Multi-agent cooperation method based on autonomous reflection and reverse attention

Abstract

The invention specifically discloses a multi-agent cooperation method based on autonomous reflection and reverse attention, and relates to the technical field of artificial intelligence. According to the method, firstly, through state perception and information sharing among intelligent agents, a global information graph is constructed to obtain task dynamics; and then, through an autonomous reflection mechanism, according to indexes such as task completion degree, resource consumption and decision execution time, evaluating a historical decision behavior value, marking and optimizing a low-efficiency decision, introducing meta-learning to generate an optimization rule, and meanwhile, detecting an environment change optimization strategy. Then, a reverse attention mechanism is used for modeling the dynamic dependency relationship between the agents, a weight matrix is constructed, resources are allocated to low-priority tasks, and resource allocation is managed and controlled; and finally, the agent executes the task and evaluates task completion efficiency, resource consumption and collaborative performance, and an evaluation result is used for optimizing a next round of strategy. By adopting the method, the collaboration efficiency of the intelligent agent is improved, the resource allocation is optimized, and the autonomous decision-making capability of the intelligent agent is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-agent collaboration method based on autonomous reflection and reverse attention. Background Art

[0002] Multi-agent systems are widely used in fields such as autonomous driving, robotic swarms, and intelligent transportation. A core challenge lies in the efficiency of collaboration between agents. Traditional approaches rely on centralized control or distributed algorithms based on fixed policies to achieve collaboration. However, these approaches exhibit significant shortcomings in dynamic and uncertain environments. This is particularly true in the following areas:

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

[0004] (2) Insufficient coordination efficiency: Current attention mechanisms mainly focus on local optimization and fail to fully consider the dynamic dependencies between agents. For example, in intelligent transportation systems, the coordination between different vehicles requires real-time dynamic adjustment, but traditional methods only focus on local traffic flow optimization and ignore global traffic mobility.

[0005] However, in the dynamic allocation of task priorities, traditional methods lack effective mechanisms and insufficient resources are allocated to low-priority tasks, which easily causes bottleneck tasks to be delayed, thereby affecting the overall system's collaborative efficiency.

[0006] (3) Uneven resource allocation: In heterogeneous task scenarios, the resource capabilities of different agents may vary significantly, but traditional methods often adopt an average allocation strategy, ignoring the specific resource requirements of the task. For example, in a drone swarm, a task may require some drones to have additional computing or energy support, but improper resource allocation can lead to mission failure or inefficiency. Insufficient dynamic resource reallocation capabilities limit the system's resilience and adaptability under high load conditions. Summary of the Invention

[0007] The purpose of this invention is to propose a multi-agent collaboration method based on autonomous reflection and reverse attention to solve the problems of traditional multi-agent collaboration methods such as insufficient autonomous thinking ability, low collaboration efficiency, and uneven resource allocation, 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 objectives, the present invention proposes a multi-agent collaboration method based on autonomous reflection and reverse attention, comprising the following steps:

[0009] Step S1, state perception and information sharing among agents: The agents obtain state information of themselves and the environment through sensors, construct state vectors and environment states E(t), and use communication networks to share information in real time to build a global information graph G(t) = (V, E(t)), where V = {1, 2, ..., N}, is the set of agents, and N is a positive integer;

[0010] Step S2: Build an autonomous reflection mechanism: The agent regularly evaluates the behavioral value of historical decisions and marks the inefficient part of the strategy according to the behavioral value. i ), through the reflection module R i For strategy π i Perform adaptive optimization and introduce meta-learning capabilities to generate strategy optimization rules, detecting environmental changes in real time to optimize the current strategy;

[0011] Step S3: Construct an inverse attention mechanism: globally model the dynamic dependency relationship, optimize it through the graph neural network (GNN), use the inverse attention mechanism to construct the weight matrix α(t), and dynamically adjust the resource allocation;

[0012] Step S4, task execution and evaluation: The agent performs the task according to the optimized strategy. Execute the task and feed back the status of the execution process to the global information graph G(t). The task completion efficiency, resource consumption and agent collaboration performance are recorded through the task evaluation module. The evaluation results are used to optimize the next round of strategies.

