An unmanned delivery task allocation method based on multi-agent cooperation

Abstract

The application discloses an unmanned distribution task allocation method based on multi-agent cooperation, and specifically comprises the following steps: step 1, obtaining distribution task data and unmanned distribution body running data; step 2, constructing a multi-agent state vector set; step 3, generating a task cost disturbance variable set and an execution time delay disturbance variable set; step 4, establishing a disturbance variable parameter coupling relationship, constructing a non-tensor polynomial chaos expansion model, and generating a disturbance propagation risk vector; step 5, constructing a parameter coupling manifold based on the expansion coefficient, and screening candidate parameters by using a DIRECT algorithm; and step 6, generating a task allocation score matrix based on the candidate parameters and the disturbance propagation risk vector, and outputting a multi-agent task matching result. The application introduces parameter coupling modeling to improve the stability of cooperative allocation.

Description

Technical Field

[0001] This invention relates to the field of multi-agent unmanned delivery task allocation technology, and in particular to an unmanned delivery task allocation method based on multi-agent collaboration. Background Technology

[0002] With the development of unmanned delivery technology, the collaborative completion of large-scale delivery tasks by multiple unmanned delivery vehicles has become a common application scenario. In actual operation, delivery tasks are typically characterized by a large number of tasks, dispersed distribution, and complex time window constraints. Furthermore, the unmanned delivery vehicles are affected by various factors such as battery status, payload capacity, travel path, and environmental changes when executing tasks. Therefore, achieving a reasonable match between delivery tasks and delivery vehicles in multi-unmanned delivery vehicle collaborative scenarios has become a key technical problem in unmanned delivery systems.

[0003] Existing unmanned delivery task allocation methods mostly employ rule-based allocation strategies or centralized optimization models, making task allocation decisions by simply modeling task distance, time constraints, or load conditions. However, these methods typically assume that task costs and execution delays are deterministic, making it difficult to characterize the uncertainties introduced during delivery by path changes, task conflicts, or execution deviations. In scenarios where multiple unmanned delivery vehicles execute tasks in parallel, these uncertainties can easily accumulate and propagate within the system, leading to insufficient stability in the allocation results.

[0004] Furthermore, some methods attempt to introduce global optimization algorithms to search for task allocation parameters. However, when the parameter dimensionality is high or there are correlations between parameters, the search space expands rapidly, and the computational complexity increases accordingly, making it difficult to meet the needs of real-time collaborative decision-making among multiple unmanned delivery vehicles. Especially when there are coupling relationships between parameters, traditional independent parameter search strategies cannot effectively utilize parameter structure features, easily leading to a decrease in search efficiency.

[0005] Therefore, there is an urgent need for a multi-agent unmanned delivery task allocation method that can simultaneously characterize the propagation characteristics of task allocation uncertainty and complete efficient search under the premise of considering parameter coupling relationship, so as to improve the rationality and stability of collaborative decision-making among multiple unmanned delivery agents. Summary of the Invention

[0006] One objective of this invention is to propose a multi-agent collaborative unmanned delivery task allocation method. This invention introduces a parameter-coupled sensing non-tensified polynomial chaotic expansion modeling mechanism and a parameter-coupled manifold constraint DIRECT search computation model to jointly model and fold the search space for cost perturbations and execution delay perturbations in multi-agent delivery tasks. Within the coupled manifold space, task allocation parameters are selected and scores are matched to construct a continuously optimized multi-agent collaborative task allocation process. This method has the advantages of strong uncertainty propagation characterization ability, high search efficiency, and stable task allocation results.

[0007] According to an embodiment of the present invention, an unmanned delivery task allocation method based on multi-agent cooperation includes the following:

[0008] Step 1: Acquire delivery task data and unmanned delivery vehicle operation data. Delivery task data includes task location components, time window components, and load components. Unmanned delivery vehicle operation data includes current location components, battery level components, and load capacity components. Step 2: Construct a multi-agent state vector set. The state vector includes current location components, battery level components, load state components, and executable task index components. Step 3: Generate a set of task cost perturbation variables and a set of execution delay perturbation variables. Step 4: Establish parameter coupling relationships between task cost perturbation variables and execution delay perturbation variables. Based on these parameter coupling relationships, construct a non-tensified polynomial chaotic expansion model. Step 5: Generate joint polynomial basis functions and calculate polynomial chaotic expansion coefficient vectors within a defined coupled subspace. Form a perturbation propagation risk vector based on the polynomial chaotic expansion coefficient vectors. Step 6: Generate a parameter correlation matrix based on the polynomial chaotic expansion coefficient vectors. Construct a parameter coupled manifold based on the parameter correlation matrix. Perform manifold mapping and search space folding on the original parameter space. Perform rectangular partitioning and candidate parameter screening using the DIRECT algorithm within the parameter coupled manifold space to form a candidate parameter set. Step 7: Generate a task assignment scoring matrix based on the candidate parameter set and the perturbation propagation risk vector. Output the multi-agent task matching results based on the task assignment scoring matrix.

[0009] Optionally, the delivery task data in step 1 includes task location components, time window components, and load components, specifically including:

[0010] Read the set of delivery task identifiers, read the corresponding geographic coordinate values ​​for each delivery task identifier, and write the geographic coordinate values ​​into the task location component; read the allowed start time and allowed end time values ​​for each delivery task identifier, and write the allowed start time and allowed end time values ​​into the time window component in a fixed order; read the load requirement value for each delivery task identifier, and write the load requirement value into the load component; establish an index mapping relationship between the task location component, time window component, and load component according to the delivery task identifier to form a delivery task data set.

[0011] Optionally, the construction of the multi-agent state vector set in step 2 includes the following steps:

[0012] The system reads the set of unmanned delivery vehicle identifiers, reads the current location coordinates for each identifier, and writes the current location coordinates into the current location component; reads the remaining battery power for each identifier and writes the remaining battery power into the battery power component; reads the current load and maximum load capacity for each identifier, performs a difference calculation on the current load and maximum load capacity, and writes the calculation result into the load status component; reads the delivery task data set, performs an executability judgment operation on each delivery task identifier in the delivery task data set, filters delivery task identifiers that meet the constraints of the load status component, and writes the filtering results into the executable task index component in index order; and concatenates the current location component, battery power component, load status component, and executable task index component according to the unmanned delivery vehicle identifier to form a multi-agent state vector set.

