A low-cost express parcel coverage collection method

Abstract

The present invention proposes a low-cost express parcel coverage collection method, which is as follows: establish a low-cost parcel coverage collection problem model, then design a low-cost parcel coverage collection problem model based on the attention model algorithm of the encoder-decoder architecture, then train the low-cost parcel coverage collection problem model, and finally obtain a trained low-cost parcel coverage collection problem model; the trained low-cost parcel coverage collection problem model receives the express parcel information to be sent and the courier's location information submitted by the sender within a period of time, and outputs the final courier collection path and the user's parcel delivery point allocation plan, so that the courier can collect the parcel according to the courier's collection path, and the user can send the parcel according to the parcel delivery point allocation plan and pay the courier fee. The present invention covers the collection method, minimizes the courier's collection cost, and benefits the user.

Description

Technical Field

[0001] The present invention belongs to the technical field of logistics optimization and optimization algorithm, and specifically relates to a low-cost express parcel coverage collection method. Background Art

[0002] The explosive growth of the e-commerce market has also driven the rapid development of China's express delivery industry in recent years. The booming development of online shopping has accelerated the evolution of the express delivery industry, which delivers goods to customers after online shopping, so the express delivery business contains huge business opportunities. In the traditional express delivery network, users who need to deliver packages apply to the express delivery platform online, and the platform will assign couriers to pick up the packages one by one according to the user's location. At the same time, the express delivery platform will calculate the optimal pickup route for each courier based on the packages assigned by the platform. Then the user will be charged the express delivery fee according to the express company's package billing rules. Generally speaking, for each package, the express delivery company's income is the difference between the user's express delivery fee and the courier's pickup fee. For the courier, the collection fee for each package is fixed. The problem that follows is that couriers need to pick up parcels door to door, and it is common for senders or couriers to wait for each other for a long time at the order location, which is time-consuming and easy to waste a lot of time. It is worth noting that the relatively fixed shipping price is not attractive to users, does not give play to the subjective initiative of users, and does not formulate an attractive pricing mechanism specifically for scattered users. In addition, the transportation used by couriers in the process of picking up parcels will also produce a lot of carbon emissions. How to make logistics transportation meet the requirements of green, low-carbon, timely and safe while reducing user costs is an urgent problem to be solved in the logistics field.

[0003] If we can diversify the delivery methods, encourage users to actively deliver packages by giving them "profits", reduce the operating costs and resource consumption in the collection process, we can ensure that the express collection process meets the requirements of low-carbon, green, efficient and timely. How to achieve low-cost, green and low-carbon in the collection process is an interesting question with practical application value. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a low-cost express parcel coverage collection method in view of the shortcomings of the above-mentioned prior art. The method plans an optimal Hamiltonian circuit for couriers to collect parcels. Through this method, couriers only need to visit some senders when collecting parcels, and the remaining users are within the coverage range of the users in the circuit, which is called the coverage radius. The covered users take the initiative to deliver the parcels to the surroundings of the designated users in the circuit. By covering the collection, the courier's collection cost is minimized, which benefits the users and solves the problems raised in the above-mentioned background technology.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a low-cost express parcel coverage collection method, characterized in that the method comprises:

[0006] S1. Establish a low-cost package coverage collection problem model, with maximizing the courier's income as the optimization goal of the low-cost package coverage collection problem model, and all senders in the courier's collection path can only be covered once, and the courier's collection path is acyclic as the constraint conditions of the low-cost package coverage collection problem model;

[0007] S2. Convert the low-cost parcel coverage problem model with maximizing the courier's income as the optimization goal into a low-cost parcel coverage problem model with minimizing the courier's coverage cost path as the optimization goal, then design the low-cost parcel coverage problem model based on the attention model algorithm of the encoder-decoder architecture, then train the low-cost parcel coverage problem model, and finally obtain the trained low-cost parcel coverage problem model;

[0008] S3. The trained low-cost parcel coverage collection problem model receives the express parcel information to be sent submitted by the sender and the courier's location information within a period of time. According to the received information, the trained low-cost parcel coverage collection problem model outputs the final courier collection path and user parcel delivery point allocation plan, so that the courier can collect the parcel according to the courier collection path, and the user can send the parcel according to the parcel delivery point allocation plan and pay the courier fee.

