A distributed multi-agent task allocation method, system, terminal and storage medium

Abstract

The application relates to the technical field of task allocation, and discloses a distributed multi-agent task allocation method, a distributed multi-agent task allocation system, a terminal and a storage medium.The method comprises the following steps: acquiring a global task, performing decomposition processing to obtain an atomic task set, and acquiring the capability portraits of a plurality of agents; calculating the allocation probability distribution of a plurality of atomic tasks according to all the capability portraits and the atomic task set, and performing task evaluation on all the atomic tasks; performing conflict detection on the agent matching result obtained by matching the atomic tasks to obtain a conflict detection result; if there is a competition conflict, generating a global candidate scheme set, and obtaining a global task allocation scheme according to the global candidate scheme set; performing allocation processing on the global task according to the global task allocation scheme, and generating a target language explanation report according to all the decision tracks of the obtained task allocation result.The application can autonomously adjust the allocation strategy according to environmental changes, and improves the efficiency and accuracy of task allocation.

Description

Technical Field

[0001] This invention relates to the field of task allocation technology, and in particular to a distributed multi-agent task allocation method, system, terminal, and computer-readable storage medium. Background Technology

[0002] In task allocation scenarios (e.g., logistics and warehousing), existing task allocation methods include traditional task allocation methods based on centralized optimization, autonomous task allocation methods based on deep reinforcement learning, and distributed task allocation methods based on consensus auction mechanisms. Among these, traditional task allocation methods based on centralized optimization typically rely on a central coordinator or global planner in traditional multi-agent systems. For example, in a classic centralized allocation architecture, the system establishes a single coordinating agent responsible for collecting the state information (such as location, load, and capabilities) of all sub-agents and calculating the global optimum using predefined rules or optimization algorithms (such as linear programming), subsequently issuing task instructions to each executing agent. Autonomous task allocation methods based on deep reinforcement learning employ an Actor-Critic network architecture, allowing agents to learn strategies through trial and error in the environment. In multi-agent reinforcement learning scenarios, agents adjust policy network parameters based on environmental reward signals by sharing experience replay buffers or learning independently to achieve autonomous allocation of multi-step tasks. The distributed task allocation method based on the consensus auction mechanism involves each agent evaluating (bidding) a task based on local information, and resolving the competition conflict among multiple agents for the same task by exchanging evaluation information with neighbors and performing consensus negotiation, ultimately converging to a conflict-free allocation scheme, without relying on a central coordinating node throughout the entire process.

[0003] However, centralized allocation methods suffer from single points of failure, poor scalability, and an inability to autonomously adjust strategies according to environmental changes; traditional deep learning methods suffer from the problem that when the system makes an incorrect task allocation, users cannot understand the underlying logic (i.e., lack of interpretability), and cannot explain the multi-step inference trajectory and the interaction logic between agents over time; consensus algorithms suffer from a lack of decision interpretability.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a distributed multi-agent task allocation method, system, terminal, and storage medium, aiming to solve the problems of existing technologies being unable to autonomously adjust strategies according to environmental changes and lacking interpretability in allocation decisions, resulting in low efficiency and accuracy of task allocation.

[0006] To achieve the above objectives, the present invention provides a distributed multi-agent task allocation method, which includes the following steps: The global tasks of the logistics and warehousing scenario are obtained, the global tasks are decomposed and processed to obtain a set of atomic tasks, and the capability profiles of multiple intelligent agents are obtained. Based on the capability profiles of all the agents and the set of atomic tasks, the action probability distribution is calculated to obtain the allocation probability distribution of multiple atomic tasks. Based on all the allocation probability distributions, the atomic tasks are evaluated to obtain the target evaluation result. Based on the target evaluation results, agent matching is performed on the atomic task to obtain agent matching results, and conflict detection is performed on the agent matching results to obtain conflict detection results. If the conflict detection result indicates the existence of a competitive conflict, then candidate schemes are generated based on the conflict detection result to obtain a global candidate scheme set. Instructions are generated on the global candidate scheme set to obtain the allocation results of all atomic tasks. Finally, a global task allocation scheme is obtained based on all the allocation results. The global tasks are allocated according to the global task allocation scheme to obtain task allocation results. A report is generated based on all decision trajectories of the task allocation results to obtain a target language interpretation report.

[0007] Optionally, the distributed multi-agent task allocation method, wherein the step of calculating the action probability distribution based on the capability profiles of all agents and the set of atomic tasks to obtain the allocation probability distribution of multiple atomic tasks, and evaluating all atomic tasks based on all the allocation probability distributions to obtain the target evaluation result, specifically includes: Based on the capability profiles of all the agents, the local observation state of each agent is obtained, and all the local observation states and the set of atomic tasks are input into the policy network. Based on the policy network, task scores are calculated for all local observation states and the set of atomic tasks to obtain task scores for multiple atomic tasks. Then, all task scores are normalized to obtain multiple allocation probability distributions. The state value of all local observation states is evaluated based on the value network to obtain multiple target state values. Evaluation scores are calculated for all atomic tasks according to all the assigned probability distributions and all the target state values ​​to obtain multiple evaluation scores. All the evaluation scores are then compared with an evaluation threshold to obtain a comparison result. If the comparison result indicates that there is an evaluation score greater than the evaluation threshold, then the evaluation information of the agent for the atomic task is obtained based on the comparison result, and the evaluation information is broadcast to neighboring agents to obtain the target evaluation result.

[0008] Optionally, the distributed multi-agent task allocation method, wherein evaluating the state value of all the local observation states based on a value network to obtain multiple target state values ​​specifically includes: All the local observation states are input into the value network, and the state value of all the local observation states is evaluated based on the value network to obtain multiple state values; Error calculations are performed on all the stated state values ​​to obtain multiple temporal difference errors. Value updates are then performed on all the stated state values ​​based on all the stated temporal difference errors to obtain multiple target state values.

[0009] Optionally, the distributed multi-agent task allocation method, wherein the step of performing agent matching on the atomic tasks based on the target evaluation result to obtain agent matching results, and performing conflict detection on the agent matching results to obtain conflict detection results, specifically includes: Based on the target evaluation result, the atomic task is matched with an agent to obtain the agent matching result, and all target atomic tasks in the agent matching result are obtained. Conflict detection is performed on all the target atomic tasks, wherein the target atomic task is the atomic task that receives the evaluation information from the agent. If the target atomic task receives evaluation information from multiple agents, the conflict detection result is determined to be a competitive conflict, and the conflict type is determined, wherein the conflict type includes resource competition conflict, spatial conflict and temporal conflict; If it is detected that the target atomic task has not received evaluation information from multiple agents, then the conflict detection result is determined to be that there is no competition conflict.