[0013] Optionally, step 3 specifically includes:

[0014] The system reads the delivery task data set and the multi-agent state vector set. It reads the task location component and time window component corresponding to each delivery task identifier, and the current location component and battery component corresponding to each unmanned delivery vehicle identifier. For each unmanned delivery vehicle identifier and each delivery task identifier, it constructs candidate pairing identifiers. It calculates the Euclidean distance between the current location component and the task location component according to the candidate pairing identifiers, and writes the Euclidean distance value into the path distance component. It performs a traffic speed constant division operation on the path distance component, and writes the result into the travel time component. It calculates the difference between the allowed end time value and the current time value in the time window component, and writes the difference calculation result into the remaining time component. It performs a difference calculation between the travel time component and the remaining time component, and writes the difference calculation result into the time window margin component. It performs vector concatenation operation on the path distance component and the time window margin component according to the candidate pairing identifiers to form a task cost perturbation variable set. It performs vector concatenation operation on the travel time component and the time window margin component according to the candidate pairing identifiers to form an execution delay perturbation variable set.

[0015] Optionally, step 4, constructing a non-tensified polynomial chaotic expansion model between the task cost perturbation variable and the execution delay perturbation variable, includes the following steps: reading the set of task cost perturbation variables and the set of execution delay perturbation variables; for each candidate pairing, reading the corresponding path distance component, time window margin component, and travel time component; performing numerical standardization on the path distance component, time window margin component, and travel time component to form a standardized perturbation variable vector; performing parameter correlation calculation on the standardized perturbation variable vector to generate a perturbation variable correlation matrix; and filtering existing pairs based on the perturbation variable correlation matrix. Non-zero correlation perturbation variable index pairs form a parameter coupling index set; within the scope of the parameter coupling index set, a non-tensified polynomial index structure is constructed, where each index term corresponds to a set of coupled perturbation variables; a joint polynomial basis function set is generated according to the polynomial index structure, which only contains the perturbation variable components corresponding to the parameter coupling index set; sampling point substitution operations are performed on the joint polynomial basis function set to construct a polynomial chaotic expansion linear equation system; coefficient solving operations are performed on the polynomial chaotic expansion linear equation system, and the solution is written into the polynomial chaotic expansion coefficient vector.

[0016] Optionally, step 4, forming a disturbance propagation risk vector based on the polynomial chaotic expansion coefficient vector, includes the following steps: reading the polynomial chaotic expansion coefficient vector; performing position analysis operations on each coefficient component in the coefficient vector according to the corresponding joint polynomial basis function index; calculating the square value of the coefficient component corresponding to each joint polynomial basis function, and writing the square value into the coefficient energy component; performing group accumulation operations on the coefficient energy component according to the parameter coupling index set to form a coupled disturbance energy vector; performing normalization operations on the coupled disturbance energy vector, and writing the normalization result into the disturbance influence weight vector; reading the task cost disturbance variable set and the execution delay disturbance variable set, and performing a weighted combination operation on the disturbance influence weight vector and the corresponding disturbance variable component; writing the weighted combination operation result into the disturbance propagation risk vector according to the candidate pairing identifier pair; and establishing a candidate pairing identifier index mapping relationship for the disturbance propagation risk vector to form a disturbance propagation risk vector set.

[0017] Optionally, step 5, generating the parameter correlation matrix and constructing the parameter coupling manifold based on the polynomial chaotic expansion coefficient vector, includes the following steps: reading the polynomial chaotic expansion coefficient vector; performing parameter position analysis on each coefficient component in the coefficient vector according to the corresponding joint polynomial basis function index to form a parameter participation identifier set; counting the number of times any two parameter indices in the parameter participation identifier set appear simultaneously in the corresponding joint polynomial basis function, and writing the statistical result into the parameter co-occurrence count matrix; performing symmetric normalization on the parameter co-occurrence count matrix, and writing the normalization result into the parameter correlation matrix; performing threshold comparison on each matrix element in the parameter correlation matrix, filtering parameter index pairs that meet the threshold conditions, and forming a parameter coupling index set; constructing a parameter coupling relationship graph structure based on the parameter coupling index set, with parameter indices as node identifiers and parameter coupling index pairs as edge identifiers; performing connected component analysis on the parameter coupling relationship graph structure to generate several parameter coupling subsets; defining a local parameter coordinate system in each parameter coupling subset, performing dimension reduction and embedding operations on the local parameter coordinate system to generate a parameter coupling manifold representation; and writing each parameter coupling manifold representation into the parameter coupling manifold set according to the parameter index.

[0018] Optionally, step 5, which involves using the DIRECT algorithm to perform rectangular partitioning and candidate parameter filtering within the parameter-coupled manifold space, includes the following steps:

[0019] Read the parameter value ranges corresponding to each parameter index in the original parameter space to form a set of original parameter ranges;

[0020] Based on the manifold representation of each parameter coupling subset in the set of parameter coupling manifolds, perform a manifold mapping operation on the original set of parameter intervals to map the original parameter intervals to the parameter coupling manifold space;

[0021] Perform dimension compression and interval merging operations on the mapped parameter intervals to form a folded manifold parameter interval set;

[0022] Initialize the DIRECT algorithm search structure within the set of manifold parameter intervals. The search structure includes the initial hyperrectangular region identifier.

[0023] For each hyperrectangular region, calculate the region center parameter vector, and perform objective function sampling operation on the region center parameter vector;

[0024] Based on the DIRECT algorithm rules, perform an optimality judgment operation on the hyperrectangular region and filter out hyperrectangular regions that meet the judgment conditions;

[0025] For hyperrectangular regions that meet the criteria, perform proportional rectangular segmentation operations along the coordinate axes of the parametrically coupled manifold to generate a set of subrectangular regions.

[0026] Repeat the region center sampling and optimality determination operation for each sub-rectangular region to filter and form a set of candidate parameters.

[0027] Optionally, step 6 includes the following steps:

[0028] Read the candidate parameter set and the disturbance propagation risk vector set. Construct a scoring unit identifier for each unmanned delivery vehicle identifier and each delivery task identifier. Read the corresponding candidate parameter vector components and disturbance propagation risk vector components according to the scoring unit identifier. Perform a weighted summation operation on the candidate parameter vector components and a combination operation with the disturbance propagation risk vector components to generate a task allocation score value. Write the task allocation score value into a two-dimensional scoring matrix according to the unmanned delivery vehicle identifier and the delivery task identifier to form a task allocation score matrix. Perform an optimal value selection operation on the task allocation score matrix according to the unmanned delivery vehicle identifier, record the delivery task identifier corresponding to the optimal value, and establish a matching mapping relationship between the delivery task identifier and the unmanned delivery vehicle identifier to form a multi-agent task matching result and output it.

[0029] The beneficial effects of this invention are:

[0030] (1) This invention establishes a parameter coupling relationship between the task cost disturbance variable and the execution delay disturbance variable, and introduces a non-tensified polynomial chaotic expansion model to jointly model the uncertainty disturbance in the multi-agent unmanned delivery task allocation process. This enables the disturbance propagation characteristics to be explicitly represented in the form of coefficient vectors, thereby avoiding information loss caused by treating the task cost and execution delay as independent variables and enhancing the ability to characterize the disturbance accumulation and propagation structure in the task allocation process.

