A distributed multi-agent task allocation method based on PI algorithm in a communication-restricted environment

Abstract

This invention discloses a distributed multi-agent task allocation method based on the PI algorithm in communication-constrained environments, belonging to the field of multi-agent cooperative control technology. The method employs a task-granularity-based scalar consensus mechanism, utilizing hard-lock priority rules and inertia threshold rules to achieve lightweight information consistency; it establishes a physical-information two-layer binding mechanism, forcibly broadcasting the hard-lock state when agents execute tasks to prevent remote bidding and preemption; it introduces inertia-based stability control, suppressing frequent task switching through inertia thresholds and congestion penalties, and combines hierarchical spatiotemporal heuristic planning to prioritize urgent tasks. This invention enables conflict-free and highly stable real-time task allocation for agent clusters under weak communication conditions with highly dynamic and non-fully connected network topologies.

Description

Technical Field

[0001] This invention relates to the field of multi-agent systems and cooperative control technology, and more specifically, to a distributed task allocation method based on the Performance Impact (PI) algorithm improved in environments with dynamic changes in network topology and limited communication bandwidth. Background Technology

[0002] Task allocation in multi-agent systems (MAS) is a core problem in cooperative control, essentially an NP-hard combinatorial optimization problem. The goal is to rationally allocate a complex set of heterogeneous tasks to a swarm of agents while satisfying constraints such as time windows, payload capacity, and kinematic constraints, ultimately optimizing the global system objectives (e.g., minimum total time, shortest distance, and lowest energy consumption). In recent years, UAV swarms and ground robot formations have been widely used in disaster relief, environmental monitoring, and military reconnaissance, placing extremely high demands on the real-time performance, robustness, and adaptability to communication environments of distributed task allocation algorithms. In the development of multi-agent task allocation algorithms, market-based auction algorithms were the first to be introduced. Early sequential single-item auction methods were computationally simple and had low communication overhead; however, because agents only used a greedy strategy to bid for tasks one by one, ignoring the spatiotemporal coupling between tasks, this algorithm often easily got trapped in local optima and struggled to handle complex task coordination constraints. To address task allocation conflicts under complex constraints, the Consensus-Based Bundle Algorithm (CBBA) was developed. CBBA introduces the concept of "task bundles," allowing agents to construct expected task execution sequences based on their own states. It resolves ownership conflicts for the same task among neighbors using vector-based consensus rules, achieving significant progress in conflict-free allocation. Building on this, the Performance Impact Algorithm (PI) further refines the optimization objective. In CBBA, each agent is self-interested, choosing the most advantageous task through a "maximum bid" strategy. PI, on the other hand, optimizes the global objective. It introduces the concept of "significance," evaluating the marginal contribution of a task's existence or absence to global performance metrics (such as average completion time, average number of rescues) to determine task inclusion or removal. Compared to CBBA, PI more directly approximates the optimal solution to the global mathematical objective and performs better in static programming scenarios.

[0003] However, the aforementioned mainstream algorithms (including CBBA and the original PI) are typically designed based on the following idealized assumptions: The assumption of full network connectivity or state stability: Existing algorithms generally assume that there are ideal communication links between agents, or that the network topology remains statically stable during the negotiation period, which can support the consensus phase of the algorithm to synchronize information across the entire network in multiple rounds until the algorithm fully converges.

[0004] Static segmentation model of allocation before execution: Existing algorithms typically treat task allocation as an independent time phase before execution. The agent must complete the allocation negotiation for all tasks and reach a consensus before locking the scheme and starting physical execution.

[0005] However, in real-world field operations (such as long-distance drone inspections and scenarios involving damaged base stations in disaster areas), these assumptions are often difficult to uphold. Real-world environments exhibit typical weak communication characteristics: extremely low communication bandwidth, limited communication distance, and highly dynamic network topology that changes with the movement of the agent. Applying existing PI algorithms to such environments reveals the following significant shortcomings: Conflict between physical execution and information-based decision-making: In the dynamic process of allocation and execution, existing PI algorithms lack awareness of the physical state. When an agent has physically begun executing a task, due to the "information lag" caused by the delay in consensus in a non-fully connected network, other remote agents may still believe, based on mathematical models, that the task can be preempted at a lower cost. This leads to the logical error of "the task being executed being remotely forcibly taken away," seriously threatening the security of the physical system.

