A distributed material transportation allocation method based on average monotonicity strategy

Abstract

The application discloses a kind of distributed material transport distribution methods based on average monotonicity strategy, it is related to distributed decision-making technical field.The method includes: establishing agent material transport distribution model, integrating capacity constraint and time window constraint;Design the task insertion and removal mechanism based on average monotonicity strategy, the task coordination benefit is smoothly handled, to ensure that the bid sequence is not decreasing;Average monotonicity correction is introduced in distributed auction-consensus framework, realize task conflict resolution and distribution consistency.The application solves the problem that traditional distributed algorithm cannot converge and cyclic distribution in logistics scheduling and other non-submodular scenes, significantly improves the task allocation success rate and system stability, while maintaining excellent scheduling efficiency, suitable for large-scale, dynamic agent collaborative scheduling scene.

Description

Technical Field

[0001] This invention relates to the field of distributed decision-making technology, and in particular to a distributed material transportation and allocation method based on an average monotonicity strategy under non-submodular constraints. Background Technology

[0002] Against the backdrop of the rapid development of modern logistics systems such as global e-commerce, instant delivery, and emergency material dispatch, the problem of large-scale, dynamic, and multi-constrained resource allocation is becoming increasingly prominent. Whether it is real-time order allocation in urban warehousing and distribution networks or the coordinated dispatch of multiple teams and materials in disaster relief, both require efficient, stable, and adaptive allocation of transportation tasks under conditions of dispersed information and uncertain environments.

[0003] Distributed material transportation allocation methods are a technical approach to address the aforementioned challenges. However, existing auction-consensus-based allocation algorithms are generally based on the assumption that task benefits are independently additive, making it difficult to handle task synergy effects arising from path overlap and temporal coupling in real-world scenarios. This "non-submodular" characteristic leads to bidding oscillations and allocation loops during the iteration process, severely compromising convergence and thus limiting the reliable application of existing methods in complex logistics scheduling.

[0004] Therefore, overcoming the convergence challenge of distributed task allocation in non-sub-modal environments and developing a new collaborative decision-making mechanism that can both ensure system stability and maintain scheduling efficiency is of great significance for promoting intelligent logistics and improving the reliability and adaptability of resource allocation systems. Summary of the Invention

[0005] This invention aims to provide a distributed material transportation and allocation method based on the average monotonicity strategy to solve the convergence problem of multi-agent task allocation under non-submodular constraints, and improve the system robustness and scheduling efficiency.

[0006] This invention relates to a distributed material transportation and allocation method based on an average monotonicity strategy, comprising:

[0007] S1. Establish a distributed multi-agent task allocation model, construct an optimization model for task set T, agent set V, agent capacity C, and task time window constraints, with the objective of minimizing the total travel time.

[0008] S2. Each agent independently calculates the performance impact value of inserting unassigned tasks, processes the bidding sequence using the average monotonicity strategy, generates a monotonically non-decreasing benefit evaluation sequence, selects the optimal task to insert into its own task list, and updates the path.

[0009] S3. Agents exchange RPI lists and agent lists through a communication network, update the global task attribution according to preset rules, and resolve conflicts among multiple agents competing for the same task.

[0010] S4. Each agent identifies the tasks that need to be removed due to consensus, calculates the performance impact of their removal, processes them using an average monotonicity strategy, and removes the tasks that cause the least damage to the overall system benefits in turn.

[0011] S5. Repeat steps S2 to S4 until the task lists and task ownership information of all agents no longer change, and the algorithm converges.

[0012] According to the distributed material transportation and allocation method based on the average monotonicity strategy provided by the present invention, in step S1, a distributed intelligent agent task allocation model is established, and the construction process includes:

[0013] S11. Define the task set For each task Define its attribute as a tuple ,in Indicates task The starting position, Indicates task The destination location Indicates task The earliest possible start time is the release time. Indicates task The latest deadline that must be completed;

[0014] S12, Define the set of intelligent agents For each intelligent agent Its maximum cargo capacity is defined as The initial position is ;

[0015] S13. Define the set of all location points as... ;

[0016] Define variables Its initial value is set to , representing intelligent agents Arrival Location Time;

[0017] Define function , representing intelligent agents Upon arrival The number of loaded tasks, and satisfying the constraint: for any and any ,have ;

