefficients;

[0026] Based on historical task execution data, the Q value is updated and the strategy is adjusted. In the Q-learning algorithm, the Q value represents the expected total reward that can be obtained after taking a certain action and continuing to execute it in a certain state.

[0027] ,

[0028] in, In state Next, execute the action Q value; α is the learning rate; γ is the discount factor; r t is the current reward; t is the current state; t is the current action; s t+1 is the next state; ɑ' is the next action; For the next state All possible actions The maximum Q value in the equation represents the expectation of the maximum possible return in the future;

[0029] The decision-making strategy is the rule or function for an agent to select actions in a given state. The agent will adjust its behavior based on the size of the Q value and select the optimal task allocation plan, thereby making decisions on the task resources in the cluster and the corresponding task allocation to obtain the cluster decision-making strategy. The larger the Q value, the better the task allocation plan has performed in historical experience, and the better the agent decision-making effect.

[0030] Preferably, the simulation evaluation unit realizes the coordination and collaboration between different task resources, specifically including:

[0031] Step 401: The simulation evaluation unit matches the task resources with the task objectives according to the status of the task resources and the requirements of the task objectives.

[0032] Step 402: The simulation evaluation unit globally coordinates and collaboratively controls task resources.

[0033] The cluster algorithm training and evolution method based on the system architecture model definition is applicable to the cluster algorithm training and evolution system based on the system architecture model definition, and includes the following steps:

[0034] Step 1: The data collection unit collects task objectives and task resources;

[0035] Step 2: The system architecture unit analyzes the task resources according to the task objectives and outputs the system architecture model;

[0036] Step 3: The algorithm training unit performs algorithm training based on the system architecture model and outputs cluster decision-making strategies and task allocation plans;

[0037] Step 4: The simulation evaluation unit simulates and deduces the cluster decision-making strategy, evaluates the effect of the task allocation scheme, and outputs a corresponding evaluation report;

[0038] Step 5: The result output unit outputs the evaluation report.

[0039] (III) Beneficial effects: The present invention is based on a cluster algorithm training and evolution system and method defined by an architecture model. First, by uniformly defining the functions and interfaces of various task resources in an architecture model, the access of new equipment or new requirements is simpler and more efficient. Secondly, by systematically defining and describing task resources and task objectives through an architecture model, the compatibility, scalability and interoperability between different task resources can be ensured, and they can be quickly adapted to different task objective scenarios without the need for large-scale modifications to existing models. Furthermore, the combination of the architecture model and the cluster decision algorithm not only optimizes the decision-making efficiency of the system, but also improves its traceability and transparency, avoiding the difficulties of understanding and debugging brought about by traditional "black box" algorithms. Finally, the standardized model description and scalable algorithm training framework can quickly adapt to new task resources or task objectives, reducing the difficulty of large-scale manual adjustments to algorithms and models during system upgrades and scenario switching, and achieving efficient maintenance and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a structural diagram of the cluster algorithm training and evolution system defined based on the system architecture model of the present invention;

[0041] Figure 2 It is a flowchart of the steps of the cluster algorithm training evolution method defined based on the system architecture model of the present invention;

[0042] Figure 3 It is a schematic diagram of the framework of the system architecture pre-model of the present invention;

[0043] Figure 4 It is an operation flow chart of the system architecture optimization model of the present invention;

[0044] Figure 5 It is an operation flow chart of the cluster algorithm training evolution method defined based on the system architecture model of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described in detail below in conjunction with preferred embodiments. More details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can obviously be implemented in a variety of other ways different from the description. Those skilled in the art can make similar generalizations and deductions based on actual application situations without violating the connotation of the present invention. Therefore, the scope of protection of the present invention should not be limited by the content of this specific embodiment.

[0046] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that the drawings are merely examples and are not drawn to scale, and should not be used to limit the actual scope of protection claimed in the present invention.

[0047] Cluster algorithm training evolution system defined based on architecture model, such as Figure 1 As shown, the system includes a data acquisition unit, an architecture unit, an algorithm training unit, a simulation evaluation unit, and a result output unit. The data acquisition unit collects task objectives and task resources. The architecture unit analyzes task resources based on the task objectives and outputs a system architecture model. The algorithm training unit trains the algorithm based on the system architecture model and outputs a cluster decision strategy and task allocation plan. The simulation evaluation unit simulates and deduces the cluster decision strategy, evaluates the effectiveness of the task allocation plan, and outputs a corresponding evaluation report. The result output unit outputs the evaluation report.

[0048] The cluster algorithm training evolution method based on the system architecture model definition is applicable to the cluster algorithm training evolution system based on the system architecture model definition, such as Figure 2 、 Figure 5 As shown, the specific steps include:

[0049] Step 1: The data collection unit collects task objectives and task resources.

[0050] Step 2: The system architecture unit analyzes the task resources according to the task objectives and outputs the system architecture model.

[0051] Step 3: The algorithm training unit performs algorithm training based on the system architecture model and outputs cluster decision-making strategies and task allocation plans.

[0052] Step 4: The simulation evaluation unit simulates and deduces the cluster decision-making strategy, evaluates the effectiveness of the task allocation plan, and outputs a corresponding evaluation report.

[0053] Step 5: The result output unit outputs the evaluation report.

[0054] The data acquisition unit in step 1 includes a resource acquisition module and a target input module; the resource acquisition module acquires relevant information data of the task resources, including the quantity, type, status, current location, energy, load, etc. of the task resources, initializes the relevant information data of the task resources, and obtains the corresponding task resource vector. Specifically, it includes:

[0055] Step 1011: Obtain relevant information data of task resources from the task resource management system or sensors.