[0009] Preferably, a low-cost package coverage collection problem model is established in S1, with maximizing the courier's income as the optimization goal, and all senders in the courier's collection path can only be covered once, and the courier's collection path is acyclic as constraints; specifically including:

[0010] S201, using the following formula to calculate the express fee that the user should pay for sending the item;

[0011]

[0012] In the formula, p i Indicates the courier fee that the user should pay for sending the item;

[0013] ε represents the discount factor, and its value range is [0,1]. The purpose of using the discount factor is to benefit users and motivate them to send parcels;

[0014] p f Indicates the first weight price stipulated by the express company;

[0015] p c Indicates the additional weight price stipulated by the express company;

[0016] H f Indicates the additional weight specified by the express company;

[0017] w i Indicates the weight of the user's package;

[0018] S202. The courier's income is defined as the difference between the courier fees of all senders and the courier's collection cost. The courier's income is calculated using the following formula:

[0019]

[0020] In the formula, π represents the courier's pickup path;

[0021] N represents the set of all senders in the model, i and j represent sender i and sender j respectively, and the number of senders is represented by n;

[0022] x i represents the position coordinates of sender i in two-dimensional space, x j represents the position coordinates of sender j in two-dimensional space;

[0023] c(π) represents the courier’s income;

[0024] χ represents the unit moving cost of the courier;

[0025] dist(x i ,x j ) represents the shortest path length between sender i and sender j;

[0026] β i,j is a decision variable. When sender i and sender j are points passed by the courier in the path of picking up the parcel, β i,j =1, otherwise β i,j =0;

[0027] S203: Define two decision variables c according to the acceptable moving distance r of the user. i,j and The specific definition is as follows:

[0028]

[0029]

[0030] S204, map the low-cost package coverage collection problem model to a complete graph G = (X, E), X = (x0, x1, x2, ..., x n ), where x0 represents the position coordinates of the courier in two-dimensional space, x1,…,x nrepresents the location coordinates of the sender in two-dimensional space, and E represents the set of edges;

[0031] S205. If the sender in the low-cost package coverage model is a point that the courier passes through in the collection path, it is called LP; if the sender in the low-cost package coverage model is centered on LP and has a user's acceptable moving distance length r as a radius, it is covered by LP, which is called DP;

[0032] In the low-cost package coverage collection problem model, all senders must be either LP or DP. Establish the constraint To ensure that each DP user can only be overwritten once, establish constraints To ensure that each LP user can only be accessed once, establish constraints To ensure that there are no loops in the courier's pickup path;

[0033] Finally, taking the courier's income maximization as the optimization goal, a unified formula for the formal low-cost express parcel coverage collection problem model is constructed: max c(π);

[0034] The constraints are:

[0035]

[0036]

[0037] Preferably, the low-cost parcel coverage collection problem model with maximizing the courier's income as the optimization goal in S2 is transformed into a low-cost parcel coverage collection problem model with minimizing the courier's coverage collection cost path as the optimization goal; specifically including:

[0038] The function of the courier's path to cover the cost of picking up a package is defined as:

[0039]

[0040] Taking maximizing the courier’s income as the optimization goal, it becomes

[0041] In the optimization objective c(π), is a function that is linearly related to N, so when N is determined, is a constant;

[0042] Therefore, the maximization of the courier’s income c(π) is transformed into minimizing the courier’s coverage collection cost path L(π). The optimization goal is to minimize the courier’s coverage collection cost path L(π), and a unified transformed low-cost package coverage collection problem model with the courier’s coverage collection cost path minimized is constructed: min L(π).

[0043] Preferably, the attention model algorithm based on the encoder-decoder architecture described in S2 designs a low-cost package coverage collection problem model, specifically including:

[0044] Construct an attention model based on the encoder-decoder architecture to determine the final courier pickup path and user package delivery point allocation plan. The attention model is divided into an encoding stage and a decoding stage, which helps to effectively mine more hidden structural information between nodes; the nodes in the attention model represent the senders in the low-cost package coverage pickup problem model;

[0045] The process at the encoding stage is as follows:

[0046] First, the location coordinate set of the sender in the two-dimensional space is input into the attention model, and then the dimension is increased, that is, the two-dimensional information is increased to a dimension of d h High-dimensional vector The specific operations are as follows:

[0047]

[0048] Among them, W, W0, b, b0 are the parameters that need to be trained for the attention model; since the encoder consists of L identical attention sub-layers, the vector h output by each sub-layer (L) It can be obtained by the following formula:

[0049] h temp =Norm (L) (h (L-1) +MHA (L) (h (L-1) ))

