A multi-agent collaboration method based on autonomous reflection and reverse attention

Abstract

The application discloses a kind of multi-agent collaborative method based on autonomous reflection and reverse attention, it is related to artificial intelligence technical field.The method is first by inter-agent state perception and information sharing, constructs global information graph to obtain task dynamics.Next, by autonomous reflection mechanism, according to task completion degree, resource consumption and decision execution time and other index evaluation historical decision behavior value, mark and optimize inefficient decision, introduce meta-learning to generate optimization rule, while detecting environmental change optimization strategy.Then, by reverse attention mechanism, the dynamic dependence between agents is modeled, a weight matrix is constructed, and resources are allocated and controlled for low-priority tasks.Finally, the agent executes the task and evaluates the task completion efficiency, resource consumption and collaboration performance, and the evaluation results are used to optimize the next round of strategy.The application uses the above method, improves the collaboration efficiency of the agent, optimizes the resource allocation, and enhances the autonomous decision-making ability of the agent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a multi-agent collaboration method based on autonomous reflection and reverse attention. BACKGROUND

[0002] Multi-agent systems are widely used in unmanned driving, robot swarm, intelligent transportation and other fields. The core problem is the collaboration efficiency between agents. In traditional methods, agents collaborate through centralized control or distributed algorithms based on fixed strategies, but these methods have obvious shortcomings in dynamic and uncertain environments. Especially in the following aspects:

[0003] (1) Lack of autonomous thinking ability: Traditional multi-agent methods are usually based on fixed rules or single optimization objectives, and lack the ability to reflect on historical decisions. When the environment changes or the task requirements are dynamically adjusted, the agent cannot effectively adjust its strategy. For example, when there is a long-term inefficient path planning in a robot swarm, traditional methods are difficult to effectively identify and adjust. In addition, due to the lack of feedback mechanism of historical decisions, agents cannot learn from mistakes, leading to the same inefficient decisions being repeated multiple times.

[0004] (2) Insufficient collaboration efficiency: Current attention mechanisms mainly focus on local optimization and fail to fully consider the dynamic dependency relationship between agents. For example, in an intelligent transportation system, the collaboration between different vehicles needs to be dynamically adjusted in real time, but traditional methods only focus on local traffic optimization, ignoring the global traffic flow.

[0005] And in the dynamic allocation of task priorities, traditional methods lack effective mechanisms and are insufficient in resource inclination to low-priority tasks, which can cause bottleneck task delays and affect the overall system collaboration efficiency.

[0006] (3) Uneven resource allocation: In heterogeneous task scenarios, the resource capabilities of different agents may differ greatly, but traditional methods usually use an average allocation strategy, ignoring the specific needs of tasks for resources. For example, in a UAV swarm, tasks may require some UAVs to have additional computing or energy support, but unreasonable resource allocation can lead to task failure or low efficiency. The lack of dynamic resource reallocation capability limits the system's flexible adaptation ability under high load conditions. SUMMARY

[0007] The purpose of the present application is to propose a multi-agent collaboration method based on autonomous reflection and reverse attention to solve the problems of insufficient autonomous thinking ability, low collaboration efficiency, and uneven resource allocation in traditional multi-agent collaboration methods, improve the collaboration efficiency of agents in dynamic and uncertain environments, optimize resource allocation, and enhance the autonomous decision-making ability of agents.

[0008] To achieve the above object, the application provides a multi-agent collaborative method based on autonomous reflection and reverse attention, comprising the following steps:

[0009] Step S1, state perception and information sharing between agents: the agent obtains state information of itself and the environment through a sensor, constructs a state vector and an environment state E(t), shares information in real time by using a communication network, and constructs a global information graph G(t)=(V, E(t)), wherein V={1, 2,..., N}, which is an agent set, and N is a positive integer;

[0010] Step S2, constructing an autonomous reflection mechanism: the agent periodically evaluates the behavior value of historical decisions, marks the low-efficiency part of the strategy according to the behavior value MarkedLowEfficiency(π i ), optimizes the strategy π i by using a reflection module R i , and introduces a meta-learning ability to generate a strategy optimization rule, and detects environmental changes in real time to optimize the current strategy;

[0011] Step S3, constructing a reverse attention mechanism: globally modeling dynamic dependencies, optimizing by using a graph neural network GNN, and constructing a weight matrix α(t) by using a reverse attention mechanism to dynamically adjust the resource allocation amount;

[0012] Step S4, task execution and evaluation: the agent executes a task according to the optimized strategy , and feeds back the state in the execution process to the global information graph G(t), and records the task completion efficiency, resource consumption and agent collaboration performance by using a task evaluation module, and the evaluation result is used for optimizing the next round of strategy.