[0056] Specifically, this includes the number of mission resources, which may be a scalar value, such as the number of drones or vehicles. The type of mission resource, such as drones or ground platforms. The current status of the mission resource, such as "standby," "on mission," or "damaged." The current geographic location of the mission resource, such as GPS coordinates (longitude and latitude). The remaining energy or battery level of the mission resource, such as a percentage. The carrying capacity of the mission resource, such as the payload of a drone or the ammunition capacity of a weapon system.

[0057] Step 1012: Initialize the relevant information data of each task resource to obtain corresponding normalized task resource data.

[0058] Specifically, the quantity data of mission resources is retained as a scalar value so that it can be used directly when converted to a vector. Each type of mission resource is assigned a unique integer code, for example, a drone is coded as 1, a ground platform is coded as 2, and so on, and then converted into a numeric variable. The status information of mission resources can be represented by a category code, for example, "standby" is coded as 0, "mission in progress" is coded as 1, and "damaged" is coded as 2. Normalization can be performed based on the geographic coordinates (longitude and latitude) of the mission resources. For example, longitude and latitude can be normalized to the range [0, 1]. The remaining energy value of the mission resource is normalized to the range [0, 1], typically expressed as the ratio of the remaining power of the resource to the maximum power. Normalization is performed based on the payload capacity of the mission resource, typically expressed as the ratio of the payload weight to the maximum load capacity.

[0059] Step 1013: Convert the normalized resource data into a unified task resource vector. The task resource vector is represented as [quantity, type, status, normalized longitude, normalized latitude, energy, load].

[0060] For example, consider a drone resource with the following data: Quantity = 3. Type = drone (coded as 1). Status = executing mission (coded as 1). Current location = (45.0, 85.0) (longitude and latitude). Energy = 80% (converted to 0.8). Payload = 50% (converted to 0.5). The corresponding mission resource vector is [Quantity, Type, Status, Longitude Normalized, Latitude Normalized, Energy, Payload] = [3, 1, 1, 0.75, 0.55, 0.8, 0.5].

[0061] The target input module collects relevant information data of the task target, including the task requirement priority, task requirement location, etc., initializes the relevant information data of the task target, and obtains the corresponding task target vector. Specifically including:

[0062] Step 1021: Obtain relevant information data of the task target from the task management system.

[0063] Specifically, it includes: the urgency (priority) of the task target, which can be an integer value, such as 1 for high-priority tasks and 3 for low-priority tasks. The target location for the task target execution, such as GPS coordinates (longitude, latitude).

[0064] Step 1022: Initialize the relevant information data of each task target separately to obtain corresponding normalized task target data.

[0065] The mission objective priority is converted into an integer value or normalized value according to the predetermined rules. The geographical coordinates (longitude, latitude) of the mission objective's required location can be normalized in the same way as the current location of the mission resource.

[0066] Step 1023: Merge the normalized task target data into a task target vector. The task target vector is expressed as [priority, longitude normalization, latitude normalization].

[0067] For example, assuming the mission objective is: priority = 1 (high priority), required location = (50.0, 90.0) (latitude and longitude). The corresponding mission objective vector is: [priority, longitude normalization, latitude normalization] = [1, 0.80, 0.60].

[0068] By initializing and converting the relevant information data for each task resource and task objective into a unified vector form, this data can be easily input into the algorithm training unit. The vector set serves as input data for task allocation and decision optimization, supporting data processing during algorithm training and simulation deduction.

[0069] The system architecture unit in step 2 includes a model building module and a model optimization module. The model building module establishes the framework of the system architecture model based on the mission objectives and mission resources: the system architecture pre-model. The system architecture pre-model includes functional components, interface protocols, communication relationships, tactics and strategies, state transition rules, etc., covering the interface compatibility and collaborative communication relationships between mission resources.

[0070] Specifically, the scenario and specific tasks of the mission objectives are analyzed, and then the system architecture pre-model is constructed based on the mission resources, including key steps such as mission resource capability analysis, system architecture performance design, and task allocation. By conducting a detailed analysis of the activity sequence and state transition rules in different mission objectives, such as Figure 3As shown in the figure, a pre-architecture model is designed, including multiple views: the global view (AV), capability view (CV), system view (SV), project view (PV), standard view (StdV), data and information view (DIV), service view (SvcV), and simulation view (OV). The pre-architecture model is then optimized and adjusted through self-verification using the model optimization module to ensure the rationality and feasibility of each view.

[0071] The global view AV includes the architectural content of all views and is used to describe the conception, scope, purpose, actions, events, and environment of the mission objectives.

[0072] The capability view (CV) refers to the capability requirements, delivery time, and capability deployment view. Based on the system's mission objectives, it analyzes the top-level capabilities of the system architecture. For example, if the system architecture needs to perform target perception tasks, it should possess target perception capabilities such as radar and optical perception. For target strike missions, it should possess strike capabilities such as air strike and surface strike. These capabilities can be gradually refined. For example, optical perception can be divided into more detailed sub-capabilities such as visible light perception and infrared perception. Through capability analysis, a complete system capability view model can be constructed to demonstrate the system's capability conception, capability decomposition, and capability dependencies.

[0073] The system view (SV) refers to mission resources, processes, activities, and their interrelationships. Based on the mission objectives and activities derived from the system architecture process analysis, after selecting a weapon ratio plan, consider how different weapon systems will cooperate to achieve the corresponding mission objectives. By designing activities and time sequences, the state transition rules, interface relationships, and interaction information of each weapon system are sorted out, leading to the functional requirements and mission analysis of the weapon system.

[0074] The project view PV refers to the relationship between task resources, task objectives and implementation projects.

[0075] The standard view StdV refers to policy standards, guidelines, restrictions and expectations. The data and information view DIV refers to the data relationship between system architectures.

[0076] The service view SvcV includes executed task resources, activities, services and their relationships.

