Redundant agent task allocation method based on posterior correction and greedy exchange

Abstract

The invention relates to the technical field of multi-agent distributed task allocation, in particular to a redundant agent task allocation method based on posterior correction and greedy exchange, which comprises the following steps: for each agent, performing communication interaction with a neighbor agent to obtain agent configuration information; a priori knowledge item is constructed based on the agent configuration information, iteration is carried out on the priori knowledge item, and the conditional probability that each agent selects each task is obtained; determining the task selected by each agent based on the conditional probability that each agent selects each task; performing communication interaction between each agent and a neighbor agent, obtaining a task selected by the neighbor agent, and updating the task selected by each agent according to whether task exchange between the agent and the neighbor agent can reduce task cost; intelligent agent task allocation is completed based on the updated tasks selected by the intelligent agents; according to the invention, the dependence on the sample number can be reduced, the parameter adjustment complexity is reduced and the distribution efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-agent distributed task allocation, and in particular to a redundant agent task allocation method based on posterior correction and greedy exchange. Background Art

[0002] Redundant agent system is a multi-agent system that has shown significant advantages in task allocation in various fields. Its deployment overcomes the shortcomings of a single agent and can perform complex tasks faster than a single agent in a centralized or distributed manner.

[0003] Distributed task allocation is an application of the Distributed Constraint Optimization Problem (DCOP). In distributed task allocation, the system needs to allocate a series of tasks to a set of resources under the premise of distributed computing and limited communication topology to maximize certain performance indicators such as efficiency, balance or cost-effectiveness.

[0004] According to previous research on DCOP and task allocation in multi-agent systems, search-based algorithms guarantee optimality, but for multi-agent systems with redundant numbers, such as unmanned bee colonies, they are often not the main optimization goal. A large number of agents ensures that the optimization results meet the given task constraints even if they are not optimal. Sampling-based algorithms are highly dependent on the number of samples and have the problem of high complexity in parameter adjustment. Although various improvements to the auction algorithm can be used for multi-robot task allocation after problem decomposition, they are very inefficient when facing large-scale problems. Summary of the invention

[0005] In view of the above problems, the present invention provides a redundant intelligent agent task allocation method based on posterior correction and greedy exchange, which solves the technical problems in the existing distributed constraint optimization and task allocation methods that are highly dependent on the number of samples, have high parameter adjustment complexity and very low efficiency.

[0006] The present invention provides a redundant agent task allocation method based on posterior correction and greedy exchange, comprising the following steps:

[0007] Step S1: For each agent, communicate and interact with neighboring agents to obtain agent configuration information, wherein the agent configuration information includes: the total number of neighboring agents, the total number of all agents, and the number of agents required for each task; the neighboring agents are the set of agents within a specified spatial distance from the current agent;

[0008] Step S2: constructing a priori knowledge items based on the agent configuration information, iterating the priori knowledge items, and obtaining the conditional probability of each agent selecting each task;

[0009] Step S3, determining the task selected by each agent based on the conditional probability of each agent selecting each task;

[0010] Step S4: Each agent communicates and interacts with its neighboring agents to obtain the tasks selected by the neighboring agents, and updates the tasks selected by each agent based on whether the task exchange between the agent and the neighboring agents can reduce the task cost;

[0011] Step S5: completing the agent task allocation based on the updated tasks selected by each agent.

[0012] Preferably, step S2 specifically includes:

[0013] Step S2-1, initializing the prior knowledge items based on the number of agents required for each task;

[0014] Step S2-2, obtaining the information entropy of each agent based on the total number of neighboring agents and the total number of all agents;

[0015] Step S2-3, obtain iterative configuration information, including learning rate, bias term and number of redundant groups, iterate the prior knowledge items based on the information entropy, iterative configuration information and ability to complete tasks of each agent, and obtain the conditional probability of each agent selecting each task.

[0016] Preferably, in step S2-1, the initialization prior knowledge item is obtained by the following expression:

[0017]

[0018] in, represents the prior knowledge item of agent n in selecting task μ at the first iteration, is the number of agents required for task μ.

[0019] Preferably, in step S2-2, the information entropy of each agent is obtained by the following expression:

[0020]

[0021] in, is the information entropy of agent n at the i-th iteration, N is the total number of agents, d n is the total number of neighboring agents of agent n.

