A logistics sorting method integrating geographic graph network under actual sorting constraints

Abstract

The present invention discloses a logistics sorting method that integrates a geographic graph network under actual sorting constraints, including: S1. defining a distance function between package flows and modeling the placement constraints of adjacent grids; S2. defining a peak capacity index and an average capacity index, and constructing an objective function for maximizing sorting capacity; S3. defining a maximization capacity index function under actual package constraints; S4. based on a digital twin system, conducting a simulation test identical to the actual sorting environment to obtain the peak capacity index and average capacity index for the entire shift; S5. based on a Monte Carlo tree search method, constructing a package geographic graph network, limiting each action space to the effective package flow range selected by the geographic graph network, and performing package selection that fully complies with the actual package constraints. By constructing a geographic graph network structure between packages and embedding it into a Monte Carlo tree search framework, the present invention obtains an optimal sorting plan that satisfies a reasonable placement strategy for adjacent grids under the goal of maximizing capacity.

Description

Technical Field

[0001] The present invention relates to the field of logistics sorting, and in particular to a logistics sorting method integrating a geographic map network under actual sorting constraints. Background Art

[0002] In the small-item sorting scenario at a logistics transfer station, the core of the sorting optimization strategy lies in establishing a dynamic mapping relationship between package flow and slots (i.e., a sorting plan) to maximize sorting capacity indicators. However, practical applications in this field face multiple technical bottlenecks: First, due to differences in the physical layout and operational characteristics of transfer stations, there are differential placement constraints on the geographical relationship between the package flow directions mapped to adjacent slots, which makes it difficult for traditional standardized sorting models to adapt across scenarios; second, existing optimization algorithms mostly focus on the design of a single capacity objective function and fail to incorporate spatial constraints into a systematic modeling framework. In practice, they rely on manual experience for local adjustments, which can easily lead to suboptimal solutions and lack global strategy stability; more importantly, for the complex optimization problem of simultaneously satisfying capacity maximization and multi-dimensional spatial constraints, the current academic community has not yet established a complete theoretical system, and related research is still in the exploratory stage. Summary of the Invention

[0003] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a logistics sorting method that integrates a geographic graph network under actual sorting constraints. By constructing a geographic graph network structure between packages and embedding it into a Monte Carlo tree search framework, it is possible to effectively search for the optimal sorting plan that meets the reasonable placement strategy of adjacent grids under the goal of maximizing production capacity.

[0004] The object of the present invention is achieved through the following technical solution: a logistics sorting method integrating a geographic map network under actual sorting constraints, comprising the following steps:

[0005] S1. Define the distance function between two package flows and model the placement constraints of adjacent slots;

[0006] S2. For input sorting plan , based on the number of packages sorted per minute, define the peak capacity indicator and average capacity index , construct the objective function of maximizing sorting capacity;

[0007] S3. Based on the defined maximum sorting capacity objective function and the constraints of adjacent slot placement, define the maximum capacity index function under the actual package constraints;

[0008] S4. Based on the digital twin system, sorting plan Conduct simulation tests that are exactly the same as the actual sorting environment to obtain peak production capacity indicators for the entire shift and average production capacity index ;

[0009] S5. Based on the Monte Carlo tree search method, a package geographic graph network is constructed, and each action space is restricted to the valid package flow range selected by the geographic graph network, effectively performing package selection that fully complies with the actual package constraints.

[0010] The beneficial effects of the present invention are as follows: the present invention conducts in-depth modeling of the geographical relationship between packages mapped to adjacent slots in a logistics transfer yard, integrates this geographical relationship into the sorting optimization problem, and constructs a multi-objective sorting optimization model; based on this model, by constructing a geographical graph network structure between packages and embedding it into a Monte Carlo tree search framework, it is achieved to effectively search for the optimal sorting plan that meets the reasonable placement strategy of adjacent slots under the goal of maximizing production capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0012] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0013] like Figure 1 As shown, a logistics sorting method integrating geographic graph network under actual sorting constraints includes the following steps:

[0014] S1. Define the distance function between two package flows and model the placement constraints of adjacent slots;

[0015] During the actual sorting process at the logistics transfer station, the sorting machine relies on the conveyor belt rollers to throw the parcels from the sorting machine into the corresponding sorting grid. However, due to the differences in friction and inertia between different cargo packaging and types, some parcels may not fall accurately into the predetermined grid, but mistakenly enter the adjacent grid. Usually, these missorted parcels will be discovered when they are mistakenly transported to the next transfer station, which will incur significant additional transportation costs and logistics time losses. In order to reduce the additional transportation costs and time delays caused by correcting missorting, the transfer station usually restricts the parcel flow in adjacent grids during actual sorting, that is, it requires that the parcels placed in adjacent grids be as close as possible in the cities they flow to. Specifically, the adjacency priorities are as follows:

[0016] 1. The destination city is in the same province; 2. The destination city is in an adjacent province; 3. The destination city is in a non-adjacent province.

