RESEARCH ON LOGISTICS DISTRIBUTION

ROUTE OPTIMIZATION BASED ON DEEP

LEARNING MODEL AND BLOCK CHAIN

TECHNOLOGY

Xiaoshan Yang*

Student Work Division, Nantong Institute of Technology, Nantong, Jiangsu,

226000, China

yxs15262754327@163.com

Weiwei Guan

College of Business, Nantong Institute of Technology, Nantong, Jiangsu, 226000,

China

Reception: 28/10/2022 Acceptance: 29/12/2022 Publication: 24/01/2023

Suggested citation:

Y., Xiaoshan and G., Weiwei (2023). Research on logistics distribution route

optimization based on deep learning model and block chain technology .

3C Empresa. Investigación y pensamiento crítico, 12(1), 68-85. https://doi.org/

10.17993/3cemp.2023.120151.68-85

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

ABSTRACT

The growing data age is reflected in all aspects of today's society. In the field of

logistics, especially when the road conditions in urban areas are complex, how to

select the optimal distribution path and reduce the distribution time is a problem

worthy of attention. Aiming at the problems faced by traditional algorithms in solving

the distribution of logistics vehicles in urban areas, however, the method based on

regional chain technology can better solve the path optimization problem. A deep

reinforcement learning algorithm based on attention mechanism and LSTM model is

designed and applied to the distribution path planning of logistics vehicles. The

distribution optimization path of logistics vehicles is obtained through sample training

experiments, Thus, it provides a new idea for the optimization of logistics distribution

path.

KEYWORDS

Regional chain technology; Logistics distribution route; Optimization; Attention

mechanism; LSTM model

PAPER INDEX

ABSTRACT

KEYWORDS

1. PREFACE

2. ANALYSIS OF LOGISTICS DISTRIBUTION ROUTE

OPTIMIZATION

2.1. Logistics distribution path optimization

2.2. Complexity of logistics distribution route optimization

3. RESEARCH ON OPTIMIZATION OF LSTM LOGISTICS

DISTRIBUTION PATH BASED ON BLOCKCHAIN

3.1. Introduction to regional chain technology

3.2. Introduction to LSTM model

3.3. Logistics distribution vehicle routing planning model

3.4. Reinforcement learning algorithm

4. EXPERIMENTAL RESULTS AND ANALYSIS

5. IN CONCLUSION

REFERENCES

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

1. PREFACE

With the rapid development of modern information technology and the arrival of the

era of big data, a large amount of scientific and technological information has different

values, which can promote the development of human science and technology to a

certain extent [1-4]. With the in-depth development of cross-border e-commerce,

people's demand for logistics efficiency has increased rapidly, and modern information

technology has gradually entered the logistics field in this region. As one of the

cutting-edge modern information technologies, blockchain technology is also trying to

enter, but its applicability needs in-depth research. At the same time, economic

cooperation has widened the gap between supply and demand of regional logistics

services. In order to alleviate this contradiction, exploratory research is carried out on

the applicability of blockchain technology in this field. From the component

perspective, the regional logistics subject and the current situation of logistics demand

make the blockchain technology applicable at the theoretical level. Based on the

application scheme of each component, the applicability of blockchain technology at

the practical level is further verified with the combination of simulation and analytic

hierarchy process. This technology helps to improve the transparency of regional

logistics services and build an efficient logistics mechanism [5-6].

At the same time, with the progress of science and technology society, people also

put forward higher requirements for logistics distribution services, the most important

of which is the punctuality of logistics distribution [7-8]. Therefore, for this requirement,

it is required that the logistics distribution path should be optimized as much as

possible. However, at present, specific urban areas such as communities and schools

have complex characteristics such as dense customer nodes and fixed areas, so the

optimization of distribution path is in an important research position. However, from

the perspective of domestic research, there is little research on solving methods

based on artificial intelligence. At present, domestic scholars mainly focus on heuristic

algorithm, meta heuristic algorithm and its improved algorithm. Wei Xiaodi et al [9]

used an improved algorithm of discrete flower pollination algorithm and discrete flower

pollination algorithm, the flower pollination operator is redefined and combined with

the improved genetic operator. Su Xinxin et al [10] relaxes the vehicle capacity and

customer time window, adds the penalty to the objective function, uses the greedy

algorithm to generate the initial solution, and then uses the tabu search algorithm to