[0050] h (L) =Norm (L) (h temp +FF (L) (h temp ));

[0051] Among them, Norm (L) (·) indicates that the L-th layer node result output is normalized;

[0052] MHA (L) (·) indicates that Multi-Head Attention operation is performed on the L-th layer nodes;

[0053] FF (L) (·) indicates feed-forward operation on the L-th layer nodes;

[0054] h temp It is an intermediate variable generated during the operation process;

[0055] Then, according to the vector of node i output by each sublayer To construct the key used in the decoding phase, the specific operations are as follows:

[0056]

[0057] In this way, more hidden structural information between nodes can be mined;

[0058] The processing process at the decoding stage is as follows:

[0059] According to the key output in the encoding phase and the partial solution π obtained in the decoding phase 1:t-1 To determine the selection in the next stage; when calculating the node selection, not only the previously selected nodes should be considered, but also the compatibility of the selected nodes with the remaining nodes; generally speaking, when determining the next node to be visited, the node that is closer to the most recently selected node may be given priority, so the GRU function f is first used GRU (·) to construct the hidden state information in the decoder GRU network That is, the node information visited before. The specific construction process is as follows:

[0060]

[0061] t represents the tth step in the decoding phase, Represents the hidden state information in the decoder GRU network;

[0062] in,

[0063]

[0064] Then we use the Scaled Dot-Product Attention mechanism to construct the query value q t , the specific construction process is as follows:

[0065]

[0066] Where: K1, V1 represents h (L) The linear mapping of represents the vector after the transpose operation of vector K1,

[0067] d kis a scaling factor. Generally speaking, M is determined by the Multi-Head Attention mechanism used, and softmax(·) means that the softmax(·) function is used to normalize the results;

[0068] Through the above operations, we get the key-query combination (q t ,k i ), then the probability of the selected node i being selected at the decoding step t is calculated by Scaled Dot-ProductAttention

[0069] Where: π 1～t-1 Represents the set of nodes that have been selected in the t-1 step before the decoding phase;

[0070] The probability of selecting a node that has been visited again is set to In case you are chosen again;

[0071] In order to facilitate calculation, the softmax(·) function is used to normalize the probability value. After normalization, the probability of node i being selected at decoding step t is The specific operations are as follows:

[0072]

[0073] Get the probability of each node being selected in this decoding process, and then use the greedy strategy to select the node with the highest probability to add to the courier's collection path;

[0074] Repeat the above steps until all nodes are selected, and obtain the courier pickup path π and the user package delivery point allocation plan.

[0075] 5. A low-cost express parcel coverage collection method according to claim 4, characterized in that the training of the low-cost parcel coverage collection problem model in S2 specifically comprises:

[0076] For the trainable parameters W K ,W,W0,W d ,b,b0, are uniformly defined as θ. Next, the deep reinforcement learning algorithm REINFORCE is used to train the above parameters and learn an optimal strategy pro θ (π|X) outputs the optimal objective function value with a greater probability. The strategy is modeled as:

[0077]

[0078] Among them; the state is represented by the coordinate set π of the sender that has been visited 1:t-1 ;

[0079] The state transition is expressed as the probability of the sender user to be visited next

[0080] The action space is represented by the sender π to be selected in step t t ;

[0081] The reward is expressed as the inverse of the shortest path - L(π);

[0082] The strategy is expressed as the probability pro of selecting the sender to visit θ (π|X);

[0083] n represents the number of senders input by the system, k represents the number of senders in the final courier pick-up path,

[0084] Next, for the input historical courier location information, historical sender location information, and express package weight information, the update strategy is specifically defined as follows:

[0085]

[0086] θ←θ+▽ θ J(θ|S)

[0087] Where b(S) is a benchmark function, and the objective function is constructed according to the greedy strategy, that is, according to the selection probability of each node output, the node with the highest probability is selected; by setting the benchmark function, the computational cost can be effectively reduced and the convergence can be accelerated; until θ is trained through the REINFORCE algorithm, a trained low-cost package coverage collection problem model is obtained.

[0088] Preferably, the express package information in S3 includes the location coordinates of the sender and the weight of the express package.

[0089] Preferably, the location coordinates are longitude and latitude coordinates.

[0090] Compared with the prior art, the present invention has the following advantages:

[0091] 1. The present invention designs a healthy and green delivery mechanism, which guides users to deliver parcels by covering the delivery collection method, greatly shortens the delivery route length of couriers, reduces the delivery collection cost of couriers, and at the same time gives play to the subjective initiative of users; it gives benefits to users by discounting express fees to attract users to deliver parcels.