[0018] For any pair of positions Define function , indicating from position Drive to the location The shortest time required, and for any ,have ;

[0019] Assuming for each task There exists at least one intelligent agent. Within its time window, it can be started from the beginning. Transport to destination ;

[0020] S14, Definition For intelligent agents According to its optimal route transportation The minimum total travel time for all tasks is calculated using the following formula: Indicates from position Drive to the location The time required It is a path The first in Node index;

[0021] in To be assigned to intelligent agents The set of transportation tasks; For intelligent agents Transport its task set The optimal driving route for all tasks;

[0022] S15. Construct an optimization model with the objective of minimizing the total travel time, and solve for the objective function:

[0023]

[0024] S16. Define model constraints:

[0025] (1) , ;

[0026] (2) ;

[0027] (3) ;

[0028] (4) ;

[0029] (5) .

[0030] According to the distributed material transportation and allocation method based on the average monotonicity strategy provided by the present invention, the average monotonicity smoothing process in steps 2 and 4 includes:

[0031] Divide the bidding sequence into multiple contiguous blocks;

[0032] If the average value of the next block is less than that of the previous block, then the two blocks are merged.

[0033] Replace all bid values ​​in each block with the average value of that block to form a non-decreasing sequence;

[0034] In step S2, the process by which each agent independently calculates the performance impact of inserting unassigned tasks includes:

[0035] S21. Calculate the minimum cost increment for the current path;

[0036]

[0037] S22, for each candidate task Construct a sequence containing the scores of currently assigned tasks and the original scores of new tasks, and apply mean monotonicity processing to the sequence.

[0038] Construct sequence ;

[0039] Smoothing ;

[0040] Extract smoothed IPI values ;

[0041] S23. Check the smoothed scores of all candidate tasks and select the best one to insert;

[0042] Find the best task ;

[0043] Decision conditions ;

[0044] Perform the insertion:

[0045] The task The start and end points are inserted into the agent. path Optimal position The position is determined by the minimum value in S21.

[0046] According to the distributed material transportation and allocation method based on the average monotonicity strategy provided by the present invention, in step S3, the process of exchanging information among intelligent agents to eliminate conflicts during the consensus phase includes:

[0047] For the task intelligent agent Intelligent agents with neighbors After communication, information received The update rules are as follows:

[0048] a) If Then update ;

[0049] b) If Then keep constant;

[0050] Through multiple rounds of interaction, all agents are made aware of each task. and Achieve global consensus.

[0051] According to the distributed material transportation and allocation method based on the average monotonicity strategy provided by the present invention, in step S4, the process of each agent independently performing the average monotonicity task removal includes:

[0052] S41. After global consensus is reached in the consensus phase (S3), each agent... This will result in a current, consistent list of agent affiliations. Based on this list, the agent needs to identify its current task list. In this context, the tasks that fail in the global competition are those of the intelligent agent. It will iterate through every task in its list. Check the global attribution results: if This means that in the global consensus, the task Determined to be caused by another intelligent agent (i.e. This makes execution more efficient. Such tasks constitute intelligent agents. The set of tasks to be removed in this round is denoted as . ;

[0053] S42, For sets Each task in Calculate the cost savings that can be achieved by removing it from the current path; ;

[0054] The higher the RPI value, the greater the cost savings from removing the task, meaning the heavier the "burden" the task currently imposes on the path.

[0055] S43. Smooth the RPI list of tasks to be removed;

[0056] Similar to the task insertion phase (S2), directly using the original RPI values ​​for removal decisions may still lead to oscillations in non-submodal scenarios. Therefore, this step performs average monotonicity smoothing on the list of RPI values ​​corresponding to the current set of tasks to be removed, ensuring that the removal benefit evaluation used for decision-making is smooth and non-decreasing, and avoiding the removal decision from getting stuck in a loop due to the complex cost coupling relationship between tasks;

[0057] Construct the RPI list: ;

[0058] Smoothing: ;

[0059] S44. Loop through and remove the least efficient task for the current agent;

[0060] Find the task that should be removed: ;

[0061] Decision-making conditions: ;

[0062] Perform the remove and update operation from the path. and list Remove task Update the remaining task set. Recalculate new RPI values ​​for the remaining tasks and re-smoothly generate new ones. ;

[0063] S45. Repeat S43-S44 until there are no more tasks to remove;

[0064] According to the distributed material transportation and allocation method based on the average monotonicity strategy provided by the present invention, the iterative convergence judgment process in step S5 includes:

[0065] After two consecutive complete iterations, the following condition is met:

[0066] (a) ;

[0067] (b) Consistent and non-empty across all agents;

[0068] The algorithm terminates and outputs when the condition is met. Thus, the intelligent agent... This round of task removal based on the average monotonicity strategy has been completed, and its task list and path have been optimized a second time based on global consensus and local efficiency.