[0031] (2) The present invention constructs a parameter correlation matrix based on the coefficient vector of polynomial chaos expansion and further forms a parameter coupling manifold. On this basis, manifold mapping and search space folding are performed on the original parameter space, so that the parameter search process can be carried out in a low-dimensional space that preserves the parameter coupling structure, effectively reducing the size of the parameter space and reducing redundant computation in the search process. It is suitable for task allocation problems with high parameter dimensions in multi-agent collaborative scenarios.

[0032] (3) This invention introduces the DIRECT algorithm to perform rectangular partitioning and candidate parameter screening in the parameter-coupled manifold space, and generates a task allocation scoring matrix by combining the perturbation propagation risk vector, so as to realize the matching decision between multiple agents and delivery tasks, and complete the task allocation process under a unified scoring framework, thereby enhancing the consistency and stability of the task allocation results of multiple agents, and making it suitable for continuous execution in a multi-unmanned delivery vehicle parallel operation environment. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0034] Figure 1 This is a flowchart of a method for allocating unmanned delivery tasks based on multi-agent collaboration proposed in this invention.

[0035] Figure 2 This is a schematic diagram of the non-tensified polynomial chaotic expansion process of an unmanned delivery task allocation method based on multi-agent cooperation proposed in this invention.

[0036] Figure 3 This is a schematic diagram of the parameter coupling manifold and DIRECT search process of an unmanned delivery task allocation method based on multi-agent cooperation proposed in this invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0038] refer to Figures 1-3 A method for allocating unmanned delivery tasks based on multi-agent collaboration includes the following steps:

[0039] Step 1: Acquire delivery task data and unmanned delivery vehicle operation data. Delivery task data includes task location components, time window components, and load components. Unmanned delivery vehicle operation data includes current location components, battery level components, and load capacity components. Step 2: Construct a multi-agent state vector set. The state vector includes current location components, battery level components, load state components, and executable task index components. Step 3: Generate a set of task cost perturbation variables and a set of execution delay perturbation variables. Step 4: Establish parameter coupling relationships between task cost perturbation variables and execution delay perturbation variables. Based on these parameter coupling relationships, construct a non-tensified polynomial chaotic expansion model. Step 5: Generate joint polynomial basis functions and calculate polynomial chaotic expansion coefficient vectors within a defined coupled subspace. Form a perturbation propagation risk vector based on the polynomial chaotic expansion coefficient vectors. Step 6: Generate a parameter correlation matrix based on the polynomial chaotic expansion coefficient vectors. Construct a parameter coupled manifold based on the parameter correlation matrix. Perform manifold mapping and search space folding on the original parameter space. Perform rectangular partitioning and candidate parameter screening using the DIRECT algorithm within the parameter coupled manifold space to form a candidate parameter set. Step 7: Generate a task assignment scoring matrix based on the candidate parameter set and the perturbation propagation risk vector. Output the multi-agent task matching results based on the task assignment scoring matrix.

[0040] In this embodiment, the delivery task data in step 1 includes task location component, time window component, and load component, specifically including:

[0041] Read the set of delivery task identifiers, read the corresponding geographic coordinate values ​​for each delivery task identifier, and write the geographic coordinate values ​​into the task location component; read the allowed start time and allowed end time values ​​for each delivery task identifier, and write the allowed start time and allowed end time values ​​into the time window component in a fixed order; read the load requirement value for each delivery task identifier, and write the load requirement value into the load component; establish an index mapping relationship between the task location component, time window component, and load component according to the delivery task identifier to form a delivery task data set.

[0042] In this embodiment, step 2, constructing the multi-agent state vector set, includes the following steps:

[0043] The system reads the set of unmanned delivery vehicle identifiers, reads the current location coordinates for each identifier, and writes the current location coordinates into the current location component; reads the remaining battery power for each identifier and writes the remaining battery power into the battery power component; reads the current load and maximum load capacity for each identifier, performs a difference calculation on the current load and maximum load capacity, and writes the calculation result into the load status component; reads the delivery task data set, performs an executability judgment operation on each delivery task identifier in the delivery task data set, filters delivery task identifiers that meet the constraints of the load status component, and writes the filtering results into the executable task index component in index order; and concatenates the current location component, battery power component, load status component, and executable task index component according to the unmanned delivery vehicle identifier to form a multi-agent state vector set.

[0044] In this embodiment, the feasibility assessment calculation in step 2 specifically includes:

[0045] Read the load status component corresponding to the current unmanned delivery vehicle. The load status component is calculated from the difference between the maximum load capacity value and the current load value. Read the load component corresponding to the current delivery task identifier. The load component represents the load requirement value of the delivery task. Perform a size comparison operation between the load status component value and the load component value. When the load status component value is not less than the load component value, generate a load executable mark. When the load status component value is less than the load component value, generate a load non-executable mark. Only write the executable task index component to the delivery task identifier that generates the load executable mark, and do not write the executable task index component to the delivery task identifier that generates the load non-executable mark.

[0046] In this embodiment, step 3 specifically includes:

[0047] The system reads the delivery task data set and the multi-agent state vector set. It reads the task location component and time window component corresponding to each delivery task identifier, and the current location component and battery component corresponding to each unmanned delivery vehicle identifier. For each unmanned delivery vehicle identifier and each delivery task identifier, it constructs candidate pairing identifiers. It calculates the Euclidean distance between the current location component and the task location component according to the candidate pairing identifiers, and writes the Euclidean distance value into the path distance component. It performs a traffic speed constant division operation on the path distance component, and writes the result into the travel time component. It calculates the difference between the allowed end time value and the current time value in the time window component, and writes the difference calculation result into the remaining time component. It performs a difference calculation between the travel time component and the remaining time component, and writes the difference calculation result into the time window margin component. It performs vector concatenation operation on the path distance component and the time window margin component according to the candidate pairing identifiers to form a task cost perturbation variable set. It performs vector concatenation operation on the travel time component and the time window margin component according to the candidate pairing identifiers to form an execution delay perturbation variable set.

[0048] In this embodiment, step 3, performing the traffic speed constant division operation, specifically includes:

[0049] Read the path distance component value corresponding to the candidate pairing identifier. The path distance component value is calculated from the distance between the current position component and the task position component. Read the traffic speed constant value from the preset parameter area. The traffic speed constant value remains constant throughout the calculation process. Perform a division operation using the path distance component value as the dividend and the traffic speed constant value as the divisor to obtain the corresponding travel time calculation result. When the path distance component value is zero, the travel time calculation result is recorded as zero. Write the travel time calculation result obtained from the division operation into the travel time component and establish an index mapping relationship according to the candidate pairing identifier.