[0092] 2. The present invention can make express delivery meet the timeliness and safety requirements while reducing the user's mailing costs, reduce the operating costs and carbon emissions in the collection process, and create a green, low-carbon, intelligent logistics and transportation system, which has certain innovation and characteristics. The present invention is highly practical, safe and reliable, and can be widely used in large-scale parcel coverage collection and distribution tasks that require quick decision-making in the process of urban scattered express delivery.

[0093] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 The present invention is a flowchart of a low-cost express parcel coverage collection method disclosed in Example 1 of the present invention.

[0095] Figure 2 This is a schematic structural diagram of the specific positions of 25 mailing users in two-dimensional space disclosed in Example 1 of the present invention.

[0096] Figure 3 The courier picks up 25 mailing users and the user package delivery point allocation scheme disclosed in Example 1 of the present invention. DETAILED DESCRIPTION

[0097] Example 1

[0098] like Figure 1 As shown, the present invention provides a technical solution: a low-cost express parcel coverage collection method, comprising the following steps:

[0099] S1. Establish a low-cost package coverage collection problem model, with maximizing the courier's income as the optimization goal of the low-cost package coverage collection problem model, and all senders in the courier's collection path can only be covered once, and the courier's collection path is acyclic as the constraint conditions of the low-cost package coverage collection problem model;

[0100] The specific process of establishing a low-cost parcel coverage collection problem model includes:

[0101] S201. Formulate a parcel delivery charge standard based on the actual delivery price of the courier company. The charge standard fully considers the relationship between the actual parcel weight and the first weight and the additional weight: when the weight of the parcel is not higher than the first weight, it is charged according to the first weight price. When the weight of the parcel is higher than the first weight, it is charged according to the product of the additional weight price and the part exceeding the first weight. Specifically, the following formula is used to calculate the courier fee that the user should pay for delivery;

[0102]

[0103] In the formula, p iIndicates the courier fee that the user should pay for sending the item;

[0104] ε represents the discount factor, and its value range is [0,1]. The purpose of using the discount factor is to benefit users and motivate them to send parcels;

[0105] p f Indicates the first weight price stipulated by the express company;

[0106] p c Indicates the additional weight price stipulated by the express company;

[0107] H f Indicates the additional weight specified by the express company;

[0108] w i Indicates the weight of the user's package;

[0109] S202. The courier's income is defined as the difference between the courier fees of all senders and the courier's collection cost. The courier's income is calculated using the following formula:

[0110]

[0111] In the formula, π represents the courier's pickup path;

[0112] N represents the set of all senders in the model, i and j represent sender i and sender j respectively, and the number of senders is represented by n;

[0113] x i represents the position coordinates of sender i in two-dimensional space, x j represents the position coordinates of sender j in two-dimensional space;

[0114] c(π) represents the courier’s income;

[0115] χ represents the unit moving cost of the courier;

[0116] dist(x i ,x j ) represents the shortest path length between sender i and sender j;

[0117] β i,j is a decision variable. When sender i and sender j are points passed by the courier in the path of picking up the parcel, β i,j =1, otherwise β i,j =0;

[0118] S203, define two decision variables c according to the acceptable moving distance r of the mailing user i,j and The specific definition is as follows:

[0119]

[0120]

[0121] S204, map the low-cost package coverage collection problem model to a complete graph G = (X, E), X = (x0, x1, x2, ..., x n ), where x0 represents the position coordinates of the courier in two-dimensional space, x1,…,x n represents the location coordinates of the sender in two-dimensional space, and E represents the set of edges;

[0122] S205. If the sender in the low-cost package coverage model is a point that the courier passes through in the collection path, it is called LP; if the sender in the low-cost package coverage model is centered on LP and has a user's acceptable moving distance length r as a radius, it is covered by LP, which is called DP;

[0123] In the low-cost package coverage collection problem model, all senders must be either LP or DP. Establish the constraint To ensure that each DP user can only be overwritten once, establish constraints To ensure that each LP user can only be accessed once, establish constraints To ensure that there are no loops in the courier's pickup path;

[0124] Finally, taking the courier's income maximization as the optimization goal, a unified formula for the formal low-cost express parcel coverage collection problem model is constructed: max c(π);

[0125] The constraints are:

[0126]

[0127]

[0128] S2. First, the low-cost parcel coverage problem model with the courier's income maximization as the optimization goal is transformed into a low-cost parcel coverage problem model with the courier's coverage cost path minimization as the optimization goal. Then, the low-cost parcel coverage problem model is designed based on the attention model algorithm of the encoder-decoder architecture, and then the low-cost parcel coverage problem model is trained to finally obtain the trained low-cost parcel coverage problem model.