[0069] This invention provides a distributed resource transportation and allocation method based on an average monotonicity strategy. By introducing this strategy, the core lies in smoothing the task bidding sequence, forcing it to form a monotonically non-decreasing sequence, thereby fundamentally eliminating the increasing benefit cycle caused by positive synergy between tasks. This innovative mechanism enables the traditional distributed auction-consensus framework to effectively adapt to non-submodular constraint scenarios, solving the key problem that existing algorithms cannot guarantee convergence in such scenarios, and significantly improving the convergence and stability of the distributed task allocation process.

[0070] By combining the aforementioned average monotonicity strategy with multi-task time window constraints, this invention achieves synergistic optimization of task allocation efficiency and overall system performance. The algorithm enables each agent to autonomously and efficiently optimize its task package based on local information and global consensus. This not only significantly improves the success rate of large-scale task allocation and reduces allocation failures caused by conflicts and loops, but also minimizes the total travel time of all agents by optimizing path planning, thereby reducing operating costs and improving overall transportation efficiency. The distributed architecture of this method avoids single points of failure, and its built-in convergence guarantee mechanism further enhances the robustness and reliability of the system in complex and dynamic environments. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 A schematic diagram of the overall process of a distributed material transportation and allocation method based on an average monotonicity strategy provided in an embodiment of the present invention;

[0073] Figure 2 A schematic diagram of the average monotonicity smoothing process in a distributed material transportation and allocation method based on the average monotonicity strategy provided in an embodiment of the present invention;

[0074] Figure 3 This is a flowchart illustrating the generation of decision schemes in a distributed material transportation and allocation method based on an average monotonicity strategy, provided in an embodiment of the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0076] The following is combined with Figures 1-3 This invention describes a distributed material transportation and allocation method based on an average monotonicity strategy.

[0077] Figure 1 This is a schematic diagram illustrating the overall process of a distributed material transportation and allocation method based on an average monotonicity strategy, provided as an embodiment of the present invention. It includes:

[0078] S1. Based on the agent capacity C and task time window constraints, establish a distributed multi-agent task allocation model;

[0079] S11. Formalize the set of tasks that need to be transported. Let there be a total of... There are a number of transportation tasks, and their index set is denoted as... Each task A quadruple A complete description of its key attributes: Among them Indicates task The starting location (i.e., the pickup point). Indicates task The destination location (i.e., the delivery point). Indicates task The release time is the earliest time the agent can begin executing this task. If the agent arrives earlier, it must wait until this time. This indicates a task. The latest delivery time, i.e. the deadline by which the task must be completed;

[0080] S12. Define the set of intelligent agents and their attributes. Let there be a total of... A set of indices for an intelligent agent performing a transportation task is denoted as . For each agent in the set Define two key attributes: one is the maximum cargo capacity. This represents the maximum number of tasks the agent can carry at one time, and is the core operational constraint; the second is the initial position. This usually represents the location of the warehouse or site where the agent is before it begins to execute the task;

[0081] S13. Define spatiotemporal relationships and basic functions. To accurately model the movement and task execution process of the intelligent agent, a series of basic variables and functions need to be established. First, summarize all relevant locations and define them as a set. It covers the start and end points of all tasks and the starting positions of all agents. For each agent... and each position Introducing variables To represent intelligent agents Arrival Location The specific time is usually initialized to -1 to indicate that it has not yet arrived. At the same time, a function is defined. To represent intelligent agents Upon arrival At that moment, the number of tasks it is carrying; this function must always satisfy a hard constraint throughout the process: for any intelligent agent and at any time They all This means that the cargo load will never exceed its maximum capacity. Furthermore, for any two locations in the environment... Define function To represent the agent's position Move directly to the location The minimum required travel time, by definition, for the same location, is... Finally, we assume that the transportation problem itself is feasible, that is, for every task... There must be at least one intelligent agent. It is possible to find a path that, while satisfying capacity constraints, falls within the time window. Internally, it will start from the beginning. Successfully delivered to the destination ;