[0050] In this embodiment, step 4, constructing a non-tensified polynomial chaotic expansion model between the task cost perturbation variable and the execution delay perturbation variable, includes the following steps: reading the set of task cost perturbation variables and the set of execution delay perturbation variables; for each candidate pairing, reading the corresponding path distance component, window margin component, and travel time component; performing numerical standardization operations on the path distance component, window margin component, and travel time component to form a standardized perturbation variable vector; performing parameter correlation calculation operations on the standardized perturbation variable vector to generate a perturbation variable correlation matrix; and filtering existing pairs based on the perturbation variable correlation matrix. Non-zero correlation perturbation variable index pairs form a parameter coupling index set; within the scope of the parameter coupling index set, a non-tensified polynomial index structure is constructed, where each index term corresponds to a set of coupled perturbation variables; a joint polynomial basis function set is generated according to the polynomial index structure, which only contains the perturbation variable components corresponding to the parameter coupling index set; sampling point substitution operations are performed on the joint polynomial basis function set to construct a polynomial chaotic expansion linear equation system; coefficient solving operations are performed on the polynomial chaotic expansion linear equation system, and the solution is written into the polynomial chaotic expansion coefficient vector.

[0051] In this embodiment, step 4, the calculation of the correlation between execution parameters, specifically includes:

[0052] Read the standardized perturbation variable vector corresponding to each candidate pair identifier. The standardized perturbation variable vector contains standardized values ​​corresponding to the path distance component, time window margin component, and travel time component. Perform pairwise combination on each perturbation variable component in the standardized perturbation variable vector according to the component index order to form a set of perturbation variable component index pairs. For each perturbation variable component index pair, read the corresponding two sets of standardized value sequences along the dimension of the candidate pair identifier. Perform mean calculation on the two sets of standardized value sequences to obtain the mean value of the corresponding component. Perform difference calculation on each standardized value and the corresponding mean value to form a mean-reduced value sequence. Perform element-wise product operation on the two sets of mean-reduced value sequences and perform accumulation operation on the product results to obtain the co-variance value. Perform squaring and accumulation operations on each mean-reduced value sequence to obtain the normalization factor value of the corresponding component. Perform division operation on the co-variance value and the corresponding normalization factor value to obtain the correlation value corresponding to the perturbation variable component index pair. Write the correlation value into the corresponding position of the matrix according to the perturbation variable component index pair to form the perturbation variable correlation matrix.

[0053] In this embodiment, step 4, constructing the non-tensified polynomial index structure, specifically includes:

[0054] The parameter coupling index set is read, where each index entry represents a set of correlated perturbation variable component indices. For each index entry in the parameter coupling index set, a corresponding index subset is generated according to the number of perturbation variable component indices contained in the index entry. The perturbation variable component indices in each index subset are arranged in a fixed order to form a perturbation variable combination index. The perturbation variable combination index is used as an index entry in a polynomial index structure, and the corresponding perturbation variable component index set is recorded in the index entry. The above generation and recording process is repeated for all index entries in the parameter coupling index set to form a polynomial index structure composed of several index entries. In the polynomial index structure, each index entry only contains the perturbation variable component indices corresponding to the same parameter coupling index entry, and does not contain perturbation variable component indices outside the parameter coupling index set. Through the above index entry generation and recording process, the construction of the non-tensified polynomial index structure is completed.

[0055] In this embodiment, step 4, generating the set of joint polynomial basis functions, specifically includes:

[0056] The process involves reading a polynomial index structure, which consists of several index entries, each recording a set of perturbation variable component indices. For each index entry, the set of perturbation variable component indices recorded in the index entry is read. Based on the perturbation variable component indices, the corresponding perturbation variable component values ​​are sequentially read from the standardized perturbation variable vector. A component-wise mapping operation is performed on each read perturbation variable component value according to the order recorded in the index entry, generating a corresponding set of component mapping results. The component mapping results within the same index entry are combined to form a joint polynomial basis function. The generated joint polynomial basis function is written into a joint polynomial basis function set, and its correspondence with the current index entry is recorded in the set. This reading, mapping, and combination process is repeated for all index entries in the polynomial index structure to form a joint polynomial basis function set. In the joint polynomial basis function set, each joint polynomial basis function only contains the perturbation variable components corresponding to the parameter coupling index set, and does not contain perturbation variable components outside the parameter coupling index set.

[0057] In this embodiment, step 4, performing the sampling point substitution operation on the set of joint polynomial basis functions, specifically includes:

[0058] Read the polynomial index term structure corresponding to each joint polynomial basis function in the joint polynomial basis function set, determine the perturbation variable components contained in the joint polynomial basis function and the power values ​​corresponding to each perturbation variable component; read the standardized perturbation variable vector corresponding to each sampling point along the sampling point index order, extract the perturbation variable component values ​​that match the current joint polynomial basis function index term from the standardized perturbation variable vector; replace the corresponding variable symbols with the extracted perturbation variable component values ​​according to the positional relationship in the joint polynomial basis function index structure, and complete the numerical substitution of the sampling points into the joint polynomial basis function; After the numerical substitution is completed, exponentiation is performed on each perturbation variable component in the joint polynomial basis function, and the exponentiation value is determined by the exponent parameter recorded in the polynomial index structure. The exponentiation results of each perturbation variable component belonging to the same polynomial term are multiplied to generate the term value corresponding to the polynomial term. The term values ​​of each polynomial term contained in the joint polynomial basis function are accumulated, and the accumulated result is used as the function value of the joint polynomial basis function under the current sampling point condition. The function values ​​of the joint polynomial basis function obtained under each sampling point condition are written into the calculation result vector in the order of the sampling point index.

[0059] In this embodiment, step 4, which involves performing coefficient calculations on the polynomial chaotic expansion of the linear equation system, specifically includes:

[0060] Read the set of joint polynomial basis functions and determine the index order of each basis function in the set; read the perturbation variable values ​​corresponding to each sampling point, and perform sampling point substitution and numerical calculation operations sequentially on the set of joint polynomial basis functions under the condition of each sampling point to generate the basis function value vectors for the corresponding sampling points; write the basis function value vectors corresponding to each sampling point into the coefficient matrix row by row according to the sampling point index order, with each column of the coefficient matrix maintaining the same index order as the joint polynomial basis functions; read the perturbation response values ​​corresponding to each sampling point and write the perturbation response values ​​into the response vector according to the sampling point index order; and write the coefficients... The matrix and response vector are combined to form a polynomial chaotic expansion linear equation system; a row and column consistency check operation is performed on the coefficient matrix to confirm that the number of rows in the coefficient matrix is ​​consistent with the length of the response vector; a numerical stability preprocessing operation is performed on the coefficient matrix, which includes scaling normalization of each column of the matrix; after the preprocessing is completed, a least squares solution operation is performed on the polynomial chaotic expansion linear equation system to generate a polynomial chaotic expansion coefficient vector; the solved polynomial chaotic expansion coefficient vector is written into the coefficient storage area in the order of the joint polynomial basis function index, and a correspondence between the coefficient components and the joint polynomial basis function index terms is established.