[0129] Furthermore, the low-cost package coverage collection problem model described in S2 with the courier's income maximization as the optimization goal is transformed into a low-cost package coverage collection problem model with the courier's coverage collection cost path minimization as the optimization goal; specifically including:

[0130] The function of the courier's path to cover the cost of picking up a package is defined as:

[0131]

[0132] Taking maximizing the courier’s income as the optimization goal, it becomes

[0133] In the optimization objective c(π), is a function that is linearly related to N, so when N is determined, is a constant;

[0134] Therefore, the maximization of the courier’s income c(π) is transformed into minimizing the courier’s coverage collection cost path L(π). The optimization goal is to minimize the courier’s coverage collection cost path L(π), and a unified transformed low-cost package coverage collection problem model with the courier’s coverage collection cost path minimized is constructed: min L(π).

[0135] Furthermore, the attention model algorithm based on the encoder-decoder architecture described in S2 designs a low-cost package coverage collection problem model, specifically including:

[0136] Construct an attention model based on the encoder-decoder architecture to determine the final courier pickup path and user package delivery point allocation plan. The attention model is divided into an encoding stage and a decoding stage, which helps to effectively mine more hidden structural information between nodes; the nodes in the attention model represent the senders in the low-cost package coverage pickup problem model;

[0137] The process at the encoding stage is as follows:

[0138] First, the location coordinate set of the sender in the two-dimensional space is input into the attention model, and then the dimension is increased, that is, the two-dimensional information is increased to a dimension of d h High-dimensional vector The specific operations are as follows:

[0139]

[0140] Among them, W, W0, b, b0 are the parameters that need to be trained for the attention model; since the encoder consists of L identical attention sub-layers, the vector h output by each sub-layer (L) It can be obtained by the following formula:

[0141] htemp =Norm (L) (h (L-1) +MHA (L) (h (L-1) ))

[0142] h (L) =Norm (L) (h temp +FF (L) (h temp ));

[0143] In the formula, Norm (L) (·) indicates that the L-th layer node result output is normalized;

[0144] MHA (L) (·) indicates that Multi-HeadAttention operation is performed on the L-th layer nodes;

[0145] FF (L) (·) indicates feed-forward operation on the L-th layer nodes;

[0146] h temp It is an intermediate variable generated during the operation process;

[0147] Then, according to the vector of node i output by each sublayer To construct the key used in the decoding phase, the specific operations are as follows:

[0148]

[0149] In this way, more hidden structural information between nodes can be mined;

[0150] The processing process at the decoding stage is as follows:

[0151] According to the key output in the encoding phase and the partial solution π obtained in the decoding phase 1:t-1 To decide the next stage of selection; when calculating node selection, not only the previously selected nodes should be considered, but also the compatibility of the selected nodes with the remaining nodes; generally speaking, when determining the next node to be visited, the node that is closer to the most recently selected node may be given priority, so the GRU function f is first used GRU (·) to construct the hidden state information in the decoder GRU network That is, the node information visited before. The specific construction process is as follows:

[0152]

[0153] t represents the tth step in the decoding phase, Represents the hidden state information in the decoder GRU network;

[0154] in,

[0155]

[0156] Then, the Scaled Dot-Product Attention mechanism is used to construct the query. The specific construction process is as follows:

[0157]

[0158] Where: K1, V1 represents h (L) The linear mapping of represents the vector after the transpose operation of vector K1,

[0159] d k is a scaling factor. Generally speaking, M is determined by the Multi-Head Attention mechanism used, and softmax(·) means that the softmax(·) function is used to normalize the results;

[0160] Through the above operations, we get the key-query combination (q t ,k i ). Next, the probability of the selected node i being selected at the decoding step t will be calculated by Scaled Dot-ProductAttention.

[0161] Where: π 1～t-1 Represents the set of nodes that have been selected in the t-1 step before the decoding phase;

[0162] The probability of selecting a node that has been visited again is set to In case you are chosen again;

[0163] In order to facilitate calculation, the softmax(·) function is used to normalize the probability value. After normalization, the probability of node i being selected at decoding step t is The specific operations are as follows:

[0164]

[0165] Get the probability of each node being selected in this decoding process, and then use the greedy strategy to select the node with the highest probability to add to the courier's collection path;

[0166] Repeat the above steps until all nodes are selected, and obtain the courier pickup path π and the user package delivery point allocation plan.