[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] Among them, 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 sign;

[0016] The environment state E(t) is represented as the set of all ordered pairs (i, j) that satisfy the conditions that the i-th agent and the j-th agent can cooperate at time t, where i = 1, 2...N, j = 1, 2...N.

[0017] Preferably, in step S1, in the global information graph, nodes represent agents, edges represent collaborative relationships between agents, and the timeliness of information is guaranteed by the timestamp mechanism τ to form 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] Among them, s m (τ m ) is the mth agent at time τ m State information at time τ m The timestamp of the state update of the mth agent, m = 1, 2...N, E(τ E ) is the environment at time τ E State information at time τ E Timestamp for when the environment status was updated.

[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] Among them, V i (h) is the behavior value of the i-th agent, w1, w2, w3 are weight coefficients, TaskComletion 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;

[0024] Inefficient part MarkedLowEfficiency(πi ) is represented as strategy π i Zhong and h|V i (h) is associated and less than θ V part;

[0025] Among them, θ V is the value threshold; through the reflection module R i The policy π for the agent i Perform adaptive optimization, and the calculation formula is as follows:

[0026]

[0027] in, is the new strategy after reflection of the i-th agent, The old policy before reflection for the i-th agent.

[0028] Preferably, the meta-learning capability is introduced to use the successful experience from previous tasks to generate generalizable strategy optimization rules. The formula is as follows:

[0029]

[0030] in, is the new strategy parameter obtained by the i-th agent after reflection, is the old policy parameter of the i-th agent before reflection, M i is the meta-learning module of the i-th agent.

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

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

[0033] Among them, D ij (t) is the dynamic dependency relationship between the i-th and j-th agents, 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 requirement of the nth task at time t;

[0034] It is optimized through the graph neural network GNN, and the node embedding update formula of GNN is:

[0035]

[0036] Among them, W1 and W2 are weight matrices, is the embedding vector of the i-th node at the k+1th layer at time t, is the embedding vector of the i-th node at the k-th layer at time t, is the embedding vector of the jth node at the kth layer at time t, σ is the activation function, b is the bias term, and k is the number of layers of the GNN;

[0037] The node embedding vector is continuously updated to achieve dynamic adjustment. The weight matrix is ​​constructed using the inverse attention mechanism. The formula is as follows:

[0038] α(t)=[α ij (t)] N×N ;

[0039] Among them, α ij (t) is the inverse attention weight of agent i to agent j, and the formula is as follows:

[0040]

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

[0042] Preferably, the resource allocation strategy is dynamically adjusted by inverse attention weight and resource demand. The resource allocation amount is calculated as follows:

[0043]

[0044] Among them, R ij (t) is the amount of resources allocated by the i-th agent to the j-th agent, α il (t) is the reverse attention weight of agent i to agent l, R total_available (i, t) is the total available resources of the i-th agent at time t, R demand (j, t) is the resource requirement of the jth task at time t, R demand (l,t) is the resource requirement of the lth task at time t.

[0045] Preferably, in step S4, the agent uses the optimized strategy During the task execution, the state vector and environment state update formula 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] Among them, s i (t+1) is the state vector of the i-th agent at time t+1, E(t+1) is the state of the environment at time t+1, and f state (·) and f environment (·) is the state update function, a i (t) is the action of the i-th agent at time t.

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

[0050]

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

[0052]

[0053] Among them, Eff is the efficiency of task completion, Cres is the resource consumption, Ccollab is the collaborative performance of the intelligent agent, and E eficiency (·) is the efficiency function, E resource (·) is the resource consumption function, E colaboration (·) is the collaborative performance function, is a set of task performance indicators, is a set of collaborative behavior patterns of intelligent agents;

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

[0055]

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

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

[0058] (1) Improving autonomous decision-making capabilities: Through autonomous reflection mechanisms, intelligent agents can evaluate and optimize historical decisions, learn from mistakes, reduce unnecessary repetitions, and improve the rationality and adaptability of their own decisions. In complex and changing environments, intelligent agents can adjust their strategies in a timely manner according to environmental changes to better complete tasks.