[0017] Under this constraint, we first model the constraint: First, the grid of the sorting machine is defined as , is the total number of sorting machine slots. The mapped package flow is defined as For different sorting areas in the sorter, we define ,in, Belong to two different sorting areas. For the parcel flow directions i and o, we define the distance function between the two flow directions as :

[0018]

[0019] in, is the province between package flows i and o. Under the defined distance function, we model the adjacent grid placement constraint as:

[0020]

[0021] in, and Number the first and last slots in each sorting area.

[0022] S2. For input sorting plan , based on the number of packages sorted per minute, define the peak capacity indicator and average capacity index , construct the objective function of maximizing sorting capacity;

[0023] Considering the sorting plan input within the parcel sorting shift ,Based on the number of packages sorted per minute, we define two capacity indicators: peak capacity index , which measures the maximum sorting capacity in the entire sorting shift; average capacity index , which measures the average sorting capacity within the entire sorting shift. Under the two capacity indicators defined, the objective function of maximizing sorting capacity is defined as the weighted maximum of the two capacity indicators:

[0024]

[0025] in, and is a weighted coefficient, and each transfer station assigns corresponding priorities to peak capacity and average capacity.

[0026] S3. Based on the defined maximum sorting capacity objective function and the constraints of adjacent slot placement, define the maximum capacity index function under the actual package constraints;

[0027] Based on the defined objective function of maximizing sorting capacity and the constraint of adjacent slot placement, we add the adjacent slot placement constraint as a soft constraint (taking into account the situation that in actual sorting, there may be a situation where a single package flow is far away from other package flows, such as packages sent to Urumqi are not adjacent to other flows) into the objective function. The maximum capacity indicator function defined under the actual package constraint is:

[0028]

[0029] Among them, the weight is the balance coefficient, and each transfer station assigns corresponding priorities to maximize production capacity and the emphasis on adjacent grid constraints.

[0030] S4. Based on the digital twin system, sorting plan Conduct simulation tests that are exactly the same as the actual sorting environment to obtain peak production capacity indicators for the entire shift and average production capacity index ;

[0031] In the process of solving the maximum capacity index function under the actual package constraints, considering that the maximum capacity index has great deployment test costs and personnel scheduling difficulties in field tests, we have established a digital twin system to sort the plan. Conduct simulation tests that are exactly the same as the actual sorting environment to obtain peak production capacity indicators for the entire shift and average production capacity index .

[0032] S5. Based on the Monte Carlo tree search method, a package geographic graph network is constructed, and each action space is restricted to the valid package flow range selected by the geographic graph network, effectively performing package selection that fully complies with the actual package constraints.

[0033] However, considering that existing algorithms (dynamic programming, genetic algorithms, simulated annealing algorithms, reinforcement learning, etc.) cannot effectively avoid exploring allocations that do not meet the actual package constraints when solving the maximum capacity indicator function under the actual package constraints, it is very difficult to solve efficiently and obtain effective solutions. Therefore, based on the Monte Carlo tree search method, we construct a package geographic graph network, restricting each action space to the effective package flow range selected by the geographic graph network, and effectively performing package selection that fully meets the actual package constraints. The specific implementation is as follows:

[0034] First, we construct a parcel geographic graph network. For the parcel flow within each sorting shift, we construct the graph network as , in The set of vertices representing the flow of all packages, Represents the set of edges connecting all packages flowing to vertices. For each package flowing to vertex Contains package flow Flow city information , timeliness information , and the proportion of the sorting shift , which is Based on the constructed parcel geographic graph network, we design the state-action set ,state Represents the package flow direction and action corresponding to the current logistics sorting grid Represents the flow of packages allowed to be placed at the next slot, where the action The action space of is constrained by the constructed package geographic graph network, that is, only the package flow vertices that are legally connected under the package flow adjacent constraints are allowed to be selected, so that in each state Select the action This fully complies with the actual wrapping constraints. Under this state-action set definition, we construct the Monte Carlo tree node:

[0035]