[0082] S14. In this model, the core decision of the algorithm is to assign a subset of tasks to each agent and plan a driving path for it, using... To represent the final allocation to the intelligent agent A set of tasks. For such a given set of tasks... ,definition For intelligent agents The optimal travel path for executing all tasks is an ordered sequence of location visits that must satisfy the following conditions: the agent must visit the start and end points of each assigned task in sequence; when visiting any task's start point, the agent must have remaining capacity to load that task; and the execution of the entire path must satisfy the time window constraints of each task. The cost of this path is also considered. Defined as the total time the agent travels along this path, its calculation formula is: ,in The first in the path One location, This represents the total path length. The cost of this path is a core metric for measuring the scheduling efficiency of a single agent.

[0083] S15. Construct an optimization model with the objective of minimizing the total travel time:

[0084] This method aims to optimize the overall performance of the multi-agent system. Its global optimization objective is set as minimizing the total system operating cost, specifically the total travel time spent by all agents to complete their assigned tasks. This objective can be formally expressed as the following optimization problem: finding a set of task assignments while satisfying all constraints. and the corresponding set of optimal paths. Solve for the objective function: ;

[0085] S16. Define model constraints;

[0086] To ensure that the generated solution is both effective and feasible, the entire model must strictly satisfy the following five types of constraints:

[0087] Task mutual exclusion constraint means that the task sets assigned to any two different agents cannot overlap. This ensures that each task can only be executed by one agent. ;

[0088] The task coverage constraint requires that the union of the tasks assigned to all agents must equal the complete task set. ,Right now To ensure that no task is missed;

[0089] The agent capacity constraint, as mentioned earlier, requires that at any given time, the number of tasks loaded by any agent must not exceed its maximum capacity. ,Right now ;

[0090] The latest time constraint requires that for each task The intelligent agent reaches its destination time It must not be later than the latest delivery time specified for the task. ,Right now ;

[0091] Time window and travel logic constraint, this is a composite constraint, its mathematical expression is as follows: This constraint ensures the temporal rationality of task execution: the agent performs the task... starting point The actual start time of service (i.e., the later of its arrival time and mission release time), plus the minimum travel time required from the origin to the destination, must be no later than its arrival time at the destination. Time;

[0092] In logistics transportation, task efficiency depends on factors such as fuel consumption or required travel distance, and synergistic effects between tasks often lead to a significant increase in the efficiency of subsequent tasks. Therefore, these efficiency gains may exceed those of earlier tasks, resulting in potential task loops and infinite loops in task allocation. This poses a challenge to distributed material transportation algorithms, as they cannot guarantee convergence in non-decreasing marginal scenarios. To address this issue, this invention provides a distributed material transportation allocation method based on an average monotonicity strategy.

[0093] Based on the above Figure 2 The average monotonicity smoothing process described in steps 2 and 4 ensures that the benefit of a task does not exceed the benefit of previous tasks, thereby avoiding cyclical task allocation. Its core idea is to calculate bids in a manner that maintains a non-decreasing order of task importance. The average monotonicity smoothing process described in steps 2 and 4 includes the following steps:

[0094] 1. Group tasks into blocks, each block representing a group of tasks to be processed together; 2. If the average benefit of a subsequent block is less than that of the current block, merge these blocks to ensure that the bidding sequence is non-decreasing; 3. Calculate the average of each block and update the sequence to ensure that no task's benefit exceeds that of any preceding task.

[0095] S2, the agent independently performs the average monotonicity task insertion;

[0096] S21. In the initial stage of distributed task allocation, each agent needs to independently evaluate the potential benefits of adding unassigned tasks to its own task list. Therefore, Insertion Performance Impact (IPI) is introduced as a core metric. For the current agent, its planned optimal path is denoted as... This path corresponds to the assigned task set. And the total operating cost is Now consider a candidate unassigned task. Its starting point and ending point are respectively and In order to calculate the task Insert existing path The additional cost we incur needs to be evaluated across all possible combinations of insertion positions: starting point Insert at a certain position in the path. Then the finish line Insert to a later position ( IPI value That is, defined as the combination of all legal insertion positions. Insert task Post-new path total cost The minimum difference between the cost of the original path and the cost of the original path;

[0097] Right now:

[0098] This difference reflects the increase in marginal time cost due to accepting new tasks and is the basis for the agent's task selection decision.

