Multi-robot task allocation method and system based on multi-task optimization
By employing a multimodal, multi-objective gene-targeted differential evolution algorithm based on multi-task optimization, combined with event triggering mechanisms and knowledge transfer, the adaptability problem of multi-robot task allocation algorithms in dynamic environments is solved. This achieves collaborative optimization of main and auxiliary tasks, improving task execution efficiency and system robustness.
Patent Information
- Application Number
- CN202511972003.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-24
AI Technical Summary
Existing multi-robot task allocation algorithms are not adaptable enough to multimodal and multi-objective optimization and dynamic task environments, making it difficult to cope with robot failures and sudden environmental changes, resulting in insufficient task execution efficiency and environmental robustness.
A multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization is adopted. By constructing mathematical models of the main task and auxiliary tasks, the improved differential evolution algorithm is used for task allocation. Combined with event triggering mechanism and knowledge transfer mechanism, information sharing and collaborative optimization between the main task and auxiliary tasks are realized.
It significantly improves the adaptability and robustness of multi-robot systems in task allocation under complex environments, enabling them to quickly switch to equivalent solutions in dynamic scenarios, reducing computational consumption, and improving task execution efficiency and system stability.
Smart Images

Figure CN121920729A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-robot collaboration and intelligent optimization technology, and in particular to a multi-robot task allocation method and system based on multi-task optimization. Background Technology
[0002] Multi-robot cooperative systems (MRCSs) offer greater fault tolerance and robustness compared to single robots, enabling them to effectively execute complex tasks through parallel operations. This has led to widespread interest across various fields, with successful applications in complex scenarios such as disaster emergency response, high-risk environment operations, intelligent logistics transportation, and wireless sensor network deployment. Multi-robot task allocation (MRTA), the core task of MRCSs, aims to allocate tasks at the upper level and plan paths at the lower level for each dispatched robot, maximizing overall system performance through efficient scheduling. Therefore, optimizing the MRTA algorithm to maximize the overall performance of MRCSs remains a crucial research topic in the field.
[0003] Existing MRTA research has paid little attention to multimodal multi-objective optimization algorithms, and the system has poor adaptability in uncertain or dynamic task environments.
[0004] As shown in the multi-robot task allocation method for nuclear accident rescue disclosed in Chinese invention patent application CN120146490A, it constructs an optimization model with the goal of minimizing the longest execution time, the maximum radiation accumulation, and the overall waiting cost by initializing scenario parameters such as the number of robots, task positions, and radiation source information. It adopts a two-stage evolutionary algorithm for task allocation optimization: the first stage generates offspring through binary competition selection and PMX crossover operators to enhance the convergence of the solution; the second stage performs 2-opt local search on the solution with the largest crowding distance to improve the diversity of the solution set. The solution is encoded by combining the task order vector and the entry and exit matrix. However, this method is insufficient in integrating multimodal and multi-objective optimization. On the one hand, its evolutionary strategy focuses on the convergence of a single Pareto front and the improvement of local diversity, failing to generate an equivalent set of optimal solutions covering the characteristics of different rescue scenarios. When faced with sudden diffusion of radiation sources or adjustment of task priorities, it lacks alternative allocation schemes that adapt to dynamic constraints. On the other hand, the response mechanism to dynamic changes in the task environment is limited. The fixed evolutionary logic of the two stages makes it difficult to achieve flexible transfer of optimization knowledge under different rescue scenarios. The encoding method of the solution is also not designed for the diversity of multimodal solutions, resulting in insufficient adaptability in rescue scenarios with dynamic coupling of multiple constraints, ultimately affecting the task execution efficiency and environmental robustness of the multi-robot system. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a multi-robot task allocation method and system based on multi-task optimization.
[0006] The objective of this invention can be achieved through the following technical solutions: A multi-robot task allocation method based on multi-task optimization, the method comprising: A mathematical model for multi-robot task allocation is constructed, wherein the optimization objective of the mathematical model is to minimize the total path length of the scheduled robots performing tasks and to minimize the overall task completion time. Based on the mathematical model for multi-robot task allocation, an improved differential evolution algorithm is used to allocate tasks among the multiple robots; wherein, The improved differential evolution algorithm is a multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization and event triggering mechanism. The tasks include a main task and an auxiliary task. The main task is a multi-robot task allocation problem that can be solved without returning to the original base station, and the auxiliary task is a multi-robot task allocation problem that returns to the base station. The improved differential evolution algorithm achieves information sharing between the main and auxiliary tasks of multiple robots through knowledge transfer, combines a gene-targeting-based correction algorithm to optimize the non-dominated solution of robot task allocation, and adopts an event-triggered mechanism to allocate computing resources to control knowledge transfer and auxiliary task optimization.
[0007] Furthermore, the mathematical model for multi-robot task allocation is expressed as follows: in, The system efficiency objective is the total path length. The goal is to achieve the optimal time, i.e., the longest path. For the first A sequence of tasks executed by a robot; For the task point and mission points The Euclidean distance between them; This is a binary variable, representing whether there is a direct route between task points; The set of all tasks; To be assigned to the A set of tasks for a robot; , Number the robot base station; To return the length of the base station number; A binary variable representing a base station Is there a direct route between this point and task point V? A binary variable representing a task point. With base station Is there a direct route between them?
[0008] Furthermore, the specific process of the improved differential evolution algorithm includes: Step A1: Based on the actual task and the robot, establish a main task population and an auxiliary task population; perform individual encoding and decoding for the main task population and the auxiliary task population; initialize population parameters, the main task population, and the auxiliary task population; the population parameters include the population size. and maximum number of iterations ; Step A2: A binary tournament selection mechanism is used to select from the primary task population and the secondary task population respectively. A number of superior individuals are used to generate the intermediate population for the main task and the intermediate population for the auxiliary task. Step A3: Based on the intermediate population of the main task, a multimodal multi-objective gene-targeted differential evolution algorithm is used to generate a subpopulation of the main task; based on the intermediate population of the auxiliary task, a multimodal multi-objective evolution algorithm is used to generate a subpopulation of the auxiliary task. Step A4: Calculate the fitness function value and population renewal rate for each subpopulation. If the current population update rate Population update rate less than historical archive If the goal is to generate a migration population through knowledge transfer, the original population, sub-population, and migration population are merged, and a new primary task population or auxiliary task population is obtained through non-dominated CSCD sorting and environmental selection; otherwise, individuals are directly selected from the original population and sub-population to update the population. Step A5: Repeat steps A2-A4 until the maximum number of iterations is reached, and output the non-dominated solution of the main task as the optimal task allocation scheme.