solve it. Four operators are used to search the solution of the neighborhood. In order

to further expand the search range, they use the perturbation operator. At present,

foreign research on VRP has successively emerged artificial intelligence methods

such as pointer network and decoder network to optimize vehicle route. Mao et al [11]

proposed a more practical mathematical model of vehicle routing problem with pick-up

and delivery, and used the double termination criterion to generate a new solution by

adding storage function and using intra line and inter line exchange.Cordeau et al [12]

adopts a unified tabu search heuristic method: periodic and multi site vehicle routing

problem with time window. The performance of the heuristic algorithm is evaluated by

comparing it with the alternative method of benchmark instance with the

characteristics of VRPTW problem.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

We mainly elaborate the problem of logistics distribution path optimization, clarify

that the main form of path optimization in logistics distribution is the problem of

determining the starting and ending points, and introduce the method of deep learning

in view of the shortcomings of the implementation process of traditional evolutionary

algorithm in path optimization and the difficulties of logistics distribution path

optimization in complex road conditions. The attention mechanism model based on

deep reinforcement learning algorithm is improved, and the model is applied to the

logistics distribution vehicle distribution path optimization problem. The strategy

network model is built through model components such as delivery length memory

neural network, decoder and attention mechanism module, and the logistics

distribution path optimization problem is handled through numerical experimental

training, verification, test and analysis, so as to promote the development of the

logistics field.

2. ANALYSIS OF LOGISTICS DISTRIBUTION ROUTE

OPTIMIZATION

2.1. LOGISTICS DISTRIBUTION PATH OPTIMIZATION

Logistics distribution path optimization is a complex problem, which is mainly to

solve the problem of path selection in logistics distribution. Usually, the logistics

distribution path optimization problem needs to consider the transportation capacity of

vehicles, limited time, transportation cost and other conditions. In the actual logistics

distribution activities, the proportion of these factors will change. Therefore, the

research on the logistics distribution path optimization problem is also divided into

many aspects, and the algorithms used are also various. When selecting the

algorithm, we should combine the cost constraints such as distribution capacity and

time, and fully consider the problems encountered in the actual distribution, so that the

algorithm can get an excellent distribution route solution [13-14].

(1)Concept of logistics distribution route optimization

The rapid economic development of our country also brings the leapfrog

development of animal logistics industry, especially the rocket development of e-

commerce, which has played a great role in promoting the development of logistics

industry. The research on the optimization of logistics distribution has become

increasingly important [15-16]. However, the current research mainly focuses on the

small cars of trucks or express brothers. In the urban transportation network, the

optimal route is solved according to the set algorithm model. The optimal route usually

needs to consider five conditions [17-19]. The first is the means of transport, which

needs to consider the types of means of transport and the cargo carrying capacity of

the means of transport. The appropriate selection and allocation of means of transport

is an important premise in the rationalization of transport. The second is the

transportation link. At the beginning of transportation, goods need to be sorted and

loaded. When the type and quantity of goods are large, this work will take up a lot of

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

operation time and increase labor and packaging costs. Therefore, the distribution link

shall be minimized in the distribution, so as to reduce the cost. The third is

transportation time. In today's logistics, transportation time becomes more and more

important. There are strict time requirements for the distribution of many items, such

as perishable fresh food, items that must be delivered within a specified time limit, etc.

Distribution time is not only the requirement of distribution, but also the intuitive

embodiment of the distribution ability of logistics companies, which affects the

satisfaction of customers. Reducing transportation time is very important to the

development of logistics companies. How to reduce the delivery time is the focus of

logistics distribution path optimization. The fourth is the transportation distance.

Compared with the above factors, the transportation distance has less and less

impact on the optimization of logistics distribution path in today's fast traffic

environment. However, the increase of transportation distance will also lead to the

increase of uncontrollable factors. Now the high incidence of traffic accidents and the

increase of road obstacles make the risk cost greatly increase when the transportation

distance becomes longer. Therefore, in the optimization of logistics distribution path,

we should also try to shorten the transportation distance. The fifth is the transportation

cost, which is the key of logistics distribution. After all, logistics distribution is a

business activity that pursues the maximization of benefits. The transportation cost

mainly includes human cost and loss cost. Now the human cost is getting higher and

higher. This problem can be effectively solved by improving the work efficiency of

logistics personnel [20-24].