[0061] In this embodiment, step 4, forming a disturbance propagation risk vector based on the polynomial chaotic expansion coefficient vector, includes the following steps: reading the polynomial chaotic expansion coefficient vector; performing position analysis operations on each coefficient component in the coefficient vector according to the corresponding joint polynomial basis function index; calculating the square value of the coefficient component corresponding to each joint polynomial basis function, and writing the square value into the coefficient energy component; performing group accumulation operations on the coefficient energy component according to the parameter coupling index set to form a coupled disturbance energy vector; performing a normalization operation on the coupled disturbance energy vector, and writing the normalization result into the disturbance influence weight vector; reading the task cost disturbance variable set and the execution delay disturbance variable set, and performing a weighted combination operation on the disturbance influence weight vector and the corresponding disturbance variable component; writing the weighted combination operation result into the disturbance propagation risk vector according to the candidate pairing identifier pair; and establishing a candidate pairing identifier index mapping relationship for the disturbance propagation risk vector to form a disturbance propagation risk vector set.

[0062] In this embodiment, step 4, performing positional analysis on each coefficient component in the coefficient vector according to the corresponding joint polynomial basis function index, specifically includes:

[0063] Establish a positional correspondence between each coefficient component and the perturbation variable index and power parameter contained in the corresponding joint polynomial basis function, according to the coefficient vector index order.

[0064] In this embodiment, step 5, generating the parameter correlation matrix and constructing the parameter coupling manifold based on the polynomial chaotic expansion coefficient vector, includes the following steps: reading the polynomial chaotic expansion coefficient vector; performing parameter position analysis on each coefficient component in the coefficient vector according to the corresponding joint polynomial basis function index to form a parameter participation identifier set; counting the number of times any two parameter indices in the parameter participation identifier set appear simultaneously in the corresponding joint polynomial basis function, and writing the statistical result into the parameter co-occurrence count matrix; performing symmetric normalization on the parameter co-occurrence count matrix, and writing the normalization result into the parameter correlation matrix; performing threshold comparison on each matrix element in the parameter correlation matrix, filtering parameter index pairs that meet the threshold conditions, and forming a parameter coupling index set; constructing a parameter coupling relationship graph structure based on the parameter coupling index set, with parameter indices as node identifiers and parameter coupling index pairs as edge identifiers; performing connected component analysis on the parameter coupling relationship graph structure to generate several parameter coupling subsets; defining a local parameter coordinate system within each parameter coupling subset, performing dimension reduction and embedding operations on the local parameter coordinate system to generate a parameter coupling manifold representation; and writing each parameter coupling manifold representation into the parameter coupling manifold set according to the parameter index.

[0065] In this embodiment, step 5, performing the symmetric normalization operation, specifically includes:

[0066] Read the parameter co-occurrence count matrix, and calculate the total count value of each parameter index in the corresponding row direction and column direction. Read the co-occurrence count values ​​in the parameter co-occurrence count matrix one by one in matrix position order, and simultaneously read the total count value in the row direction and the total count value in the column direction of the co-occurrence count value. Divide the co-occurrence count value by the product of the square roots of the total count value in the corresponding row direction and the total count value in the column direction, respectively, to generate normalized values. Write the normalized values ​​into the matrix cells with the same positions as the original parameter co-occurrence count matrix to form the parameter correlation matrix.

[0067] In this embodiment, step 5, performing the threshold comparison operation, specifically includes:

[0068] Read the parameter correlation matrix and the preset correlation threshold value; read the values ​​of each matrix element in the parameter correlation matrix one by one according to the matrix storage order; perform a numerical comparison operation between the read matrix element values ​​and the correlation threshold values; for matrix elements that meet the threshold comparison conditions, read the row direction parameter index and column direction parameter index corresponding to the matrix element to form a set of parameter index pairs; write each parameter index pair into the parameter coupling index set in the generation order to form the index result of parameter coupling relationship.

[0069] In this embodiment, step 5, constructing the parameter coupling relationship diagram structure, specifically includes:

[0070] Read the parameter coupling index set and extract all parameter index values ​​appearing in the parameter coupling index set; perform deduplication processing on each extracted parameter index value to generate a parameter node identifier set; generate corresponding node record items according to each parameter index value in the parameter node identifier set, and write each node record item into the graph structure node storage area; read each set of parameter index pairs one by one in the order of the index pairs in the parameter coupling index set; for each set of parameter index pairs, locate the corresponding two parameter node record items in the graph structure node storage area; generate a connection relationship record between the corresponding two parameter node record items, and write the connection relationship record into the graph structure edge storage area; combine the parameter node record items in the node storage area with the connection relationship record in the edge storage area to form a parameter coupling relationship graph structure.

[0071] In this embodiment, step 5, performing the connected component parsing operation, specifically includes:

[0072] Read the parameter coupling graph structure and all parameter node records and connection records in the graph structure; mark all parameter node records as unvisited and initialize the connected component record structure; read each parameter node record in the order of its storage; for parameter node records in the unvisited state, read the connection records directly associated with that parameter node record; add the parameter node records reachable through the connection records to the current connected component record structure in sequence, and mark the added parameter node records as visited; repeat the connection reading and node marking operations for newly added parameter node records in the current connected component record structure until no new parameter node records are generated; write the expanded connected component record structure to the parameter coupling subset storage area; continue reading the next parameter node record in the unvisited state and repeat the above connected component parsing process until all parameter node records are marked as visited; output the set of parameter nodes recorded in the parameter coupling subset storage area as the parameter coupling subset set.

[0073] In this embodiment, step 5, performing dimension reduction embedding operations to generate a parameter-coupled manifold representation, specifically includes:

[0074] Read the parameter indices contained in the parameter coupling subset and construct a parameter coordinate vector set according to the parameter index order; perform decentralization processing on each parameter coordinate vector in the parameter coordinate vector set to form a decentralized parameter coordinate set; perform cooperative change statistical operation on the decentralized parameter coordinate set to generate a parameter cooperative change matrix; perform eigenvalue decomposition operation on the parameter cooperative change matrix to generate a feature value sequence and a corresponding feature direction sequence; eigenvalue decomposition is a matrix decomposition operation performed on the parameter cooperative change matrix to extract the corresponding feature value sequence and feature direction sequence and establish a one-to-one correspondence; select a preset number of feature direction sequences from the sorted result of the feature value sequence; perform projection operation on the parameter coordinate vector set along the selected feature direction sequence to generate an embedded parameter representation; write the embedded parameter representation into the parameter coupling manifold storage structure according to the parameter index order to form a parameter coupling manifold representation.