[0010] Optionally, the distributed multi-agent task allocation method, wherein if the conflict detection result indicates the existence of a competition conflict, then candidate schemes are generated based on the conflict detection result to obtain a global candidate scheme set, instructions are generated on the global candidate scheme set to obtain the allocation results of all atomic tasks, and a global task allocation scheme is obtained based on all the allocation results, specifically including: If the conflict detection result indicates the existence of a competitive conflict, a local candidate solution set is generated based on the conflict detection result, and the local candidate solution set is exchanged and merged based on the agent to obtain a global candidate solution set. For each candidate allocation scheme in the global candidate scheme set, the team utility is calculated to obtain the total team utility, and the total team utility is decomposed into individual values ​​to obtain multiple marginal contribution values; Instructions are generated based on all the marginal contribution values ​​to obtain multiple task allocation instructions. Tasks are allocated to all agents based on all the task allocation instructions to obtain the allocation results of all atomic tasks. All the allocation results are then aggregated to obtain a global task allocation scheme.

[0011] Optionally, the distributed multi-agent task allocation method, wherein the step of allocating global tasks according to the global task allocation scheme to obtain task allocation results, and generating a target language interpretation report based on all decision trajectories of the task allocation results, specifically includes: The global tasks are allocated according to the global task allocation scheme to obtain the task allocation result, and all decision trajectories recorded in the task allocation result are obtained. The decision trajectories include state sequences, action sequences, reward sequences, tool call logs, inter-agent communication records, and conflict detection and elimination records. Nodes are extracted from all the decision trajectories to obtain multiple key decision nodes. A causal chain is constructed for each key decision node to obtain multiple causal chains, wherein the causal chains include a task allocation causal chain and an anomaly causal chain. All the causal chains are input into a large language model. Reports are generated for all the causal chains based on the large language model to obtain a language explanation report. The language explanation report is then validated for consistency to obtain a target language explanation report.

[0012] Optionally, in the distributed multi-agent task allocation method, the step of performing consistency verification on the language interpretation report to obtain the target language interpretation report specifically includes: The key entities in the language explanation report are compared item by item with the data of all the causal chains to obtain the data comparison results; If the data comparison results show inconsistencies, the language explanation report will be adjusted to obtain the target language explanation report. If the data comparison result shows no data inconsistency, then a target language interpretation report is obtained based on the data comparison result.

[0013] Optionally, the distributed multi-agent task allocation method, wherein the distributed multi-agent task allocation system includes: The task decomposition module is used to obtain global tasks in the logistics and warehousing scenario, decompose the global tasks to obtain a set of atomic tasks, and obtain capability profiles of multiple intelligent agents. The task evaluation module is used to calculate the action probability distribution based on the capability profiles of all the agents and the set of atomic tasks, obtain the allocation probability distribution of multiple atomic tasks, and evaluate all the atomic tasks based on all the allocation probability distributions to obtain the target evaluation result. The conflict detection module is used to perform agent matching on the atomic task based on the target evaluation result, obtain agent matching result, and perform conflict detection on the agent matching result to obtain conflict detection result; The task allocation module is used to generate candidate solutions based on the conflict detection result if the conflict detection result indicates that there is a competition conflict, obtain a global candidate solution set, generate instructions on the global candidate solution set to obtain the allocation results of all atomic tasks, and obtain a global task allocation scheme based on all the allocation results. The report generation module is used to allocate and process the global tasks according to the global task allocation scheme, obtain the task allocation results, and generate a report based on all decision trajectories of the task allocation results, thereby obtaining a target language interpretation report.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a distributed multi-agent task allocation program stored in the memory and executable on the processor, wherein when the distributed multi-agent task allocation program is executed by the processor, it implements the steps of the distributed multi-agent task allocation method as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a distributed multi-agent task allocation program, which, when executed by a processor, implements the steps of the distributed multi-agent task allocation method as described above.

[0016] In this invention, a global task for a logistics and warehousing scenario is acquired, decomposed to obtain a set of atomic tasks, and capability profiles of multiple agents are obtained. Based on the capability profiles of all agents and the set of atomic tasks, action probability distributions are calculated to obtain allocation probability distributions for the multiple atomic tasks. All atomic tasks are evaluated based on these allocation probability distributions to obtain target evaluation results. Agent matching is performed on the atomic tasks based on the target evaluation results to obtain agent matching results, and conflict detection is performed on the agent matching results to obtain conflict detection results. If the conflict detection results indicate competition or conflict, candidate solutions are generated based on the conflict detection results to obtain a global candidate solution set. Instructions are generated on the global candidate solution set to obtain allocation results for all atomic tasks, and a global task allocation scheme is obtained based on all allocation results. The global tasks are allocated according to the global task allocation scheme to obtain task allocation results, and a report is generated based on all decision trajectories of the task allocation results to obtain a target language interpretation report. This invention can autonomously adjust allocation strategies according to environmental changes, and the allocation decisions are interpretable, improving the efficiency and accuracy of task allocation. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the distributed multi-agent task allocation method of the present invention; Figure 2 This is a schematic diagram of the overall system architecture corresponding to the distributed multi-agent task allocation method of the present invention; Figure 3 This is a schematic diagram of the overall process of the distributed multi-agent task allocation method of the present invention; Figure 4 This is a schematic diagram of the Actor-Critic network structure and the genetic algorithm optimization structure in a preferred embodiment of the present invention; Figure 5 This is a schematic diagram of the strategy optimization process enhanced by the genetic algorithm in a preferred embodiment of the present invention; Figure 6 This is a schematic diagram of the process for generating an explanation report based on a decision trajectory in a preferred embodiment of the present invention; Figure 7 This is a structural diagram of a preferred embodiment of the distributed multi-agent task allocation system of the present invention; Figure 8 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0020] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0021] The distributed multi-agent task allocation method described in the preferred embodiment of the present invention, such as... Figure 1 As shown, the distributed multi-agent task allocation method includes the following steps: Step S10: Obtain the global tasks of the logistics and warehousing scenario, decompose the global tasks to obtain a set of atomic tasks, and obtain the capability profiles of multiple intelligent agents.