According to different constraints, distribution problems can be roughly divided into

the following categories: traveling salesman problem (TSP), collection and forwarding

problem, time limited path optimization, multi vehicle path optimization, path

optimization with loading capacity constraints, path optimization with compatibility

constraints, etc [25-29]. Some literatures are divided into static path optimization

problem SVRP and dynamic path optimization problem DVRP according to the

treatment methods of the changes of influencing factors in logistics distribution. In

today's logistics distribution links, most of the logistics distribution path optimization

problems can be regarded as the traveling salesman problem and the problem of

determining the starting and ending points. The continuous and rapid growth of traffic

flow leads to complex and changeable traffic conditions. Therefore, in the logistics

distribution path optimization problem, the research on the logistics distribution path

optimization problem based on dynamic road conditions has very important research

significance and application value.

(2)Logistics distribution path optimization

In modern logistics distribution, goods are often sent directly to customers from

distribution centers or stores. At this time, the form of logistics distribution path

optimization is: knowing the departure and arrival places, selecting the best

transportation route in the urban transportation network, delivering goods in the

shortest time and improving distribution efficiency. The specific form is shown in

Figure 1.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

Figure 1 Schematic diagram of starting and ending points of distribution

As can be seen from Figure 1, distribution vehicles start from the starting point,

each line represents a road section, and nodes represent intersections. The

optimization process is to select the route with the lowest weight value in the urban

transportation network to reach the distribution terminal.

2.2. COMPLEXITY OF LOGISTICS DISTRIBUTION ROUTE

OPTIMIZATION

In recent years, great changes have taken place in the way, environment and

requirements of logistics distribution in China, such as the intelligent management of

modern logistics, the changes of logistics transportation tools, customers' higher

requirements for logistics distribution standards and increasingly complex traffic

conditions. It also puts forward new requirements for the algorithm of logistics

distribution path optimization. The traditional algorithm mainly exists in today's

logistics distribution path optimization: the solution mode is single, the research on the

change of path in path optimization is not accurate enough, and the distribution route

cannot be changed flexibly when the traffic conditions change [30].

(1) Complexity of road condition changes

Road conditions many accidents will have a serious impact on road traffic. For

example, accidents, road construction, infrastructure construction, abnormal weather,

holidays and other events will reduce the road capacity or make it impassable. The

traditional algorithm does not consider or deal with these influencing factors. The

method is simple and rough, and the calculation accuracy of the influence value on

the road condition is very low. In the route optimization, congestion and section

gradient are taken into account, and the equivalent consumption is transformed into a

flat road with a certain length, which is optimized by the traditional algorithm [31-33].

There is also a logistics distribution route optimization algorithm based on travel time

prediction by using the historical average method to predict the road travel time

[34-35]. The processing of these algorithms is relatively simple, and the complexity of

road conditions is not fully considered.

(2) Complexity of logistics distribution

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

In the actual logistics distribution link, the driver in the distribution process, the

route selection mainly depends on past experience or colleagues' suggestions, which

can deliver the goods quickly and effectively [36]. Or in large logistics companies,

provide traffic flow guidance services to drivers through intelligent logistics software.

These methods have achieved good results in practical application, but there are also

some problems. In a new distribution area, drivers need to spend a lot of time getting

familiar with the road, which seriously affects the efficiency of distribution. And

listening to the experience of others can not guarantee the correct and timely choice

in the complex road. In addition, it is difficult for the distribution personnel to

understand and analyze the huge and updated information such as the surge of

logistics business volume and the expansion of service area, the large-scale

construction of cities and the new planning of road construction. The huge logistics

distribution network has difficulty in memorizing information.

3. RESEARCH ON OPTIMIZATION OF LSTM LOGISTICS

DISTRIBUTION PATH BASED ON BLOCKCHAIN

3.1. INTRODUCTION TO REGIONAL CHAIN TECHNOLOGY

Blockchain technology is a decentralized distributed database. A continuous record

storage structure is formed on the block with time stamps. The block contains various

recording applications, such as clearing, smart contract, etc. the data recording node

calculates the hash through a specific algorithm, and the current hash, previous block

hash, data record, etc. are recorded in the block. Such a data system is credible, so

blockchain is a tool for manufacturing credit, and the irreversibility and irreversibility of

blockchain technology are all highlighted.