[0077] The simulation view (OV) includes task resources, processes, task resource activities, and the resource requirements for each task objective. Simulation process analysis is based on the system's task objectives, the results of simulation task analysis, and the results of task resource capacity analysis. It links task objectives with various task resources, considers how different task resources will cooperate to achieve the corresponding task objectives, and designs the system's simulation process. Considering adversarial factors in the simulation process makes it more realistic and meaningful, ultimately deriving activity, timing, interface relationships, and state transition rules.

[0078] The text input of the task scenario, composition, and goal of the system architecture pre-model is static. The system architecture pre-model is used to achieve the coordination and management of multiple task resources and task goals.

[0079] The system architecture pre-model complies with international standards and provides a standardized framework to ensure compatibility and integration between multiple task resources. The system architecture pre-model ensures operational consistency and standardized interfaces, providing the necessary framework support for subsequent cluster algorithm task allocation and collaborative control.

[0080] The system architecture pre-model preferably adopts standards such as DoDAF and UPDM; however, other architectural frameworks such as MODAF, TOGAF, SysML, or UML Profile can also be used for system modeling and interface definition. The specific choice is not limited as long as it can provide a standardized description of the functional structure, interface relationships, and business processes of the multi-weapon system and enable data exchange with the algorithm training and simulation platforms.

[0081] The model optimization module dynamically verifies the architecture pre-model, including the correctness of task resources, interface compatibility, and optimization of tactical logic, to obtain an optimized architecture model: the architecture optimization model. The model optimization module optimizes based on the feedback results to improve the adaptability of the architecture model in complex environments.

[0082] Self-verification is the main logic verification method of the model optimization module of the present invention. Figure 4 As shown, algorithm training is performed based on the architecture pre-model to obtain a cluster decision-making strategy and task allocation plan. The simulation evaluation unit then evaluates the effectiveness of the allocation plan, and the evaluation results are fed back to the model optimization module to adaptively train and adjust the architecture pre-model. Finally, through updates to the adaptive computational training process, the correctness of the architecture pre-model's decisions can be continuously adjusted based on real-time feedback from new task resources and objectives, forming a closed-loop feedback mechanism and ultimately achieving the architecture optimization model.

[0083] The self-verification described in this embodiment is sufficient to standardize the standardization and logical correctness of the input data used for algorithm training, improve the interpretability of the algorithm model, etc. According to the actual use process, the model optimization module includes a variety of optimization and adjustment methods, which are not specifically limited in the present invention.

[0084] The architecture optimization model serves as the sole data source for the entire algorithm training and simulation evaluation process, defining and managing entities, attributes, constraints, and more throughout the entire process. The design and verification of the architecture model provides a unified data source and standardized interface for subsequent algorithm training and simulation deduction.

[0085] The architecture optimization model is capable of modeling the logical model of the assurance requirements architecture, supporting the verification of the logical correctness of the assurance requirements design logic views using activity diagrams, state machine diagrams, and sequence diagrams. Through in-depth analysis of behavioral and parametric diagram model elements, and the ability to solve and analyze parametric equations, the software ensures the rigorous and reasonable underlying logic, the clear and smooth simulation execution, and the authenticity and reliability of the simulation analysis results during simulation analysis.

[0086] Through the system architecture unit, it is ensured that tasks and resources can be uniformly managed and scheduled at the global level, ensuring that multiple task resources can be efficiently coordinated and executed.

[0087] The algorithm training unit in step 3 trains different task resources, continuously optimizing the behavior strategies of each task resource. By continuously optimizing the behavior strategies of each task resource in the training cluster, it outputs a cluster decision-making strategy and task allocation plan that meets the task objectives. This invention uses deep reinforcement learning (DRL) technology to enhance the intelligence and adaptability of the algorithm training unit. Through multi-agent reinforcement learning, each task resource acts as an independent agent to perform tasks in different environments, and optimizes the cluster decision-making strategy through interaction with the environment.

[0088] The algorithm training unit uniformly defines the functional composition, interface relationships, and operational logic of task resources. It treats different types of task resources as different intelligent agents, and through repeated interactions with different environments of task resources, it learns to make optimal decisions in complex environments.

[0089] The algorithm training unit provides a training framework, including reward mechanisms, experience replay, and policy updates, enabling agents to continuously optimize their behavioral strategies during training. During the training phase, the algorithm training unit utilizes a distributed training and centralized evaluation approach to ensure that agents can gradually converge over multiple rounds of training and form an efficient cluster collaborative decision-making strategy.

[0090] The specific steps include:

[0091] Step 301: Conduct agent modeling for different task resources.

[0092] Each different type of mission resource (such as drones and ground platforms) is defined as an independent agent. Each agent selects and executes an appropriate action based on its local observations (such as its position and enemy and friendly situation). Agent actions include, but are not limited to, movement, attack, and reconnaissance. Through agent modeling, agents can dynamically adjust their strategies based on real-time battlefield conditions, enhancing the flexibility and responsiveness of swarm decision-making strategies and task allocation.

[0093] Step 302: Train the agent's behavior through a reward mechanism.

[0094] Each agent receives rewards or penalties based on the actions it takes and the feedback from the environment. Positive rewards are given when the agent takes the right action, and penalties are given when the agent makes an incorrect decision, thereby guiding the agent to continuously adjust its decision-making strategy. This is shown in Formula (1):

[0095] R t =λ1·Task completion reward -λ2·Resource consumption penalty -λ3·Task failure penalty (1),

[0096] Among them, λ1 is the reward coefficient, which is the feedback the agent receives from the environment after taking an action. It helps the agent understand the pros and cons of the current behavior; λ2 and λ3 are penalty coefficients. The task completion reward and resource consumption penalty are dynamically adjusted according to the task execution status.