[0075] In this embodiment, step 5, which involves using the DIRECT algorithm to perform rectangular partitioning and candidate parameter filtering within the parameter-coupled manifold space, includes the following steps:

[0076] Read the parameter value ranges corresponding to each parameter index in the original parameter space to form a set of original parameter ranges;

[0077] Based on the manifold representation of each parameter coupling subset in the set of parameter coupling manifolds, perform a manifold mapping operation on the original set of parameter intervals to map the original parameter intervals to the parameter coupling manifold space;

[0078] Perform dimension compression and interval merging operations on the mapped parameter intervals to form a folded manifold parameter interval set;

[0079] Initialize the DIRECT algorithm search structure within the set of manifold parameter intervals. The search structure includes the initial hyperrectangular region identifier.

[0080] For each hyperrectangular region, calculate the region center parameter vector, and perform objective function sampling operation on the region center parameter vector;

[0081] Based on the DIRECT algorithm rules, perform an optimality judgment operation on the hyperrectangular region and filter out hyperrectangular regions that meet the judgment conditions;

[0082] For hyperrectangular regions that meet the criteria, perform proportional rectangular segmentation operations along the coordinate axes of the parametrically coupled manifold to generate a set of subrectangular regions.

[0083] Repeat the region center sampling and optimality determination operation for each sub-rectangular region to filter and form a set of candidate parameters.

[0084] In this embodiment, step 5, performing the manifold mapping operation, specifically includes:

[0085] Read the parameter value vectors corresponding to each parameter index in the original parameter space and construct the original parameter vector set according to the parameter index order; read the embedded parameter representations recorded in the parameter-coupled manifold representation and establish the correspondence between the original parameter indexes and the embedded parameter indices; for each parameter vector in the original parameter vector set, extract the embedded parameter representation associated with it according to the correspondence; perform combination operations on the parameter components belonging to the same parameter coupling subset in the original parameter vectors to generate subset parameter vectors; input the subset parameter vectors into the corresponding embedded parameter representations to perform mapping calculations to generate manifold parameter vectors; write the generated manifold parameter vectors into the manifold parameter storage structure according to the parameter index order to form the manifold mapping result.

[0086] In this embodiment, step 5, which performs an optimality determination operation on the hyperrectangular region according to the DIRECT algorithm rules, specifically includes:

[0087] Read all hyperrectangular region identifiers in the current manifold parameter interval set, and read the region center parameter vector corresponding to each hyperrectangular region; perform objective function numerical calculation on each region center parameter vector to generate corresponding region evaluation values; statistically analyze the region scale information corresponding to each hyperrectangular region, wherein the region scale information consists of the interval length of the hyperrectangular region in each parameter dimension direction; combine the region evaluation values ​​and region scale information to form region determination data pairs; perform grouping and sorting operations on all region determination data pairs according to region scale information; perform extreme value screening operations on the region evaluation values ​​within each scale group to determine candidate hyperrectangular regions in the corresponding scale group; merge the selected hyperrectangular regions in each scale group to form an optimal region set; output the optimal region set as the input region set for subsequent rectangle segmentation operations.

[0088] In this embodiment, step 6 includes the following steps:

[0089] The system reads the candidate parameter set and the disturbance propagation risk vector set. For each unmanned delivery vehicle identifier and each delivery task identifier, a scoring unit identifier is constructed. The corresponding candidate parameter vector components and disturbance propagation risk vector components are read according to the scoring unit identifier. A weighted summation operation is performed on the candidate parameter vector components, and a combination operation is performed with the disturbance propagation risk vector components to generate a task allocation score value. The combination operation involves synthesizing the candidate parameter evaluation values ​​and the corresponding disturbance propagation risk values ​​according to preset numerical combination rules to generate a single task allocation score value. The task allocation score values ​​are written into a two-dimensional scoring matrix according to the unmanned delivery vehicle identifier and the delivery task identifier to form a task allocation score matrix. An optimal value selection operation is performed on the task allocation score matrix according to the unmanned delivery vehicle identifier, recording the delivery task identifier corresponding to the optimal value. A matching mapping relationship is established between the delivery task identifier and the unmanned delivery vehicle identifier to form a multi-agent task matching result, which is then output.

[0090] In this embodiment, step 6, performing the optimal value selection operation, specifically includes:

[0091] Read the task allocation scoring matrix and determine the corresponding score value set according to the unmanned delivery vehicle identifier; for each unmanned delivery vehicle identifier, read the task allocation score value of each delivery task identifier corresponding to it one by one; perform a numerical comparison operation on the score value set according to the preset scoring comparison rules, and determine the extreme value score value that satisfies the comparison rules as the optimal value; read the delivery task identifier corresponding to the optimal value in the task allocation scoring matrix; establish a matching mapping relationship between the delivery task identifier and the corresponding unmanned delivery vehicle identifier to form a multi-agent task matching result and output it.

[0092] Example 1:

[0093] In multi-agent unmanned delivery collaborative operation scenarios, there are often situations involving a dense number of delivery tasks, discrete task distribution, significant differences in task time window constraints, and dynamic changes in the operating status of unmanned delivery vehicles. In such scenarios, if task allocation is based solely on fixed rules or static cost models, problems such as concentrated task loads on some unmanned delivery vehicles, delayed execution of some tasks, or frequent fluctuations in allocation results can easily occur, thus affecting the stability and continuity of the overall operation. This embodiment addresses the above problems by providing a specific application method and a complete description of the multi-agent unmanned delivery task allocation process.

[0094] In this scenario, the system first receives multiple delivery task data entries. Each task data entry includes a corresponding task location component, a time window component, and a load component. Simultaneously, it collects operational data from multiple unmanned delivery vehicles, which includes current location, battery level, and load capacity components. By uniformly organizing this data, a basic input set covering both the task and execution sides can be formed, providing data support for subsequent collaborative allocation.

[0095] In practical applications, the multi-agent state vector set consists of the current state of each unmanned delivery vehicle. Each state vector comprehensively reflects the spatial location, power level, and load occupancy of the delivery vehicle. By associating delivery task data with the multi-agent state vector set, the range of tasks that each delivery vehicle can perform in its current state can be obtained, thereby avoiding situations where load mismatch or execution conditions are not met during the task allocation phase.

[0096] In the task allocation calculation, a set of task cost perturbation variables and a set of execution delay perturbation variables are introduced. The task cost perturbation variables mainly reflect the cost changes caused by factors such as delivery path length and remaining task time margin, while the execution delay perturbation variables are used to characterize the time offsets that may occur during the delivery process. In this embodiment, the above two types of perturbation variables are not treated independently, but are modeled in a unified manner through parameter coupling, so that the correlation between path changes and time window changes can be reflected in the calculation process.