[0009] Furthermore, in step A1, the main task population individuals include chromosome segment I and chromosome segment II, where chromosome segment I is for task allocation and has a dimension of [missing information]. The chromosome fragment II is the returned base station number, with a length of [missing information]. The auxiliary task population includes task allocation information, and the individual length is [missing information]. .
[0010] Furthermore, the specific process of encoding and decoding the main task population includes: The robot is assigned a main task through rounding, the execution order of the main tasks is determined by gene locus sorting, and chromosome segment I is edited based on the main task assignment result and the main task execution order; chromosome segment II is edited based on the specified return base station number; The specific process of encoding and decoding the individual members of the auxiliary task population includes: Auxiliary tasks are assigned to the robot through rounding operations, the execution order of the auxiliary tasks is determined by sorting by gene size, and the genes of the individuals in the auxiliary task population are edited based on the auxiliary task assignment results and the auxiliary task execution order; wherein, in the individuals in the auxiliary task population, the robot returns to the starting base station after completing the task.
[0011] Furthermore, the knowledge transfer process includes: Knowledge transfer from the main task to the auxiliary task: Delete chromosome segment II from the individual in the main task population, keeping only chromosome segment I; transfer this individual from the main task population to the auxiliary task population; Knowledge transfer from auxiliary task to main task: Combine chromosome segment II of an individual in the auxiliary task population with that of a non-dominant individual in the main task population; transfer this individual from the auxiliary task population to the main task population.
[0012] Furthermore, in step A3, the process of generating the main task subpopulation using a multimodal, multi-target gene-targeted differential evolution algorithm includes: The location of the bottleneck dimension is determined probabilistically: A Monte Carlo-based detection method is used to randomly sample from the dimension of individuals in the intermediate population of the main task, and the dimensionality of the individuals in the intermediate population of the main task is obtained from a univariate Gaussian distribution. Sampling probability in each dimension And generate a random value from [0,1] for this dimension. If this random value is less than the probability... If so, then this dimension will be taken as the bottleneck dimension for the current individual; Constructing homologous targeting vectors: Constructing homologous targeting vectors for individuals in the intermediate population of the main task using the DE / best / 1 mutation strategy or the DE / r-best / 1 mutation strategy; Insert the homologous target vector into the corresponding individual: Replace the bottleneck dimension in the individual of the main task intermediate population with the corresponding dimension in the homologous target vector, while keeping other dimensions in the individual unchanged, and generate the main task sub-population.
[0013] Furthermore, the process of generating the auxiliary task subpopulation using a multimodal, multi-objective evolutionary algorithm includes: Based on the pre-trained deep reinforcement learning model, the fitness function values of all individuals in the intermediate population of the auxiliary task are calculated, and the individuals in the population are sorted using the non-dominated CSCD sorting method. The DBESM method is used to select role model individuals from the intermediate population of the auxiliary task, and the differential evolution algorithm is combined to generate a mutation and crossover operation to generate the offspring population. Based on each individual in the offspring population, the task execution path of the corresponding robot is planned using an end-to-end DRL model; Update the offspring population, and when the preset number of iterations approaches the maximum value, extract all non-dominated solutions in the population to form a solution set PS; The individuals in the solution set PS are decomposed into sub-individuals corresponding to each robot. The number of intersections of the task paths of each robot is calculated. If there are path intersections, the large neighborhood search algorithm is used to optimize the path of the corresponding robot. The optimized solution set PS is updated and merged with the descendant population. The merged population is then sorted again using the non-dominated CSCD sorting method. Redundant individuals in the merged population are removed by a redundancy removal method, and then NP individuals are selected from them by an environment selection method to form a new population for the next iteration. Finally, the latest solution set PS and Pareto front PF in the population are output as results. The Pareto front PF is the mapping set of all non-dominated solutions in the multi-robot task assignment scenario.
[0014] Furthermore, the triggering conditions of the event triggering mechanism include: When the population renewal rate of the main task population When the population update rate decreases, auxiliary task optimization and knowledge transfer are triggered; when the population update rate... When the number of tasks remains unchanged or increases, pause auxiliary task optimization to save computing resources; The population renewal rate The calculation expression is: in, It refers to population size. It represents the number of individuals updated in the current generation.
[0015] A multi-robot task allocation system based on multi-task optimization, the system comprising a model building module, an algorithm building module, and a multi-task collaborative allocation module.
[0016] The model building module is used to construct a mathematical model for multi-robot task allocation, and to define the optimization objectives and constraints for the total path length and the longest path. The algorithm construction module is used to construct a multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization and event triggering mechanism. The main task is set as a multi-robot task allocation problem that can be solved without returning to the original base station, and the auxiliary task is set as a multi-robot task allocation problem that returns to the original base station. The multi-task collaborative allocation module is used to complete the multi-robot task allocation based on the multi-modal, multi-objective gene-targeted differential evolution algorithm constructed by the mathematical model built by the model building module and the algorithm building module based on multi-task optimization and event triggering mechanism. The improved differential evolution algorithm realizes information sharing between the main and auxiliary tasks of multiple robots through knowledge transfer, optimizes the non-dominated solution of robot task allocation by combining gene-targeting-based correction algorithm, and uses an event triggering mechanism to allocate computing resources to control knowledge transfer and auxiliary task optimization.
[0017] Compared with the prior art, the beneficial effects of the present invention include: 1. This invention proposes a multimodal, multi-objective, multi-robot task allocation method based on multi-task optimization. It designs a multimodal, multi-objective gene-targeted differential evolution algorithm that integrates multi-task optimization, incorporating it into the multi-robot task allocation model to achieve collaborative optimization of primary and auxiliary tasks. This invention introduces cross-scenario multi-task optimization concepts into multi-robot task allocation by assigning non-return-to-base-station tasks as primary tasks and return-to-base-station tasks as auxiliary tasks. Through knowledge transfer, it achieves information sharing and collaborative evolution among tasks, significantly improving the adaptability of task allocation in complex environments and providing multiple sets of non-dominated solutions for diverse scenarios, allowing decision-makers to choose according to actual needs.
[0018] 2. This invention addresses the MRTA problem from a multimodal, multi-objective optimization perspective, enhancing its adaptability to uncertain or dynamic task environments. Traditional MRTA research often focuses on finding a single optimal solution, which is insufficient to handle dynamic scenarios such as robot failures and sudden environmental changes. This invention, through a multimodal, multi-objective optimization framework, converges to the Pareto front in the objective space while simultaneously uncovering numerous equivalent optimization schemes in the decision space. When a preset scheme fails due to unforeseen circumstances, it can quickly switch to an equivalent solution to maintain task continuity, significantly improving system robustness.