[0022] Preferably, in step S2-3, the conditional probability of each agent selecting each task is expressed as:

[0023]

[0024]

[0025] in, is the normalized conditional probability of agent n selecting task μ based on prior knowledge and a posteriori knowledge at the i-th iteration, where A and B are prior knowledge and a posteriori knowledge respectively. is the conditional probability of agent n selecting task μ based on prior knowledge and posterior knowledge at the i-th iteration, M is the total number of tasks, m is the task index, δ is the bias term, α is the learning rate, R is the number of redundant groups, represents the total number of agents that select task μ among the neighboring agents of agent n, ReLU(·) represents the linear rectification function, e nμ Indicates whether agent n is capable of completing task μ, represents the prior knowledge item of agent n in selecting task μ at the i-th iteration, is the normalized conditional probability of agent n selecting task μ based on prior knowledge and posterior knowledge at the i-1th iteration.

[0026] Preferably, step S4 specifically includes:

[0027] Step S4-1, based on the communication results between the current agent and the neighboring agent, obtain the task selected by the neighboring agent;

[0028] Step S4-2, determining whether the current agent and the neighboring agents can complete the exchanged tasks and reduce the task cost after the tasks selected by the agents are exchanged;

[0029] If the task after the exchange can be completed and the task cost can be reduced, the tasks selected by the current agent and the current neighbor agent are exchanged; after multiple iterations, the tasks selected by each agent are updated.

[0030] Preferably, step S4-2 specifically includes:

[0031] After multiple iterations, the tasks selected by each agent are updated. For the jth iteration:

[0032] For agent n, the neighbor agent set neighbors n , traverse neighbors n For each neighbor η in the network, if agent n can complete the task x of neighbor η η , and the neighbor η can complete the task of agent n in the jth iteration And meet Then the tasks of agent n and neighbor η are exchanged; where f n(·) is the objective function of the task cost required to complete the task of agent n, f η (·) is the objective function of the task cost required to complete the task of neighbor η.

[0033] Compared with the prior art, the present invention has at least the following beneficial effects:

[0034] (1) The present invention provides an entropy-driven EPC-GS method for decomposing a distributed constrained optimization problem (DCOP) into constraint satisfaction and optimization. The constraint on the number of agents for each task can be satisfied by iteration only through local communication. After the task allocation method provided by the present invention is completed, each agent is a completely independent decision node, and can perform distributed multi-agent dynamic task allocation learning.

[0035] (2) The present invention provides a posterior correction method for estimating group information based on neighborhood information. By comprehensively considering prior knowledge and neighbor information, a reasonable selection probability is calculated to meet the quantity constraint in multiple iterations. It can improve the efficiency and rationality of task allocation, enhance the decision-making ability of intelligent agents in complex environments, and make task allocation more flexible and adaptable.

[0036] (3) The present invention provides a greedy exchange method based on the greedy idea. Without changing the constraints that are satisfied, a solution that is not far from the theoretical optimal solution can be obtained through simple interaction rules. The flexibility and adaptability of task allocation are improved, so that intelligent agents can collaborate more effectively in a dynamic environment, and ultimately achieve a significant improvement in overall performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings are only for the purpose of illustrating particular embodiments and are not to be construed as limiting the invention.

[0038] Figure 1 Flowchart of the redundant agent task allocation method based on posterior correction and greedy exchange provided by the present invention.

[0039] Figure 2 This is a schematic diagram of the EPCA architecture provided by the present invention.

[0040] Figure 3 This is a schematic diagram of the GSA architecture provided by the present invention.

[0041] Figure 4 This is a schematic diagram of the EPC-GS architecture provided by the present invention.

[0042] Figure 5 The quantity change curve diagram provided by the present invention.

[0043] Figure 6 A schematic diagram of the change of selection probability information entropy provided by the present invention.

[0044] Figure 7 This is a schematic diagram of the cost change curve provided by the present invention.

[0045] Figure 8 Box plot of the ratio limit changes provided by the present invention.

[0046] Fig. 9 This is a schematic diagram comparing the convergence time of EPC-GS and CBAA provided by the present invention.

[0047] Fig.10 This is a schematic diagram of the EPCA effect of a non-connected graph provided by the present invention.