However, regional chain technology has many advantages. Blockchain adopts

distributed accounting and storage, and there is no centralized hardware or

management organization. Therefore, the rights and obligations of any node are

equal. Moreover, the blockchain system is open in nature. In addition to the private

information of the trading parties being encrypted, the data of the blockchain is open

to all. In the blockchain, any human intervention will not work, changing the trust in

"people" to the trust in machines [37]. It enables all nodes in the whole system to

exchange data freely and safely in the untrusted environment. In addition, once the

information is verified and added to the block in the blockchain, it will be permanently

stored and cannot be modified, which improves the corresponding security. In

addition, there are regular changes in the changes of urban traffic conditions in a

certain period of time. There must be the characteristics of road condition information

in this large amount of traffic data. Compared with the traditional neural network, the

regional chain technology has a stronger ability to extract the characteristics of traffic

data through the setting of multi hidden layer parameters. After the self coding

learning and training of road information features, the internal correlation of road

information features is closer, so that the prediction of future road conditions is more

accurate. Provide accurate road condition parameters for logistics distribution path

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

optimization. Compared with traditional algorithms, the optimization results in actual

distribution are more accurate.

3.2. INTRODUCTION TO LSTM MODEL

LSTM is an improved version of sample RNN. RNN will be affected by short-term

memory, but LSTM can remember longer sequences. During back propagation, RNN

will disappear gradient. If the gradient value becomes very small, it will not continue to

learn. LSTM can solve this series of problems. Figure 2 shows a unit structure of

LSTM model σ And tanh represent the feedforward network layer, where σ

The

activation function in is sigmoid.

Figure 2 Structure diagram of LSTM model unit

The LSTM model has the characteristics of "long-term and short-term memory",

which can be analyzed in this way. It can be seen from Figure 2 that the part shown

by the red line is long-term memory. There is only a small amount of linear interaction

on this line, which can realize the passage of information from the whole cell structure

without change. The part shown by the blue line is short-term memory, which has

three parts. The first part is the forgetting gate, It will use ht-1 and Xt to determine

which information in the previous unit state Ct-1 is removed. The function formula

used is (1):

， (1)

The second part is the input gate, which determines which information is put into

the unit state, mainly including two steps. The first is that the tanh layer generates the

candidate value that can be added to the state, where is shown in formula (2).

The second is that the sigmoid layer generates the activation value it of the input gate

based on ht-1 and Xt, where it is shown in formula (3):

， (2)

, (3)

)(

1ftfhtfxt

bhWXWsigmoidf ++=

−

)tanh(

ctchtcxt

bhWXWC ++=

−

)(

1itihtixt

bhWXWsigmoidi ++=

−

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

The results of the first and second parts jointly determine the new cell state Ct,

where Ct is shown in formula (4):

, (4)

The third part is the output gate, which determines which information in the unit

state is used for production output. The function used is shown in formula (5) (6):

, (5)

. (6)

3.3. LOGISTICS DISTRIBUTION VEHICLE ROUTING PLANNING

MODEL

All learning algorithms are learned through data. By analyzing the data

characteristics of customer requirements, customer location information and demand

information, we quantify the location information and demand information of customer

nodes, and use the deep reinforcement learning method to design an end-to-end

framework to solve the logistics vehicle routing problem [38]. In this method, the

training strategy network and value network model only observe the reward signal and

follow the feasibility rules to find the near optimal solution for the problem samples

sampled from the given distribution. It can be simply explained as a parameterized

probability estimation model based on attention mechanism or actor strategy network

[39]. The structure of strategy network model is shown in Figure 3.

Figure 3 Attention mechanism model based on deep learning reinforcement

learning algorithm

1ttttt CiCfC  += −

)(

1GtGhtGxt

bhWXWsigmoido ++=

−

)tanh( ttt Coh =

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

This model includes three parts: customer state information fusion module,

attention mechanism module and LSTM module. The customer state information

fusion module can transform the structured local and global data information into high-

dimensional vectors and fuse the information. The attention mechanism module can

estimate the probability of each node I. The LSTM module can remember a longer

sequence of customer nodes and improve the solution quality of the model to deal

with the customer scale of unmanned fleet and distribution path.

(1)

The first part of the model is to map the local and global state information of

reinforcement learning into a high-dimensional vector space through one-

dimensional convolution operation. In this process, the abstract features of customer

location time interval, customer goods demand and unmanned vehicle loading are

extracted through deep neural network, as shown in Figure 4.