[0097] In a task allocation scenario, when an agent (such as a drone) successfully completes a task, it receives a positive reward; if it fails or consumes too many resources, it receives a penalty signal. The reward signal tells the agent which actions are effective and which are disadvantageous.

[0098] Reward signals are core information in reinforcement learning, helping agents learn which behaviors contribute to achieving their goals. The design of the reward function is crucial, determining the system's learning efficiency and decision-making optimization effectiveness. By designing a well-designed reward mechanism, agents can continuously adjust their behavior to maximize their mission objectives while avoiding becoming trapped in local optima in complex environments.

[0099] Step 303: Optimize the behavior of each agent.

[0100] Through continuous interaction between the agent and the environment, the action strategy of each agent is optimized; then, through experience replay and target network in deep reinforcement learning, the agent is helped to extract valuable information from historical experience and accelerate the convergence of decision-making strategies.

[0101] Valuable information refers to historical data, experience, or feedback that improves the agent's decision-making and optimization strategies. In deep reinforcement learning, valuable information typically refers to various data that contribute to strategy optimization and can drive the agent's learning and decision-making. This includes reward signals, historical experience, state information, Q-values, and policy update information. This information, derived from the agent's interaction with the environment, accumulates over time and provides effective learning signals for the agent, helping it gradually improve its decision-making strategy and thus learn to make better decisions in complex and changing environments. Through techniques such as experience replay and target networks, the agent can draw valuable information from historical experience, reduce randomness, and accelerate policy convergence.

[0102] The target network is a technique used in deep reinforcement learning to stabilize the training process. It is a delayed version of the current training network and is typically used to calculate target Q-values, reducing fluctuations during training. In a Deep Q-Network (DQN), the agent uses the target network to calculate the target Q-value. The target network is updated at intervals, avoiding the high-frequency fluctuations that can occur when updating network parameters at every step. The target network provides a stable training objective, helping to reduce fluctuations in policy updates during training and ensuring the stability and effectiveness of the learning process.

[0103] Historical experience refers to the sequence of states, actions, and rewards experienced by an agent during past training. If an agent successfully completed a task in a previous attempt by selecting a specific path, it would record the state transitions and rewards along that path. When faced with a similar situation, the agent can search for similar state-action pairs from this historical experience and make decisions based on this past experience. Through experience replay, the agent can leverage past experience to improve current decisions. The accumulation of historical experience helps improve decision-making efficiency and avoid repetitive exploration of inefficient strategies.

[0104] This invention uses Q-learning or the Deep Q Network (DQN) algorithm to select the optimal action based on the agent's state and task status. Based on historical task execution data, the Q value is updated and the strategy is adjusted. The Q value is the expected cumulative reward for the agent choosing a specific action in a certain state. It is a quantitative measure of the quality of an action in a specific state.

[0105] Specifically, the calculation of Q values ​​uses historical task execution data to determine different time mechanism strategies. These time mechanism strategies split the time required to calculate the Q values ​​of different agents into total computation time and agent computation time, depending on the importance of the task objectives. This allows for the optimization of the total computation time when accelerating the optimization of different agents.

[0106] Total computation time is the total time it takes for all agents to calculate Q-values ​​for each task objective. Agent computation time is the time it takes for each agent to calculate Q-values ​​for a single task objective. The total computation time is the sum of the agent computation times.

[0107] For example, when the task target is one-to-one with the graphics processing unit (GPU), that is, the number of task targets is less than or equal to the number of GPUs, the CPU allocates different task targets and the task resources (i.e., agents) under the task target to different graphics processing units (GPUs) to calculate the Q values ​​of different agents for the task target, thereby achieving a total computation time equal to the agent computation time, thereby achieving an optimized acceleration of the total computation time. However, when the task target is many-to-one with the graphics processing unit (GPU), that is, the number of task targets is greater than the number of GPUs, the same GPU may need to process two or more computational tasks. When calculating the Q value, the CPU determines the importance of the current task target based on different historical task execution data, marks the computational tasks with high importance as accelerated, and assigns them to the GPU. The GPU will then prioritize the assigned computational tasks, so that the GPU will prioritize the Q values ​​of the computational tasks corresponding to the agents with high importance. By allocating the computation time of different agents, the total computation time is optimized and accelerated. For example, an agent has 10 tasks to calculate the Q value of a task objective, but there are only three GPUs. Therefore, the first GPU is assigned four tasks, meaning it needs to perform four different Q-value calculations. The computation time for the first GPU is the sum of the computation times of the four agents. Meanwhile, the other three GPUs are each assigned three tasks, so the computation time for the other three GPUs is equal to the sum of the computation times of the three agents. Because these four GPUs can perform computations simultaneously, the total computation time is equal to the computation time of the longest GPU (the sum of the computation times of the four agents), which is much less than the CPU's computation time for all agents in sequence (the sum of the computation times of all 10 agents).

[0108] By allocating the time for Q-value calculation through the time mechanism strategy, the time for calculating Q-value for the intelligent agent in the cluster decision strategy can be greatly shortened, thereby achieving the effect of accelerating the calculation of Q-value in the entire calculation process.

[0109] In the Q-learning algorithm, the Q value represents the expected total reward obtained by taking a certain action and continuing to execute it in a certain state. The Q value can provide a basis for the agent to choose actions and guide the agent to make decisions that can maximize the cumulative reward. As shown in formula (2):

[0110] (2),

[0111] in, In state Next, execute the action Q value (action value); α is the learning rate; γ is the discount factor; r t is the current reward; t is the current state; t is the current action; s t+1 is the next state; ɑ' is the next action; For the next state All possible actions The maximum Q value in the equation represents the expectation of the maximum possible return in the future.