[0059] (2) Enhanced collaborative efficiency: The reverse attention mechanism fully considers the dynamic dependencies between agents. By rationally allocating resources and prioritizing bottleneck tasks, it reduces task waiting time and improves the overall collaborative efficiency of the system. Whether in intelligent transportation systems or smart factories, more efficient task execution can be achieved.

[0060] (3) Optimizing resource allocation: For heterogeneous task scenarios, the method proposed in this invention can dynamically adjust the resource allocation strategy according to the specific resource requirements of the task and the resource capabilities of the intelligent agent, avoiding the average distribution and unreasonable use of resources, improving resource utilization, and enhancing the elastic adaptability of the system under high load conditions.

[0061] (4) Broad application prospects: The multi-agent collaboration method based on autonomous reflection and reverse attention proposed in this invention is applicable to a variety of multi-agent system application scenarios, such as unmanned driving, robot clusters, intelligent transportation, smart factories, 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 solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is an overall flow chart of a multi-agent collaboration method based on autonomous reflection and reverse attention in the present invention. DETAILED DESCRIPTION

[0064] To make the technical solutions, advantages, and purposes of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are part of the embodiments of the present invention, not all of them. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0065] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0066] Example 1

[0067] like Figure 1 FIG. 1 is a flowchart of a multi-agent collaboration method based on autonomous reflection and reverse attention according to the present invention, comprising the following steps:

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

[0069] The state vector formula is as follows:

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

[0071] Among them, 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 sign;

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

[0073] In the global information graph, nodes represent agents, edges represent collaborative relationships between agents, and the timeliness of information is guaranteed by the timestamp mechanism τ, forming the global state information S(t). The formula is as follows:

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

[0075] Among them, s m (τ m ) is the mth agent at time τ m State information at time τ m The timestamp of the state update of the mth agent, m = 1, 2...N, E(τ E ) is the environment at time τ E State information at time τ E Timestamp for when the environment status was updated.

[0076] 2. Build an autonomous reflection mechanism: The agent regularly evaluates the behavioral value of historical decisions and marks the inefficient parts of the strategy according to the behavioral value. i ), through the reflection module R i For strategy π i Perform adaptive optimization and introduce meta-learning capabilities to generate strategy optimization rules, detect environmental changes in real time and optimize the current strategy.

[0077] The formula for calculating behavior value is as follows:

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

[0079] DecisionTime i (h);

[0080] Among them, V i (h) is the behavior value of the i-th agent, w1, w2, w3 are weight coefficients, 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] Inefficient part MarkedLowEfficiency(π i ) is represented as strategy π i Zhong and h|V i (h) is associated and less than θ V part;

[0082] Among them, θ V is the value threshold; through the reflection module R i The policy π for the agent i Perform adaptive optimization, and the calculation formula is as follows:

[0083]

[0084] in, is the new strategy after reflection of the i-th agent, The old policy before reflection for the i-th agent.

[0085] By introducing meta-learning capabilities and leveraging successful experience from previous tasks, we can generate generalizable policy optimization rules. The formula is as follows:

[0086]

[0087] in, is the new strategy parameter obtained by the i-th agent after reflection, is the old policy parameter of the i-th agent before reflection, M i is the meta-learning module of the i-th agent.

[0088] 3. Build an inverse attention mechanism: Globally model dynamic dependencies, optimize through graph neural network (GNN), use the inverse attention mechanism to build the weight matrix α(t), and dynamically adjust resource allocation.