[0019] 3. This invention innovatively introduces a gene-targeted correction mechanism into the multimodal, multi-objective differential evolution algorithm. It uses Monte Carlo detection to locate bottleneck dimensions and combines mutation strategies such as DE / best / 1 to construct targeted vectors to optimize non-dominated solutions. This mechanism specifically addresses the dimensionality imbalance problem in algorithm evolution, effectively avoiding getting trapped in local optima and significantly improving the quality and diversity of non-dominated solutions.
[0020] 4. This invention designs an event-triggered computational resource allocation mechanism, which dynamically regulates knowledge transfer and auxiliary task optimization through the population update rate (UR): when the UR decreases, migration is triggered to inject new information; when the UR stabilizes, auxiliary tasks are paused to save resources. This dynamic adaptation strategy reduces ineffective computational consumption and significantly improves algorithm efficiency while ensuring optimization effectiveness, making it particularly suitable for resource-constrained real-time scheduling scenarios involving multiple robots.
[0021] 5. This invention innovatively designs a differentiated encoding and decoding method: the main task adopts a two-stage encoding of task allocation + return to the base station (length s + q), while the auxiliary task adopts a single task allocation encoding (length s). Task-robot matching and execution order planning are achieved through rounding operations and gene sorting. This encoding method accurately adapts to the scenario differences between the main and auxiliary tasks, improves the expressive power and decoding efficiency of the solution, and ensures the feasibility of solutions under complex task constraints.
[0022] 6. This invention constructs a bidirectional knowledge transfer mechanism. The main task migrates to the auxiliary task by extracting task allocation fragments, while the auxiliary task migrates back by splicing base station information from the non-dominated solutions of the main task, achieving efficient sharing of optimization knowledge across tasks. This mechanism breaks down the information barriers of traditional single-task optimization, expands the algorithm search space, improves the convergence speed of non-dominated solutions of the main task, and significantly improves the uniformity of solution distribution.
[0023] 7. This invention introduces a multimodal multi-objective evolutionary algorithm (MMOEA-DL) into the optimization of the auxiliary task. It uses weighted vector decomposition of the multi-objective problem and Chebyshev scalarization functions to guide the optimization direction. Combined with deep reinforcement learning (DRL) path planning and large neighborhood search (LNS) local optimization, it enhances the quality of the auxiliary task solution. High-quality solutions from the auxiliary task are fed back into the main task through knowledge transfer, forming a closed-loop optimization of "main and auxiliary tasks working together and mutually benefiting each other," further improving the system's robustness under dynamic constraints.
[0024] 8. The optimization objectives of this invention directly address the core requirements of multi-robot operations: optimizing the total path length reduces overall system energy consumption, optimizing the longest path shortens the upper limit of task completion time, and achieving optimal system efficiency and time through dual-objective collaborative balancing. Simultaneously, the model strictly constrains full task coverage and conflict-free allocation, ensuring that all tasks are executed appropriately, providing a practical and reliable task allocation scheme for real-world scenarios such as industrial logistics and disaster relief. Attached Figure Description
[0025] Figure 1 This is a schematic diagram illustrating the steps of the multi-robot task allocation method based on multi-task optimization in this invention; Figure 2 This is an architecture diagram of the multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization and event triggering mechanisms in this invention; Figure 3 This is a schematic diagram of the multi-task optimization knowledge transfer process in the embodiment; Figure 4 This is a schematic diagram of an MRTA instance scenario in the embodiment; Figure 5 This is a schematic diagram showing the HV performance ranking of the five comparison algorithms in the embodiment; Figure 6 This is a schematic diagram showing the IGDX performance ranking of the five comparison algorithms in the embodiment; Figure 7 The image shows a comparison of the approximate power parameters (PFs) obtained by the five comparison algorithms in the embodiment on MRTA instance 2-2-20. Detailed Implementation
[0026] The technical solutions of the embodiments of the present 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 the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0027] Example 1 This embodiment discloses a multi-robot task allocation method based on multi-task optimization. The method is as follows: Figure 1 As shown, the process specifically includes steps S1-S3, each described in detail below: S1. Construct a mathematical model for multi-robot task allocation. The optimization objective of the mathematical model is to minimize the total path length of the scheduled robots performing tasks and to minimize the overall task completion time. S2. Construct a multi-task optimized multimodal multi-objective gene-targeted differential evolution algorithm, with multi-robot task allocation that can avoid returning to the original base station as the main task and the problem of returning to the base station as the auxiliary task.
[0028] S3. Using an improved differential evolution algorithm, information sharing between the main task and auxiliary tasks is achieved through knowledge transfer. The non-dominated solution is optimized by combining a gene-targeting-based correction algorithm. An event-triggered mechanism is used to allocate computing resources to control knowledge transfer and complete the multi-robot task allocation.
[0029] Multi-task optimization (MTA) is a recently emerging and effective method for solving complex optimization problems. It can solve multiple optimization problems simultaneously and achieve knowledge sharing between tasks. Compared to single-task optimization, MTA performs better in solving complex problems. Furthermore, traditional MRTA research often focuses on finding a single optimal solution, making it difficult to handle dynamic scenarios such as robot failures and sudden environmental changes. Therefore, to effectively solve the multi-robot task allocation problem in complex environments, this paper proposes a multimodal, multi-objective gene-targeted differential evolution algorithm (MMOGTDE-MTO) based on MTA. This algorithm assigns multi-robot tasks that do not need to return to the original base station as the main task, and tasks that return to the base station as auxiliary tasks. A modified algorithm is used to rationally allocate computational resources, and co-evolution is achieved through knowledge transfer between the main and auxiliary tasks.
[0030] In step S1, the mathematical model for multi-robot task allocation is expressed as follows: in, The system efficiency objective is the total path length. The goal is to achieve the optimal time, i.e., the longest path. For the first A sequence of tasks executed by a robot; For the task point and mission points The Euclidean distance between them; This is a binary variable, representing whether there is a direct route between task points; The set of all tasks; To be assigned to the A set of tasks for a robot; , Number the robot base station; To return the length of the base station number; A binary variable representing a base station Is there a direct route between this point and task point V? A binary variable representing a task point. With base station Is there a direct route between them?
[0031] The flowchart of the multi-modal, multi-objective gene-targeted differential evolution algorithm based on multi-task optimization is as follows: Figure 2 As shown, the specific operating steps are as follows: S201, Initialize Population and Parameters: Set the population size to... The maximum number of iterations is Initialize population update rate Iteration counter The generated population size is The main task population Support mission population Due to population Individuals in the population have two different codes. During population initialization, the population... Chromosomal segment I of an individual and the population Each individual in the population is initialized according to a specific number of dispatch plans, while the population... Chromosomal segment II of the individual is randomly initialized.