[0006] System oscillations in dynamic topologies: In dynamically changing local topologies, due to the lack of inertial protection for the assigned state, even small cost disturbances (such as small distance changes caused by GPS positioning errors) can lead to frequent switching of task ownership between different agents, causing the system to fall into a dead loop or fail to converge to a stable solution within a finite time.

[0007] Therefore, there is an urgent need for a distributed task allocation method that can support simultaneous allocation and execution in environments with limited communication and dynamic topology changes, effectively resolve conflicts between physical execution and information allocation, and possess high stability. Summary of the Invention

[0008] The purpose of this invention is to provide a distributed multi-agent task allocation method based on the PI algorithm in communication-constrained environments. It aims to solve the technical problems caused by the existing PI algorithm under conditions of non-fully connected networks and highly dynamic topologies, such as communication bandwidth congestion, physical execution and information decision-making conflicts, and system oscillations and non-convergence, which are due to the reliance on fully connected networks in the consensus phase and the static mode of "allocation before execution".

[0009] To achieve the above objectives, this invention provides a distributed multi-agent task allocation method based on the PI algorithm in communication-constrained environments. This method includes the following steps: S1. Situational Awareness and Information Maintenance: The agent monitors the distribution of neighboring nodes within the communication radius in real time, and determines whether the current state is in a state of communication congestion based on the number of neighbors, the distance to neighbors, and ID priority. At the same time, the agent periodically checks the timestamps of tasks in the local database and actively cleans up outdated task information that exceeds the preset failure threshold to prevent "ghost tasks" caused by node offline or communication interruption from blocking system allocation. S2. Scalar Consensus Processing: Abandoning vector clocks that grow linearly with the number of agents, a lightweight, task-granular scalar timestamp is used for network-wide information exchange. When processing neighbor broadcast state information, the "physical authority priority" principle is followed, meaning that the "hard-locked" state at the physical layer has the highest priority and unconditionally overrides mathematical bidding. For regular mathematical bidding updates, an inertia-based stability control mechanism is introduced, allowing a change in task ownership only when the significant difference in benefits from neighboring schemes exceeds a preset inertia threshold, thus effectively suppressing system oscillations caused by minor cost disturbances. S3, Task Removal Planning: The agent periodically evaluates the rationality of the assigned tasks. When calculating the marginal benefit of removing a task, a congestion penalty mechanism is introduced; if the agent is in the congested area detected in step S1, a penalty factor is added to the cost calculation function to artificially reduce the willingness to retain tasks, prompting tasks to flow from high-density areas to agents in non-congested areas, thus achieving implicit load balancing. S4. Hierarchical Spatiotemporal Heuristic Planning: In the task inclusion phase, the greedy strategy based solely on distance cost is abandoned, and a "time-space" hierarchical screening mechanism is established. The system prioritizes the first priority task with a remaining deadline less than the emergency threshold, followed by the second priority task with the closest spatial distance; the agent calculates the marginal salience of candidate tasks according to priority and selects the best to add, in order to maximize the completion rate of tasks on the verge of timeout; S5. Physical-Information Two-Layer Binding: Establishes a strong binding relationship between the physical execution state and the bidding state in the information layer. When an agent physically reaches the target and begins to execute a task, it triggers an atomic state-locking operation, forcibly marking the task as a "hard-locked" state in the information layer (assigning a very small significance value), and refreshing the latest timestamp and broadcasting it to the current neighbors. This directly suppresses the bidding behavior of other agents for the task at the protocol level, preventing remote preemption conflicts. S6. Iterative Loop: The agent continuously loops through steps S1 to S5, adjusting the allocation scheme in real time according to the dynamically changing environment and task status, until all tasks are marked as completed or unexecutable.