[0089] The calculation formula for dynamic dependencies is as follows:

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

[0091] Among them, D ij (t) is the dynamic dependency relationship between the i-th and j-th agents, 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 requirement of the nth task at time t;

[0092] It is optimized through the graph neural network GNN, and the node embedding update formula of GNN is:

[0093]

[0094] Among them, W1 and W2 are weight matrices, is the embedding vector of the i-th node at the k+1th layer at time t, is the embedding vector of the i-th node at the k-th layer at time t, is the embedding vector of the jth node at the kth layer at time t, σ is the activation function, b is the bias term, and k is the number of layers of the GNN;

[0095] The node embedding vector is continuously updated to achieve dynamic adjustment. The weight matrix is ​​constructed using the inverse attention mechanism. The formula is as follows:

[0096] α(t)=[[α ij (t)] N×N ;

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

[0098]

[0099] Among them, D il (t) is the dynamic dependency relationship between the i-th and l-th agents.

[0100] By inversely analyzing attention weights and resource requirements, the resource allocation strategy is dynamically adjusted. The resource allocation amount is calculated as follows:

[0101]

[0102] Among them, R ij (t) is the amount of resources allocated by the i-th agent to the j-th agent, α il (t) is the reverse attention weight of agent i to agent l, R total_available (i, t) is the total available resources of the i-th agent at time t, R demand (j, t) is the resource requirement of the jth task at time t, R demand (l,t) is the resource requirement of the lth task at time t.

[0103] 4. Task execution and evaluation: The agent performs tasks according to the optimized strategy. Execute the task and feed back the status of the execution process to the global information graph G(t). The task completion efficiency, resource consumption and agent collaboration performance are recorded through the task evaluation module. The evaluation results are used to optimize the next round of strategies.

[0104] During the task execution, the current state is continuously fed back to the global information graph. The state vector and environment state update formula are 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] Among them, s i (t+1) is the state vector of the i-th agent at time t+1, E(t+1) is the state of the environment at time t+1, and f state (·) and f environment (·) is the state update function, a i(t) is the action of the i-th agent at time t.

[0108] The task evaluation module E records the efficiency of task completion, resource consumption, and agent collaboration performance. The formula is as follows:

[0109]

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

[0111]

[0112] Among them, Eff is the efficiency of task completion, Cres is the resource consumption, Ccollab is the collaborative performance of the intelligent agent, and E eficiency (·) is the efficiency function, E resource (·) is the resource consumption function, E colaboration (·) is the collaborative performance function, is a set of task performance indicators, is a set of collaborative behavior patterns of intelligent agents;

[0113] The evaluation results {Ef, Cres} serve as the input of the autonomous reflection module and are used to optimize the next round of strategies. The formula is as follows:

[0114]

[0115] in, is the next round strategy of the i-th agent, is the current strategy of the ith agent.

[0116] Example 2

[0117] This example uses a drone swarm performing an inspection mission as an example. Before applying the self-reflection mechanism, relevant metrics were recorded: the total mission duration for the swarm was 100 minutes, the average energy consumption of a single drone was 80% of its battery life, the swarm selected inefficient paths 25 times during the mission, the average path planning decision time was 1.5 seconds per attempt, and the mission completion score was 95 out of 100.

[0118] During operation, the drone regularly evaluates the behavioral value of historical decisions using an autonomous reflection mechanism. Based on a weighted calculation of task completion, resource consumption, and decision execution time, it marks inefficient paths in the strategy. For example, if a path planning results in a long task completion time and high energy consumption, its behavioral value is lower than the set value threshold θ. V , then the path planning part is marked as an inefficient decision.

[0119] Then, the strategy is optimized through the reflection module. Based on historical inefficient paths, the weights of path planning are adjusted so that subsequent path planning gives priority to efficient paths. At the same time, meta-learning capabilities are used to learn from previous successful mission experiences and generate better strategy optimization rules. After a period of operation and reflection learning, relevant indicators were recorded again: the total time for the cluster to complete the same task was shortened to 90 minutes, a 10% reduction in task completion time; the average energy consumption of a single drone was reduced to 75% of the power, a 6.25% reduction in energy consumption; the number of inefficient path selections was significantly reduced to 10 times, a 60% reduction; the average decision time for path planning was also shortened to 1.3 seconds per time, a 13.3% reduction; and the task completion score was improved to 97 points, an increase of 2.1%.