[0013] Preferably, in step S1, the state vector formula is as follows:

[0014] s i (t)=[p i (t),v i (t),e i (t),…] T ;

[0015] Wherein s i (t) is the state vector of the i-th agent at t time, p i (t) is the position of the i-th agent at t time, v i (t) is the speed of the i-th agent at t time, e i (t) is the energy consumption of the i-th agent at t time, and T is a vector transpose symbol.

[0016] The environmental state E(t) is represented as a set of all ordered pairs (i, j) that satisfy the condition that the ith agent and the jth agent can cooperate at time t, where i = 1, 2, …, N, and j = 1, 2, …, N.

[0017] Preferably, in step S1, in the global information graph, the nodes represent the agents, the edges represent the cooperation relationship between the agents, and the time stamp mechanism τ is used to ensure the timeliness of the information, thereby forming the global state information S(t), which is as follows:

[0018] S(t) = (s1(τ1), τ1), (s2(τ2), τ2), …, (s N (τ N ), τ N ), (E(τ E ), τ E );

[0019] where s m (τ m ) is the state information of the mth agent at time τ m , τ m is the time stamp of the state update of the mth agent, m = 1, 2, …, N, E(τ E ) is the state information of the environment at time τ E , and τ E is the time stamp of the state update of the environment.

[0020] Preferably, in step S2, the behavior value calculation formula is as follows:

[0021] V i (h) = w1·TaskCompletion i (h) - w2·ResourceConsumption i (h) - w3·

[0022] DecisionTime i (h);

[0023] where V i (h) is the behavior value of the ith agent, w1, w2, and w3 are weight coefficients, TaskComletion i (h) is the task completion degree of the ith agent, ResourceConsumption i (h) is the resource consumption of the ith agent, DecisionTime i (h) is the decision execution time of the ith agent, and h is the historical decision index.

[0024] The low-efficiency part is MarkedLowEfficiency(πi ) represents the policy π i in the middle and h|V i (h) is associated and less than θ V part;

[0025] wherein, θ V is the value threshold; the policy π i of the agent is adaptively optimized by the reflection module R i , and the calculation formula is as follows:

[0026]

[0027] wherein, is the new policy of the i-th agent after reflection, is the old policy of the i-th agent before reflection.

[0028] Preferably, the meta-learning ability is introduced, the successful experience Experience in the past tasks is used to generate a generalizable policy optimization rule, and the formula is as follows:

[0029]

[0030] wherein, is the new policy parameter of the i-th agent after reflection, is the old policy parameter of the i-th agent before reflection, and M i is the meta-learning module of the i-th agent.

[0031] Preferably, in step S3, the calculation formula of the dynamic dependency relationship is as follows:

[0032] D ij (t)=β1·P task (n,t)+β2·R demand (n,t);

[0033] wherein, D ij (t) is the dynamic dependency relationship of the i-th and j-th agent, β1 and β2 are weight coefficients, P task (n,t) is the priority of the n-th task at time t, and R demand (n,t) is the resource requirement of the n-th task at time t;

[0034] is optimized by the graph neural network GNN, and the node embedding update formula of the GNN is as follows:

[0035]

[0036] wherein, W1 and W2 are weight matrices, is the embedding vector of the i-th node at time t in the k+1 layer, Let be the embedding vector of the i-th node at time t in layer k. Let σ be the embedding vector of the j-th node at time t in layer k, where σ is the activation function, b is the bias term, and k is the number of layers in the GNN.