[0032] S202, Individual Fitness Assessment: Based on the different MRTA questions corresponding to the primary and secondary tasks, the population is assessed separately. and All individuals are evaluated, their fitness function values are calculated, and the feasibility of task-robot matching and execution sequence planning is verified. S203, Iteration Main Loop: Enter the iteration main loop under the condition that... .
[0033] S204, Population Update Rate Archive: Saves the current population update rate. The value is archived in the temporary population rate. middle.
[0034] In this embodiment, the population renewal rate As an event triggering condition, it effectively allocates computational resources between the main task and auxiliary tasks, reduces unnecessary knowledge transfer, and saves computational resources. Specifically, when the main task population is in the current generation... When the workload is reduced, auxiliary tasks are optimized and knowledge is transferred. The calculation method is expressed as follows: in, It refers to population size. This represents the number of individuals updated in the current generation. When When it remains unchanged or increases, it indicates that the individual evolution efficiency in the main task is relatively high. During descent, auxiliary tasks are used to promote the evolution of the main task.
[0035] S205, Intermediate Population Generation: A binary tournament selection mechanism is used, selecting from the parent population... and Selected Individuals, generating the intermediate population for the main task. and the intermediate population of auxiliary tasks .
[0036] S206, Subpopulation Generation: The intermediate population was generated using the Multimodal Multi-Objective Gene-Targeted Differential Evolutionary Algorithm (MMOGTDE). After completing the iterative optimization, extract the set of non-dominated solutions from the output and randomly select from it. Individuals form subpopulations ,make sure It possesses both convergence and diversity.
[0037] In this embodiment, the basic steps of MMOGTDE (Multimodal Multi-Target Gene-Targeted Differential Evolution Algorithm) are as follows: Step 1: Initialization: Generate the initial population And set the number of iterations. ; Step 2: Calculate fitness: Calculate the fitness function value for all individuals in the initial population; Step 3: Non-dominated sorting: For the initial population Perform non-dominated CSCD sorting to obtain the population. ; Step 4: Role Model Selection: Using DBESM for selection A role model for every individual; Step 5: Non-dominated solution screening: Find Non-dominated solutions in; Step 6: Individual Evolution: On the Population Each individual in ( , (for population size), if If it is a non-dominant solution, then for Offspring individuals are obtained by executing a correction algorithm based on targeted genes. Otherwise, for The offspring individuals are obtained using the DE algorithm. ; Step 7: New Population Construction: A new population is formed by all offspring individuals. ; Step 8: Population Merging and Sorting: ... and Merge the populations and perform non-dominated CSCD sorting on the merged populations. Step 9: Redundant Solution Removal and Environment Selection: Redundant solutions are removed from the merged population using the redundancy removal method, and then the environment selection method is used to select from the processed population. Each individual receives a new Number of iterations Add 1; Step 10: Repeat Steps 4-10 until the maximum number of iterations is reached. ; Step 11: Output Results: After the loop ends, output the final population. Selecting non-dominated individuals as the Pareto optimal set And calculate the Pareto frontier Finally, output and .
[0038] In this embodiment, the non-dominated CSCD sorting in Step 3 first divides the solutions into layers through non-dominated sorting, and then calculates the special crowding distance for the solutions in the same layer by combining clustering information. Solutions with low non-dominated levels and large crowding distances are selected first, thereby filtering out well-distributed solutions.
[0039] In this embodiment, the distance-based elite selection mechanism (DBESM) in Step 4 refers to the process where, for each individual in the population, DBESM evaluates the distribution of solutions around the individual by calculating the distance relationship between the individual and other individuals in the decision space and the target space. Individuals that are close to each other may have similar characteristics, and DBESM selects representative individuals from these neighbors as learning samples for the current individual, thereby prompting the population to converge toward the Pareto front while maintaining diversity.
[0040] In this embodiment, the gene-targeting correction algorithm in Step 6 mainly consists of the following three steps: (1) Probabilistic location of bottleneck dimensions: A Monte Carlo-based detection method is used to randomly sample and determine whether each dimension of the optimal individual is a bottleneck dimension. The first bottleneck dimension is derived from the univariate Gaussian distribution. Sampling probability in each dimension and from the current dimension A random value is generated. If this value is less than the probability... Then the first One dimension serves as the bottleneck dimension; (2) Constructing homologous target vectors: Use the DE / best / 1 mutation strategy or the DE / r-best / 1 mutation strategy to construct the homologous target vectors of the optimal individuals; (3) Insert the same target vector into the optimal individual: The bottleneck dimension in the optimal individual is changed by the corresponding dimension in the same target vector, while the other dimensions in the individual remain unchanged, resulting in a new individual.
[0041] In this embodiment, the redundant solution removal method in Step 9 is to remove solutions that may consume computational resources and adversely affect the convergence of the algorithm during the update iteration. Redundant solutions are determined by comparing the similarity between individuals in the objective function value and the decision variable value. After identifying redundant solutions, one or more of them are removed from the population. The formula for calculating individual similarity is as follows: in, Represents an individual Similarity; It represents the total number of robots; It is the first The number of tasks performed by each robot in an individual; It is a coefficient that measures the importance of different task assignments, and is usually set to 1.
[0042] S207, Calculate fitness and renewal rate: Calculate subpopulation The fitness function value is calculated, and the population update rate UR is updated.
[0043] S208, Auxiliary task trigger judgment: If This triggers the auxiliary task process; if If the above is true, then jump directly to (1.15).
[0044] S209, Auxiliary subpopulation generation (when triggered): When At that time, the intermediate population was analyzed using the MMOEA-DL algorithm. Processing is performed to generate a size of subpopulation .