[0167] Furthermore, the training of the low-cost package coverage collection problem model described in S2 specifically includes:

[0168] For the trainable parameters W K ,W,W0,W d ,b,b0, are uniformly defined as θ. Next, the deep reinforcement learning algorithm REINFORCE is used to train the above parameters and learn an optimal strategy pro θ (π|X) outputs the optimal objective function value with a greater probability. The strategy is modeled as:

[0169]

[0170] Among them; the state is represented by the coordinate set π of the sender that has been visited 1:t-1 ;

[0171] The state transition is expressed as the probability of the sender user to be visited next

[0172] The action space is represented by the sender π to be selected in step t t ;

[0173] The reward is expressed as the inverse of the shortest path - L(π);

[0174] The strategy is expressed as the probability pro of selecting the sender to visit θ (π|X);

[0175] n represents the number of senders input by the system, and k represents the number of senders in the final courier pickup path;

[0176] Next, the gradient descent algorithm is used to update θ based on the input historical sender location information and user package weight information. The update strategy is specifically defined as follows:

[0177]

[0178] θ←θ+▽ θ J(θ|S)

[0179] Where b(S) is a benchmark function, and the objective function is constructed according to the greedy strategy, that is, according to the selection probability of each node output, the node with the highest probability is selected; by setting the benchmark function, the computational cost can be effectively reduced and the convergence can be accelerated; until θ is trained through the REINFORCE algorithm, a trained low-cost package coverage collection problem model is obtained.

[0180] When training θ through the REINFORCE algorithm, the historical courier location information, historical sender location information, and sender package weight information are used for multiple trainings. There are 3 historical couriers in different locations, 130 historical senders in different locations, and the information of the express packages to be sent is also different. In each training, one is randomly selected from the 3 historical couriers and 20 are selected from the 130 historical senders. Due to the huge amount of data, only the data used in one training is given here, as shown in the following table:

[0181] Table 1. Some historical data information used in training

[0182]

[0183]

[0184] S3. The trained low-cost parcel coverage collection problem model receives the express parcel information to be sent submitted by the sender and the courier's location information within a period of time. According to the received information, the trained low-cost parcel coverage collection problem model outputs the final courier collection path and user parcel delivery point allocation plan, so that the courier can collect the parcel according to the courier collection path, and the user can send the parcel according to the parcel delivery point allocation plan and pay the courier fee.

[0185] Furthermore, the express package information includes the location coordinates of the sender and the weight of the package, and the location coordinates of the sender are expressed in longitude and latitude coordinates. The time period after receiving the express package information can be set according to the current express delivery and receiving situation, such as 12 hours or 24 hours.

[0186] Specifically, assume that the set of senders received by the trained low-cost package coverage collection problem model within a period of time is N = {1, 2, ..., 25}, and each sender i∈N submits their own express information B i =(w i ,x i ) to the low-cost parcel coverage collection problem model, the specific information submitted is shown in the following table:

[0187] Table 2 Express package information table of the sender

[0188]

[0189]

[0190] The specific location of the sender in two-dimensional space is as follows: Figure 2 shown.

[0191] The courier fees that the 25 mailers should pay are shown in Table 2, where the discount factor ε is set to 0.1 and the first weight price p specified by the courier company is f Set to 12; the additional weight price p stipulated by the courier company c Set to 1; the additional weight H specified by the courier company f Set to 0.5.

[0192] Table 3 Express delivery fees payable by users

[0193]

[0194]

[0195] The trained low-cost package coverage collection problem model receives the location coordinates of the 25 senders and the courier fee determined by the courier weight, and calculates the courier collection path π and the user package delivery point allocation plan. The specific results are as follows Figure 3 As shown in the figure, the red package is LP and the black package is DP. The calculated shortest path length is L(π) = 28.71 km. The courier picks up the parcels in the corresponding order according to the assigned LP and NP (i.e. the parcel delivery point allocation plan). After picking up the parcels, the user can pay the courier fee online.