[0037] The node embedding vector is continuously updated to achieve dynamic adjustment. A weight matrix is ​​constructed using an inverse attention mechanism, as shown in the following formula:

[0038] α(t)=[α ij (t)] N×N ;

[0039] Where, α ij (t) represents the reverse attention weight of agent i to agent j, as shown in the following formula:

[0040]

[0041] Among them, D il (t) represents the dynamic dependency relationship between the i-th and l-th agents.

[0042] Preferably, the resource allocation strategy is dynamically adjusted by inverse attention weights and resource requirements. The formula for calculating the resource allocation amount is as follows:

[0043]

[0044] Among them, R ij (t) represents the resource allocation amount from the i-th agent to the j-th agent, α il (t) represents the reverse attention weights of agent i to agent l, R total_available (i,t) represents the total available resources of the i-th agent at time t, R demand (j,t) represents the resource requirement of the j-th task at time t, R demand (l,t) represents the resource requirements of the l-th task at time t.

[0045] Preferably, in step S4, the agent follows the optimized strategy. During task execution, the formulas for updating the state vector and environment state are as follows:

[0046] s i (t+1)=f state (s i (t),a i (t),E(t));

[0047] E(t+1)=f environment (E(t),{a1(t),a2(t),…,a N (t)});

[0048] where s i (t+1) is the state vector of the i-th agent at time t+1, E(t+1) is the environment state at time t+1, f state (·) and f environment (·) are state update functions, a i (t) is the action of the i-th agent at time t.

[0049] Preferably, after the completion of the task, the efficiency, resource consumption and agent collaboration performance of the completed task are recorded by the task evaluation module E, and the formula is as follows:

[0050]

[0051] Cres=E resource ({s i (t)}i∈V);

[0052]

[0053] where Eff is the efficiency of task completion, Cres is the resource consumption, Ccollab is the agent collaboration performance, E eficiency (·) is the efficiency function, E resource (·) is the resource consumption function, E colaboration (·) is the collaboration performance function, is a set of task performance indicators, is a set of agent collaboration behavior patterns;

[0054] The evaluation results {Eff, Cres} are used as the input of the autonomous reflection module to optimize the next round of strategy, and the formula is as follows:

[0055]

[0056] where is the next round of strategy of the i-th agent, is the current strategy of the i-th agent.

[0057] Therefore, the present application proposes a multi-agent collaboration method based on autonomous reflection and reverse attention, which has the following beneficial effects:

[0058] (1) Improve autonomous decision-making ability: through the autonomous reflection mechanism, the agent can evaluate and optimize the historical decision-making, learn from mistakes, reduce unnecessary repeated errors, and improve the rationality and adaptability of its own decision-making. In a complex and variable environment, the agent can adjust the strategy in time according to the environmental changes, and better complete the task.

[0059] (2) Enhancing coordination efficiency: The reverse attention mechanism fully considers the dynamic dependency relationship between agents, and by reasonably allocating resources, it prioritizes processing bottleneck tasks, reduces task waiting time, and improves the overall coordination efficiency of the system. Whether in intelligent transportation systems or intelligent factories, etc. scenarios, more efficient task execution can be achieved.

[0060] (3) Optimizing resource allocation: For heterogeneous task scenarios, the method proposed in the present application can dynamically adjust the resource allocation strategy according to the specific resource requirements of the task and the resource capabilities of the agent, avoiding the average allocation and unreasonable use of resources, improving the resource utilization rate, and enhancing the system's flexible adaptive ability under high load conditions.

[0061] (4) Wide application prospects: The multi-agent collaboration method based on autonomous reflection and reverse attention proposed in the present application is applicable to various multi-agent system application scenarios such as unmanned driving, robot swarm, intelligent transportation, intelligent factory, etc. It can provide more efficient and intelligent solutions for the development of these fields, and has broad application prospects and economic value.

[0062] The technical solutions of the present application will be further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The flowchart of the multi-agent collaboration method based on autonomous reflection and reverse attention of the present application. DETAILED DESCRIPTION

[0064] In order to make the technical solutions, advantages and purposes of the present application clearer, the technical solutions of the embodiments of the present application will be described clearly and completely below. The described embodiments are part of the embodiments of the present application, not all embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0065] Unless otherwise defined, the technical terms or scientific terms used in the present application shall be understood as their usual meanings understood by those skilled in the art in the field of the present application.