[0120] This example fully verifies the effectiveness of the autonomous reflection mechanism in improving the efficiency of drone cluster missions, reducing resource consumption, optimizing path planning, and improving mission completion.

[0121] Example 3

[0122] This example uses a scenario where different robots collaborate to complete assembly tasks in a smart factory. Before applying the inverse attention mechanism, relevant system metrics were recorded: the system's average task waiting time was 5 minutes, the system's overall task completion time was 60 minutes, the average robot idle time accounted for 20% of the time, the resource allocation balance score was 70 points (out of 100), and the high-priority task delay rate reached 10%.

[0123] When applying the inverse attention mechanism, the dynamic dependencies between robots are first globally modeled. These dependencies are quantified based on task priorities and resource requirements, and then dynamically adjusted using a graph neural network to optimize node embedding vectors.

[0124] A weight matrix is ​​constructed and resources are allocated to low-priority tasks based on the weight matrix. For example, more robots are assigned to tasks with long wait times to reduce system bottlenecks. Regarding resource allocation management, resource allocation strategies are dynamically adjusted based on inverse attention weights and real-time task requirements, such as prioritizing tasks with high resource requirements.

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

[0126] This example demonstrates that the inverse attention mechanism can effectively reduce task waiting time in smart factory systems, improve resource utilization, and optimize resource allocation balance, thereby significantly improving the overall operational efficiency of smart factories. It is worth noting that any content not elaborated in detail in this invention is prior art and is well known to those skilled in the art.

[0127] Therefore, the present invention provides a multi-agent collaboration method based on autonomous reflection and reverse attention. Through the collaborative operation of state perception and information sharing among agents, autonomous reflection mechanism, reverse attention mechanism, and task execution and evaluation, it solves the problems of traditional multi-agent collaboration in dynamic environments, such as lack of autonomous thinking ability, low collaboration efficiency and uneven resource allocation, and improves the collaborative performance of multi-agent systems in complex scenarios.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-agent collaboration method based on autonomous reflection and reverse attention, characterized by: The following steps are involved: Step S1, state perception and information sharing among agents: The agents obtain state information of themselves and the environment through sensors, construct state vectors and environment states E(t), and use communication networks to share information in real time to build a global information graph G(t) = (V, E(t)), where V = {1, 2, ..., N}, is the set of agents, and N is a positive integer; Step S2: Build an autonomous reflection mechanism: The agent regularly evaluates the behavioral value of historical decisions and marks the inefficient part of the strategy according to the behavioral value. i ), through the reflection module R i For strategy π i Perform adaptive optimization and introduce meta-learning capabilities to generate strategy optimization rules, detecting environmental changes in real time to optimize the current strategy; Step S3: Construct an inverse attention mechanism: globally model the dynamic dependency relationship, optimize it through the graph neural network (GNN), use the inverse attention mechanism to construct the weight matrix α(t), and dynamically adjust the resource allocation; Step S4, task execution and evaluation: The agent performs the task according to the optimized strategy. Execute the task and feed back the status of the execution process to the global information graph G(t). The task completion efficiency, resource consumption and agent collaboration performance are recorded through the task evaluation module. The evaluation results are used to optimize the next round of strategies.

2. A multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 1, characterized in that: In step S1, the state vector formula is as follows: s i (t)=[p i (t),v i (t),e i (t),…] T ; Among them, 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 sign; The environment state E(t) is represented as the set of all ordered pairs (i, j) that satisfy the conditions that the i-th agent and the j-th agent can cooperate at time t, where i = 1, 2...N, j = 1, 2...N.

3. The multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 1 is characterized in that: In step S1, in the global information graph, nodes represent agents, edges represent the collaborative relationships between agents, and the timeliness of information is guaranteed by the timestamp mechanism τ, forming the global state information S(t). The formula is as follows: S(t)=(s1(τ1),τ1),(s2(τ2),τ2),…,(s N (t N ),t N ),(E(τ E ),t E ); Among them, s m (τ m ) is the mth agent at time τ m State information at time τ m The timestamp of the state update of the mth agent, m = 1, 2...N, E(τ E ) is the environment at time τ E State information at time τ E Timestamp for when the environment status was updated.