[0045] In this embodiment, the basic steps of MMOEA-DL (Deep Reinforcement Learning-Assisted Multimodal Multi-Objective Evolutionary Algorithm) are as follows: Step 1: Initialization: Generate the initial population Set the number of iterations At the same time, initialize the parameters required for the algorithm (such as population size). Maximum number of iterations wait); Step 2: DRL Model Training: Train the DRL model using the Actor-Critic training algorithm. 1,000,000 TSP instances with 20 task points were uniformly distributed as training data. The Adam optimizer (learning rate 5E-4), batch size 256, and number of training epochs were used to obtain the trained DRL model, and the weight vector was determined. and ; Step 3: Calculate fitness: Based on the minimum total path length problem of the MRTA problem. and minimizing the longest path These two objectives are used to calculate the initial population. The fitness function values of all individuals in the dataset; Step 4: Non-dominated sorting: For the initial population Perform non-dominated CSCD sorting to select solutions that are well distributed in both the decision space and the objective space, thus obtaining the population. ; Step 5: Role Model Selection: Use DBESM to select A role model for every individual; Step 6: Offspring Population Generation: Using the DE algorithm and the role models determined in Step 5, the offspring population is generated through the DE / current-to-exemplar / 1 mutation strategy and crossover operator. ; Step 7: DRL Path Planning: For offspring populations Each individual in the process is broken down into... For each sub-individual, the DRL model trained in step 2 is used to perform end-to-end path planning for each sub-individual to obtain the initial task execution sequence of each robot and update the individual path information. Step 8: Local optimization of LNS: Determine if the number of iterations is sufficient. If so, then from Select all non-dominated solutions to form a set. ;Will Each individual is split into For each sub-individual, calculate the number of intersections for each sub-individual. ,like Then, the LNS algorithm is used to perform local optimization on the sub-individual, adjusting the task point order through a destruction-repair mechanism and updating. The path information of the individual; Step 9: Population Merging and Sort: Merge and sort the original population Offspring population And the optimized Merge the populations and perform non-dominated CSCD sorting on the merged populations again to ensure the convergence and diversity of solutions. Step 10: Redundancy Removal and Environment Selection: The merged population is processed using the redundancy removal method, and then an environment selection method is used to select from the processed population. Each individual, resulting in a new population. Number of iterations Add 1; Step 11: Repeat the iteration: Repeat steps 5-10 until the specified number of iterations is reached. achieve ; Step 12: Output Results: After the loop ends, output the final population. Selecting non-dominated individuals as the Pareto optimal set And calculate the Pareto frontier Finally, output and .
[0046] S210, Calculate the fitness of the auxiliary subpopulation (when triggered): Calculate the subpopulation The fitness function value.
[0047] S211, Knowledge Transfer Population Construction (Triggered): Knowledge sharing between two tasks is achieved by transferring individuals. Based on the knowledge transfer method designed in this invention, a migration population is generated. and .
[0048] Knowledge transfer from the main task to the auxiliary task: The individual in the main task consists of two parts, chromosome segment I and chromosome segment II, while the individual in the auxiliary task only has one part. Therefore, by deleting chromosome segment II from the main task, the individual in the main task can be transferred to the auxiliary task.
[0049] Knowledge transfer from auxiliary task to main task: Combining individuals in the auxiliary task with chromosome segment II in the main task creates a transferred individual. To provide high-quality UAV return-to-base information to individuals in the auxiliary task, chromosome segment II from a non-dominant individual in the main task is selected for transfer.
[0050] from Figure 3 It can be seen that by deleting a portion of chromosome fragments, the optimization of the primary task in assisting the secondary task can be achieved. Combining a portion of chromosome fragments with the secondary task's chromosomes can provide more information for optimizing the primary task.
[0051] S212, Migrating Population Assessment (When Triggered): Assess the migrating population based on the different MRTA problems in the main and auxiliary tasks. and .
[0052] S213, Population Merge (When Triggered): Merge , and Building a population ,merge , and Building a population .
[0053] S214, Environment Selection (when triggered): Selecting from the population using non-dominated CSCD sorting and environment selection respectively. and NP individuals were selected to form the next generation population. and Then, proceed to the next iteration.
[0054] S215, Single-task environment selection (when not triggered): When At that time, directly from In the process, NP individuals are selected through non-dominated CSCD sorting and environmental selection to update the main task population. Auxiliary tasks will not be included in the iteration for the time being.
[0055] S216, Iteration Termination and Result Output: When the iteration counter... achieve The algorithm terminates its iteration. From the obtained main task population... To obtain the Pareto optimal set, non-dominated individuals are selected. and Pareto Front and output the obtained and Repeat steps S203-S216 until the termination condition is met.
[0056] Example 2 This embodiment, based on Embodiment 1 above, discloses a specific implementation example of a multi-robot task allocation method based on multi-task optimization, in order to verify the performance of the proposed algorithm.
[0057] In this embodiment, to verify the performance of the proposed algorithm, multiple robot stations, robots, and task points of different sizes were set up. The number of robot stations was set to 2 and 3 respectively, with 2 and 3 identical robots schedulable in each station. The task sizes were 20, 30, 40, and 50. The distribution of task points and robot stations in the MRTA instance is as follows: Figure 4 As shown.
[0058] Four advanced algorithms were compared. IMMODE_SA is a hybrid algorithm combining differential evolution and simulated annealing, exhibiting differentiated design in areas such as population quality improvement, redundant solution control, and local search enhancement. MMODE_CSCD can efficiently locate multiple Pareto optimal solution sets for multimodal multi-objective problems. SS-MOPSO does not require pre-setting subpopulation parameters, its species partitioning closely matches the real distribution, and it employs dual-space collaborative optimization, efficiently locating multiple equivalent Pareto sets for MMOPs while ensuring a uniform distribution of the Pareto front (PF) in the target space, providing an effective tool for complex optimization problems requiring multiple alternatives. DSC-MOAGDE overcomes the limitations of traditional single-space optimization, achieving a triple balance between convergence accuracy, dual-space diversity, and engineering constraint adaptation.
[0059] The four algorithms described above demonstrate good performance in solving multi-agent task allocation and path planning problems. All comparison algorithms were tested using MATLAB (R2022a) on a Windows 11 (64-bit) computer. In addition to using HV to evaluate the performance of all comparison algorithms in the target space, the inverted generation distance of the decision space, termed IGDX, was used to assess the quality of the PS obtained in the decision space. A smaller IGDX value indicates better algorithm performance. For all comparison algorithms, the total number of fitness function evaluations was set to 40,000, and the population size was set to 200. To effectively evaluate the search performance of the algorithms, each algorithm was run independently 22 times on all MRTA instances, and the mean and standard deviation of the algorithm searches were statistically analyzed. The Wilcoxon test was used to assess whether there were significant performance differences between the algorithms; "+", "−", and "=" indicate that the proposed algorithm's results are significantly better than, significantly worse than, and statistically similar to the comparison algorithms, respectively. Furthermore, the Friedman test was used to rank the performance of all comparison algorithms.
[0060] In this embodiment, Python (3.9) and PyTorch (2.0.0) are used to train the DRL model via the CPU. TSP instances with 50 task points are generated using a uniform distribution of [0,1] to train the DRL. The Adam optimizer is used, and the hyperparameters during the training phase are shown in Table 1. The model trained on the 50 task-point TSP instances is used as the solver to solve for all MRTA instances from 20 to 50 task points.