[0048] Fig.11 This is a schematic diagram of the EPC-GS dynamic task allocation simulation provided by the present invention. DETAILED DESCRIPTION

[0049] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0050] An intelligent agent is an entity with computing capabilities that can autonomously perform tasks, make decisions, and interact with the environment or other intelligent agents. An intelligent agent can be a drone, a robot, or other forms of automated systems. Distributed task allocation by intelligent agents refers to the process in which the intelligent agents autonomously allocate and coordinate tasks in a system composed of multiple intelligent agents, such as allocating multiple delivery tasks to multiple drones in a drone delivery system, and allocating different production tasks to multiple robots in an intelligent manufacturing factory. The present invention provides an efficient task allocation method to address the above problems.

[0051] In order to illustrate the effectiveness of the method proposed by the present invention, the above technical solution of the present invention is described in detail below through a specific embodiment. Figure 1 As shown, a redundant agent task allocation method based on posterior correction and greedy exchange is disclosed, and the specific implementation steps are as follows:

[0052] Step S1, each agent communicates and interacts with its neighboring agents to obtain agent configuration information, wherein the agent configuration information includes: the total number of neighboring agents, the total number of all agents, and the number of agents required for each task; the neighboring agents are agents within a specified spatial distance from the current agent.

[0053] Step S2: construct a priori knowledge items based on the agent configuration information, iterate the prior knowledge items, and obtain the conditional probability of each agent selecting each task.

[0054] Step S3: Determine the task selected by each agent based on the conditional probability of each agent selecting each task.

[0055] In a multi-agent system, agents that are close to the current agent in space can share information through communication and collaboration, thereby jointly completing complex goals. The set of agents that are close to the current agent in space is referred to as the agent's neighbors. In the present invention, the agent's neighbors include the agent itself.

[0056] In this step, based on the idea of ​​estimating group information using neighbor information, the constraints are satisfied in multiple iterations by comprehensively considering prior knowledge and neighbor information. The architecture of this step is as follows Figure 2 As shown in Figure 1, the selection probability is adjusted through multiple iterations to obtain the task selected by each agent.

[0057] Meeting the constraints on the number of agents required for the task is actually a matter of increasing the number when it is insufficient and not reducing it when it is sufficient. This is similar to what ReLU (Rectified Linear Unit) can achieve.

[0058] For distributed task allocation, the random variable X represents the set of task choices of all agents in the agent group, which contains the tasks selected by each agent. When the choices of the neighbors of agent n are known, the choices of non-neighbors need to be estimated. n represents the task selected by agent n, P(X n ) indicates X n The probability of occurrence is estimated as frequency. The information entropy of agent n can be calculated by the probability, so that the information entropy of agent n represents the uncertainty of its neighbors' information.

[0059] In the present invention, the total number of tasks is M, the number of agents is N, m is the task index, and n is the agent index. The number of agents that select task m from the neighbors of agent n is The number of neighbors of agent n is d n , In general, the number of elements in X should be close to Calculate the magnitude of each element X in X k The frequency cost of (k is the index of the element in X) is very large, so suppose that each X k There are no repetitions, that is, their frequencies are all 1; it is also assumed that the agent has no ability constraints, so the probability of each task selection is The number of elements in X is

[0060] In the calculation of information entropy, the logarithm is usually based on 2 and the unit is bit. However, other logarithmic bases can also be used. For simplicity, based on M, the information entropy of agent n is It is approximately as follows:

[0061]

[0062] in, is the information entropy of agent n at the i-th iteration. This expression shows that It can be approximated by the number of non-neighbors of agent n at iteration i. In other words, the more non-neighbors an agent has, the less reliable information it has and the greater the entropy.