Figure 4 Reinforcement learning state input information fusion module

For customer i, its local information can be expressed by formula (7) and mapped

into vector with

dimension through one-dimensional convolution. In order to

reduce the amount of parameters, all customers share the weight of one-dimensional

convolution. Another one-dimensional convolution layer is used to map the global

variable to the vector space of dimension

, where the global variable is shown in

equation (8).The dynamic and static elements contained in local information and

global information learn their characteristics in two different convolution layers, and

finally fuse them, just like multi-layer information fusion in supervised learning

resnet50, and and

are combined in a weighted linear way through the ReLU

activation function, and finally the fusion information

of customer node i is

obtained. As shown in equation (9), it contains both global information and local

information.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

. (7)

. (8)

. (9)

The attention mechanism can effectively deal with the expansion of the scale of

customer nodes. When decoding customer node i, we pay more attention to the part

of customers entering the customer node. The processing ability of the attention

mechanism in the face of the expansion of the scale of customer nodes is used to

calculate the access probability of each node i. Firstly, the attention weight is

calculated. Firstly, the similarity between the hidden state ht of the recurrent neural

network and the fusion information of each node I is calculated to obtain . Then,

after obtaining the similarity between each customer node i and the hidden state, the

attention weight

is obtained by softmax normalization transformation, and the

obtained attention weight is used to calculate the weighted sum of the input

information

to obtain ct, and then the probability distribution of each customer

node I is obtained [40]. The calculation diagram of attention mechanism for solving the

route planning of unmanned logistics distribution fleet is shown in Figure 5, and the

calculation process is shown in formulas (10) - (16).

Figure 5 Calculation diagram of attention mechanism for solving logistics

distribution path planning

, (10)

, (11)

, (12)

here ht is the hidden state output of recurrent neural network LSTM，

is the i-th

term of vector , tanh is a nonlinear activation function，

are a trainable

variable， is the number of customer nodes，among them are:

),,,,( t

iiiii

idleyxX =

},{

ttt

cηG=

)

(Re 21

iGθXθLUµ +=

]),[(

iuv

hµθyanhθv=

)max( tt

ivSofta =

∑

tµac

uv θθ ,

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

, (13)

, (14)

Then calculate the probability estimate of each customer, where

is the i-th

element of ， are trainable variables：

, (15)

. (16)

(3) In Section 3.1, the long-term and short-term memory neural network is

introduced. Its input has three parts. In the specific application here, the hidden state

output ht-1 at the previous time, the cell state ct-1 at the previous time. The input of

LSTM is representing the information of customer node yt and the hidden state

ht-1 of the previous time. The output hidden state ht is used as the input of attention

mechanism to calculate attention weight.

3.4. REINFORCEMENT LEARNING ALGORITHM

Deep reinforcement learning algorithm based on round trajectory reward is used to

train strategy network and value network model [41-43], the main algorithms are as

follows: first initialize the network weight parameters θ,j, generate N VEPTW training

instances and iterate circularly epoch→∞，regenerate into a batch of training samples

(M training samples from N) , loop n=1,2,...M, select according to

the model probability distribution , until all node requirements are

0，then calculate track reward R [44-45]. The training results are obtained according

to equations (17) - (18):

, (17)

. (18)

4. EXPERIMENTAL RESULTS AND ANALYSIS

This experiment considers the practical problems of urban logistics distribution,

such as the loading capacity of logistics vehicles. For different sizes of distribution

tasks, use distribution vehicles with different loading sizes to carry out numerical

simulation experiments on the distribution tasks with customer sizes of 5, 15 and 25 in

the end area of the city, and select the distribution vehicle models with unmanned

−

=)tanh(

∑

xSoft )max(

cg θθ ,

]),[tanh( t

icg

icµθθg=

)max( tt

igSoftp =

][]2[]1[ ,...,, M

XXX

][

),,( ][][][][

1nnnn

nθYGXyP

)(log)]()([

][][

][][ ii

XYPXvYR

θϕ

θ ∇−=∑

[ ]

∑

−∇=

ii XvYR

][][ )()(

ϕϕ

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

vehicle loading capacity of 15, 25 and 35 respectively. The experimental results will

give the optimized distribution path.