[0112] Step 304: Optimize and adjust the behavior of each agent in the cluster to obtain a cluster decision strategy.

[0113] The training process utilizes a distributed reinforcement learning framework, where each agent operates independently within a simulation environment but shares global information for policy coordination. Through repeated training of different agents, the agents continuously optimize their decision-making strategies, ultimately forming an optimal collaborative decision-making strategy for the swarm.

[0114] A decision-making policy is a rule or function that an agent uses to select an action in a given state. These policies can be deterministic (a specific action is assigned to each state) or stochastic (an action is selected based on a probability given a state). In swarming tasks, these policies might include selecting actions such as reconnaissance, attack, or defense at a given moment. These policies are continuously optimized through training to maximize cumulative rewards. By continuously optimizing their policies, agents can make more appropriate decisions in complex environments, thereby improving task completion efficiency.

[0115] Based on the current state, the decision strategy obtained through training is used to determine the tasks to be performed by each agent. The priority of task allocation is determined based on the size of the Q value, thereby making decisions on the task resources in the cluster and the corresponding task allocation to obtain the cluster decision strategy. As shown in formula (3),

[0116] (3),

[0117] Among them, ɑ* is the optimal action, that is, the best task allocation decision.

[0118] A larger Q-value indicates a higher expected reward for the state-action pair under the current strategy. This represents the total potential reward (i.e., expected reward) that the agent will receive if it takes a certain action in a given state. This indicates that this action is superior to other actions in that state, and the agent will be more inclined to choose it. In task allocation models, a larger Q-value indicates that the task allocation solution has historically performed well, resulting in higher cumulative rewards (e.g., task completion, resource efficiency, etc.), and thus more effective agent decisions. Therefore, the agent adjusts its actions based on the Q-value to select the optimal task allocation solution, thereby improving the efficiency and success rate of task execution.

[0119] The simulation evaluation unit in step 4 uses a cluster decision-making strategy to perform simulations, execute tasks, and plan paths. During the simulation, the platform evaluates the performance of the tasks (such as task success rate, path planning quality, and resource consumption) and generates a corresponding evaluation report.

[0120] Specifically, the simulation evaluation unit generates simulation data, including force deployment and tactical strategies, based on the system architecture model, simulating a real-world environment to evaluate the effectiveness of the swarm decision-making strategy. During the simulation, task allocation, path planning, and collaborative behavior among mission resources are adjusted through algorithmic optimization to ensure high levels of coordination among mission resources.

[0121] The present invention realizes the coordination and collaboration between different task resources through the simulation evaluation unit. The specific operation steps are as follows:

[0122] Step 401: The simulation evaluation unit matches the task resources with the task objectives according to the status of the task resources and the requirements of the task objectives.

[0123] The simulation evaluation unit calculates and generates a task allocation plan in real time based on the system architecture model and cluster decision-making algorithm. This plan is dynamically adjusted based on the status and location of task resources, including location matching, resource matching, and priority matching. Each task resource performs the corresponding operations based on the assigned tasks and objectives. Through dynamic task allocation, the simulation evaluation unit ensures efficient coordination among all task resources, avoiding task conflicts and resource waste.

[0124] To design the operational steps for dynamically adjusting task allocation based on information such as the status and location of task resources, as shown in Table 1, the following process can be described. Each step ensures that task allocation can be adjusted in real time based on the current status and location of task resources and task objectives, and effectively avoids task conflicts and resource waste.

[0125] Step 402: The simulation evaluation unit globally coordinates and collaboratively controls task resources.

[0126] During swarm algorithm training, each task resource must ensure global coordination while executing its own tasks. The simulation evaluation unit uses a global information sharing mechanism to enable agents to coordinate and cooperate with each other during the training process. By accessing global information, agents can adjust their strategies and ensure coordinated training with other agents, ultimately improving the overall training efficiency of the swarm system.

[0127] Simulation results are analyzed using metrics such as mission completion rate, resource consumption, and tactical execution effectiveness, providing feedback for iterative algorithm optimization. The evaluation results are fed back to the training platform, forming a closed-loop feedback mechanism of "training-deduction-evaluation-optimization."

[0128] The simulation evaluation unit is responsible for simulating complex mission resource environments and conducting real-time simulations. Combining the system architecture model with the output of cluster algorithm training, this unit constructs multi-scenario, multi-dimensional simulation tasks, adjusts the actions of mission resources in real time, and provides feedback based on real-time data. The simulation evaluation unit evaluates the effectiveness of cluster decisions and, through dynamic updates, helps the system continuously optimize strategies and decisions. It not only simulates enemy and friendly situations and tactical rules, but also considers physical characteristics (such as flight dynamics and weapon range) to ensure that the simulation results are highly realistic and valuable.

[0129] Table 1 Task assignment steps

[0130]

[0131] The simulation evaluation unit includes two parts: simulation analysis of behavioral graph model elements and simulation analysis of parameter graph model elements. The simulation analysis of behavioral graph model elements mainly constructs a simulation model by parsing the elements and relationships in the behavioral graph, and then uses the simulation engine to run the model. During the operation process, the software will dynamically monitor the state changes of the model, record the operation data of the model, and generate a simulation report. The simulation report records the behavior and performance of the model during the simulation process in detail, helping users understand and evaluate the behavior of the system. The simulation analysis of parameter graph model elements is to construct a simulation model by parsing the elements and relationships in the parameter graph, and then use the simulation engine to run the model. During the operation process, the software will dynamically monitor the state changes of the model, record the operation data of the model, and generate a simulation report. These reports record the behavior and performance of the model during the simulation process in detail, helping users understand and evaluate the performance of the system.