[0036] in is the evaluation value and evaluation times of the node in the Monte Carlo tree search tree for the production capacity, is the reward value and number of selections under the expansion strategy of this node in the Monte Carlo tree search tree, is the weighted mean value of the node:

[0037]

[0038] Unlike traditional Monte Carlo tree search trees, state transitions no longer rely on a predefined similarity probability matrix, as the similarity probability matrix can still lead to the selection of package flows that do not conform to the actual package constraints. Furthermore, if a probability threshold is set for action screening, it is impossible to accurately quantify the similarity probability threshold for each package flow under the actual package constraints according to different logistics scenarios. Considering the algorithm's universality, this invention strictly uses a geographic graph network as the action space for action selection. On this basis, the Monte Carlo tree expansion still follows the process of selection, expansion, evaluation, and backtracking to expand the nodes of the Monte Carlo tree:

[0039] Selection steps:

[0040] At the root node, explore the child node information and select the current optimal child node according to the polynomial confidence selection rule. The strategy is:

[0041]

[0042] in, The coefficient is the coefficient for balancing exploration and utilization of existing high-quality nodes. is the sum of the number of explorations of the node's parent nodes. Unlike traditional Monte Carlo tree search, this selection strategy uses the classic UCT algorithm, considering that the similarity probability matrix is no longer used, to ensure that each node is explored once before further selection.

[0043] Expansion steps:

[0044] Directly use the constructed parcel geographic graph network to expand the Monte Carlo tree search:

[0045] In the expansion phase of the Monte Carlo tree search, the specified number of child nodes to be selected for each expansion is first determined. Based on the constructed parcel geographic graph network, the specified number of expandable nodes are selected as the expanded child nodes of the current node in the order of distance function values from small to large.

[0046] The number of child nodes selected for each expansion is user-defined and should be much smaller than the number of expandable nodes; for example, 5 or 10.

[0047] Specifically, first calculate the distance function value between the current node and all expandable nodes; then select a specified number of expandable nodes in the order of distance function value from small to large. For example, first select the flow direction nodes in the same province, that is, the distance function value The expandable node, then select the adjacent province flow node, that is, the distance function value The expandable node, and then select the distance function value until the number of expandable nodes selected is equal to the specified number;

[0048] In the traditional Monte Carlo tree search, all expandable nodes are selected during package expansion. Then, in this application, in the geographic map network architecture, the child nodes of other provinces are directly excluded through the distance function of the package map network, avoiding the waste of computing resources caused by the expansion simulation of useless nodes; the constructed geographic map network effectively prunes the Monte Carlo tree for invalid actions, so that all computing resources are concentrated on high-quality, constrained nodes, effectively accelerating the convergence performance of the Monte Carlo tree search.

[0049] Evaluation steps:

[0050] During the Monte Carlo tree expansion process, starting from the leaf node and continuing until the Monte Carlo tree is fully expanded, the reward value of the current branch selection is obtained based on the parcel grid distribution of the current branch and the reward function. The reward value is completely based on the weighted sum of the two production capacity indicators returned by the established digital twin system: , and assign the weighted sum to the reward value under the expansion strategy of the node in the Monte Carlo tree search tree, that is .

[0051] Backtracking steps:

[0052] After the Monte Carlo tree is fully expanded, the number of selections is updated upward from the current leaf node of the Monte Carlo tree. And the corresponding node evaluation value and the reward value under the expansion strategy , through the backtracking process, the weighted value mean of the node can be calculated

[0053]

[0054] After multiple rounds of Monte Carlo tree expansion and exploration interactions, when all the slots have been selected in the selection phase, the expansion path under the current root node is the optimal path, that is, the optimal slot sorting strategy.

[0055] The foregoing description is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Instead, the present invention can be used in other combinations, modifications, and environments and can be modified within the scope of the concept described herein through the above teachings or techniques or knowledge in the relevant field. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be protected by the appended claims.

Claims

1. A logistics sorting method integrating geographic graph networks under actual sorting constraints, characterized by: The following steps are involved: S1. Define the distance function between two package flows and model the placement constraints of adjacent slots; The step S1 comprises: Define the sorter slot as , is the total number of sorting machine slots, for the slots The mapped package flow is defined as ; For different sorting areas in the sorter, it is defined as ,in, Belonging to two different sorting areas, for the package flow directions i and o, the distance function between the two flow directions is defined as : in, The province between which the package flows is i and o; Under the defined distance function, the adjacent grid placement constraints are modeled as: in, and Number the first and last slots in each sorting area; S2. For input sorting plan , based on the number of packages sorted per minute, define the peak capacity indicator and average capacity index , construct the objective function of maximizing sorting capacity; S3. Based on the defined maximum sorting capacity objective function and the constraints of adjacent slot placement, define the maximum capacity index function under the actual package constraints; S4. Based on the digital twin system, sorting plan Conduct simulation tests that are exactly the same as the actual sorting environment to obtain peak production capacity indicators for the entire shift and average production capacity index ; S5. Based on the Monte Carlo tree search method, a package geographic graph network is constructed, and each action space is restricted to the valid package flow range selected by the geographic graph network, effectively performing package selection that fully complies with the actual package constraints.