[0022] Specifically, existing technologies cannot autonomously adjust strategies according to environmental changes, and the allocation decisions lack interpretability, resulting in low efficiency and accuracy in task allocation. Therefore, this invention proposes a distributed multi-agent task allocation method. The implementation of this method is based on a system such as... Figure 2As shown, this system abandons the traditional single central scheduler model and adopts a closed-loop architecture of perception-planning-execution-interpretation. The system mainly consists of five core modules, which work together to achieve interpretable autonomous task allocation. These five core modules are the Perception Module, Planning Module, Execution Module, Communication Module, and Interpretation Generation Module (XAI Module). The Perception Module is responsible for building an internal model of the environment, receiving external sensor data or API (Application Programming Interface) information, calculating the environmental state, and identifying the states of other intelligent agents and environmental constraints (such as obstacles and communication range). The corresponding algorithm logic is as follows: ; in, For the first An intelligent agent at time The local observation status, The state transition function serves as the dynamic law of the encapsulated environment. For the first An intelligent agent at time The local observation status, These are the action commands that the intelligent agent ultimately outputs; among them, The input is and The output is the state at the next time step. The planning module employs a belief-wish-intention architecture combined with a large language model. This planning module receives the global task. The process is broken down into a set of executable atomic tasks, generating a preliminary action plan. The planning module also deploys an Actor-Critic (policy-value) network and a genetic algorithm optimizer, responsible for outputting the optimal action based on the current state. The execution module is responsible for calling specific tools to execute the task. The execution module provides real-time feedback on execution results (e.g., success, failure). It acts as the "limbs" of the agent, translating the action instructions output by the planning module into physical movements or logical operations. The communication module facilitates agent-to-agent communication (Agent Communication Language) based on a shared ontology, exchanging intentions and states. The interpretation generation module generates interpretations based on execution trajectories, rather than relying on static feature attribution.

[0023] The specific methods and procedures are as follows: Figure 3As shown, step one, distributed task decomposition and capability broadcasting, specifically involves obtaining the global tasks for the logistics and warehousing scenario. The global tasks within the logistics and warehousing scenario... The global task can be defined as "to sort and deliver 50 packages in the warehouse within 2 hours." This global task clearly defines the time constraints, task objects, and expected goals. The global task is then decomposed into a set of atomic tasks. Specifically, the planning module uses a large language model to process the global task. Decomposed into a set of atomic tasks ,in, For the first atomic task, For the second atomic task, For the first Atomic tasks; for example, this global task can be decomposed into: To sort package A onto the conveyor belt, To sort package B to the packing area, The task is to deliver package C to the door. The purpose of decomposition is to transform complex global tasks into independently assignable and executable atomic operation units, providing a task pool for subsequent distributed bidding. Afterwards, capability profiles of multiple agents are obtained, and each agent broadcasts its own capability profile through a communication module. (Including hardware configuration, current location, remaining battery power, and skill tags), not just broadcast status. Capability Profile Used to evaluate the agent's suitability for atomic tasks during subsequent bidding processes.

[0024] Step S20: Calculate the action probability distribution based on the capability profiles of all the agents and the set of atomic tasks to obtain the allocation probability distribution of multiple atomic tasks, and evaluate all the atomic tasks based on all the allocation probability distributions to obtain the target evaluation result.

[0025] Specifically, such as Figure 3 As shown, in step two, strategy optimization (bidding calculation) based on Actor-Critic and genetic algorithms, in this embodiment of the invention, an improved multi-agent reinforcement learning algorithm is used for task bidding and allocation. The specific processing procedure is as follows: Figure 4 As shown, the local observation state of each agent is obtained based on the capability profiles of all the agents. The system then inputs all the local observation states and the set of atomic tasks into the Actor network (policy network), and outputs the probability distribution of each atomic task. ,in, The parameters of the Actor network are assigned; the process of obtaining the probability distribution is as follows: the Actor network includes an online policy network and a target policy network; the online policy network receives the local observation states. Taking a set of atomic tasks as input, each atomic task is computed through a multi-layer neural network (including an attention mechanism). Score Then, it is normalized to a probability distribution using the Softmax function, and the corresponding expression is: ; in, For the first The scores of each atomic task, and the probability distribution representing the agent's score. Select to execute an atomic task in the current state. The possibility of [the task being evaluated]. Then, a task evaluation mechanism is used to evaluate all the atomic tasks, where task evaluation refers to the agent expressing its willingness to undertake a particular atomic task. After obtaining the probability distribution of each atomic task, it is necessary to determine whether the agent should evaluate. Based on the Critic network (value network), the state value of all the local observed states is evaluated to obtain multiple target state values. Based on all the assigned probability distributions and all the target state values, evaluation scores are calculated for all the atomic tasks to obtain multiple evaluation scores. The corresponding expression is: ; in, The target state value is evaluated by the Critic network. These are the parameters of the Critic network (i.e., the online network parameters). The capability matching function, with a value range of [0,1], is used to evaluate the agent. ability portrait With atomic tasks The degree of matching. For example, if It requires working at heights, and If the training does not include high-altitude work skills, then... Compare all the aforementioned assessment scores with the assessment threshold. A comparison is performed to obtain a comparison result. If the comparison result indicates that an evaluation score is greater than the evaluation threshold, then the agent's evaluation information for the atomic task is obtained based on the comparison result. That is, the agent evaluates the atomic task and broadcasts the evaluation information (including task ID, evaluation score, intent description, etc.) to neighboring agents through the communication module. The evaluation threshold can be dynamically adjusted according to the system load: the threshold is increased when the system load is high to reduce bidding conflicts, and decreased when the load is low to fully utilize resources. Specifically, the threshold can be adjusted according to... Dynamic adjustments are made, among which, As a preset baseline threshold, The normalized current bidding load of the system (e.g., the ratio of the number of evaluation requests generated per unit time to the system's processing capacity). This is an adjustable sensitivity coefficient. Higher loads linearly increase the threshold, thus suppressing low-value assessments and reducing conflict probability; lower loads cause the threshold to approach a more stable value. This encourages more intelligent agents to participate in bidding in order to improve resource utilization.