[0099] S22. In traditional auction algorithms, directly using the original IPI value for decision-making may lead to bid value oscillations and cyclical allocation when there is a positive synergistic effect between tasks. To ensure the convergence of the algorithm in non-submodal scenarios, this invention introduces an average monotonicity strategy to smooth the bidding sequence in the agent's internal decision-making process;

[0100] Specifically, for current intelligent agents and candidate tasks We first construct an extended scoring sequence: this sequence consists of a list of scores corresponding to the tasks currently assigned to the agent. With the task to be inserted The original IPI value This sequence is formed by connecting the following components. The average monotonicity smoothing function is then applied to this sequence. ;

[0101] After processing, we extract the corresponding task from the end of this smoothed sequence. New rating This process is essentially a smoothing and monotonic correction of the original bid value. It suppresses sudden changes in local benefits caused by task synergy and forces the bid sequence to exhibit non-decreasing characteristics, thereby fundamentally avoiding the allocation cycle that may be triggered by the benefits of subsequent tasks surpassing those of previous tasks.

[0102] S23. Obtain smooth scores for all candidate tasks after processing for average monotonicity. Afterwards, each agent will independently make a local task insertion decision. The decision logic is as follows: from all unassigned candidate tasks, the agent selects the task that brings the greatest relative benefit to its current path, i.e., its smoothed score and the globally best score for that task known to the agent (stored in [the database]). The task with the greatest difference (in China). Specifically, the agent seeks the task. ;

[0103] Determine decision conditions Is it true? If true, it indicates that the task should be inserted into the agent. The path can generate net income;

[0104] The agent will perform the insertion operation: based on the optimal insertion position determined when calculating the original IPI in S21. and , will the task The starting point and the end point Insert sequentially into its current path The corresponding position is used to update the path. and task list At the same time, update the task rating list stored locally. and agent allocation list If the condition is not met, it means that there are currently no unassigned tasks that can bring positive benefits, and the agent will end the local task insertion phase of this round. In this way, each agent autonomously and greedily optimizes its own task load based on local information and smoothed benefit evaluation, while laying the foundation for subsequent global conflict negotiation;

[0105] S3, Consensus Phase: Agents exchange information to eliminate conflicts;

[0106] In the distributed task allocation process, since each agent makes task insertion decisions independently based on local information only in phase S2, it is inevitable that multiple agents will simultaneously compete to select the same task, resulting in task allocation conflicts. If such conflicts cannot be effectively resolved, tasks will be repeatedly allocated, undermining the feasibility of the solution. To address this, this invention designs a communication-based consensus phase, aiming to achieve global consensus among all agents regarding the final assignment of each task. This phase relies on a predetermined communication network topology, where agents can only exchange information with their direct neighbors in the network. Through multiple rounds of information propagation and iterative updates, the crucial consistency information is ultimately disseminated to all agents throughout the network.

[0107] For the task intelligent agent Intelligent agents with neighbors After communication, information received The update rules are as follows:

[0108] like This indicates that the neighboring intelligent agent Provided information about the task If the agent obtains better (lower) cost information, then the intelligent agent Updating one's own understanding, that is, updating ;

[0109] like Then it retains its original better value and maintains constant;

[0110] This means intelligent agents Task The RPI value is lower (i.e., the execution efficiency is higher). Therefore, the agent... Update, acknowledge mission Should be handled by intelligent agents Execution. Through multiple rounds of interaction, all agents are enabled to perform tasks. and Achieve global consensus;

[0111] S4, Each Intelligent Agent Identify tasks that need to be removed due to consensus, calculate their performance impact value, process them using the average monotonicity strategy, and remove the tasks that cause the least damage to the overall system benefits in sequence.