[0010] Furthermore, the determination of whether the current state is congested in step S1 is specifically as follows: when the agent determines that there are neighbor nodes within its own communication radius that are less than the preset congestion perception radius and have a higher ID priority than itself, it marks the local state as congested.

[0011] Furthermore, step S2, which involves updating the local task state based on the hard-lock priority rule and the inertia threshold rule, specifically includes: If the saliency value of the task sent by the neighbor is a hard-locked flag, the agent will forcibly accept the ownership information of the neighbor regardless of whether the local scalar timestamp is updated, and update the local saliency value of the task to the hard-locked flag. If the task saliency value sent by the neighbor is not a hard-locked identifier, and the neighbor's scalar timestamp is greater than the local scalar timestamp, then determine the ownership of the current task: If the local entity believes that the task belongs to itself, it calculates the difference between the local significance value and the neighbor's significance value. Only when the difference is greater than the preset inertia threshold will the ownership change of the neighbor be accepted. If the local entity believes that the task does not belong to it, it directly accepts the salience value, ownership, and timestamp updates from its neighbors.

[0012] Furthermore, the congestion penalty described in step S3 is implemented as follows: The agent determines the number of neighboring nodes within its communication radius that have a higher ID priority than itself. If there are neighboring nodes that meet the conditions, it is marked as a congested state. When calculating the local salience cost of retaining the task, a positive congestion penalty constant is added to the calculation result, thereby artificially reducing the willingness to retain the task and prompting the task to flow to the agent in the non-congested area.

[0013] Furthermore, the specific steps of the hierarchical screening described in step S4 are as follows: Define emergency threshold ; Calculate the remaining deadlines for candidate tasks. ,in This is the deadline for the task. The current moment; like The task is assigned to the first priority set, and then ranked within the set according to... Sort in ascending order; if The task is assigned to the second priority set and sorted in ascending order within the set according to the Euclidean distance between the agent's current position and the task position. The agent first attempts to insert tasks from the first priority set into the local task sequence, and only attempts to insert tasks from the second priority set if the first priority set is empty.

[0014] Furthermore, the implementation of the hard-locking mechanism in step S5 includes: Define the hard-locked identifier as a negative constant whose value is less than any normal saliency cost calculated based on distance and time. When the agent's physical state is executing a task, the task saliency value calculated based on algorithm logic is ignored, and the saliency value of the corresponding task is directly overwritten as the hard-locked identifier value.

[0015] Furthermore, the scalar timestamp is a floating-point time record maintained for each independent task, used to replace the vector clock for each agent in the original PI; When an agent acquires task ownership through computation or locks a task through a hard lock mechanism, it updates the scalar timestamp corresponding to the task to the current local system time.

[0016] Furthermore, the significant gain described in step S3 is calculated in the following manner: Define the current task execution sequence of the agent as: Calculate the total path cost for this task sequence. ; Calculate the total path cost of the new task sequence after removing a single task. ,in Tasks to be removed; Task significance value The change in this value represents the significant benefit after removing the task.

[0017] Furthermore, the total cost of the path The calculation formula is:

[0018] in, For the first The location of each task Represents Euclidean distance. The average moving speed of the agent. For the first The execution time of each task.

[0019] Furthermore, the formula for calculating marginal significance in step S4 is as follows:

[0020] in, This represents the total path cost of the original task sequence. To carry out the new task Insert at the optimal position in the original task sequence The total path cost of the new sequence.

[0021] Compared with the prior art, the essential features and significant effects of this invention are: (1) This invention establishes a physical-information dual-layer binding mechanism, which forcibly locks the information layer state during the physical execution phase of the agent, effectively solving the logical conflict problem caused by the separation of the "allocation" and "execution" phases in traditional algorithms. This mechanism ensures that the task being executed will not be preempted by remote nodes under network latency or topology changes, thereby guaranteeing the execution security and task continuity of the physical system.

[0022] (2) This invention introduces an inertia-based stability control and congestion penalty mechanism, which suppresses frequent task switching caused by small cost differences by setting a threshold. Experimental simulation results show that this mechanism can effectively avoid system deadlock or livelock states caused by circular negotiation in dynamic interactions among multiple agents, ensuring that the cluster can continue to advance task execution in complex conflict environments and maintaining the convergence and operational stability of the system allocation scheme.