[0026] Subsequently, the specific process of evaluating the state value of all the local observation states based on the value network is as follows: the Critic network is responsible for evaluating the current state value and calculating the temporal difference error; the Critic network takes the local observation states as input and outputs the state value through a multi-layer neural network. The state value represents the expected cumulative reward that the agent can obtain by following the current policy in the current state; error calculation is performed based on all the state values ​​to obtain multiple temporal difference errors. The corresponding expression is: ; in, For instant rewards, This is a discount factor used to measure the weight of future rewards. The evaluation value of the target value network. The target network parameters are online network parameters. A slow copy is used to provide a stable target value.

[0027] Furthermore, strategy optimization is enhanced through genetic algorithms. A genetic algorithm (GA) is introduced as the optimizer to optimize the network weights of the Actor and Critic networks, aiming to escape local optima and find the globally optimal allocation scheme. The specific process is as follows: Figure 5 As shown, population initialization: The Actor network is initialized... The probability distribution of the output of each atomic task is used as prior knowledge, from which samples are generated. There are 10 candidate action individuals (population), and each candidate action individual This represents a possible task allocation scheme. Fitness evaluation: For each candidate individual, its value is evaluated using a Critic network as a fitness function. The state value output by the Critic network after substituting the candidate action individual is the value; a higher value indicates a better allocation scheme. Selection: A roulette wheel selection method is used, selecting superior individuals for the next generation based on their fitness ratio. The selection probability... The corresponding expression is: ; in, For the first Individual candidate actions.

[0028] Crossover: A uniform crossover method is used to cross over two selected parent individuals (i.e., candidate action individuals). and With crossover probability (generally (Sectional gene exchange.) Specifically, the allocation position for each atomic task is determined with a 50% probability from the parent. The value has a 50% probability of being inherited from the parent. Values ​​are selected to generate offspring individuals. Mutation: Gaussian mutation is used to mutate the offspring produced by crossover using the mutation probability. (generally , K For each gene locus (where the number of atomic tasks is ), a mutation is performed. The mutation operation is as follows: add a value with a mean of 0 and a standard deviation of to the allocation probability at the current position. The Gaussian noise is then normalized to make the sum of the probability distributions equal to 1. The role of mutation is to introduce new gene variations, increase population diversity, and help the system escape local optima. Elite retention: In each generation, the offspring with the highest fitness is retained and directly enters the next generation (elite strategy), ensuring that the optimal solution is not lost.

[0029] go through After each iteration, the individual with the highest fitness is output as the optimized action. This is taken as the final output action of the intelligent agent. Finally, the agent will generate tuples of experience from the interactions. The data is stored in the Experience Replay Buffer. During training, a mini-batch of data is randomly sampled from the Experience Replay Buffer to break the temporal correlation between data points.

[0030] For Critic updates: Minimize the mean square loss of the time-series difference error. The corresponding expression is: ; Update parameters via backpropagation .

[0031] For Actor updates: Utilize the advantage function Gradient ascent strategy to update parameters The corresponding expression is: ; in, β Let be the Actor's learning rate; the advantage function is calculated as follows: ; in, For trace attenuation parameters, for The advantage function at time.

[0032] Step S30: Perform agent matching on the atomic task based on the target evaluation result to obtain agent matching result, and perform conflict detection on the agent matching result to obtain conflict detection result.

[0033] Specifically, such as Figure 3 As shown, step three, consensus-driven conflict resolution, addresses the need for conflict detection in a distributed environment where multiple agents may simultaneously compete for the same atomic task. The specific detection process involves agent matching based on the target evaluation result, obtaining the agent matching result, and acquiring all target atomic tasks within the matching result. Conflict detection is then performed on all target atomic tasks, where the target atomic task is the one that receives the agent's evaluation information (i.e., broadcasts its evaluation information, including the evaluated atomic task ID, evaluation score, and intent description), through the communication module. When an agent... Received from neighbor's smart agent When evaluating information, check for conflicts in target tasks; if neighboring intelligent agents... Also for the evaluation of atomic tasks (i.e. Evaluation objectives If the target atomic task receives evaluation information from multiple agents, it is determined that a competition conflict exists. Specifically, if the target atomic task receives evaluation information from multiple agents, the conflict detection result is determined to be a competition conflict, and the conflict type is determined. The conflict types include resource competition conflict, spatial conflict, and temporal conflict. Resource competition conflict refers to multiple agents evaluating the same atomic task. Spatial conflict refers to the intersection or overlap of the movement paths of multiple agents, which may lead to collisions. Temporal conflict refers to multiple agents planning to perform mutually exclusive operations at the same time step. If the target atomic task does not receive evaluation information from multiple agents, the conflict detection result is determined to be no competition conflict.

[0034] Step S40: If the conflict detection result indicates the existence of a competitive conflict, then candidate schemes are generated based on the conflict detection result to obtain a global candidate scheme set. Instructions are generated on the global candidate scheme set to obtain the allocation results of all atomic tasks. A global task allocation scheme is obtained based on all the allocation results.

[0035] Specifically, if the conflict detection result indicates the existence of a competing conflict, a local candidate solution set is generated based on the conflict detection result, i.e., for each agent. Based on the local bidding results, a local candidate solution set is generated. , where each candidate solution It contains a set of triples (agent ID, atomic task ID, evaluation score) representing the task allocation scheme proposed by the agent; then, based on the agent, the local candidate scheme set is exchanged and merged to obtain the global candidate scheme set, that is, the agent exchanges local candidate schemes with neighboring agents through the communication module. Collect candidate solutions from neighboring agents and merge them into a global candidate solution set. Among them, for For multiple assessments of the same atomic task, they are sorted in descending order of assessment scores, retaining the top scores. One scheme is selected as the candidate assignment for this task.

[0036] A reasoning layer is introduced to calculate the team utility of each candidate allocation scheme in the global candidate scheme set, thereby obtaining the total team utility. The corresponding expression is: ; in, A set of intelligent agents competing for atomic tasks. The priority weights for atomic tasks. For intelligent agents Evaluation score for atomic tasks, Feasibility factors (considering spatial and temporal conflicts; if conflicts exist, feasibility is reduced).