4. The multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 1 is characterized in that: In step S2, the behavior value calculation formula is as follows: V i (h)=w1·TaskCompletion i (h)-w2·ResourceConsumption i (h)-w3· DecisionTime i (h); Among them, V i (h) is the behavior value of the i-th agent, w1, w2, w3 are weight coefficients, 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; Inefficient part MarkedLowEfficiency(π i ) is represented as strategy π i Zhong and h|V i (h) is associated and less than θ V part; Among them, θ V is the value threshold; through the reflection module R i The policy π for the agent i Perform adaptive optimization, and the calculation formula is as follows: in, is the new strategy after reflection of the i-th agent, The old policy before reflection for the i-th agent.

5. A multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 4, characterized in that Meta-learning capability uses successful experience from previous tasks to generate generalizable policy optimization rules. The formula is as follows: in, is the new strategy parameter obtained by the i-th agent after reflection, is the old policy parameter of the i-th agent before reflection, M i is the meta-learning module of the i-th agent.

6. The multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 1 is characterized in that: In step S3, the calculation formula of the dynamic dependency is as follows: D ij (t)=β1·P task (n,t)+β2·R demand (n,t); Among them, D ij (t) is the dynamic dependency relationship between the i-th and j-th agents, 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 requirement of the nth task at time t; It is optimized through the graph neural network GNN, and the node embedding update formula of GNN is: Among them, W1 and W2 are weight matrices, is the embedding vector of the i-th node at the k+1th layer at time t, is the embedding vector of the i-th node at the k-th layer at time t, is the embedding vector of the jth node at the kth layer at time t, σ is the activation function, b is the bias term, and k is the number of layers of the GNN; The node embedding vector is continuously updated to achieve dynamic adjustment. The weight matrix is ​​constructed using the inverse attention mechanism. The formula is as follows: α(t)=[α ij (t)] N×N ; Among them, α ij (t) is the inverse attention weight of agent i to agent j, and the formula is as follows: Among them, D il (t) is the dynamic dependency relationship between the i-th and l-th agents.

7. A multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 6, characterized in that: By inversely analyzing attention weights and resource requirements, the resource allocation strategy is dynamically adjusted. The resource allocation amount is calculated as follows: Among them, R ij (t) is the amount of resources allocated by the i-th agent to the j-th agent, α il (t) is the reverse attention weight of agent i to agent l, R total_available (i, t) is the total available resources of the i-th agent at time t, R demand (j, t) is the resource requirement of the jth task at time t, R demand (l,t) is the resource requirement of the lth task at time t.

8. The multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 1 is characterized in that: In step S4, the agent uses the optimized strategy During the task execution, the state vector and environment state update formula are as follows: s i (t+1)=f state (s i (t),a i (t),E(t)); E(t+1)=f environment (E(t),{a1(t),a2(t),…,a N (t)}); Among them, s i (t+1) is the state vector of the i-th agent at time t+1, E(t+1) is the state of the environment at time t+1, and f state (·) and f environment (·) is the state update function, a i (t) is the action of the i-th agent at time t.

9. The multi-agent collaboration method based on autonomous reflection and reverse attention according to claim 8, characterized in that: After the task is completed, the task evaluation module E records the efficiency, resource consumption, and agent collaboration performance of the task. The formula is as follows: Cres=E resource ({s i (t)} i∈V ); Among them, Eff is the efficiency of task completion, Cres is the resource consumption, Ccollab is the collaborative performance of the intelligent agent, and E eficiency (·) is the efficiency function, E resource (·) is the resource consumption function, E colaboration (·) is the collaborative performance function, is a set of task performance indicators, is a set of collaborative behavior patterns of intelligent agents; The evaluation results {Eff, Cres} serve as the input of the self-reflection module and are used to optimize the next round of strategies. The formula is as follows: in, is the next round strategy of the i-th agent, is the current strategy of the ith agent.

Images