[0112] S41, to The RPI list of tasks in the middle is smoothed out;

[0113] Construct the RPI list: ;

[0114] Smoothing: ;

[0115] S42. Based on the global allocation list reached after the consensus phase. Find the tasks you lost in your task list. ;

[0116] S43, For sets Each task in Calculate the cost savings that can be achieved by removing it from the current path. ;

[0117] S44. Loop through and remove the least efficient task for the current agent;

[0118] Find the task that should be removed: ;

[0119] Decision-making conditions: ;

[0120] Perform remove and update: from path and list Remove task Update the remaining task set. Recalculate new RPI values ​​for the remaining tasks and re-smoothly generate new ones. ;

[0121] S45. Repeat S43-S44 until there are no more tasks to remove;

[0122] S5. Iterative convergence judgment;

[0123] S2 to S4 together constitute a complete distributed task allocation and optimization cycle: First, each agent independently inserts a task based on smoothness benefit evaluation (S2). Then, a global consensus on task ownership is reached through the communication network to resolve conflicts (S3). Finally, each agent removes a task based on smoothness evaluation according to the consensus result to optimize the local load (S4). However, a single execution of this cycle is often insufficient to obtain a globally stable and optimal allocation scheme, because local insertion and removal decisions may change again after a new round of consensus is reached. Therefore, the algorithm is designed as an iterative process, requiring repeated execution of S2 to S4 until the state of the entire multi-agent system satisfies the convergence condition.

[0124] The ultimate goal of convergence is the task list of all agents. And the global task attribution information they perceive. No further changes, that is:

[0125] Each agent Task allocation set The set is exactly the same as the set after the previous iteration, that is ;

[0126] All agents must maintain completely consistent global attribution information for each task, and this consistent information must indicate that each task has a definite attributor. That is... Consistent and non-empty across all agents;

[0127] The algorithm terminates and outputs when the condition is met. .

[0128] To illustrate the implementation process of this invention in more detail, the following uses a small-scale scenario to demonstrate the complete process of the algorithm from the initial state to convergence step by step:

[0129] Assume there are 2 agents in the system. There are 3 transportation tasks. The relevant parameter settings for the intelligent agent and the task are shown in Tables 1 and 2. The coordinates are Cartesian coordinates, the distance unit is kilometers, the time unit is minutes, and the speed is set to 1 km / min.

[0130] Table 1 Task Parameter Table

[0131]

[0132] Table 2. Agent Parameter Table

[0133]

[0134] S2 Task Insertion Phase, Intelligent Agent The raw insertion performance impact (IPI) of each task is calculated based on local information, and the results are obtained. , , The IPIs were approximately 10.0, 17.07, and 4.14, respectively. Meanwhile, It is also calculated independently, because its starting point is closer, and thus... The IPI is approximately 9.66. Both agents choose the task with the lowest IPI. Inserting into its own task list results in a task conflict. It was assigned repeatedly;

[0135] During the S3 consensus phase, the two agents exchange their respective RPI lists and attribution information through a communication network. right It has a rating of 10.0. The score is 9.66. According to the consensus update rules, the one with the lower score wins. Confirmed. Should be Execute and update the global ownership list;

[0136] S4 task removal phase, intelligent agent Checking its own list, it was found The overall owner is not oneself, therefore... Marked as pending removal. It calculates the removal. The cost savings, i.e., its original IPI of 10.0, are compared with the global consensus score of 9.66. Since the removal benefit exceeds the known best execution efficiency for this task, the decision condition is met. Will Remove from its path;

[0137] Thus, the first iteration ended, but the task... and Not yet allocated; system has not converged.

[0138] The algorithm automatically starts the second iteration. In phase S2, Reassess the remaining tasks. At this point, due to... Already carried The path has changed. Calculate insertion or The IPI may decrease due to potential path cooperation effects. The two agents independently choosing the optimal task insertion again may generate new conflicts, such as simultaneous competition. The subsequent S3 consensus phase will resolve conflicts again, and the S4 phase will remove tasks that failed to reach a consensus.

[0139] based on Figure 3 The flowchart of the decision-making scheme describes a repetitive insertion-consensus-removal cycle. Under the influence of the average monotonicity strategy, the benefit score used for each decision is smoothed to ensure it is monotonically non-decreasing, effectively avoiding score oscillations and allocation cycles caused by positive synergistic effects between tasks. After several rounds of adjustments, the system eventually reaches a stable state.