[0037] Based on reward attribution technology, the total team reward is decomposed into individual contributions, conflicts are resolved, and the final allocation instruction is generated. Specifically, the total team utility is decomposed into individual contributions to obtain multiple marginal contribution levels. The corresponding expression is: ; in, To exclude intelligent agents A subset of intelligent agents other than For including intelligent agents Team effectiveness at that time For those that do not include intelligent agents Team effectiveness at that time A collection of intelligent agents participating in the bidding for a certain atomic task. Let f(x) be the factorial symbol, where f(x) is the factorial symbol. Quantify the intelligent agent Overall contribution to the team. It should be noted that this applies to subsets. The computational cost of an exact traversal increases factorially with the number of agents participating in the bidding for the task, but because... Decomposition is performed only within the local conflict scope where multiple agents simultaneously bid for the same atomic task. The number of agents participating in the conflict is usually much smaller than the total number of agents in the system (the total number of agents can reach thousands, but the set of bidders for a single atomic task is generally only a few dozen to a dozen). Therefore, precise calculation is feasible in engineering. When the number of bidders for individual tasks is large, this invention can also use the Monte Carlo sampling approximation method, which approximates the task by randomly sampling the order of agents and estimating the mean of the marginal contribution. While ensuring the rationality of the attribution, the computational complexity is kept within an acceptable range, thus maintaining consistency with the overall goal of large-scale scalability of this invention.

[0038] Conflict resolution rule: For atomic tasks that have competing conflicts, assign the task to... The highest-priority agent; agents that do not receive a task will be reorganized and will bid for the next highest priority task from the remaining atomic tasks.

[0039] Conflict resolution determination: When all atomic tasks have been allocated (i.e., each atomic task can be successfully bid on by at most one agent), and no agent obtains a better task through re-bidding, the conflict is determined to be resolved, and the system outputs the final allocation instruction.

[0040] Subsequently, instructions are generated based on all the marginal contributions to obtain multiple task allocation instructions. Tasks are then allocated to all agents according to these instructions, resulting in the allocation of all atomic tasks. Finally, all allocation results are aggregated to obtain a global task allocation scheme. ,in, Each agent is assigned to perform an atomic task; after receiving the assignment instruction, each agent enters the execution phase.

[0041] Step S50: Allocate the global tasks according to the global task allocation scheme to obtain the task allocation result, and generate a report based on all decision trajectories of the task allocation result to obtain the target language interpretation report.

[0042] Specifically, after obtaining the global task allocation scheme, the global tasks are allocated according to the scheme to obtain the task allocation result; such as Figure 3 As shown, step four, full-link trajectory recording and interpretation generation, involves the following specific processing steps: Figure 6As shown, during task execution, the system records decision trajectories in real time and obtains all decision trajectories recorded in the task allocation results. The decision trajectories include state sequences, action sequences, reward sequences, tool call logs (records of each execution module calling API or hardware interface), inter-agent communication records (evaluation information, intent exchange, consensus negotiation process), and conflict detection and resolution records (conflict type, participating agents, resolution result).

[0043] Next, a target language interpretation report needs to be generated. This report is automatically generated by the interpretation generation module (XAIModule) based on an LLM (Large Language Model). Specifically, nodes are extracted from all decision trajectories to obtain multiple key decision nodes. These key decision nodes include task allocation decision points (which agent is assigned to which task), conflict occurrence and resolution points (which agents compete and how to resolve the conflict), and anomaly or anomaly recovery points (task failure, path replanning, etc.). A causal chain is constructed for each key decision node to obtain multiple causal chains. These causal chains include a task allocation causal chain (e.g., global task → LLM decomposition into...). → Bidding (Evaluation score = 0.85) → Consensus resolution → Allocation results: → ) and abnormal causal chains (e.g., task failure ← path planning error ← perception module misses dynamic obstacles ← obstacle not in (as embodied in the text) All the causal chains are input into a large language model. Based on the large language model, combined with the task context and the agent's capability profile, a report is generated for all the causal chains to obtain a language explanation report. The language explanation includes the following parts: Selection criteria: For example, select Agent A to execute atomic tasks. (Sorting Package C) received the highest evaluation score (score=0.87) because it was closest to the task point (d=2.3m) and had high-altitude operation skills (match=0.95). Exclusion criteria: For example, excluding Agent B from performing atomic tasks. Because its current load is too high (remaining power = 12%, below the threshold of 20%), and it does not have the skills to work at height (match = 0.3). Constraint effects: For example, the path planning detour is due to the detection of a dynamic obstacle (type: moving person, location: (5.2, 3.1)), making the original shortest path impassable; Conflict resolution basis: for example, atomic tasks It was simultaneously bid on by Agent A (evaluation score 0.82) and Agent C (evaluation score 0.76), and the Shapley score (i.e., ...) was used to determine the best bid. The marginal contribution of Agent A is greater (Agent A's marginal contribution). The marginal contribution of Agent C is 0.65. The value is 0.35, therefore, Therefore, it is assigned to Agent A; Global optimization explanation: For example, the overall team utility of this allocation scheme is U=24.7, which is 12% higher than the previous round. This is mainly due to Agent D successfully taking over the low-priority tasks after replanning.

[0044] Subsequently, since the natural language generation process relies on LLM, to avoid the LLM generating "illusionary" content that contradicts the actual decision-making process due to a lack of evidence, this invention further introduces a consistency verification step after obtaining the target language interpretation report. This step verifies the consistency of the target language interpretation report, resulting in a new target language interpretation report. Specifically, the key entities in the language interpretation report (agent ID, task ID, evaluation score, Shapley value, etc.) are compared item by item with the data in all the causal chains to obtain data comparison results. If the data comparison results show inconsistencies, the consistency verification fails, and the language interpretation report is adjusted (e.g., triggering regeneration or using structured templates to replace corresponding statements) to obtain the target language interpretation report. This ensures that the final interpretation report (i.e., the target language interpretation report) faithfully reflects the actual decision-making trajectory of the system at the factual level, rather than being a rationalized description freely generated by the LLM. If the data comparison results show no inconsistencies, the target language interpretation report is obtained based on the data comparison results. The role of the target language interpretation report is as follows: Function 1: Auditing and Compliance: Provides users with a complete decision audit trail, meets inspection and compliance requirements, and allows users to review the reasons for each task assignment item by item to ensure the fairness and rationality of the assignment decision; Function 2: Troubleshooting and Debugging: When the system malfunctions (e.g., task failure, unreasonable allocation), users can quickly locate the root cause of the problem by interpreting the report in the target language. For example, if the report indicates "task failure ← path planning error ← sensing module missed detection", users can make targeted improvements to the sensing module. Compared to traditional black-box systems that require manual checking of logs one by one, this can shorten the fault location time by 60% to 70%. Thirdly, establishing human-machine trust: By providing transparent "reasons for selection" and "reasons for exclusion," the system eliminates users' doubts about the autonomous system's decision-making and builds trust in the system. Fourthly, online learning and self-improvement: Users can provide feedback to explain the unreasonable aspects of the report (e.g., Agent C should not be ignored). The system uses this feedback to make online fine-tuning, so that the system performance continues to improve over time.