[0196] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent change made to the above embodiment according to the technical essence of the invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A low-cost express parcel coverage collection method, characterized in that: The method includes: S1. Establish a low-cost package coverage collection problem model, with maximizing the courier's income as the optimization goal of the low-cost package coverage collection problem model, and all senders in the courier's collection path can only be covered once, and the courier's collection path is acyclic as the constraint conditions of the low-cost package coverage collection problem model; S2. Convert the low-cost parcel coverage problem model with maximizing the courier's income as the optimization goal into a low-cost parcel coverage problem model with minimizing the courier's coverage cost path as the optimization goal, then design the low-cost parcel coverage problem model based on the attention model algorithm of the encoder-decoder architecture, then train the low-cost parcel coverage problem model, and finally obtain the trained low-cost parcel coverage problem model; S3. The trained low-cost parcel coverage collection problem model receives the express parcel information to be sent submitted by the sender and the courier's location information within a period of time. According to the received information, the trained low-cost parcel coverage collection problem model outputs the final courier collection path and the user's parcel delivery point allocation plan, so that the courier can collect the parcel according to the courier's collection path, and the user can send the parcel according to the parcel delivery point allocation plan and pay the courier fee; In S1, a low-cost package coverage collection problem model is established, with the courier's income maximization as the optimization goal, and all senders in the courier's collection path can only be covered once, and the courier's collection path has no loops as constraints; specifically, it includes: S201. Calculate the express fee payable by the user for sending the item using the following formula; In the formula, p i Indicates the courier fee that the user should pay for sending the item; ε represents the discount factor, and its value range is [0,1]. The purpose of using the discount factor is to benefit users and motivate them to send parcels. p f Indicates the first weight price stipulated by the express company; p c Indicates the additional weight price stipulated by the express company; H f Indicates the additional weight specified by the express company; w i Indicates the weight of the user's package; S202. The courier's income is defined as the difference between the courier fees of all senders and the courier's collection cost. The courier's income is calculated using the following formula: In the formula, π represents the courier's pickup path; N represents the set of all senders in the model, i and j represent sender i and sender j respectively, and the number of senders is represented by n; x i represents the position coordinates of sender i in two-dimensional space, x j represents the position coordinates of sender j in two-dimensional space; c(π) represents the courier’s income; χ represents the unit moving cost of the courier; dist(x i ,x j ) represents the shortest path length between sender i and sender j; β i,j is a decision variable. When sender i and sender j are points passed by the courier in the path of picking up the parcel, β i,j =1, otherwise β i,j =0; S203, define two decision variables c according to the acceptable moving distance r of the mailing user i,j and The specific definition is as follows: S204, map the low-cost package coverage collection problem model to a complete graph G = (X, E), X = (x0, x1, x2, ..., x n ), where x0 represents the position coordinates of the courier in two-dimensional space, x1,…,x n represents the location coordinates of the sender in two-dimensional space, and E represents the set of edges; S205. If the sender in the low-cost package coverage model is a point that the courier passes through in the collection path, it is called LP; if the sender in the low-cost package coverage model is centered on LP and has a user's acceptable moving distance length r as a radius, it is covered by LP, which is called DP; In the low-cost package coverage collection problem model, all senders must be either LP or DP. Establish the constraint To ensure that each DP user can only be overwritten once, establish constraints To ensure that each LP user can only be accessed once, establish constraints To ensure that there are no loops in the courier's pickup path; Finally, taking the courier's income maximization as the optimization goal, a unified formula for the formal low-cost express parcel coverage collection problem model is constructed: max c(π); The constraints are: The attention model algorithm based on the encoder-decoder architecture described in S2 designs a low-cost package coverage collection problem model, specifically including: Construct an attention model based on the encoder-decoder architecture to determine the final courier pickup path and user package delivery point allocation plan. The attention model is divided into an encoding stage and a decoding stage, which helps to effectively mine more hidden structural information between nodes; the nodes in the attention model represent the senders in the low-cost package coverage pickup problem model; The process at the encoding stage is as follows: First, the location coordinate set of the sender in the two-dimensional space is input into the attention model, and then the dimension is increased, that is, the two-dimensional information is increased to a dimension of d h High-dimensional vector The specific operations are as follows: Among them, W, W0, b, b0 are the parameters that need to be trained for the attention model; since the encoder consists of L identical attention sub-layers, the vector h output by each sub-layer (L) It can be obtained by the following formula: h temp =Norm (L) (h (L-1) +MHA (L) (h (L-1) )) h (L) =Norm (L) (h temp +FF (L) (h temp )); In the formula, Norm (L) (·) indicates that the L-th layer node result output is normalized; MHA (L) (·) indicates that Multi-HeadAttention operation is performed on the L-th layer nodes; FF (L) (·) indicates feed-forward operation on the L-th layer nodes; h temp It is an intermediate variable generated during the operation process; Then, the key vector K used by node i in the decoding phase is constructed based on the vector of node i output by each sub-layer i , the specific operations are as follows: In this way, more hidden structural information between nodes can be mined; The processing process at the decoding stage is as follows: According to the key output in the encoding phase and the partial solution π obtained in the decoding phase 1:t-1 To determine the selection in the next stage; when calculating the node selection, not only the previously selected nodes should be considered, but also the compatibility of the selected nodes with the remaining nodes; generally speaking, when determining the next node to be visited, the node that is closer to the most recently selected node may be given priority, so the GRU function f is first used GRU (·) to construct the hidden state information in the decoder GRU network That is, the node information visited before. The specific construction process is as follows: t represents the tth step in the decoding phase, Represents the hidden state information in the decoder GRU network; in, Then we use the Scaled Dot-Product Attention mechanism to construct the query value q t , the specific construction process is as follows: Where: K1, V1 represents h (L) The linear mapping of represents the vector after the transpose operation of vector K1, d k is a scaling factor. Generally speaking, M is determined by the Multi-Head Attention mechanism used, and softmax(·) means that the softmax(·) function is used to normalize the results; Through the above operations, we get the key-query combination (q t ,k i ), then the probability of the selected node i being selected at the decoding step t is calculated by Scaled Dot-ProductAttention Where: π 1～t-1 Represents the set of nodes that have been selected in the t-1 step before the decoding phase; The probability of selecting a node that has been visited again is set to In case you are chosen again; In order to facilitate calculation, the softmax(·) function is used to normalize the probability value. After normalization, the probability of node i being selected at decoding step t is The specific operations are as follows: Get the probability of each node being selected in this decoding process, and then use the greedy strategy to select the node with the highest probability to add to the courier's collection path; Repeat the above steps until all nodes are selected, and obtain the courier pickup path π and the user package delivery point allocation plan.