[0140] To verify the effectiveness of the method of this invention, we conducted Monte Carlo simulations to evaluate the algorithm's performance. Different numbers of agents were configured in a 100m × 100m simulation environment. Number of tasks and agent capacity Higher agent capacity This typically reduces the total travel time. This can be attributed to having a higher capacity. The agent can handle more tasks simultaneously, thus achieving better feasible route planning. However, as the capacity increases, the feasibility of agent task planning improves, but conflicts between agent plans also increase, which in turn increases the possibility of cyclic assignment failures. Furthermore, the task assignment method using average monotonicity smoothing has fewer unfinished assignments. Tables 3 to 5 describe the number of unfinished assignments under different capacities, with and without average monotonicity smoothing.

[0141] Table 3 Capacity Unallocated Number

[0142]

[0143] Table 4 Capacity Unallocated Number

[0144]

[0145] Table 5 Capacity Unallocated Number

[0146]

[0147] This method smooths out the synergistic effect between tasks, effectively solves the non-submodular problem caused by the synergistic effect of positive tasks in the traditional auction consensus framework, and improves the convergence speed and stability of distributed task allocation.

[0148] In summary, by introducing the ingenious correction mechanism of "average monotonicity strategy", this invention completely solves the convergence problem of distributed task allocation under non-submodular constraints without significantly sacrificing the optimization quality of the allocation scheme. This greatly improves the practicality, robustness and reliability of multi-agent scheduling systems, and is particularly suitable for dynamic and complex scenarios such as modern logistics and emergency delivery.