[0132] In addition, the simulation evaluation unit also supports the solution and analysis of parametric equations. During this process, the simulation evaluation unit will analyze the parametric equations and then build a calculation model based on the analysis results. After the model is built, the simulation evaluation unit will start the solution engine and run the solution model according to the logic defined in the equation. During the solution operation process, the simulation evaluation unit will dynamically monitor the state changes of the model, record the model's operation data, and generate a solution report.

[0133] The simulation engine integrates a fUML (basic UML) engine and a parameter solver. The fUML engine parses and executes SysML-based behavioral models. The parameter solver processes parametric models, parsing and solving models containing mathematical equations and expressions. It provides the ability to calculate model parameter metrics using built-in and external solvers (such as MATLAB, JavaScript, and Python). After setting appropriate constraints, calculations can be used to evaluate whether the parameter results meet requirements. Parametric diagrams make system parameters computable and simulable. This supports requirements verification and system performance demonstration, enabling reliable system design.

[0134] The result output unit in step 5 outputs and displays the evaluation report, including the task execution evaluation report, the optimized task allocation and path planning scheme, and the analysis and evaluation of the simulation results.

[0135] The optimized task allocation and path planning scheme includes adjustments to weights, learning rates, and hyperparameters. The analysis and evaluation of the simulation results include metrics such as task success rate, resource consumption, and path quality. These metrics enable a comprehensive analysis of the effectiveness of cluster decision-making and task allocation, assessing the cluster system's performance in different task environments, identifying potential optimization opportunities, and providing data support for subsequent optimization.

[0136] The following is a detailed description with reference to specific embodiments:

[0137] Assuming there are 5 drones and 3 mission targets, the allocation will be optimized based on the status of each drone (such as battery level, distance to the mission) and mission requirements (such as priority and time consumption), ensuring that high-priority tasks are executed by drones first and minimizing the time and resource consumption for mission completion.

[0138] Task resource information:

[0139] Mission Resource 1: UAV, Current Status: Mission Executing, Current Location: Longitude (45.0), Latitude (85.0), Energy: 80%, Payload: 50%. Mission Resource 2: UAV, Current Status: Standby, Current Location: Longitude (46.0), Latitude (86.0), Energy: 60%, Payload: 30%. Mission Resource 3: Ground Platform, Current Status: Mission Executing, Current Location: Longitude (47.0), Latitude (87.0), Energy: 90%, Payload: 70%. Mission Resource 4: Ground Platform, Current Status: Damaged, Current Location: Longitude (48.0), Latitude (88.0), Energy: 40%, Payload: 60%. Mission Resource 5: UAV, Current Status: Mission Executing, Current Location: Longitude (49.0), Latitude (89.0), Energy: 70%, Payload: 40%. The mission resources are converted into corresponding mission resource vectors through initialization processing, as shown in Table 2.

[0140] Table 2 Task resource vector matrix

[0141]

[0142] Mission objective information:

[0143] Mission Objective 1: Priority: High, Location: Longitude (50.0), Latitude (90.0). Mission Objective 2: Priority: Medium, Location: Longitude (51.0), Latitude (91.0). Mission Objective 3: Priority: Low, Location: Longitude (52.0), Latitude (92.0). The mission objective information is converted into the corresponding mission objective vector through initialization processing, as shown in Table 3.

[0144] Table 3 Task target vector matrix

[0145]

[0146] Initialize the state vectors for all drones: the mission resource vector, and the mission state vector: the mission goal vector. Each drone's state includes its position, speed, energy, mission status, and distance to other drones. Mission status includes its location, priority, and required resources (battery charge). The overall state space of the swarm encompasses the states of all drones, as shown in Table 4. Drone actions include all possible drone actions, including movement direction, speed adjustment, and mission selection.

[0147] Table 4 Status of drones

[0148]

[0149] Next, we define a reward function based on task completion, resource consumption, and penalties for task failure. If the task is completed, the reward is based on task completion; assuming 100% completion, the reward is 1. If resource consumption (power) is excessive, a penalty is applied; for example, a penalty factor of 0.2 is applied for every unit of power consumed. If the task is delayed or interrupted, a penalty is applied; assuming a penalty factor of -1.

[0150] Initialize each drone's starting position, target position, obstacle information, flight speed, and energy limit. Use the A* or Dijkstra algorithm to calculate the shortest path for each drone, taking into account possible obstacles and the positions of other drones during flight. For each drone, a path planning algorithm calculates a path from the starting point to the target point.

[0151] As shown in formula (3), the heuristic evaluation function in the A* algorithm is:

[0152] f(x)=g(x)+h(x)(3)

[0153] Among them, g(x) is the actual cost from the starting point to the current node, and h(x) is the estimated cost (heuristic function) from the current node to the goal node.

[0154] Dynamic planning is used to adjust the path and update the path in real time to respond to environmental changes, such as enemy activity or changes in weather conditions. When the battlefield environment changes (such as enemy interference, obstacles, etc.), the path planning needs to be dynamically adjusted. Dynamic planning is used to update the path, as shown in formula (4):

[0155] Pnew=Pcurrent+ΔP(4)

[0156] Where: Pnew is the new path after adjustment, and ΔP is the adjustment amount of the path.

[0157] Calculate the resources consumed during the execution of the path (such as battery power, time, etc.). As shown in formula (5), the resource consumption formula is:

[0158] R consume =λ7×path length+λ8×flight time (5),

[0159] Where: λ7 and λ8 are the weight coefficients of path length and flight time, path length is the total distance flown, and flight time is the time required to complete the path.

[0160] If a drone needs to fly from starting point A to target B, the A* algorithm will calculate the shortest path, avoid collisions with obstacles, and dynamically adjust the path based on environmental information to ensure that time and resource consumption are minimized.