[0063] The prior knowledge item for redundant multi-agent task allocation is set as:

[0064]

[0065] in, represents the prior knowledge item of agent n in selecting task μ at the i-th iteration; is the number of agents required for task μ, A and B are prior knowledge and a posteriori knowledge respectively, is the normalized conditional probability of agent n selecting task μ based on prior knowledge and posterior knowledge at the i-1th iteration. In the first iteration, is set to the number of agents required for task μ

[0066] In order to avoid the first agent making decisions having a greater impact on the group than other agents when making sequential decisions, parallel decision making is adopted to Set it to the selection probability of task μ in the previous iteration, so that it can inherit the correct probability of the previous iteration. Then the probability expression of agent n selecting task μ in the i-1th iteration is:

[0067]

[0068] in, is the conditional probability of agent n selecting task μ at the i-th iteration under the condition of prior knowledge and posterior knowledge. A represents the environmental knowledge required for the number of task agents in the first iteration. In subsequent iterations, A also includes the knowledge learned in the previous iterations, collectively referred to as prior knowledge. δ is a bias term, which is used to prevent the algorithm from running when the agent's neighbors are the entire group, and can be set to a small positive real number. α is the learning rate, which is a positive real number greater than 1. R represents the number of redundant groups, which can be expressed as calculate. represents the total number of agents that choose task μ among the neighbors of agent n, d n represents the number of neighbors of agent n, which can be obtained by calculate. Compute the difference between the proportion of selected tasks μ in the neighbors of agent n and the proportion required by the environment knowledge. The agent will use this difference to correct its own selection probability of the corresponding task with a learning rate of α, and the probability of selecting the task will not increase when the number of agents in the neighborhood that select task μ satisfies the environment knowledge constraint. ReLU(·) represents the linear rectification function. is the conditional probability of agent n selecting task m at the i-th iteration based on prior knowledge and posterior knowledge. In formula (4), normalization is achieved by dividing by the sum of the probabilities of selecting each task.

[0069] According to the formula, the larger α is, the stronger the ability of the agent to make corrections based on posterior knowledge. The probability of all non-optimal choices of each agent will approach 0 faster, and the agent's decision will be more firm. Therefore, a larger learning rate will not only increase the convergence speed, but also reduce the amplitude after convergence. However, if the learning rate is too large, the denominator in formula (4) will be 0, and the iterative process will not work, so it is necessary to choose a suitable learning rate.

[0070] The pseudo code of this step is shown in Table 1.

[0071] Table 1

[0072]

[0073]

[0074] Among them, e nμ Indicates whether agent n is capable of completing task μ. If agent n is capable of completing task μ, then e nμ =1, Tasks is the collection of all tasks.

[0075] Step S4: Each intelligent agent communicates and interacts with its neighboring intelligent agents to obtain the tasks selected by the neighboring intelligent agents, and updates the tasks selected by each intelligent agent based on whether the task exchange between the intelligent agent and the neighboring intelligent agents can reduce the task cost.

[0076] Step S5: completing the agent task allocation based on the updated tasks selected by each agent.

[0077] After completing the decision-making process of the previous steps, only one of the solutions that satisfies the constraints is obtained. However, it is not a solution that minimizes the objective function, and the present invention deals with this problem.

[0078] For each agent, evaluate whether exchanging tasks with all its neighbors will lead to a reduction in the objective function representing the cost. If a swap is found that reduces the overall cost, perform the swap. Repeat this process for all agents until no swap can further reduce the overall cost. This step requires that both agents can complete the swapped tasks. The architecture of this step is as follows: Figure 3 As shown, the total cost is reduced by exchanging tasks through multiple iterations, and finally the tasks for assigning agents are obtained.

[0079] The pseudo code of this step is shown in Table 2.

[0080] Table 2

[0081]

[0082]

[0083] in Indicates whether agent n can complete the task x of its neighbor η η , Indicates whether the neighbor η can complete the task of the current iteration round of agent n f n (·) is the objective function of the task cost required to complete the task of agent n, f η (·) is the objective function of the task cost required to complete the task of neighbor η. The method in this step only performs the exchange task operation, so the number constraints satisfied will not change. This step selects the current best decision at each step.

[0084] like Figure 4 As shown, the entropy-driven posterior correction and greedy swapping method (EPC-GS) provided by the present invention first uses the entropy-driven posterior correction algorithm (EPCA) to enable individuals to learn the current decisions that need to be made through interactions with their neighbors, so that the group's decision meets the constraints; then, under the premise of satisfying the constraints, the greedy swap algorithm (GSA) is used to exchange with neighbors to reduce the group cost, thereby completing the distributed task allocation.

[0085] The present invention also evaluates the effect of the EPC-GS method provided by the present invention.