[0045] Furthermore, to ensure that the natural language feedback from users to the target language interpretation report (e.g., Agent C should not be ignored) can be practically used for online fine-tuning, this invention further sets up feedback parsing rules: users submit feedback through a preset structured feedback interface, which provides "acceptance / questioning" options and optional brief explanations for specific task allocation decisions; when a user marks an allocation as "questioning," the system penalizes the team's total utility corresponding to that allocation by a preset margin (e.g., reducing it by a fixed percentage or mapping it to a negative reward increment), and backpropagates the corrected utility value, transforming it into a correction term for the immediate reward used in training the Critic network; when a user marks an allocation as "accepting," a positive reward increment is given to the action on the corresponding decision-making chain. This mechanism transforms unstructured human language feedback into structured numerical signals that can be used for gradient updates, making the online fine-tuning process feasible.

[0046] Furthermore, the present invention will be illustrated through a complete embodiment: (1) Global task input: The system receives global tasks. =Complete the sorting and delivery of 50 packages in the warehouse within 2 hours; (2) Task decomposition and broadcasting: The LLM in the planning module will Decomposed into a set of atomic tasks, which includes =Sort packages C to the packing area (high-altitude operation required); Agents A, B, and C broadcast their respective capability profiles; (3) Strategy optimization and bidding: Agent A's Actor network pairs The output shows a high selection probability. After evaluation by the Critic network and population optimization by the genetic algorithm, the evaluation score is 0.87, which meets the evaluation threshold. Therefore, for Assessment; Agent B's capability profile does not include the "high-altitude operation" skill, the match rate is match=0.3, the assessment score is below the assessment threshold, and no assessment is conducted; (4) Consensus-driven conflict resolution: Agent A and Agent C simultaneously... The evaluation (scores were 0.87 and 0.76 respectively) showed that the communication module detected a resource contention conflict; the inference layer calculated the total team utility of the candidate allocation schemes and calculated the marginal contribution of Agent A based on the Shapley value method. The marginal contribution of Agent C is 0.65. The conflict resolution rule will be set to 0.35. Assigned to Agent A; (5) Trajectory Recording and Interpretation Generation: The system records the complete decision trajectory of this allocation, and the interpretation generation module extracts key nodes and constructs a causal chain. →Agent A bids (0.87) → Conflict with Agent C → Shapley resolves ( → Assigned to Agent A”, after LLM natural language generation and consistency verification, the final interpretation report is output: Select Agent A to execute. Agent C received the highest evaluation score due to its proximity to the mission location and its high-altitude work skills; Agent C was excluded because of its marginal contribution ( ) lower than Agent A ( ); (6) Human-machine feedback closed loop: After the user reviews the explanation report, he marks it as approved. Based on this, the system gives Agent A a positive reward increment for this decision-making link, which is used for the subsequent online fine-tuning of the Critic network.

[0047] Furthermore, such as Figure 7 As shown, based on the above-described distributed multi-agent task allocation method, the present invention also provides a distributed multi-agent task allocation system, wherein the distributed multi-agent task allocation system includes: The task decomposition module 51 is used to obtain global tasks in the logistics and warehousing scenario, decompose the global tasks to obtain a set of atomic tasks, and obtain capability profiles of multiple intelligent agents. The task evaluation module 52 is used to calculate the action probability distribution based on the capability profiles of all the intelligent agents and the set of atomic tasks, obtain the allocation probability distribution of multiple atomic tasks, and evaluate all the atomic tasks based on all the allocation probability distributions to obtain the target evaluation result. The conflict detection module 53 is used to perform agent matching on the atomic task based on the target evaluation result, obtain agent matching result, and perform conflict detection on the agent matching result to obtain conflict detection result; The task allocation module 54 is used to generate candidate schemes based on the conflict detection results if the conflict detection results indicate that there is a competition conflict, obtain a global candidate scheme set, generate instructions on the global candidate scheme set to obtain the allocation results of all atomic tasks, and obtain a global task allocation scheme based on all the allocation results. The report generation module 55 is used to allocate the global tasks according to the global task allocation scheme, obtain the task allocation results, and generate a report based on all decision trajectories of the task allocation results to obtain a target language interpretation report.

[0048] Furthermore, such as Figure 8 As shown, based on the above-mentioned distributed multi-agent task allocation method, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 8 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0049] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a distributed multi-agent task allocation program 40, which can be executed by the processor 10 to implement the distributed multi-agent task allocation method of this application.

[0050] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the distributed multi-agent task allocation method.

[0051] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0052] In one embodiment, when the processor 10 executes the distributed multi-agent task allocation program 40 in the memory 20, the following steps are performed: The global tasks of the logistics and warehousing scenario are obtained, the global tasks are decomposed and processed to obtain a set of atomic tasks, and the capability profiles of multiple intelligent agents are obtained. Based on the capability profiles of all the agents and the set of atomic tasks, the action probability distribution is calculated to obtain the allocation probability distribution of multiple atomic tasks. Based on all the allocation probability distributions, the atomic tasks are evaluated to obtain the target evaluation result. Based on the target evaluation results, agent matching is performed on the atomic task to obtain agent matching results, and conflict detection is performed on the agent matching results to obtain conflict detection results. If the conflict detection result indicates the existence of a competitive conflict, then candidate schemes are generated based on the conflict detection result to obtain a global candidate scheme set. Instructions are generated on the global candidate scheme set to obtain the allocation results of all atomic tasks. Finally, a global task allocation scheme is obtained based on all the allocation results. The global tasks are allocated according to the global task allocation scheme to obtain task allocation results. A report is generated based on all decision trajectories of the task allocation results to obtain a target language interpretation report.