[0061] Table 1. Hyperparameter settings for DRL Unlike existing multi-robot task allocation algorithms, this scheme employs a strategy combining multi-task optimization and gene-targeted correction to improve the algorithm's search performance and solution quality in complex task environments. The main idea is to treat "MRTA problems that do not require returning to the original base station" as the primary task and "MRTA problems that require returning to the original base station" as auxiliary tasks. A bidirectional knowledge transfer mechanism facilitates information exchange and sharing between the primary and auxiliary tasks, while an event-triggered mechanism regulates the allocation of computational resources and the timing of knowledge transfer, achieving the goal of co-evolution of the primary and auxiliary tasks. A gene-targeted correction algorithm is used to assist the multimodal, multi-objective differential evolution algorithm, performing local optimization on the non-dominated solutions of the primary task. This overcomes the dimensionality imbalance problem in algorithm evolution, effectively avoiding getting trapped in local optima, thereby enhancing the algorithm's ability to solve complex MRTA problems. Furthermore, this embodiment proposes 16 complex MRTA scenarios, which have different numbers of robot stations (2-3), different numbers of schedulable robots (2-3 / station), and different scales of task points (20-50). The task point distribution includes both uniform and non-uniform types. This can effectively verify the effectiveness and robustness of the proposed MMOGTDE-MTO algorithm in diverse and complex scenarios.
[0062] HV and IGDX were used to evaluate the performance of all compared algorithms. Table 2 shows the mean and standard deviation of HV values obtained by all algorithms. Additionally, the statistical analysis results obtained through the Wilcoxon rank-sum test are also presented in Table 2. The results show that MMOGTDE-MTO outperforms other algorithms in HV values across 16 MRTA instances. In other words, this algorithm has better optimization capabilities in the target space than the other compared algorithms. The average HV value obtained for each MRTA instance was subjected to a Friedman test, and the performance ranking of all compared algorithms is as follows: Figure 5 As shown, the proposed algorithm ranks first in performance among all compared algorithms. This is likely primarily because the improved MMOGTDE algorithm enhances its optimization capability by further optimizing the non-dominated solutions found in each generation to find more high-quality solutions. Furthermore, knowledge transfer in multi-task optimization enhances the algorithm's search ability, and the event-triggered mechanism rationally allocates computational resources, resulting in excellent search efficiency for all algorithms.
[0063] Table 2 HV values of all comparison algorithms on different MRTA instances In this embodiment, the IGDX metric is used to evaluate the performance of all comparison algorithms in the decision space. Table 3 shows the IGDX values of the comparison algorithms on all MRTA instances, as well as the statistical results of the Wilcoxon rank-sum test. As shown in Table 3, the IGDX value of MMOGTDE-MTO on all MRTA instances is significantly better than the other four comparison algorithms, meaning that the proposed algorithm can find more high-quality solutions. Figure 6 The ranking of the average IGDX values obtained by all compared algorithms is shown in the Friedman test. Clearly, the proposed algorithm demonstrates a significant performance advantage.
[0064] Table 3. IGDX values of all comparison algorithms on different MRTA instances. To visually demonstrate the effectiveness of the proposed algorithm in solving the MRTA problem, Figure 7 This demonstrates the approximations obtained by all the comparison algorithms on MRTA instance 2-2-20. We can see from Figure 7 It was observed that, under the condition of the same total fitness evaluation, MMOGTDE-MTO can find more high-quality solutions and obtain approximate results. It outperforms other comparison algorithms in terms of convergence and diversity.
[0065] Based on the above analysis, we can conclude that MMOGTDE-MTO can generate multiple sets of high-quality, uniformly distributed non-dominated solutions in complex multi-robot task allocation scenarios, and is an effective method for solving various complex MRTA problems, especially MRTA problems that do not require returning to the original base station.
[0066] In summary, this scheme presents a multimodal, multi-objective, multi-robot task allocation method based on multi-task optimization. The main steps are as follows: Multi-task optimization combined with a gene-targeted correction strategy is used to improve the algorithm's search performance and solution quality in complex task environments. The main idea is to treat the "MRTA problem without returning to the original base station" as the primary task and the "MRTA problem returning to the original base station" as the auxiliary task. A bidirectional knowledge transfer mechanism is used—the primary task extracts task allocation fragments and migrates them to the auxiliary task, while the auxiliary task splices base station information from the primary task's non-dominated solution and migrates it in reverse—to achieve information interaction and sharing between the primary and auxiliary tasks. Simultaneously, an event-triggered mechanism is introduced, based on the population update rate. Dynamic judgment, when The descent triggers knowledge transfer and auxiliary task optimization. When the system is stable, auxiliary tasks are paused to conserve resources, thereby regulating computational resource allocation and achieving the goal of co-evolution of primary and auxiliary tasks. A gene-targeted correction algorithm is used, which locates the bottleneck dimension through Monte Carlo detection and combines it with mutation strategies such as DE / best / 1 to construct targeted vector optimization for non-dominated solutions. This assists the multimodal, multi-objective differential evolution algorithm, overcoming the dimensionality imbalance problem in algorithm evolution and avoiding getting trapped in local optima, thus improving the algorithm's ability to solve complex MRTA problems. A multi-robot task allocation model is established, clearly defining the dual optimization objectives of "minimizing total path length" and "minimizing longest path," while constraining full task coverage and conflict-free allocation. The proposed MMOGTDE-MTO algorithm is used to model and simulate MRTA scenarios, mainly considering factors such as the number of robot stations, the number of schedulable robots, and the scale of task points. This enriches the scenarios for multi-robot task allocation and also facilitates the verification of the algorithm's allocation effect in real complex task environments.
[0067] Furthermore, this solution addresses the MRTA problem from a multimodal, multi-objective, and multi-task perspective. It mines multiple equivalent solutions in the decision space through multimodal optimization and achieves cross-scenario knowledge sharing through multi-task optimization; this provides a reference for multi-robot task allocation in complex environments. Sixteen complex MRTA scenarios are constructed, each with a different number of robot stations, a different number of schedulable robots, and different scales of task points, more realistically reflecting the actual environment of multi-robot operations. Under different scenarios, this solution provides a series of diverse robot task allocation schemes, thus offering decision-makers more high-quality solutions. Based on complex task environments, the multi-robot system can quickly switch between equivalent allocation schemes according to emergencies, maintaining task continuity and efficiently completing tasks in various fields such as industrial logistics and disaster relief.
[0068] Example 3 This embodiment, based on embodiments 1 and 2 above, discloses an effectiveness analysis of the MMOGTDE method and the multi-task optimization method.