[0149] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0150] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A distributed material transportation and allocation method based on an average monotonicity strategy, characterized in that, include: S1. Establish a distributed multi-agent task allocation model, construct an optimization model of task set T, agent set V, and task time window constraints, with the goal of minimizing the total travel time; S11, Assume there are a total There are a number of transportation tasks, and their index set is denoted as... Each task A quadruple Fully describe its key attributes: Indicates task The starting location (i.e., the pickup point). Indicates task The destination location (i.e., the delivery point). Indicates task The release time is the earliest time the agent can begin executing this task. If the agent arrives earlier, it must wait until this time. This indicates a task. The latest delivery time, i.e. the deadline by which the task must be completed; S12, Collection of Intelligent Agents ,in Represents the total number of agents, for each agent in the set. It has two key attributes: one is the maximum cargo capacity. This represents the maximum number of tasks the agent can carry at one time, and is the core operational constraint; the second is the initial position. This usually represents the location of the warehouse or site where the agent is before it begins to execute the task; S13. Summarize all relevant locations and define them as a set. It covers the start and end points of all tasks and the starting positions of all agents, for each agent and each position Introducing variables To represent intelligent agents Arrival Location The specific time is usually initialized to -1 to indicate that it has not yet arrived. Meanwhile, a function is defined. To represent intelligent agents Upon arrival At that moment, the number of tasks it was carrying; This function must always satisfy a hard constraint throughout the entire process: for any intelligent agent and at any time They all That is, the cargo capacity will never exceed its maximum capacity. Furthermore, for any two locations in the environment... Define function To represent the agent's position Move directly to the location The minimum required travel time, by definition, for the same location, is... ; S14, Let To be assigned to intelligent agents The set of transportation tasks; let's assume For intelligent agents Transport its task set The optimal driving path for all tasks; let For intelligent agents According to its optimal route transportation The minimum total travel time for all tasks is calculated using the following formula: Indicates from position Drive to the location The time required It is a path The first in Node index; S15. Construct an optimization model with the objective of minimizing the total travel time, and solve for the objective function: ； S16. Define model constraints: Task mutual exclusion constraint means that the task sets assigned to any two different agents cannot overlap. This ensures that each task can only be executed by one agent. ; The task coverage constraint requires that the union of the tasks assigned to all agents must equal the complete task set. ,Right now To ensure that no task is missed; The agent capacity constraint, as mentioned earlier, requires that at any given time, the number of tasks loaded by any agent must not exceed its maximum capacity. ,Right now ; The latest time constraint requires that for each task The intelligent agent reaches its destination time It must not be later than the latest delivery time specified for the task. ,Right now ; Time windows and schedule logic constraints ensure the temporal rationality of task execution: the agent performs tasks... starting point The later of the arrival time and the mission release time, plus the minimum travel time required from the origin to the destination, must be no later than the time the mission arrives at the destination. The time, that is ; S2. Each agent independently calculates the performance impact value of inserting unassigned tasks, processes the bidding sequence using the average monotonicity strategy, generates a monotonically non-decreasing benefit evaluation sequence, selects the optimal task to insert into its own task list, and updates the path. S21. For the current agent, its planned optimal path is denoted as... This path corresponds to the assigned task set. And the total operating cost is Now consider a candidate unassigned task. Its starting point and ending point are respectively and In order to calculate the task Insert existing path The additional cost is evaluated by considering all possible combinations of insertion positions: [The sentence fragment about starting point is incomplete and lacks context. It's unclear what this refers to.] Insert at a certain position in the path. Then the finish line Insert to a later position ( IPI value That is, defined as the combination of all legal insertion positions. Insert task Post-new path total cost The minimum difference between the cost of the original path and the cost of the original path; Right now: This difference reflects the increase in marginal time cost due to accepting new tasks and is the basis for the agent's task selection decision. S22. To ensure the convergence of the algorithm in non-submodal scenarios, this invention introduces an average monotonicity strategy to smooth the bidding sequence in the agent's internal decision-making. For the current intelligent agent and candidate tasks First, an extended scoring sequence is constructed: this sequence consists of the scoring list corresponding to the tasks currently assigned to the agent. With the task to be inserted The original IPI value Connected together; Apply the average monotonicity smoothing function to this sequence ; After processing, extract the corresponding task from the end of this smoothed sequence. New rating ; S23. Check the smoothed scores of all candidate tasks and select the best one to insert; Get a rating Subsequently, each agent will independently make a local task insertion decision, that is, its smoothed score will be compared with the task known to the agent at present and stored in the database. The task with the largest difference in global best score, specifically, the task the agent searches for. ; Determine decision conditions Is it true? If true, it indicates that the task should be inserted into the agent. The path can generate net income; The agent will perform the insertion operation: insert the task The start and end points are inserted into the agent. path Optimal position This position is determined by the minimum value in S21; S3. Agents exchange RPI lists and agent lists through a communication network, update the global task attribution according to preset rules, and resolve conflicts among multiple agents competing for the same task. For the task intelligent agent Intelligent agents with neighbors After communication, information received The update rules are as follows: a) If Then update ; b) If Then keep constant; Through multiple rounds of interaction, all agents are made aware of each task. and Achieve global consensus; S4, Each Intelligent Agent Identify tasks that need to be removed due to consensus, calculate their performance impact value, process them using the average monotonicity strategy, and remove the tasks that cause the least damage to the overall system benefits in sequence. S41, to The RPI list of tasks in the middle is smoothed out; Construct the RPI list: ; Smoothing: ; S42. Based on the global allocation list reached after the consensus phase. Find the tasks you lost in your task list. ; S43, For sets Each task in Calculate the cost savings that can be achieved by removing it from the current path. ; S44. Loop through and remove the least efficient task for the current agent; Find the task that should be removed: ; Decision-making conditions: ; Perform remove and update: from path and list Remove task Update the set of remaining tasks. Recalculate new RPI values ​​for the remaining tasks and re-smoothly generate new ones. ; S45. Repeat S43-S44 until there are no more tasks to remove; S5. Repeat steps S2 to S4 until the task lists and task ownership information of all agents no longer change, and the algorithm converges.

2. The distributed material transportation and allocation method based on the average monotonicity strategy according to claim 1, characterized in that, The iterative convergence criterion for step S5 is that, after two consecutive complete iterations, the following condition is met: (a) The task assignment set for each agent remains unchanged, i.e.: ; (b) All agents maintain consistent global task attribution information, and each task has a definite attributor, i.e.: Consistent and non-empty across all agents; The algorithm terminates and outputs when the condition is met. .

3. The distributed material transportation and allocation method based on the average monotonicity strategy according to claim 1, characterized in that, The average monotonicity smoothing process described in steps S2 and S4 includes: Divide the bidding sequence into multiple contiguous blocks; If the average value of the next block is less than that of the previous block, then the two blocks are merged. Replace all bid values ​​in each block with the average value of that block to form a non-decreasing sequence.

Images