[0053] Specifically, the step of calculating the action probability distribution based on the capability profiles of all the agents and the set of atomic tasks to obtain the allocation probability distribution of multiple atomic tasks, and then evaluating all the atomic tasks based on all the allocation probability distributions to obtain the target evaluation result, includes: Based on the capability profiles of all the agents, the local observation state of each agent is obtained, and all the local observation states and the set of atomic tasks are input into the policy network. Based on the policy network, task scores are calculated for all local observation states and the set of atomic tasks to obtain task scores for multiple atomic tasks. Then, all task scores are normalized to obtain multiple allocation probability distributions. The state value of all local observation states is evaluated based on the value network to obtain multiple target state values. Evaluation scores are calculated for all atomic tasks according to all the assigned probability distributions and all the target state values ​​to obtain multiple evaluation scores. All the evaluation scores are then compared with an evaluation threshold to obtain a comparison result. If the comparison result indicates that there is an evaluation score greater than the evaluation threshold, then the evaluation information of the agent for the atomic task is obtained based on the comparison result, and the evaluation information is broadcast to neighboring agents to obtain the target evaluation result.

[0054] Specifically, the evaluation of the state value of all the local observed states based on the value network to obtain multiple target state values ​​includes: All the local observation states are input into the value network, and the state value of all the local observation states is evaluated based on the value network to obtain multiple state values; Error calculations are performed on all the stated state values ​​to obtain multiple temporal difference errors. Value updates are then performed on all the stated state values ​​based on all the stated temporal difference errors to obtain multiple target state values.

[0055] Specifically, the step of performing agent matching on the atomic task based on the target evaluation result to obtain agent matching results, and performing conflict detection on the agent matching results to obtain conflict detection results, includes: Based on the target evaluation result, the atomic task is matched with an agent to obtain the agent matching result, and all target atomic tasks in the agent matching result are obtained. Conflict detection is performed on all the target atomic tasks, wherein the target atomic task is the atomic task that receives the evaluation information from the agent. If the target atomic task receives evaluation information from multiple agents, the conflict detection result is determined to be a competitive conflict, and the conflict type is determined, wherein the conflict type includes resource competition conflict, spatial conflict and temporal conflict; If it is detected that the target atomic task has not received evaluation information from multiple agents, then the conflict detection result is determined to be that there is no competition conflict.

[0056] Specifically, if the conflict detection result indicates the existence of a competitive conflict, then candidate solutions are generated based on the conflict detection result to obtain a global candidate solution set. Instructions are then generated from the global candidate solution set to obtain the allocation results of all atomic tasks. Finally, a global task allocation scheme is obtained based on all the allocation results. This process specifically includes: If the conflict detection result indicates the existence of a competitive conflict, a local candidate solution set is generated based on the conflict detection result, and the local candidate solution set is exchanged and merged based on the agent to obtain a global candidate solution set. For each candidate allocation scheme in the global candidate scheme set, the team utility is calculated to obtain the total team utility, and the total team utility is decomposed into individual values ​​to obtain multiple marginal contribution values; Instructions are generated based on all the marginal contribution values ​​to obtain multiple task allocation instructions. Tasks are allocated to all agents based on all the task allocation instructions to obtain the allocation results of all atomic tasks. All the allocation results are then aggregated to obtain a global task allocation scheme.

[0057] Specifically, the process of allocating global tasks according to the global task allocation scheme to obtain task allocation results, and generating a target language interpretation report based on all decision trajectories of the task allocation results, includes: The global tasks are allocated according to the global task allocation scheme to obtain the task allocation result, and all decision trajectories recorded in the task allocation result are obtained. The decision trajectories include state sequences, action sequences, reward sequences, tool call logs, inter-agent communication records, and conflict detection and elimination records. Nodes are extracted from all the decision trajectories to obtain multiple key decision nodes. A causal chain is constructed for each key decision node to obtain multiple causal chains, wherein the causal chains include a task allocation causal chain and an anomaly causal chain. All the causal chains are input into a large language model. Reports are generated for all the causal chains based on the large language model to obtain a language explanation report. The language explanation report is then validated for consistency to obtain a target language explanation report.

[0058] Specifically, the process of performing a consistency check on the language interpretation report to obtain the target language interpretation report includes: The key entities in the language explanation report are compared item by item with the data of all the causal chains to obtain the data comparison results; If the data comparison results show inconsistencies, the language explanation report will be adjusted to obtain the target language explanation report. If the data comparison result shows no data inconsistency, then a target language interpretation report is obtained based on the data comparison result.

[0059] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a distributed multi-agent task allocation program, which, when executed by a processor, implements the steps of the distributed multi-agent task allocation method as described above.

[0060] In summary, this invention provides a distributed multi-agent task allocation method, system, terminal, and storage medium. The method includes: acquiring a global task in a logistics and warehousing scenario; decomposing the global task to obtain a set of atomic tasks and acquiring capability profiles of multiple agents; calculating action probability distributions based on the capability profiles of all agents and the set of atomic tasks to obtain allocation probability distributions for multiple atomic tasks; evaluating all atomic tasks based on all allocation probability distributions to obtain target evaluation results; matching agents for the atomic tasks based on the target evaluation results to obtain agent matching results; performing conflict detection on the agent matching results to obtain conflict detection results; if the conflict detection results indicate competition, generating candidate solutions based on the conflict detection results to obtain a global candidate solution set; generating instructions for the global candidate solution set to obtain allocation results for all atomic tasks; obtaining a global task allocation scheme based on all allocation results; allocating the global tasks according to the global task allocation scheme to obtain task allocation results; and generating a report based on all decision trajectories of the task allocation results to obtain a target language interpretation report. This invention can autonomously adjust the allocation strategy according to environmental changes, and the allocation decision is interpretable, thus improving the efficiency and accuracy of task allocation.