[0023] (3) This invention adopts a hierarchical spatiotemporal heuristic planning strategy, abandoning the single greedy cost calculation and prioritizing the selection of urgent tasks that are close to the deadline. It has been verified by multiple sets of experimental configurations from low load (e.g., 6 agents and 12 tasks) to high load (e.g., 16 agents and 32 tasks). The experimental results show that even in weak communication environments with limited communication distance and dynamically changing network topology, this invention still achieves excellent task completion rates. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall algorithm of the method of the present invention; Figure 2 This is a schematic diagram illustrating the principle of the physical-information dual-layer binding mechanism in this invention; Figure 3 This is a schematic diagram of a scenario for the hierarchical spatiotemporal heuristic planning strategy in this invention. Detailed Implementation

[0025] The following is in conjunction with the appendix Figure 1-3 The present invention will be further described in detail with reference to specific embodiments.

[0026] This embodiment uses a multi-agent simulation environment built with Python to verify the algorithm. It should be noted that the specific parameters given in the embodiment (such as communication radius and threshold value) are only the best implementation methods for explaining the technical solution of this invention and should not be considered as limitations on the scope of protection of this invention.

[0027] Existing PI algorithms rely on fully connected networks and separate the allocation and execution phases, which makes them prone to information delays, allocation conflicts, and system oscillations in environments with limited communication.

[0028] To address the aforementioned problems, this invention proposes a distributed multi-agent task allocation method based on the PI algorithm in communication-constrained environments. This method employs a task-granularity-based scalar consensus mechanism, utilizing hard-lock priority rules and inertia threshold rules to achieve lightweight information consistency; it establishes a physical-information two-layer binding mechanism, forcibly broadcasting the hard-lock state when agents execute tasks to prevent remote bidding and preemption; it introduces inertia-based stability control, suppressing frequent task switching through inertia thresholds and congestion penalties, and combines hierarchical spatiotemporal heuristic planning to prioritize urgent tasks. This invention enables conflict-free and highly stable real-time task allocation for agent clusters under conditions of highly dynamic and not fully connected network topologies and weak communication.

[0029] (I) Hardware operating environment and system architecture

[0030] The distributed task allocation method described in this invention is deployed in a heterogeneous intelligent agent cluster. Each intelligent agent (e.g., a quadcopter drone, a fixed-wing drone, or a ground-based unmanned vehicle) is equipped with an independent onboard computing platform, forming the physical basis for the algorithm's operation. The specific hardware architecture includes: Perception module: Used to acquire its own three-dimensional position information in real time. Information on obstacles in the surrounding environment.

[0031] Communication module: Employs self-organizing network communication equipment. Communication is affected by physical antenna power and environmental obstructions; the effective communication radius of this module is limited. Within this communication range, agents can communicate with each other, and the communication link will exhibit dynamic connection characteristics as the agents move.

[0032] Computational processing unit: Employs an embedded processor to run the multi-task distributed allocation algorithm of this invention. This unit internally maintains a local task list, a saliency list, and a scalar timestamp list.

[0033] Execution control module: Used to receive the task sequence output by the algorithm and drive the underlying motor controller to navigate the agent to the task target point to perform specific tasks (such as material delivery, sensor scanning, etc.).

[0034] (II) Algorithm Implementation Process and Mathematical Model

[0035] like Figure 1 The complete control logic from step S1 (environmental awareness) to step S6 (iterative loop) is demonstrated.

[0036] The main process of the algorithm includes the following detailed steps and mathematical calculation model: Step S1, Environmental Perception and Information Maintenance: Intelligent Agent In each control cycle ( First, preprocess the surrounding environment and local information: Congestion detection: The agent traverses the set of neighboring nodes detected within the current communication range. Set the congestion sensing radius. (In this embodiment, a distance of 20.0 units is used). If the set There exists a distance that satisfies And the ID has a higher priority than its neighboring nodes. (Right now If the condition is met, the local state will be marked as "congested". This state will serve as a penalty switch for subsequent cost calculations.