[0097] In the perturbation modeling phase, a non-tensified polynomial chaotic expansion model is constructed to expand the perturbation subspace defined by parameter coupling relationships, yielding a polynomial chaotic expansion coefficient vector. This coefficient vector reflects the degree of influence of different combinations of perturbation variables on the task execution result and is further used to generate a perturbation propagation risk vector. This approach allows the impact of multi-source perturbations on task allocation results to be mapped onto a unified risk representation, avoiding information distortion caused by simple superposition. In the parameter search phase, the system generates a parameter correlation matrix based on the polynomial chaotic expansion coefficient vector and constructs a parameter coupling manifold accordingly. By performing manifold mapping and search space folding on the original parameter space, the high-dimensional parameter search problem can be compressed into the coupling manifold space, thereby reducing redundant search regions. Within the coupling manifold space, the DIRECT algorithm is used to perform rectangular partitioning and filtering of the parameter space, gradually narrowing the search range to obtain a stable set of candidate parameters.

[0098] During the task output phase, the system integrates the candidate parameter set and the disturbance propagation risk vector to construct a task allocation scoring matrix. Each element in the scoring matrix reflects the overall matching degree between a certain unmanned delivery vehicle and a certain delivery task. By traversing and matching the scoring matrix, the multi-agent task matching result can be obtained. This matching result maintains relatively small fluctuations in multiple consecutive runs, demonstrating good stability.

[0099] To verify the feasibility of this invention in practice, it was applied to a collaborative operation scenario involving multiple delivery tasks and multiple unmanned delivery vehicles. The task completion rate, average execution latency, and allocation result fluctuations under different methods were compared and analyzed. Experimental results show that, under the same input conditions, the task allocation results generated by this invention exhibit more stable characteristics in terms of task completion rate, latency control, and allocation consistency.

[0100] Table 1: Comparison of Task Allocation Effects for Multi-Agent Unmanned Delivery

[0101] Comparison indicators Traditional rule-based methods Static optimization method Introducing disturbance modeling methods Method of the present invention Comparison and explanation Task completion rate 0.82 0.86 0.89 0.94 The closer the value is to 1, the higher the degree of completion. Average execution latency 1.35 1.21 1.12 0.98 The smaller the value, the more stable the execution efficiency. Task allocation fluctuation range 0.27 0.21 0.16 0.09 The magnitude of fluctuation reflects the degree of distribution stability. Number of convergences in parameter search 18 15 12 8 Fewer attempts mean higher search efficiency. Average overall score 0.68 0.72 0.78 0.85 Comprehensive reflection of distribution quality

[0102] Table 1 shows the changes in several key indicators for different task allocation methods under the same input conditions. The task completion rates show that the traditional rule-based method has a completion rate of 0.82, indicating that tasks may not be allocated in a timely manner when some tasks are dense or the state changes significantly. The static optimization method improves the completion rate to 0.86, but still shows some lag when the running state changes. Introducing the perturbation modeling method results in a completion rate of 0.89, indicating that the task allocation coverage has expanded after characterizing uncertainties. The method of this invention achieves a completion rate of 0.94, approaching complete coverage, indicating that the task allocation results are more comprehensive in multi-agent collaborative scenarios.

[0103] Regarding the average execution latency, the traditional rule-based method yields a value of 1.35, reflecting significant time redundancy in path selection and task matching. The static optimization method reduces this value to 1.21, but fluctuations still occur during continuous operation. The average execution latency further decreases to 1.12 after introducing a perturbation modeling method. The method of this invention yields a value of 0.98, with a significant change in value, indicating that the overall execution rhythm is more balanced after jointly processing cost perturbations and latency perturbations during the task allocation phase.

[0104] The fluctuation range of task allocation shows that the traditional rule-based method has a fluctuation range of 0.27, with the allocation result changing significantly with the running state; the static optimization method decreases to 0.21; the method with disturbance modeling is introduced to decrease to 0.16; the method of this invention has a fluctuation range of 0.09, which is significantly reduced, reflecting that the allocation result is more consistent in the continuous running process.

[0105] Regarding the number of convergences in parameter search, traditional rule-based methods require 18 convergences, static optimization methods require 15, and methods incorporating perturbation modeling require 12, while the method of this invention requires only 8. This indicates that the parameter search process is more focused through manifold mapping and search space folding during parameter space processing. The average comprehensive score gradually increased from 0.68 to 0.85, and the numerical change is consistent with the trends of the above indicators, reflecting a stable improvement in the overall task allocation quality.

[0106] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for allocating unmanned delivery tasks based on multi-agent collaboration, characterized in that, Includes the following steps: Step 1: Acquire delivery task data and unmanned delivery vehicle operation data. Delivery task data includes task location components, time window components, and load components. Unmanned delivery vehicle operation data includes current location components, battery level components, and load capacity components. Step 2: Construct a multi-agent state vector set. The state vector includes current location components, battery level components, load state components, and executable task index components. Step 3: Generate a set of task cost perturbation variables and a set of execution delay perturbation variables. Step 4: Establish parameter coupling relationships between task cost perturbation variables and execution delay perturbation variables. Based on these parameter coupling relationships, construct a non-tensified polynomial chaotic expansion model. Step 5: Generate joint polynomial basis functions and calculate polynomial chaotic expansion coefficient vectors within a defined coupled subspace. Form a perturbation propagation risk vector based on the polynomial chaotic expansion coefficient vectors. Step 6: Generate a parameter correlation matrix based on the polynomial chaotic expansion coefficient vectors. Construct a parameter coupled manifold based on the parameter correlation matrix. Perform manifold mapping and search space folding on the original parameter space. Perform rectangular partitioning and candidate parameter screening using the DIRECT algorithm within the parameter coupled manifold space to form a candidate parameter set. Step 7: Generate a task assignment scoring matrix based on the candidate parameter set and the perturbation propagation risk vector. Output the multi-agent task matching results based on the task assignment scoring matrix.

2. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 1, characterized in that, The delivery task data in step 1 includes task location component, time window component, and load component, specifically including: Read the set of delivery task identifiers, read the corresponding geographic coordinate values ​​for each delivery task identifier, and write the geographic coordinate values ​​into the task location component; read the allowed start time and allowed end time values ​​for each delivery task identifier, and write the allowed start time and allowed end time values ​​into the time window component in a fixed order; read the load requirement value for each delivery task identifier, and write the load requirement value into the load component; establish an index mapping relationship between the task location component, time window component, and load component according to the delivery task identifier to form a delivery task data set.

3. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 2, characterized in that, Step 2, constructing the multi-agent state vector set, includes the following steps: The system reads the set of unmanned delivery vehicle identifiers, reads the current location coordinates for each identifier, and writes the current location coordinates into the current location component; reads the remaining battery power for each identifier and writes the remaining battery power into the battery power component; reads the current load and maximum load capacity for each identifier, performs a difference calculation on the current load and maximum load capacity, and writes the calculation result into the load status component; reads the delivery task data set, performs an executability judgment operation on each delivery task identifier in the delivery task data set, filters delivery task identifiers that meet the constraints of the load status component, and writes the filtering results into the executable task index component in index order; and concatenates the current location component, battery power component, load status component, and executable task index component according to the unmanned delivery vehicle identifier to form a multi-agent state vector set.

4. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 3, characterized in that, Step 3 specifically includes: The system reads the delivery task data set and the multi-agent state vector set. It reads the task location component and time window component corresponding to each delivery task identifier, and the current location component and battery component corresponding to each unmanned delivery vehicle identifier. For each unmanned delivery vehicle identifier and each delivery task identifier, it constructs candidate pairing identifiers. It calculates the Euclidean distance between the current location component and the task location component according to the candidate pairing identifiers, and writes the Euclidean distance value into the path distance component. It performs a traffic speed constant division operation on the path distance component, and writes the result into the travel time component. It calculates the difference between the allowed end time value and the current time value in the time window component, and writes the difference calculation result into the remaining time component. It performs a difference calculation between the travel time component and the remaining time component, and writes the difference calculation result into the time window margin component. It performs vector concatenation operation on the path distance component and the time window margin component according to the candidate pairing identifiers to form a task cost perturbation variable set. It performs vector concatenation operation on the travel time component and the time window margin component according to the candidate pairing identifiers to form an execution delay perturbation variable set.

5. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 4, characterized in that, Step 4, constructing a non-tensified polynomial chaotic expansion model between the task cost perturbation variable and the execution delay perturbation variable, includes the following steps: reading the sets of task cost perturbation variables and execution delay perturbation variables; for each candidate pairing, reading the corresponding path distance component, window margin component, and travel time component; performing numerical standardization on the path distance component, window margin component, and travel time component to form a standardized perturbation variable vector; performing parameter correlation calculation on the standardized perturbation variable vector to generate a perturbation variable correlation matrix; and filtering for non-zero perturbation variables based on the perturbation variable correlation matrix. Correlated perturbation variable index pairs are used to form a parameter coupling index set. Within the scope of the parameter coupling index set, a non-tensified polynomial index structure is constructed, where each index term corresponds to a set of coupled perturbation variables. A joint polynomial basis function set is generated according to the polynomial index structure, and the joint polynomial basis functions only contain the perturbation variable components corresponding to the parameter coupling index set. Sampling point substitution operations are performed on the joint polynomial basis function set to construct a polynomial chaotic expansion linear equation system. Coefficient solving operations are performed on the polynomial chaotic expansion linear equation system, and the solution is written into the polynomial chaotic expansion coefficient vector.

6. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 5, characterized in that, Step 4, forming a disturbance propagation risk vector based on the polynomial chaotic expansion coefficient vector, includes the following steps: reading the polynomial chaotic expansion coefficient vector; performing position analysis operations on each coefficient component in the coefficient vector according to the corresponding joint polynomial basis function index; calculating the square value of the coefficient component corresponding to each joint polynomial basis function, and writing the square value into the coefficient energy component; performing group accumulation operations on the coefficient energy component according to the parameter coupling index set to form a coupled disturbance energy vector; performing a normalization operation on the coupled disturbance energy vector, and writing the normalization result into the disturbance influence weight vector; reading the task cost disturbance variable set and the execution delay disturbance variable set, and performing a weighted combination operation on the disturbance influence weight vector and the corresponding disturbance variable component; writing the weighted combination operation result into the disturbance propagation risk vector according to the candidate pairing identifier; and establishing a candidate pairing identifier index mapping relationship for the disturbance propagation risk vector to form a disturbance propagation risk vector set.

7. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 6, characterized in that, Step 5, which generates a parameter correlation matrix and constructs a parameter-coupled manifold based on the polynomial chaotic expansion coefficient vector, includes the following steps: Reading the polynomial chaotic expansion coefficient vector; performing parameter position analysis on each coefficient component according to its corresponding joint polynomial basis function index to form a parameter participation identifier set; counting the number of times any two parameter indices in the parameter participation identifier set appear simultaneously in their corresponding joint polynomial basis functions, and writing the result into a parameter co-occurrence count matrix; performing symmetric normalization on the parameter co-occurrence count matrix, and writing the normalization result into the parameter correlation matrix; performing a threshold comparison operation on each element of the parameter correlation matrix, filtering parameter index pairs that meet the threshold conditions to form a parameter coupling index set; constructing a parameter coupling graph structure based on the parameter coupling index set, with parameter indices serving as node identifiers and parameter coupling index pairs as edge identifiers; performing connected component analysis on the parameter coupling graph structure to generate several parameter coupling subsets; defining a local parameter coordinate system within each parameter coupling subset, performing dimension reduction and embedding operations on the local parameter coordinate system to generate a parameter coupling manifold representation; and writing each parameter coupling manifold representation into the parameter coupling manifold set according to its parameter index.

8. The unmanned delivery task allocation method based on multi-agent collaboration according to claim 7, characterized in that, Step 5, which involves using the DIRECT algorithm to perform rectangular partitioning and candidate parameter filtering within the parameter-coupled manifold space, includes the following steps: Read the parameter value ranges corresponding to each parameter index in the original parameter space to form a set of original parameter ranges; Based on the manifold representation of each parameter coupling subset in the set of parameter coupling manifolds, perform a manifold mapping operation on the original set of parameter intervals to map the original parameter intervals to the parameter coupling manifold space; Perform dimension compression and interval merging operations on the mapped parameter intervals to form a folded manifold parameter interval set; Initialize the DIRECT algorithm search structure within the set of manifold parameter intervals. The search structure includes the initial hyperrectangular region identifier. For each hyperrectangular region, calculate the region center parameter vector, and perform objective function sampling operation on the region center parameter vector; Based on the DIRECT algorithm rules, perform an optimality judgment operation on the hyperrectangular region and filter out hyperrectangular regions that meet the judgment conditions; For hyperrectangular regions that meet the criteria, perform proportional rectangular segmentation operations along the coordinate axes of the parametrically coupled manifold to generate a set of subrectangular regions. Repeat the region center sampling and optimality determination operation for each sub-rectangular region to filter and form a set of candidate parameters.

9. A method for allocating unmanned delivery tasks based on multi-agent collaboration as described in claim 8, characterized in that, Step 6 includes the following steps: Read the candidate parameter set and the disturbance propagation risk vector set. Construct a scoring unit identifier for each unmanned delivery vehicle identifier and each delivery task identifier. Read the corresponding candidate parameter vector components and disturbance propagation risk vector components according to the scoring unit identifier. Perform a weighted summation operation on the candidate parameter vector components and a combination operation with the disturbance propagation risk vector components to generate a task allocation score value. Write the task allocation score value into a two-dimensional scoring matrix according to the unmanned delivery vehicle identifier and the delivery task identifier to form a task allocation score matrix. Perform an optimal value selection operation on the task allocation score matrix according to the unmanned delivery vehicle identifier, record the delivery task identifier corresponding to the optimal value, and establish a matching mapping relationship between the delivery task identifier and the unmanned delivery vehicle identifier to form a multi-agent task matching result and output it.

Images