2. A low-cost express package coverage collection method according to claim 1, characterized in that: The low-cost package coverage problem model described in S2, which takes maximizing the courier’s income as the optimization goal, is transformed into a low-cost package coverage problem model with minimizing the courier’s coverage cost path as the optimization goal; specifically, it includes: The function of the courier's path to cover the cost of picking up a package is defined as: Taking maximizing the courier’s income as the optimization goal, it becomes In the optimization objective c(π), is a function that is linearly related to N, so when N is determined, is a constant; Therefore, the maximization of the courier’s income c(π) is transformed into minimizing the courier’s coverage collection cost path L(π). The optimization goal is to minimize the courier’s coverage collection cost path L(π), and a unified transformed low-cost package coverage collection problem model with the courier’s coverage collection cost path minimized is constructed: min L(π).

3. A low-cost express package coverage collection method according to claim 1, characterized in that: S2 describes training a low-cost package coverage collection problem model; specifically includes: For the trainable parameters W K ,W,W0,W d ,b,b0, are uniformly defined as θ. Next, the deep reinforcement learning algorithm REINFORCE is used to train the above parameters and learn an optimal strategy pro θ (π|X) outputs the optimal objective function value with a greater probability. The strategy is modeled as: Among them; the state is represented by the coordinate set π of the sender that has been visited 1:t-1 ; The state transition is expressed as the probability of the sender user to be visited next The action space is represented by the sender π to be selected in step t t ; The reward is expressed as the inverse of the shortest path - L(π); The strategy is expressed as the probability pro of selecting the sender to visit θ (π|X); n represents the number of senders input by the system, and k represents the number of senders in the final courier pickup path; Next, the gradient descent algorithm is used to update θ based on the input historical courier location information, historical sender location information, and express package weight information. The update strategy is specifically defined as follows: Where b(S) is a benchmark function, and the objective function is constructed according to the greedy strategy, that is, according to the selection probability of each node output, the node with the highest probability is selected; by setting the benchmark function, the computational cost can be effectively reduced and the convergence can be accelerated; until θ is trained through the REINFORCE algorithm, a trained low-cost package coverage collection problem model is obtained.

4. A low-cost express package coverage collection method according to claim 1, characterized in that: The express package information in S3 includes the location coordinates of the sender and the weight of the express package.

5. A low-cost express package coverage collection method according to claim 1, characterized in that: The location coordinates are longitude and latitude coordinates.

Images