Using the training set to verify the test set data, 18000 instances are generated as

the training set, 180 instances are the verification set and 180 instances are the test

set. Each instance is a logistics vehicle distribution task problem with fixed customer

size and distribution center. The customer node data of verification set and test set

are generated in the same way as the data of training set. The information of

customer i is expressed by equation (19):

, (19)

The customer demand is randomly selected from the discrete number {1,2,3,4,5,6}

and the location

of the customer and the distribution center is randomly generated

in the surface space of

. This study uses this to simulate the urban area. For

the three customer scale distribution tasks, logistics distribution vehicles with loading

capacity of 15, 25 and 35 are selected respectively.

Table 1 Vehicle information of different customer sizes

The training set generates 18000 instances. Each iteration batch selects 180

vrptw-5 from 18000 instances as a training batch. Each training batch makes a

gradient update to the strategy network and evaluation network. After completing all

the instance data, it is used as an iterative epoch. After each epoch is completed, the

model is tested with the verification set. The verification set of this study is composed

of 180 instances. The running time of the verification set is collected, Average fleet,

total distribution mileage, average reward and other information to evaluate the model.

Set 25 epochs for iterations, and each iteration completes one epoch to verify the

verification set. When 25 iterations are completed, the model training is completed,

and the test model stage is entered. 180 instances are randomly generated from the

data of the test set and the verification set in the same way, and finally the optimal

logistics distribution path is obtained.

The global and local information studied are mapped into 128 dimensional vectors

through two different one-dimensional convolutions, and the hidden state output of

LSTM is 128 dimensions. All trainable variables begin to officially enter the training

stage after a period of pre training. In order to make the model have generalization

ability, let the agent contact more environmental conditions and diversify the situations

that the model may encounter, In training, this study adopts the strategy of random

sampling. When testing, it adopts random decoding and greedy decoding, and

compares the advantages of the two decoding strategies. Experiments are carried out

on three VRPTW problems. The scale of customer service nodes are 5, 15 and 25

),,,,( t

iiiii

idleyxX =

yx ,

]1,0[]1,0[ ×

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

respectively. For the problem of each customer scale, 180 examples are tested and

solved.

After the model training, the test set data is used to test the model. The test

decoding strategy adopts random decoding and greedy decoding. Figure 6 shows

different distribution route schemes to achieve the optimal logistics distribution route.

Due to the different tasks to be performed by different distribution vehicles and the

different capacity of distribution vehicles, the optimal route will also be different. Three

recommended routes are generated from each different logistics distribution starting

point to the destination. Due to the complexity of the road and the problems of

selection, it can be seen that figure 6 (a) shows three logistics vehicle distribution

routes under vrptw-5: blue line {1,2,3,4}, red line {5,6,7,8}, green line {9,10,11}, and

blue line is the optimal logistics vehicle distribution route; Figure 6 (b) shows three

logistics vehicle distribution paths under vrptw-15: blue line {1,2,3,4}, red line {5,6,7,8},

green line {9,10,11,12,13}. Red line is the optimal logistics vehicle distribution path;

Figure 6 (c) shows three logistics vehicle distribution paths under vrptw-25: red line

{1,2,3,4,5}, blue line {6,7,8,9}, green line {10,11,12,13}. Blue line is the optimal

logistics vehice distribution path.

Figure 6 Logistics distribution optimization path with (a)VRPTW-5; (b)VRPTW-15;

(c)VRPTW-25

5. IN CONCLUSION

With the rapid development of modern society and entering the Internet era, online

shopping has become people's daily life, which makes the logistics industry more and

more prosperous. At the same time, it will also put forward higher requirements for

logistics distribution, especially in path planning, the logistics distribution method of

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

applying science and technology has played a more and more important role. The

research results of in-depth learning and reinforcement learning have gradually

appeared in the research of artificial intelligence method in vehicle routing problem.

Compared with the traditional problem-solving algorithm, the method based on

regional chain technology to solve the path optimization problem is more attractive.

Therefore, a deep reinforcement learning algorithm based on attention mechanism

and LSTM model is designed and applied to the distribution path planning problem of

logistics vehicles. The training set is designed for sample test training. 18000

examples are generated from the training set. Experiments are carried out on three

VRPTW problems, and then iterative calculation is carried out. Each of them obtains

three recommended paths of logistics distribution vehicles, Finally, the shortest

optimized logistics distribution path is obtained, which promotes the logistics

distribution technology in the information age.