[0066] Embodiment one

[0067] As shown in Figure 1 The flowchart of the multi-agent collaboration method based on autonomous reflection and reverse attention of the present application, including the following steps:

[0068] 1. State perception and information sharing among agents: Agents obtain state information of themselves and the environment through sensors, construct state vectors and environment state E(t), share information in real time using communication networks, and construct global information graph G(t) = (V, E(t)), where V = {1, 2, …, N}, N is a positive integer, and is the set of agents.

[0069] The state vector formula is as follows:

[0070] s i (t) = [p i (t), v i (t), e i (t), …] T ;

[0071] where s i (t) is the state vector of the i-th agent at time t, p i (t) is the position of the i-th agent at time t, v i (t) is the speed of the i-th agent at time t, e i (t) is the energy consumption of the i-th agent at time t, and T is the vector transpose symbol.

[0072] The environment state E(t) is represented as a set of all ordered pairs (i, j) that satisfy the i-th agent and the j-th agent at time t can cooperate, where i = 1, 2, …, N, j = 1, 2, …, N.

[0073] In the global information graph, nodes represent agents, edges represent the cooperative relationship between agents, and the time stamp mechanism τ is used to ensure the timeliness of information, forming the global state information S(t), which is as follows:

[0074] S(t) = (s1(τ1), τ1), (s2(τ2), τ2), …, (s N (τ N ), τ N ), (E(τ E ), τ E );

[0075] where s m (τ m ) is the state information of the m-th agent at time τ m , τ m is the time stamp of the m-th agent for state update, m = 1, 2, …, N, E(τ E ) is the state information of the environment at time τ E , and τ E is the time stamp of the environment state update.

[0076] 2. Constructing self-reflective mechanism: the agent periodically evaluates the behavior value of historical decisions, marks the low efficiency part of the strategy MarkedLowEfficiency(π i ), and optimizes the strategy π i through the reflection module R i , and introduces meta-learning ability to generate strategy optimization rules, and real-time detects environmental changes to optimize the current strategy.

[0077] The behavior value calculation formula is as follows:

[0078] V i (h)=w1·TaskCompletion i (h)-w2·ResourceConsumption i (h)-w3·

[0079] DecisionTime i (h);

[0080] Wherein, V i (h) is the behavior value of the i-th agent, w1, w2, w3 is the weight coefficient, TaskCompletion i (h) is the task completion degree of the i-th agent, ResourceConsumption i (h) is the resource consumption of the i-th agent, DecisionTime i (h) is the decision execution time of the i-th agent, and h is the historical decision index.

[0081] MarkedLowEfficiency(π i ) represents the part of the strategy π i associated with h|V i (h) and less than θ V ;

[0082] Wherein, θ V is the value threshold; the strategy π i of the agent is optimized through the reflection module R i , and the calculation formula is as follows:

[0083]

[0084] Wherein, is the new strategy of the i-th agent after reflection, is the old strategy of the i-th agent before reflection.

[0085] Introduce meta-learning ability, use the success experience of past tasks to generate generalizable strategy optimization rules, the formula is as follows:

[0086]

[0087] Wherein, is the new strategy parameter of the ith agent after reflection, is the old strategy parameter of the ith agent before reflection, M i is the meta-learning module of the ith agent.

[0088] 3, Construct reverse attention mechanism: globally model dynamic dependency relationship, optimize through graph neural network GNN, construct weight matrix α(t), dynamically adjust resource allocation.

[0089] The calculation formula of dynamic dependency relationship is as follows:

[0090] D ij (t)=β1·P task (n,t)+β2·R demand (n,t);

[0091] Wherein, D ij (t) is the dynamic dependency relationship of the ith and jth agent, β1 and β2 are weight coefficients, P task (n,t) is the priority of the nth task at time t, R demand (n,t) is the resource demand of the nth task at time t;

[0092] Optimize it through graph neural network GNN, the node embedding update formula of GNN is:

[0093]

[0094] Wherein, W1 and W2 are weight matrices, is the embedding vector of the ith node at time t in the k+1 layer, is the embedding vector of the ith node at time t in the k layer, is the embedding vector of the jth node at time t in the k layer, σ is the activation function, b is the bias term, and k is the number of layers of GNN;

[0095] The embedding vector of the node is updated continuously to realize dynamic adjustment, and the weight matrix is constructed by using the reverse attention mechanism, and the formula is as follows:

[0096] α(t)=[[α ij (t)] N×N ;

[0097] Wherein, αij (t) is the reverse attention weight of agent i to agent j, and the formula is as follows:

[0098]

[0099] Where D il (t) is the dynamic dependence relationship of the ith, lth agent.