[0037] Outdated information cleanup: The agent checks all task information in its local storage. For any task... If it is currently marked as "already assigned to someone else" (i.e. and (and the last update timestamp of the task information), With current system time The difference exceeds the preset timeout threshold. (30.0 seconds in this embodiment): If the holder of the task information is found to have lost communication or gone offline, the agent will reset the local task status to "unassigned" and reset its salience value to the initial state (e.g., infinity) so that the task can be reassigned.

[0038] Step S2, scalar consensus processing: The agent processes the received neighbor broadcast messages. This embodiment abandons the vector clock in the traditional PI algorithm and uses it for each task. Maintain an independent scalar timestamp The consensus process follows this logic: Hard-lock priority rule: Check the task saliency value sent by the neighbor. .like Equal to the preset hard lock flag value (In this embodiment, it is set to -1.0). Then, regardless of whether the local timestamp is updated, the agent will forcibly accept the neighbor's ownership of the task and update the local salience value to 1.0. The timestamp is updated to the neighbor's timestamp.

[0039] Inertia threshold rule: If not in a hard-locked state, and the neighbor's timestamp is newer than the local timestamp ( If the local entity believes it owns the task, then the cost improvement brought by the neighbor's solution is calculated. Only when (The inertia threshold is set to 15.0 in this embodiment) Only when the inertia threshold is reached will the task change be accepted; otherwise, the task ownership will be retained, and only the timestamp will be updated to synchronize the information status, thereby suppressing system oscillations caused by minor cost differences. If the local system does not own the task, it will directly accept the updated information from its neighbors.

[0040] Step S3, Task Removal Planning: The agent evaluates the rationality of the current task sequence and executes the removal operation. At this stage, the core relies on the mathematical calculation of saliency. Define the agent. The current task execution sequence is .

[0041] The total path cost function is defined as follows: Let the agent's current position be... ,Task The position is The execution time is .path Total time cost The calculation is as follows:

[0042] in Represents Euclidean distance. This represents the average movement speed of the agent.

[0043] Significance calculation: a certain task For intelligent agents significance value Defined as the marginal contribution of the task's existence to the total cost:

[0044] That is: (Total path cost when the task is included) minus (Total path cost after removing the task from the sequence).

[0045] Congestion penalty application: If "congestion" is marked in step S1, a congestion penalty factor is introduced when calculating the local task retention cost. (In this embodiment, 2000.0 is used). Corrected local saliency. for:

[0046] like (The global saliency of this task) indicates that retaining this task is no longer economical in a congested environment, and the agent actively releases the task.

[0047] Step S4, Hierarchical Spatiotemporal Heuristic Planning: The agent selects unassigned tasks from the candidate task pool and adds them to the local sequence. This embodiment adopts a hierarchical strategy of "urgent priority, distance secondary": Tiered screening: Setting emergency time thresholds (This example uses 150.0 seconds). Calculate the remaining deadline for each candidate task. .like Then it is classified as the first priority (urgent level), according to... Sort in ascending order; otherwise, it is placed in the second priority (normal layer) and sorted in ascending order according to the Euclidean distance from the agent's current position to the task point.

[0048] Greedy insertion and marginal returns: The agent prioritizes traversing the first priority list. For tasks in the list... Try inserting it into the current path All possible locations Calculate the cost of the new path after insertion. Marginal significance Defined as:

[0049] The second priority list is considered only if the first priority list is empty or both are infeasible. The task that brings the greatest significant benefit is inserted into the local execution sequence.

[0050] Step S5, physical-information two-layer binding, such as Figure 2 This is a crucial step in ensuring execution security, establishing an atomic binding between physical state and information state: Physical state monitoring: The agent monitors its own physical state machine in real time. When the physical location reaches the task coordinate point and the state switches to "in execution", a two-layer binding mechanism is triggered.

[0051] Information layer locking: Agent ignores formula-based locking The calculated task saliency value is used to force the saliency value of the currently executing task to be overwritten with the hard-lock flag value. .