[0161] If two or more drones work together to complete a mission (e.g., jointly complete reconnaissance), they are rewarded. If drones collide or overlap excessively, they are penalized. Drones that complete mission objectives or optimize task allocation are rewarded. As shown in formula (6),

[0162] (6),

[0163] Among them, λ1 and λ2 are adjustment coefficients to ensure the balance between rewards and penalties.

[0164] Use Q-learning or Deep Q Network (DQN) algorithm to select the optimal action based on the state of the drone and the mission status. Based on the following assumptions:

[0165] Reward coefficient: task completion reward λ1=1, resource consumption penalty λ2=0.2, task failure penalty λ3=-1.

[0166] Learning rate: α=0.1,

[0167] Discount factor: γ=0.9,

[0168] The update formula through Q-learning is: Q(s,a)←Q(s,a)+a[R(s,a)+γ*maxQ(s',a)-Q(s,a)].

[0169] During the task allocation process, the optimal task allocation strategy is selected. Assume there are two tasks (Task 1 and Task 2) and two drones (Drone A and Drone B). Based on the drone states and the task states, the Q value is calculated and the optimal task allocation is selected.

[0170] Q(UAV A, Task 1) = 5;

[0171] Q(UAV A, Task 2) = 3;

[0172] Q(UAV B, Task 1) = 4;

[0173] Q(UAV B, Task 2) = 6;

[0174] From the perspective of Q value, Drone A has a higher Q value for choosing Task 1 (5 vs 3), while Drone B has a higher Q value for choosing Task 2 (6 vs 4). Therefore, based on the Q value decision, Task 1 is assigned to Drone A and Task 2 is assigned to Drone B.

[0175] Multi-agent reinforcement learning (MARL) is used to optimize the collaborative decision-making of swarm drones, ensuring efficient collaboration and conflict avoidance in dynamic environments. The goal of swarm algorithm training is to ensure that each drone in the swarm can make coordinated decisions based on the current environment and the state of other drones to complete the mission.

[0176] Multiple drones work together to avoid conflicts based on environmental changes and mission requirements. In this process, the agents communicate with each other about their status and adjust their strategies to coordinate cooperation. As shown in formula (7),

[0177] V(s t )=V(s t )+α·(r t +γ·V(s t+1 )-V(s t ))(7),

[0178] Among them, V (s t ) is the value function, r t It's a reward, S t and S t+1 are the current and next states, and α is the learning rate.

[0179] Based on the optimized strategy and reward function, the drones' action decisions are automatically adjusted to ensure that each drone can coordinate with other drones, ensuring efficient mission execution while avoiding conflicts and resource waste. Based on the optimization results of multi-agent reinforcement learning, the drones select the optimal collaborative decision-making strategy.

[0180] When multiple drones in a swarm perform reconnaissance missions, they share target information and path planning data to avoid repeated reconnaissance of the same area and maximize area coverage. Reinforcement learning algorithms adapt each drone's action strategy in real time, minimizing mission execution time and resource consumption, ensuring timely completion and avoiding wasted energy and other resources.

[0181] Assume that the optimal drone is selected between Task 1 and Task 2. After multiple rounds of training, the Q value gradually converges: for Task 1, the optimal drone may be Drone 3 because its status and resources are most compatible and it is closer to the task location; for Task 2, the optimal drone may be Drone 1 because it has more power and is closer to the task location.

[0182] Finally, after optimization and strategy calculation, the final output task allocation results are shown in Table 5:

[0183] Table 5 Final output task assignment results

[0184]

[0185] Based on the mission objectives and environmental information, the optimal flight path for each UAV is calculated and the corresponding evaluation report is output.

[0186] The present invention is based on a cluster algorithm training and evolution system and method defined by an architecture model. First, by uniformly defining the functions and interfaces of various task resources in an architecture model, the access of new equipment or new requirements is made simpler and more efficient. Secondly, by systematically defining and describing task resources and task objectives through the architecture model, the compatibility, scalability and interoperability between different task resources can be ensured, and they can be quickly adapted to different task objective scenarios without the need for large-scale modifications to existing models. Furthermore, the combination of the architecture model and the cluster decision algorithm not only optimizes the decision-making efficiency of the system, but also improves its traceability and transparency, avoiding the understanding and debugging difficulties brought about by traditional "black box" algorithms. Finally, the standardized model description and the scalable algorithm training framework can quickly adapt to new task resources or task objectives, reducing the difficulty of large-scale manual adjustments to algorithms and models during system upgrades and scenario switching, and achieving efficient maintenance and management.

[0187] The above content is an explanation of the preferred embodiments of the present invention, which can help those skilled in the art to more fully understand the technical solutions of the present invention. However, these embodiments are merely illustrative, and it cannot be determined that the specific implementation methods of the present invention are limited to the description of these embodiments. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions and transformations can be made, which should be deemed to fall within the scope of protection of the present invention.

Claims

1. A cluster algorithm training and evolution system defined based on an architecture model, characterized by: It includes a data acquisition unit, a system architecture unit, an algorithm training unit, a simulation evaluation unit and a result output unit; the data acquisition unit collects task objectives and task resources; The system architecture unit analyzes the task resources according to the task objectives and outputs the system architecture model; the system architecture unit includes a model building module and a model optimization module; the model building module establishes the framework of the system architecture model according to the task objectives and task resources: the system architecture pre-model; the model building module analyzes the scenarios and specific tasks of the task objectives, and then constructs the system architecture pre-model according to the task resources, including task resource capability analysis, system architecture performance design, and key steps of task allocation; through a detailed analysis of the activity sequence and state transition rules in different task objectives, a system architecture pre-model containing multiple views is designed; the system architecture pre-model includes functional composition, interface protocol, communication relationship, tactics and strategies, and state transition rules, covering the interface compatibility and collaborative communication relationship between task resources; The model optimization module dynamically verifies the system architecture pre-model, including the correctness of task resources, interface compatibility, and optimization of tactical logic, to obtain an optimized system architecture model: a system architecture optimization model; The algorithm training unit performs algorithm training based on the system architecture model and outputs a cluster decision strategy and task allocation plan. The algorithm training unit regards different types of task resources as different intelligent agents and learns to make optimal decisions in complex environments by repeatedly interacting with different environments of task resources. The simulation evaluation unit simulates and deduces the cluster decision-making strategy, evaluates the effect of the task allocation scheme, and outputs a corresponding evaluation report; The result output unit outputs an evaluation report.