[0100] Through the reverse attention weight and resource demand, the resource allocation strategy is dynamically adjusted, and the resource allocation amount calculation formula is as follows:

[0101]

[0102] Where R ij (t) is the resource allocation amount of the ith agent to the jth agent, and a il (t) is the reverse attention weight of agent i to agent l, R total_available (i, t) is the total available resource of the ith agent at time t, R demand (j, t) is the resource demand of the jth task at time t, and R demand (l, t) is the resource demand of the lth task at time t.

[0103] 4. Task execution and evaluation: the agent executes the task according to the optimized strategy , and the state in the execution process is fed back to the global information graph G(t). The task completion efficiency, resource consumption and agent cooperation performance are recorded through the task evaluation module, and the evaluation results are used to optimize the next round of strategy.

[0104] In the task execution process, the current state is continuously fed back to the global information graph, and the state vector and environment state update formula is as follows:

[0105] s i (t+1)=f state (s i (t),a i (t),E(t));

[0106] E(t+1)=f environment (E(t),{a1(t),a2(t),…,a N (t)});

[0107] Where s i (t+1) is the state vector of the ith agent at time t+1, E(t+1) is the environment state at time t+1, f state (·) and f environment (·) are state update functions, a i(t) is the action of the ith agent at time t.

[0108] The efficiency of task completion, resource consumption and agent collaboration performance are recorded by the task evaluation module E, as follows:

[0109]

[0110] Cres=E resource ({s i (t)} i∈V );

[0111]

[0112] wherein Eff is the efficiency of task completion, Cres is the resource consumption, Ccollab is the agent collaboration performance, E eficiency (·) is the efficiency function, E resource (·) is the resource consumption function, E colaboration (·) is the collaboration performance function, is a set of task performance indicators, is a set of agent collaboration behavior patterns;

[0113] The evaluation results {Ef, Cres} are used as the input of the autonomous reflection module to optimize the next round of strategy, as follows:

[0114]

[0115] wherein is the next round strategy of the ith agent, is the current strategy of the ith agent.

[0116] Embodiment Two

[0117] This embodiment takes the execution of a patrol task by a UAV cluster as an example. Before applying the autonomous reflection mechanism, the relevant indicators are recorded: the total time for the cluster to complete the task is 100 minutes, the average energy consumption of a single UAV is 80% of the power, the cluster has selected an inefficient path 25 times during the task execution process, the average decision time for path planning is 1.5 seconds / time, and the task completion degree score is 95 points (full score 100 points).

[0118] During operation, the UAVs periodically evaluate the behavioral value of historical decisions according to the autonomous reflection mechanism. According to the weighted calculation of task completion degree, resource consumption and decision execution time, the inefficient path part in the strategy is marked. For example, if a path planning results in a longer task completion time and higher energy consumption, its behavioral value is lower than the set value threshold θ V , then this path planning part is marked as an inefficient decision.

[0119] Then, the strategy is optimized by the reflection module. Based on the historical inefficient path, the weight of path planning is adjusted, so that the subsequent path planning prefers to choose the efficient path. At the same time, by using the meta-learning ability, the experience of past successful tasks is learned to generate more optimal strategy optimization rules. After a period of operation and reflection learning, the relevant indicators are recorded again: the total time for the swarm to complete the same task is shortened to 90 minutes, and the task completion time is reduced by 10%; the average energy consumption of a single unmanned aerial vehicle is reduced to 75% of the power, and the energy consumption is reduced by 6.25%; the number of inefficient path selection is significantly reduced to 10 times, which is reduced by 60%; the average decision time of path planning is also shortened to 1.3 seconds / time, which is reduced by 13.3%; the task completion score is improved to 97 points, which is improved by 2.1%.