[0069] Both the MMOGTDE method and the multi-task optimization method affect the performance of the MMOGTDE-MTO algorithm. Therefore, in this embodiment, 16 MRTA instances are set up, and all algorithms are run independently 22 times in each scenario. The total number of fitness function evaluations is set to 40,000. In this embodiment, the effectiveness of the MMOGTDE method and the multi-task optimization method is analyzed.
[0070] To verify the effectiveness of the gene-targeting-based correction algorithm, the performance of MMOGTDE-MTO was compared with that of a variant without the gene-targeting-based correction algorithm (referred to as MMODE-MTO). Table 4 shows the average HV of the two algorithms across all MRTA instances. As can be seen from Table 4, MMOGTDE-MTO outperforms MMODE-MTO on all 16 MRTA instances. Therefore, we can conclude that the gene-targeting-based correction algorithm plays a crucial role in improving algorithm performance and enhancing the search efficiency for the global optimum.
[0071] Table 4 HV values of MMOGTDE-MTO and MMODE-MTO on different MRTA instances To verify the effectiveness of multi-task optimization, 16 MRTA instances were solved using both MMOGTDE-MTO and a variant without auxiliary tasks (referred to as MMOGTDE-MTO-I). Table 5 lists the HV values obtained by the two algorithms on all MRTA instances. As can be seen from Table 5, the performance of MMOGTDE-MTO is significantly better than MMOGTDE-MTO-I on all MRTA instances, indicating that multi-task optimization can help improve the performance of the proposed algorithm. In multi-task optimization, due to the presence of auxiliary tasks, the proposed algorithm can save resources in a specific search space, greatly improving search efficiency and thus enhancing algorithm performance.
[0072] Table 5 HV values of MMOGTDE-MTO and MMOGTDE-MTO-I on different MRTA instances. In summary, to improve the task allocation adaptability and solution quality of multi-robot systems in complex task environments, this paper proposes a multimodal, multi-objective gene-targeted differential evolution algorithm (MMOGTDE-MTO) based on multi-task optimization. In the proposed algorithm, the "multi-robot task allocation problem without returning to the original base station" is treated as the primary task, while the auxiliary task is the "multi-robot task allocation problem returning to the original base station." A bidirectional knowledge transfer mechanism is used to achieve information interaction and sharing between the primary and auxiliary tasks. Simultaneously, a gene-targeted correction algorithm is introduced to optimize non-dominated solutions, and an event-triggered mechanism is used to dynamically allocate computational resources and control the timing of knowledge transfer. To test the performance of the proposed algorithm, this embodiment presents 16 MRTA scenarios of varying complexity and compares them with four high-performance multimodal, multi-objective optimization algorithms. In addition, the experiments also analyzed the effectiveness of the gene-targeted correction strategy and the multi-task optimization method. The results showed that the gene-targeted correction method can effectively help the proposed algorithm overcome the dimensionality imbalance problem and find higher quality non-dominated solutions, while the multi-task optimization method can significantly improve the search efficiency and solution diversity of the proposed algorithm in complex MRTA scenarios, and significantly improve the planning ability of the proposed algorithm in complex three-dimensional environments.
[0073] Example 4 This embodiment, based on Embodiment 1 above, discloses a multi-robot task allocation system based on multi-task optimization. The system includes a model building module, an algorithm building module, and a multi-task collaborative allocation module; wherein... The model building module is used to construct a mathematical model for multi-robot task allocation, and to define the optimization objectives and constraints for the total path length and the longest path. The algorithm building module is used to build a multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization and event triggering mechanisms. The main task is set as a multi-robot task allocation problem that can be solved without returning to the original base station, and the auxiliary task is set as a multi-robot task allocation problem that returns to the original base station. The multi-task collaborative allocation module is used to complete the multi-robot task allocation based on the mathematical model built by the model building module and the multi-modal, multi-objective gene-targeted differential evolution algorithm built by the algorithm building module, which is based on multi-task optimization and event triggering mechanism. Among them, the improved differential evolution algorithm realizes information sharing between the main and auxiliary tasks of multiple robots through knowledge transfer, combines the gene-targeted correction algorithm to optimize the non-dominated solution of robot task allocation, and uses an event triggering mechanism to allocate computing resources to control knowledge transfer and auxiliary task optimization.
[0074] For details regarding the above modules, please refer to the relevant descriptions and effects in Example 1 for further understanding.
[0075] Example 5 Based on Embodiment 1, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the multi-robot task allocation method based on multi-task optimization as described above.
[0076] At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the multi-robot task allocation method based on multi-task optimization described above. Of course, in addition to software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.
[0077] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0078] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A multi-robot task allocation method based on multi-task optimization, characterized in that, The method includes: A mathematical model for multi-robot task allocation is constructed, wherein the optimization objective of the mathematical model is to minimize the total path length of the scheduled robots performing tasks and to minimize the overall task completion time. Based on the mathematical model for multi-robot task allocation, an improved differential evolution algorithm is used to allocate tasks among the multiple robots; wherein, The improved differential evolution algorithm is a multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization and event triggering mechanism. The tasks include a main task and an auxiliary task. The main task is a multi-robot task allocation problem that can be solved without returning to the original base station, and the auxiliary task is a multi-robot task allocation problem that returns to the base station. The improved differential evolution algorithm achieves information sharing between the main and auxiliary tasks of multiple robots through knowledge transfer, combines a gene-targeting-based correction algorithm to optimize the non-dominated solution of robot task allocation, and adopts an event-triggered mechanism to allocate computing resources to control knowledge transfer and auxiliary task optimization.
2. The multi-robot task allocation method based on multi-task optimization according to claim 1, characterized in that, The mathematical model for multi-robot task allocation is expressed as follows: in, The system efficiency objective is the total path length. The goal is to achieve the optimal time, i.e., the longest path. For the first A sequence of tasks executed by a robot; For the task point and mission points The Euclidean distance between them; This is a binary variable, representing whether there is a direct route between task points; The set of all tasks; To be assigned to the A set of tasks for a robot; , Number the robot base station; To return the length of the base station number; A binary variable representing a base station Is there a direct route between this point and task point V? A binary variable representing a task point. With base station Is there a direct route between them? 3. The multi-robot task allocation method based on multi-task optimization according to claim 1, characterized in that, The specific process of the improved differential evolution algorithm includes: Step A1: Based on the actual task and the robot, establish a main task population and an auxiliary task population; perform individual encoding and decoding for the main task population and the auxiliary task population; initialize population parameters, the main task population, and the auxiliary task population; the population parameters include the population size. and maximum number of iterations ; Step A2: A binary tournament selection mechanism is used to select from the primary task population and the secondary task population respectively. A number of superior individuals are used to generate the intermediate population for the main task and the intermediate population for the auxiliary task. Step A3: Based on the intermediate population of the main task, a multimodal multi-objective gene-targeted differential evolution algorithm is used to generate a subpopulation of the main task; based on the intermediate population of the auxiliary task, a multimodal multi-objective evolution algorithm is used to generate a subpopulation of the auxiliary task. Step A4: Calculate the fitness function value and population renewal rate for each subpopulation. If the current population update rate Population update rate less than historical archive If the goal is to generate a migration population through knowledge transfer, the original population, sub-population, and migration population are merged, and a new primary task population or auxiliary task population is obtained through non-dominated CSCD sorting and environmental selection; otherwise, individuals are directly selected from the original population and sub-population to update the population. Step A5: Repeat steps A2-A4 until the maximum number of iterations is reached, and output the non-dominated solution of the main task as the optimal task allocation scheme.