[0086] A. Comparison of EPC-GS and CBAA in the case of connected graph.

[0087] The consensus-based auction algorithm (CBAA) can also solve the distributed single-task allocation problem. The simulation parameters of the present invention are shown in Table 3.

[0088] Table 3

[0089]

[0090] When the communication topology is a connected graph, EPC-GS is used for redundant agent task allocation, and CBAA is used for one-to-one task allocation to solve the multi-agent distributed task allocation problem. Figure 5 As shown, Figure 5 The results show that under the premise of agent redundancy, EPCA only needs a small number of iterations to meet the constraints of environmental information.

[0091] Figure 6 The curve showing the change of information entropy of the agent's choice probability during the iteration process in EPCA. The decrease of information entropy reflects the decrease of choice uncertainty. Figure 6 The results show that after many iterations, the information entropy of most agent selection probabilities converges to 0 or some other value.

[0092] The total cost is reduced by exchanging tasks through GSA. The total cost as a function of exchange is as follows: Figure 7 The ratio bounds obtained by simulation under different problem scales are shown in Figure 8 As shown, Figure 8 It is shown that the gap between the approximate solution produced by GSA in EPC-GS and the optimal solution is about 2%, and it is independent of the problem size.

[0093] The convergence time of the two algorithms is compared as follows: Fig. 9 The summary table of performance comparison is shown in Table 4.

[0094] Table 4

[0095]

[0096] After a finite number of iterations, both methods can satisfy the constraints, and the algorithms are very stable after repeated runs. However, without sacrificing optimality, compared with CBAA, which takes a long time to reach a consensus, EPC-GS has the advantage of using probability to make decisions in terms of efficiency.

[0097] B. EPCA in case of non-connected graph.

[0098] Using the parameters defined in Table 3, when the communication topology is a non-connected graph, the general auction algorithm cannot completely resolve the conflict through neighbor information, so the final result cannot meet the constraints. However, EPCA uses neighbor information to approximate the overall choice. When the graph is not connected, EPCA still uses neighbor selection to estimate the overall choice, and then makes the group's choice meet the constraints, such as Fig.10 As shown in Figure 2, the convergence of EPCA is only related to the average degree of the graph, and has nothing to do with whether the graph is connected or not. After several iterations, the number of agents selected for all three tasks satisfies the constraints.

[0099] C. Dynamic performance of EPC-GS.

[0100] In the simulation, the agents in the group are divided into three categories according to the capability constraint matrix, represented by squares, circles and triangles respectively. The capability constraints and numbers of agents in EPC-GS and CBAA are defined in Table 5.

[0101] Table 5

[0102]

[0103] Since each task can only be selected by a subset of agents, the shapes of agents that should appear at different task points in the simulation results are shown in Table 5.

[0104] Table 5

[0105]

[0106] The movement of the group uses the Reynolds-Boid model, and the parameters used are defined in Table 6.

[0107] Table 6

[0108]

[0109] Among them, w s 、w a 、w c are the weights of the three forces of separation, alignment, and aggregation. The positions of Task 2 and Task 3 are constant, (0,8) and (6,8) respectively. The position of Task 1 is time-varying. Its trajectory is defined as a circle with a center of (0,0) and a radius of 6m. The speed of movement is 1.2πm / s. Task 2 disappears during 3s-7s and appears at the rest of the time.

[0110] The dynamic task allocation simulation of EPC-GS is as follows Fig.11 As shown, Fig.11As shown in (a) of , after the simulation starts, the agent moves towards the assigned task point. When task 3 disappears at 3s < t < 7s, the group can dynamically adjust the assignment according to the algorithm. When t ≥ 7s and task 3 appears again, the group dynamically adjusts the assignment according to the algorithm. The results show that EPC-GS can use its fast convergence speed to perform real-time dynamic task allocation in terms of time and space.