[0120] This embodiment fully verifies the effectiveness of the autonomous reflection mechanism in improving the efficiency of unmanned aerial vehicle swarm tasks, reducing resource consumption, optimizing path planning, and improving task completion.

[0121] Embodiment Three

[0122] This embodiment takes the scenario of different robots in an intelligent factory cooperating to complete assembly tasks. Before applying the reverse attention mechanism, the relevant indicators of the system are recorded: the average task waiting time of the system is 5 minutes, the overall task completion time of the system is 60 minutes, the average idle time of the robot accounts for 20% of the time proportion, the resource allocation balance score is 70 points (full score 100 points), and the high-priority task delay rate reaches 10%.

[0123] When applying the reverse attention mechanism, the dynamic dependency relationship between robots is first globally modeled. According to the task priority and resource demand, the dynamic dependency relationship is quantified, and the node embedding vector is optimized through the graph neural network to realize the dynamic adjustment of the dynamic dependency relationship.

[0124] A weight matrix is constructed, and resources are preferentially allocated to low-priority tasks according to the weight matrix. For example, for tasks with long waiting times, more robots are allocated to execute to reduce system bottlenecks. In terms of resource allocation control, the resource allocation strategy is dynamically adjusted according to the reverse attention weight and real-time task demand, such as preferentially allocating corresponding resources to tasks with large resource demand.

[0125] After applying the reverse attention mechanism, the relevant indicators are recorded again: the average task waiting time of the system is reduced to 2 minutes, which is reduced by 60%; the overall task completion time of the system is shortened to 54 minutes, and the overall efficiency of the system is improved by 10%; the average idle time of the robot is reduced to 15% of the time proportion, which is reduced by 25%; the resource allocation balance score is improved to 80 points, which is improved by 14.3%; the high-priority task delay rate is reduced to 5%, which is reduced by 50%.

[0126] The embodiment shows that the reverse attention mechanism can effectively reduce the task waiting time in the intelligent factory system, improve the resource utilization rate, optimize the resource allocation balance, and thus significantly improve the overall operation efficiency of the intelligent factory. It should be noted that the contents not elaborated in the present application are all prior art and are well known to those skilled in the art.

[0127] Therefore, the present application provides a multi-agent collaborative method based on autonomous reflection and reverse attention. Through the collaborative operation of state perception and information sharing between agents, autonomous reflection mechanism, reverse attention mechanism, and task execution and evaluation, the problems of lack of autonomous thinking ability, low collaboration efficiency and uneven resource allocation of traditional multi-agent collaboration in dynamic environment are solved, and the collaboration performance of multi-agent system in complex scenarios is improved.