2. The cluster algorithm training and evolution system based on architecture model definition according to claim 1, characterized in that: The data acquisition unit includes a resource acquisition module and a target input module; the resource acquisition module collects relevant information data of the task resources, including the quantity, type, status, current location, energy, and load of the task resources, initializes the relevant information data of the task resources, and obtains the corresponding task resource vector; The target input module collects relevant information data of the task target, including the task requirement priority and the task requirement position, and performs initialization processing on the relevant information data of the task target to obtain a corresponding task target vector.

3. The cluster algorithm training and evolution system based on architecture model definition according to claim 2, characterized in that: The task resource vector is [quantity, type, status, longitude normalization, latitude normalization, energy, load]; The task target vector is expressed as [priority, longitude normalization, latitude normalization].

4. The cluster algorithm training and evolution system based on architecture model definition according to claim 1, characterized in that: The model optimization module adopts a self-verifying logic verification method, performs algorithm training according to the system architecture pre-model, obtains the cluster decision-making strategy and task allocation plan, and the simulation evaluation unit evaluates the effect of the allocation plan. The evaluation results are fed back to the model optimization module to perform adaptive training and adjustment on the system architecture pre-model; finally, the correctness decision of the system architecture pre-model is continuously adjusted according to the real-time feedback of task resources and task objectives, forming a closed-loop feedback mechanism to obtain the system architecture optimization model.

5. The cluster algorithm training and evolution system based on architecture model definition according to claim 1, characterized in that: The algorithm training unit specifically includes: Step 301: Conduct agent modeling for different task resources; Step 302: Train the agent's behavior through a reward mechanism; Step 303: Optimize the behavior of each agent; Step 304: Optimize and adjust the behavior of each agent in the cluster to obtain a cluster decision strategy.

6. The cluster algorithm training and evolution system based on architecture model definition according to claim 5, characterized in that: Each agent receives corresponding rewards or penalties based on the actions it takes and the feedback from the environment; when the agent takes the right action, it is given positive rewards, and when it makes an incorrect decision, it is punished. R t =λ1·Task completion reward -λ2·Resource consumption penalty -λ3·Task failure penalty, Among them, λ1 is the reward coefficient, which is the feedback obtained by the agent from the environment after taking a certain action; λ2 and λ3 are penalty coefficients; Based on historical task execution data, the Q value is updated and the strategy is adjusted. In the Q-learning algorithm, the Q value represents the expected total reward that can be obtained after taking a certain action and continuing to execute it in a certain state. ， in, In state Next, execute the action Q value; α is the learning rate; γ is the discount factor; r t is the current reward; t is the current state; t is the current action; s t+1 is the next state; ɑ' is the next action; For the next state Next, all actions Maximum Q value; The decision-making strategy is the rule or function for an agent to select actions in a given state. The agent will adjust its behavior based on the size of the Q value and select the optimal task allocation plan, thereby making decisions on the task resources in the cluster and the corresponding task allocation to obtain the cluster decision-making strategy. The larger the Q value, the better the task allocation plan has performed in historical experience, and the better the agent decision-making effect.

7. The cluster algorithm training and evolution system based on architecture model definition according to claim 1, characterized in that: The simulation evaluation unit realizes the coordination and collaboration between different task resources, including: Step 401: The simulation evaluation unit matches the task resources with the task objectives according to the status of the task resources and the requirements of the task objectives. Step 402: The simulation evaluation unit globally coordinates and collaboratively controls task resources.

8. A cluster algorithm training and evolution method based on a system architecture model definition, applicable to a cluster algorithm training and evolution system based on a system architecture model definition, characterized in that: The following steps are involved: Step 1: The data collection unit collects task objectives and task resources; Step 2: The system architecture unit analyzes the task resources according to the task objectives and outputs the system architecture model; the system architecture unit includes a model building module and a model optimization module; the model building module establishes the framework of the system architecture model according to the task objectives and task resources: the system architecture pre-model; the model building module analyzes the scenarios and specific tasks of the task objectives, and then constructs the system architecture pre-model according to the task resources, including task resource capability analysis, system architecture performance design, and key task allocation steps; through a detailed analysis of the activity sequence and state transition rules in different task objectives, a system architecture pre-model containing multiple views is designed; the system architecture pre-model includes functional composition, interface protocol, communication relationship, tactics and strategies, and state transition rules, covering the interface compatibility and collaborative communication relationship between task resources; The model optimization module dynamically verifies the system architecture pre-model, including the correctness of task resources, interface compatibility, and optimization of tactical logic, to obtain an optimized system architecture model: a system architecture optimization model; Step 3: The algorithm training unit performs algorithm training based on the system architecture model and outputs a cluster decision strategy and task allocation plan. The algorithm training unit treats different types of task resources as different intelligent agents and learns to make optimal decisions in complex environments by repeatedly interacting with different environments of task resources. Step 4: The simulation evaluation unit simulates and deduces the cluster decision-making strategy, evaluates the effect of the task allocation scheme, and outputs a corresponding evaluation report; Step 5: The result output unit outputs the evaluation report.

Images