[0111] Although the specific embodiments of the present invention depict various actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the particular order shown or in a sequential order, or requiring all the illustrated actions or steps to be performed to achieve the desired result. In certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the present disclosure. Certain features described in the context of separate embodiments may also be implemented in combination in a single implementation. Conversely, the various features described in the context of a single implementation may also be implemented separately or in any suitable sub-combination in multiple implementations. As described above, only the specific embodiments of the present invention are preferred, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

[0112] As described above, only the specific embodiments of the present invention are preferred, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A redundant agent task allocation method based on posterior correction and greedy exchange, characterized in that: The following steps are involved: Step S1, each agent communicates and interacts with its neighboring agents to obtain agent configuration information, wherein the agent configuration information includes: the total number of neighboring agents, the total number of all agents, and the number of agents required for each task; the neighboring agents are agents within a specified spatial distance from the current agent; Step S2: constructing a priori knowledge items based on the agent configuration information, iterating the priori knowledge items, and obtaining the conditional probability of each agent selecting each task; Step S3, determining the task selected by each agent based on the conditional probability of each agent selecting each task; Step S4: Each agent communicates and interacts with its neighboring agents to obtain the tasks selected by the neighboring agents, and updates the tasks selected by each agent based on whether the task exchange between the agent and the neighboring agents can reduce the task cost; Step S5: completing the agent task allocation based on the updated tasks selected by each agent.

2. The redundant agent task allocation method based on posterior correction and greedy exchange according to claim 1 is characterized in that: Step S2 specifically includes: Step S2-1, initializing the prior knowledge items based on the number of agents required for each task; Step S2-2, obtaining the information entropy of each agent based on the total number of neighboring agents and the total number of all agents; Step S2-3, obtain iterative configuration information, including learning rate, bias term and number of redundant groups, iterate the prior knowledge items based on the information entropy, iterative configuration information and ability to complete tasks of each agent, and obtain the conditional probability of each agent selecting each task.

3. The redundant agent task allocation method based on posterior correction and greedy exchange according to claim 2 is characterized in that: In step S2-1, the initialization prior knowledge item is obtained by the following expression: in, represents the prior knowledge item of agent n in selecting task μ at the first iteration, is the number of agents required for task μ.

4. The redundant agent task allocation method based on posterior correction and greedy exchange according to claim 3 is characterized in that: In step S2-2, the information entropy of each agent is obtained by the following expression: in, is the information entropy of agent n at the i-th iteration, N is the total number of agents, d n is the total number of neighboring agents of agent n.

5. The redundant agent task allocation method based on posterior correction and greedy exchange according to claim 4 is characterized in that: In step S2-3, the conditional probability of each agent selecting each task is expressed as: in, is the normalized conditional probability of agent n selecting task μ based on prior knowledge and a posteriori knowledge at the i-th iteration, where A and B are prior knowledge and a posteriori knowledge respectively. is the conditional probability of agent n selecting task μ based on prior knowledge and posterior knowledge at the i-th iteration, M is the total number of tasks, m is the task index, δ is the bias term, α is the learning rate, R is the number of redundant groups, represents the total number of agents that select task μ among the neighboring agents of agent n, ReLU(·) represents the linear rectification function, e nμ Indicates whether agent n is capable of completing task μ, represents the prior knowledge item of agent n in selecting task μ at the i-th iteration, is the normalized conditional probability of agent n selecting task μ based on prior knowledge and posterior knowledge at the i-1th iteration.

6. The redundant agent task allocation method based on posterior correction and greedy exchange according to claim 5 is characterized in that: Step S4 specifically includes: Step S4-1, based on the communication results between the current agent and the neighboring agent, obtain the task selected by the neighboring agent; Step S4-2, determining whether the current agent and the neighboring agents can complete the exchanged tasks and reduce the task cost after the tasks selected by the agents are exchanged; If the task after the exchange can be completed and the task cost can be reduced, the tasks selected by the current agent and the current neighbor agent are exchanged; after multiple iterations, the tasks selected by each agent are updated.

7. The redundant agent task allocation method based on posterior correction and greedy exchange according to claim 6 is characterized in that: Step S4-2 specifically includes: After multiple iterations, the tasks selected by each agent are updated. For the jth iteration: For agent n, the neighbor agent set neighbors n , traverse neighbors n For each neighbor η in the network, if agent n can complete the task x of neighbor η η , and the neighbor η can complete the task of agent n in the jth iteration And meet Then the tasks of agent n and neighbor η are exchanged; Among them, f n (·) is the objective function of the task cost required to complete the task of agent n, f η (·) is the objective function of the task cost required to complete the task of neighbor η.

Images