2. The logistics sorting method integrating geographic map network under actual sorting constraints according to claim 1 is characterized by: The step S2 comprises: In the parcel sorting shift, for the sorting plan entered , based on the number of packages sorted per minute, two capacity indicators are defined: Peak capacity index , which measures the maximum sorting capacity during the entire sorting shift; Average production capacity index , which measures the sorting capacity averaged over the entire sorting shift; Under the two defined capacity indicators, the objective function of maximizing sorting capacity is defined as the weighted maximum of the two capacity indicators: ； in, 、 It is the weighting coefficient of peak capacity index and average capacity index.

3. The logistics sorting method integrating geographic map network under actual sorting constraints according to claim 2 is characterized by: The step S3 comprises: Based on the defined objective function of maximizing sorting capacity and the constraint of adjacent slot placement, the constraint of adjacent slot placement is added to the objective function as a soft constraint. The maximum capacity index function under the actual package constraint is defined as: Among them, the weight is the balance coefficient, and each transfer station assigns corresponding priorities to maximize production capacity and the emphasis on adjacent grid constraints.

4. The logistics sorting method integrating geographic map network under actual sorting constraints according to claim 1 is characterized by: The step S5 comprises: Construct a parcel map network: For the parcel flow within each sorting shift, the construction graph network is , in The set of vertices representing the flow of all packages, Represents the set of edges connecting all packages flowing to the vertex; For each package flowing to a vertex Contains package flow Flow city information , timeliness information , and the proportion of sorting shifts , which is ; Design state-action sets based on the constructed parcel geographic graph network ,state Represents the package flow direction and action corresponding to the current logistics sorting grid Represents the flow of packages allowed to be placed at the next slot, where the action The action space of is constrained by the constructed package geographic graph network, that is, only the package flow vertices that are legally connected under the package flow adjacent constraints are allowed to be selected, so that in each state Select the action Full compliance with actual package constraints; Under this state-action set definition, construct the Monte Carlo tree node: in is the evaluation value and evaluation times of the node in the Monte Carlo tree search tree for the production capacity, is the reward value and number of selections under the expansion strategy of this node in the Monte Carlo tree search tree, is the weighted mean value of the node: in, is the weight coefficient of the node; Use geographic graph networks as action spaces for action selection; On this basis, the expansion of the Monte Carlo tree still follows the process of selection, expansion, evaluation, and backtracking to expand the nodes of the Monte Carlo tree: Selection steps: At the root node, explore the child node information and select the current optimal child node according to the polynomial confidence selection rule. The strategy is: in, The coefficient is the coefficient for balancing exploration and utilization of existing high-quality nodes. The strategy uses the classic UCT algorithm to ensure that each node is explored once before making further selections. Expansion steps: Directly use the constructed parcel geographic graph network to expand the Monte Carlo tree search: In the expansion phase of the Monte Carlo tree search, the specified number of child nodes to be selected for each expansion is first determined. Based on the constructed parcel geographic graph network, the specified number of expandable nodes are selected as the expanded child nodes of the current node in the order of distance function values from small to large. Evaluation steps: During the Monte Carlo tree expansion process, starting from the leaf node and continuing until the Monte Carlo tree is fully expanded, the reward value of the current branch selection is obtained based on the parcel grid distribution of the current branch and the reward function. The reward value is completely based on the weighted sum of the two production capacity indicators returned by the established digital twin system: , and assign the weighted sum to the reward value under the expansion strategy of the node in the Monte Carlo tree search tree, that is ; Backtracking steps: After the Monte Carlo tree is fully expanded, the number of selections is updated upward from the current leaf node of the Monte Carlo tree. And the corresponding node evaluation value and the reward value under the expansion strategy , through the backtracking process, the weighted value mean of the node is calculated : After multiple rounds of Monte Carlo tree expansion exploration interactions, when all the slots have been selected in the selection stage, the expansion path under the current root node is the optimal path, that is, the optimal slot sorting strategy.

Images