[0061] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0062] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0063] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A distributed multi-agent task allocation method, characterized in that, The distributed multi-agent task allocation method includes: The global tasks of the logistics and warehousing scenario are obtained, the global tasks are decomposed and processed to obtain a set of atomic tasks, and the capability profiles of multiple intelligent agents are obtained. Based on all the capability profiles, obtain the local observation state of each agent, calculate the task score for all the local observation states and the set of atomic tasks, obtain the task score of multiple atomic tasks, and normalize all the task scores to obtain multiple allocation probability distributions. The state value of all the local observation states is evaluated to obtain multiple target state values. Evaluation scores are calculated for all the atomic tasks based on all the assigned probability distributions and all the target state values ​​to obtain multiple evaluation scores. The target evaluation result is obtained based on all the evaluation scores. Based on the target evaluation results, agent matching is performed on the atomic task to obtain agent matching results, and conflict detection is performed on the agent matching results to obtain conflict detection results. If the conflict detection result indicates the existence of a competition conflict, a local candidate solution set is generated based on the conflict detection result. The local candidate solution set is then exchanged and merged based on the agent to obtain a global candidate solution set. Instructions are generated on the global candidate solution set to obtain the allocation results of all atomic tasks. Finally, a global task allocation scheme is obtained based on all the allocation results. The global tasks are allocated according to the global task allocation scheme to obtain task allocation results. A report is generated based on all decision trajectories of the task allocation results to obtain a target language interpretation report.

2. The distributed multi-agent task allocation method according to claim 1, characterized in that, The process of obtaining the target evaluation result based on all the evaluation scores specifically includes: Compare all the evaluation scores with the evaluation threshold to obtain the comparison results; If the comparison result indicates that there is an evaluation score greater than the evaluation threshold, then the evaluation information of the agent for the atomic task is obtained based on the comparison result, and the evaluation information is broadcast to neighboring agents to obtain the target evaluation result.

3. The distributed multi-agent task allocation method according to claim 1, characterized in that, The evaluation of the state value of all the local observed states to obtain multiple target state values ​​specifically includes: All the local observation states are input into the value network, and the state value of all the local observation states is evaluated based on the value network to obtain multiple state values; Error calculations are performed on all the stated state values ​​to obtain multiple temporal difference errors. Value updates are then performed on all the stated state values ​​based on all the stated temporal difference errors to obtain multiple target state values.

4. The distributed multi-agent task allocation method according to claim 1, characterized in that, The step of performing agent matching on the atomic task based on the target evaluation result to obtain agent matching results, and performing conflict detection on the agent matching results to obtain conflict detection results, specifically includes: Based on the target evaluation result, the atomic task is matched with an agent to obtain the agent matching result, and all target atomic tasks in the agent matching result are obtained. Conflict detection is performed on all the target atomic tasks, wherein the target atomic task is the atomic task that receives the evaluation information from the agent. If the target atomic task receives evaluation information from multiple agents, the conflict detection result is determined to be a competitive conflict, and the conflict type is determined, wherein the conflict type includes resource competition conflict, spatial conflict and temporal conflict; If it is detected that the target atomic task has not received evaluation information from multiple agents, then the conflict detection result is determined to be that there is no competition conflict.

5. The distributed multi-agent task allocation method according to claim 1, characterized in that, The step of generating instructions for the global candidate scheme set to obtain the allocation results of all atomic tasks, and obtaining a global task allocation scheme based on all the allocation results, specifically includes: For each candidate allocation scheme in the global candidate scheme set, the team utility is calculated to obtain the total team utility, and the total team utility is decomposed into individual components to obtain multiple marginal contribution values; Instructions are generated based on all the marginal contribution values ​​to obtain multiple task allocation instructions. Tasks are allocated to all agents based on all the task allocation instructions to obtain the allocation results of all atomic tasks. All the allocation results are then aggregated to obtain a global task allocation scheme.

6. The distributed multi-agent task allocation method according to claim 1, characterized in that, The process of allocating global tasks according to the global task allocation scheme to obtain task allocation results, and generating a target language interpretation report based on all decision trajectories of the task allocation results, specifically includes: The global tasks are allocated according to the global task allocation scheme to obtain the task allocation results, and all decision trajectories recorded in the task allocation results are obtained. The decision trajectories include state sequences, action sequences, reward sequences, tool call logs, inter-agent communication records, and conflict detection and elimination records. Nodes are extracted from all the decision trajectories to obtain multiple key decision nodes. A causal chain is constructed for each key decision node to obtain multiple causal chains, wherein the causal chains include a task allocation causal chain and an anomaly causal chain. All the causal chains are input into a large language model. Reports are generated for all the causal chains based on the large language model to obtain a language explanation report. The language explanation report is then validated for consistency to obtain a target language explanation report.

7. The distributed multi-agent task allocation method according to claim 6, characterized in that, The process of performing a consistency check on the language interpretation report to obtain the target language interpretation report specifically includes: The key entities in the language explanation report are compared item by item with the data of all the causal chains to obtain the data comparison results; If the data comparison results show inconsistencies, the language explanation report is adjusted to obtain the target language explanation report. If the data comparison result shows no data inconsistency, then a target language interpretation report is obtained based on the data comparison result.

8. A distributed multi-agent task allocation system, characterized in that, The distributed multi-agent task allocation system is used to implement the distributed multi-agent task allocation method according to any one of claims 1-7, wherein the distributed multi-agent task allocation system comprises: The task decomposition module is used to obtain global tasks in the logistics and warehousing scenario, decompose the global tasks to obtain a set of atomic tasks, and obtain capability profiles of multiple intelligent agents. The task evaluation module is used to calculate the action probability distribution based on the capability profiles of all the agents and the set of atomic tasks, obtain the allocation probability distribution of multiple atomic tasks, and evaluate all the atomic tasks based on all the allocation probability distributions to obtain the target evaluation result. The conflict detection module is used to perform agent matching on the atomic task based on the target evaluation result, obtain agent matching result, and perform conflict detection on the agent matching result to obtain conflict detection result; The task allocation module is used to generate candidate solutions based on the conflict detection result if the conflict detection result indicates that there is a competition conflict, obtain a global candidate solution set, generate instructions on the global candidate solution set to obtain the allocation results of all atomic tasks, and obtain a global task allocation scheme based on all the allocation results. The report generation module is used to allocate and process the global tasks according to the global task allocation scheme, obtain the task allocation results, and generate a report based on all decision trajectories of the task allocation results, thereby obtaining a target language interpretation report.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps of the distributed multi-agent task allocation method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program thereon, and the computer-readable storage medium stores a distributed multi-agent task allocation program, which, when executed by a processor, implements the steps of the distributed multi-agent task allocation method as described in any one of claims 1-7.

Images