4. The multi-robot task allocation method based on multi-task optimization according to claim 3, characterized in that, In step A1, the main task population includes chromosome segment I and chromosome segment II, where chromosome segment I is used for task allocation, and its dimension is... The chromosome fragment II is the returned base station number, with a length of [missing information]. The auxiliary task population includes task allocation information, and the individual length is [missing information]. .
5. A multi-robot task allocation method based on multi-task optimization according to claim 4, characterized in that, The specific process of encoding and decoding the individual population in the main task includes: The robot is assigned a main task through rounding, the execution order of the main tasks is determined by gene locus sorting, and chromosome segment I is edited based on the main task assignment result and the main task execution order; chromosome segment II is edited based on the specified return base station number; The specific process of encoding and decoding the individual members of the auxiliary task population includes: Auxiliary tasks are assigned to the robot through rounding operations, the execution order of the auxiliary tasks is determined by sorting by gene size, and the genes of the individuals in the auxiliary task population are edited based on the auxiliary task assignment results and the auxiliary task execution order; wherein, in the individuals in the auxiliary task population, the robot returns to the starting base station after completing the task.
6. A multi-robot task allocation method based on multi-task optimization according to claim 4, characterized in that, The knowledge transfer process includes: Knowledge transfer from the main task to the auxiliary task: Delete chromosome segment II from the individual in the main task population, keeping only chromosome segment I; transfer this individual from the main task population to the auxiliary task population; Knowledge transfer from auxiliary task to main task: Combine chromosome segment II of an individual in the auxiliary task population with that of a non-dominant individual in the main task population; transfer this individual from the auxiliary task population to the main task population.
7. A multi-robot task allocation method based on multi-task optimization according to claim 3, characterized in that, In step A3, the process of generating the main task subpopulation using a multimodal, multi-target gene-targeted differential evolution algorithm includes: The location of the bottleneck dimension is determined probabilistically: A Monte Carlo-based detection method is used to randomly sample from the dimension of individuals in the intermediate population of the main task, and the dimensionality of the individuals in the intermediate population of the main task is obtained from a univariate Gaussian distribution. Sampling probability in each dimension And generate a random value from [0,1] for this dimension. If this random value is less than the probability... If so, then this dimension will be taken as the bottleneck dimension for the current individual; Constructing homologous targeting vectors: Constructing homologous targeting vectors for individuals in the intermediate population of the main task using the DE / best / 1 mutation strategy or the DE / r-best / 1 mutation strategy; Insert the homologous target vector into the corresponding individual: Replace the bottleneck dimension in the individual of the main task intermediate population with the corresponding dimension in the homologous target vector, while keeping other dimensions in the individual unchanged, and generate the main task sub-population.
8. A multi-robot task allocation method based on multi-task optimization according to claim 3, characterized in that, The process of generating the auxiliary task subpopulation using a multimodal, multi-objective evolutionary algorithm includes: Based on the pre-trained deep reinforcement learning model, the fitness function values of all individuals in the intermediate population of the auxiliary task are calculated, and the individuals in the population are sorted using the non-dominated CSCD sorting method. The DBESM method is used to select role model individuals from the intermediate population of the auxiliary task, and the differential evolution algorithm is combined to generate a mutation and crossover operation to generate the offspring population. Based on each individual in the offspring population, the task execution path of the corresponding robot is planned using an end-to-end DRL model; Update the offspring population, and when the preset number of iterations approaches the maximum value, extract all non-dominated solutions in the population to form a solution set PS; The individuals in the solution set PS are decomposed into sub-individuals corresponding to each robot. The number of intersections of the task paths of each robot is calculated. If there are path intersections, the large neighborhood search algorithm is used to optimize the path of the corresponding robot. The optimized solution set PS is updated and merged with the descendant population. The merged population is then sorted again using the non-dominated CSCD sorting method. Redundant individuals in the merged population are removed by a redundancy removal method, and then NP individuals are selected from them by an environment selection method to form a new population for the next iteration. Finally, the latest solution set PS and Pareto front PF in the population are output as results. The Pareto front PF is the mapping set of all non-dominated solutions in the multi-robot task assignment scenario.
9. A multi-robot task allocation method based on multi-task optimization according to claim 1, characterized in that, The triggering conditions for the event triggering mechanism include: When the population renewal rate of the main task population When the population update rate decreases, auxiliary task optimization and knowledge transfer are triggered; when the population update rate... When the number of tasks remains unchanged or increases, pause auxiliary task optimization to save computing resources; The population renewal rate The calculation expression is: in, It refers to population size. It represents the number of individuals updated in the current generation.
10. A multi-robot task allocation system based on multi-task optimization, characterized in that, The system includes a model building module, an algorithm building module, and a multi-task collaborative allocation module; The model building module is used to construct a mathematical model for multi-robot task allocation, and to define the optimization objectives and constraints for the total path length and the longest path. The algorithm construction module is used to construct a multimodal, multi-target gene-targeted differential evolution algorithm based on multi-task optimization and event triggering mechanism. The main task is set as a multi-robot task allocation problem that can be solved without returning to the original base station, and the auxiliary task is set as a multi-robot task allocation problem that returns to the original base station. The multi-task collaborative allocation module is used to complete the multi-robot task allocation based on the mathematical model constructed by the model construction module and the multi-modal multi-objective gene-targeted differential evolution algorithm constructed by the algorithm construction module based on multi-task optimization and event triggering mechanism. Among them, the improved differential evolution algorithm realizes information sharing between the main task and auxiliary task of multiple robots through knowledge transfer, combines the gene-targeted correction algorithm to optimize the non-dominated solution of robot task allocation, and uses an event triggering mechanism to allocate computing resources to control knowledge transfer and auxiliary task optimization.
Citation Information
Patent Citations
Two-stage evolutionary multi-robot task allocation method in nuclear accident rescue and electronic equipment
CN120146490A