[0128] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: the technical solutions of the present application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A multi-agent collaborative method based on autonomous reflection and reverse attention, characterized in that, Includes the following steps: Step S1: State perception and information sharing among agents: Agents acquire their own and the environment's state information through sensors, and construct state vectors and environmental states. By utilizing communication networks to share information in real time, a global information graph can be constructed. ,in V= {1 , 2 ,...,N } represents a set of intelligent agents. N It is a positive integer; Step S2: Construct a self-reflection mechanism: The agent periodically evaluates the behavioral value of historical decisions and marks inefficient parts of the strategy based on the behavioral value. Through the reflection module R i strategy Adaptive optimization is performed, and meta-learning ability is introduced to generate strategy optimization rules, and the current strategy is optimized in real time by detecting environmental changes. Step S3: Constructing an inverse attention mechanism: Globally model the dynamic dependencies, optimize using a graph neural network (GNN), and construct a weight matrix using an inverse attention mechanism. Dynamically adjust resource allocation; Step S4, Task Execution and Evaluation: The agent performs the task according to the optimized strategy. Execute the task and report the status during the execution process to the global information graph. The task evaluation module records task completion efficiency, resource consumption, and agent collaboration performance, and the evaluation results are used to optimize the next round of strategies. In step S1, the state vector formula is as follows: ； in, For the first i An intelligent agent in t The state vector at time t, For the first i An intelligent agent in t Location at any given moment For the first i An intelligent agent in t The speed of time, For the first i An intelligent agent in t Energy consumption at any time The symbol for vector transpose; Environmental conditions Represented as all satisfying the first i The first intelligent agent and the first j An ordered pair of agents that can cooperate at time t i , j The set of ), where, , ; In step S2, the formula for calculating behavioral value is as follows: ； in, For the first i The value of an agent's behavior. , , These are the weighting coefficients. For the first i Task completion rate of each agent For the first i Resource consumption of each intelligent agent For the first i Decision execution time of an agent h For historical decision indexing; Inefficient part Represented as a strategy Zhongyu Related and less than Part of it; in, It is a value threshold; through the reflection module Policy for intelligent agents Adaptive optimization is performed, and the calculation formula is as follows: ； in, For the first i A new strategy after reflection by an intelligent agent. For the first i An agent reflects on its old strategies; In step S3, the formula for calculating dynamic dependencies is as follows: ； in, For the first i, j Dynamic dependencies between agents and These are the weighting coefficients. For the first n The task at time t priority, For the first n Each task at time t Resource requirements; Optimize it using a Graph Neural Network (GNN). The node embedding update formula for GNN is: ； in, and This is the weight matrix. For the first i The node at the th k +1 floor t Embedding vector at time step, For the first i The node at the th k layer t Embedding vector at time step, For the first j The node at the th k layer t Embedding vector at time step, For activation function, For bias terms, k The number of layers in the GNN; The node embedding vector is continuously updated to achieve dynamic adjustment. A weight matrix is ​​constructed using an inverse attention mechanism, as shown in the following formula: ； in, For intelligent agents i For intelligent agents j The inverse attention weights are calculated using the following formula: ； in, For the first i、l Dynamic dependencies of individual agents.

2. The multi-agent collaborative method based on autonomous reflection and reverse attention according to claim 1, characterized in that, In step S1, in the global information graph, nodes represent agents, edges represent the collaborative relationships between agents, and a timestamp mechanism is used. Ensure the timeliness of information and form global status information. The formula is as follows: ； in, For the first m An intelligent agent in time Status information at that time For the first m The timestamps used by each agent to update its state. , For the environment in time Status information at that time The timestamp for updating the environment state.

3. The multi-agent collaborative method based on autonomous reflection and reverse attention as described in claim 1, characterized in that, it introduces... Meta-learning capability leverages successful experiences from previous tasks to generate generalizable policy optimization rules, as shown in the following formula: ； in, For the first i The new policy parameters obtained by the agent after reflection For the first i The old policy parameters of the agent before reflection For the first i Meta-learning modules for each agent.

4. The multi-agent collaborative method based on autonomous reflection and reverse attention according to claim 1, characterized in that, By dynamically adjusting the resource allocation strategy based on inverse attention weights and resource requirements, the resource allocation amount is calculated using the following formula: ； in, For the first i The agent is assigned to the first j Resource allocation for each agent For intelligent agents i For intelligent agents l Inverse attention weights, For the first i An intelligent agent in t Total available resources at any given time For the first j The task is in t Resource requirements at all times For the first l The task is in t Resource requirements at all times.

5. The multi-agent collaborative method based on autonomous reflection and reverse attention according to claim 1, characterized in that, In step S4, the agent applies the optimized policy. During task execution, the formulas for updating the state vector and environment state are as follows: ； ； in, For the first i An intelligent agent in t The state vector at time +1 for t The environment state at time +1 and For state update functions, For the first i An intelligent agent at time t The action.

6. The multi-agent collaborative method based on autonomous reflection and reverse attention according to claim 5, characterized in that, After the task is completed, the efficiency, resource consumption, and agent collaboration performance are recorded by the task evaluation module E, as shown in the following formula: ； ； ； in, For the efficiency of task completion. For resource consumption, This describes the collaborative performance of intelligent agents. Let be the efficiency function. This is a resource consumption function. For collaborative performance functions, A set of task performance indicators, A set of collaborative behavior patterns of intelligent agents; Evaluation results As input to the self-reflection module, it is used to optimize the strategy for the next round, as shown in the following formula: ； in, For the first i The next round strategy for each agent For the first i The current policy of each agent.

Images