[0052] Broadcast suppression: This hard lock value is broadcast when sending status packets to neighbors. and the latest timestamp. Because And any saliency calculated based on distance Combined with the priority rules in step S2, this broadcast will effectively suppress the bidding behavior of other intelligent agents for this task across the entire network, ensuring that the physical execution process is not interrupted by remote logic.

[0053] Step S6, iterative loop: The agent determines whether there are any unfinished tasks in the system. If so, it returns to step S1 to begin the next control cycle iteration; if all tasks are marked as completed, the allocation process ends and the task allocation scheme converges.

[0054] (III) Demonstration of typical conflict resolution scenarios

[0055] To more intuitively illustrate the operational mechanism of "inertial stability control" and "dual-layer binding" in this invention, combined with... Figure 3 Describe a typical dynamic interaction scenario: Assume there are agents A and B in the scenario, and a task T1 to be assigned.

[0056] Initial phase: Agent A first discovers task T1, calculates its saliency cost as 50.0, adds T1 to its local list, and generates a scalar timestamp. .

[0057] Encounter and Bidding: Agent B enters A's communication range. B calculates that its cost to execute T1 is only 40.0. At this point, B broadcasts bidding information to A.

[0058] Inertial suppression (application of step S2): A receives information from B. Although B's cost (40.0) is lower than A's (50.0), the difference is 10.0. According to the inertial threshold rule of the present invention (assuming... ),At A refuses to relinquish task T1. A believes, "Although B is slightly stronger, I will continue to hold onto the task to avoid system instability." This effectively avoids meaningless task jitter.

[0059] Physical locking (application of step S5): Subsequently, A flies to position T1 and begins execution. At this time, A triggers "physical-information double binding", forcibly setting the salience of T1 to -1.0 (hard locking) and refreshing the timestamp. .

[0060] Remote conflict resolution: At this point, even if a superior agent C (costing only 30.0) enters the network and attempts to bid, when C receives a hard lock value of -1.0 broadcast by A, according to the "physical authority priority" rule, C must unconditionally give up bidding.

[0061] Parameter settings and variations: Those skilled in the art should understand that the specific parameter values ​​in the above embodiments can be adjusted according to actual application scenarios. Regarding the congestion sensing radius This value can be set to a fixed value (such as 20 meters) or it can be designed as an adaptive parameter that changes dynamically with the strength of the communication signal.

[0062] Regarding the inertia threshold In this embodiment, the threshold is set to 15.0. However, in scenarios with extremely high task density, the threshold can be appropriately increased to further enhance system stability. Conversely, in static scenarios with extremely high requirements for optimality, the threshold can be set to 0, degenerating into a greedy strategy.

[0063] Regarding the hierarchical planning strategy: In addition to the two levels of "deadline" and "spatial distance" mentioned in the example, a third priority (such as task value, load matching degree, etc.) can be extended according to actual needs.

[0064] The method framework of this invention is universally applicable to the above-mentioned parameter variations.

Claims

1. A distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment, characterized in that, Includes the following steps S1. Environmental Awareness and Information Maintenance: The agent monitors the number and distance of neighboring nodes within the communication range in real time, and determines whether it is in a congested state by combining the neighbor ID priority; at the same time, it checks all task timestamps stored locally. If a task has not been updated within the preset timeout period and has not been locked locally, the task status is reset to unassigned. S2, Scalar consensus processing: The agent receives state information broadcast by its neighbors, which includes a task saliency list, a task ownership list, and a scalar timestamp based on task granularity; the agent compares the local scalar timestamp and saliency value with those of its neighbors, and updates the local task state according to the hard-lock priority rule and the inertia threshold rule. S3, Task Removal Planning: The agent traverses the locally assigned task sequence and calculates the significant benefit of removing a task; If the agent is in a congested state, a congestion penalty value is added when calculating significant gains; If the incremental benefit from removing a task is greater than zero, then release the ownership of that task. S4. Hierarchical Spatiotemporal Heuristic Planning: The agent performs hierarchical screening of unassigned candidate tasks. First, it selects the first priority task with a deadline less than the emergency threshold, and then selects the second priority task with the closest spatial distance. Calculate the marginal significance of the filtered tasks and insert the task with the highest significance gain into the local task sequence; S5. Physical-Information Binding and Broadcasting: The agent controls physical movement according to the task sequence; when the agent's physical state changes to executing a certain task, the hard lock mechanism is triggered, the saliency value of the task is forcibly set to the hard lock flag value, the scalar timestamp of the task is updated to the current system time, and the updated task status information is broadcast to the neighbors. S6. Iteration: Repeat steps S1 to S5 until all task statuses are marked as completed, then end the task assignment process.

2. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, Step S2, which updates the local task state based on the hard-lock priority rule and the inertia threshold rule, specifically includes: If the saliency value of the task sent by the neighbor is a hard-locked flag, the agent will forcibly accept the ownership information of the neighbor regardless of whether the local scalar timestamp is updated, and update the local saliency value of the task to the hard-locked flag. If the task saliency value sent by the neighbor is not a hard-locked identifier, and the neighbor's scalar timestamp is greater than the local scalar timestamp, then determine the ownership of the current task: If the local entity believes that the task belongs to itself, it calculates the difference between the local significance value and the neighbor's significance value. Only when the difference is greater than the preset inertia threshold will the ownership change of the neighbor be accepted. If the local entity believes that the task does not belong to it, it directly accepts the salience value, ownership, and timestamp updates from its neighbors.

3. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The step S1, determining whether the current state is congested, specifically involves the agent marking its local state as congested when it determines that there are neighboring nodes within its own communication radius that are less than the preset congestion sensing radius and have a higher ID priority than itself.

4. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The congestion penalty mentioned in step S3 is implemented as follows: The agent determines the number of neighboring nodes within its communication radius that have a higher ID priority than itself. If there are neighboring nodes that meet the conditions, it is marked as a congested state. When calculating the local salience cost of retaining the task, a positive congestion penalty constant is added to the calculation result, thereby artificially reducing the willingness to retain the task and prompting the task to flow to the agent in the non-congested area.

5. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The specific steps of the hierarchical screening described in step S4 are as follows: Define emergency threshold ; Calculate the remaining deadlines for candidate tasks. ,in This is the deadline for the task. The current moment; like The task is assigned to the first priority set, and then ranked within the set according to... Sort in ascending order; if The task is assigned to the second priority set and sorted in ascending order within the set according to the Euclidean distance between the agent's current position and the task position. The agent first attempts to insert tasks from the first priority set into the local task sequence, and only attempts to insert tasks from the second priority set if the first priority set is empty.

6. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The implementation of the hard-locking mechanism in step S5 includes: Define the hard-locked identifier as a negative constant whose value is less than any normal saliency cost calculated based on distance and time. When the agent's physical state is executing a task, the task saliency value calculated based on algorithm logic is ignored, and the saliency value of the corresponding task is directly overwritten as the hard-locked identifier value.

7. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The scalar timestamp is a floating-point time record maintained for each independent task, used to replace the vector clock for each agent in the original PI; When an agent acquires task ownership through computation or locks a task through a hard lock mechanism, it updates the scalar timestamp corresponding to the task to the current local system time.

8. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The significant benefit mentioned in step S3 is calculated in the following way: Define the current task execution sequence of the agent as: Calculate the total path cost for this task sequence. ; Calculate the total path cost of the new task sequence after removing a single task. ,in Tasks to be removed; Task significance value The change in this value represents the significant benefit after removing the task.

9. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 8, characterized in that, Total path cost The calculation formula is: in, For the first The location of each task Represents Euclidean distance. The average moving speed of the agent. For the first The execution time of each task.

10. The distributed multi-agent task allocation method based on the PI algorithm in a communication-constrained environment according to claim 1, characterized in that, The formula for calculating marginal significance in step S4 is as follows: in, This represents the total path cost of the original task sequence. To carry out the new task Insert at the optimal position in the original task sequence The total path cost of the new sequence.

Images