REFERENCES

(1) Chen, J., Xu, S., Chen, H., Zhao, C., & Xue, K. (2020, May).

Research on

optimization of food cold chain logistics distribution route based on

internet of things. In Journal of Physics: Conference Series (Vol. 1544, No. 1,

p. 012086). IOP Publishing.

(2) ZHAO, Z. X., & LI, X. M. (2020).

Electric Vehicle Route Optimization for

Fresh Logistics Distribution Based on Time-varying Traffic

Congestion.

Journal of Transportation Systems Engineering and Information

Technology, 20(5), 218.

(3) Liu B. Logistics Distribution Route Optimization Model Based on Recursive

Fuzzy Neural Network Algorithm[J].

Computational Intelligence and

Neuroscience, 2021, 2021.

(4) Xiong H. Research on cold chain logistics distribution route based on ant

colony optimization algorithm[J]. Discrete Dynamics in Nature and Society,

2021, 2021.

(5) Liu W.

Route optimization for last-mile distribution of rural E-commerce

logistics based on ant colony optimization[J]. IEEE Access

, 2020, 8:

12179-12187.

(6) Zhao Z, Li X, Zhou X. Distribution route optimization for electric vehicles in

urban cold chain logistics for fresh products under time-varying traffic

conditions[J]. Mathematical Problems in Engineering, 2020, 2020.

(7) Sun R, Liu M, Zhao L. Research on logistics distribution path optimization

based on PSO and IoT[J].

International Journal of Wavelets, Multiresolution

and Information Processing, 2019, 17(06): 1950051.

(8) Yuanguo Y, Shengyu H.

Research on optimization of distribution route for

cold chain logistics of agricultural products based on traffic big data[J].

Management Review, 2019, 31(4): 240.

(9) Wei Xiaodi, Zheng Hongqing Improved discrete flower pollination algorithm

for solving vehicle routing problem with time window [J].

Practice and

understanding of mathematics, 2020, 50 (02): 195-202.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(10) Su Xinxin, Qin Hu, Wang Kai

Tabu search algorithm for solving vehicle

routing problem with time window and multiple distribution personnel [J].

Journal of Chongqing Normal University (NATURAL SCIENCE EDITION)

, 2020,

37 (1): 22-30.

(11) DENG A M, MAO C, ZHOU Y T.

Optimizing research of an improved

simulated annealing algorithm to soft time windows vehicle routing

problem with pick-up and delivery[J]. S

ystems Engineering-Theory &

Practice, 2009, 29(5):186-192.

(12) CORDEAU J F, LAPORTE G, MERCIER A.

A unified tabu search heuristic for

vehicle routing problems with time windows[J].

Journal of the Operational

Research Society, 2001, 52(8):928-936.

(13) Wang H, Li W.

Study on logistics distribution route optimization by

improved particle swarm optimization[J]. Computer Simulation

, 2012, 29(5):

243-246.

(14) Chen Y, Huang Z, Ai H, et al.

The impact of GIS/GPS network information

systems on the logistics distribution cost of tobacco enterprises[J]

Transportation Research Part E: Logistics and Transportation Review

, 2021,

149: 102299.

(15) Zhao Y, Zhou Y, Deng W.

Innovation mode and optimization strategy of B2C

e-commerce logistics distribution under big data[J]. Sustainability

, 2020,

12(8): 3381.

(16) Liu T, Yin Y, Yang X.

Research on logistics distribution routes optimization

based on ACO[C].

2020 5th International Conference on Information Science,

Computer Technology and Transportation (ISCTT). IEEE, 2020: 641-644.

(17) Liu H, Pretorius L, Jiang D.

Optimization of cold chain logistics distribution

network terminal[J].

EURASIP Journal on Wireless Communications and

Networking, 2018, 2018(1): 1-9.

(18) Ouyang F.

Research on port logistics distribution route planning based on

artificial fish swarm algorithm[J]. Journal of Coastal Research

, 2020, 115(SI):

78-80.

(19) Li J.

Logistics distribution route optimization based on chaotic cuckoo

algorithm[C]. Advanced Materials Research. Trans Tech Publications Ltd

, 2014,

1049: 1681-1684.

(20) Ding D, Zou X.

The optimization of logistics distribution route based on

Dijkstra’s algorithm and CW savings algorithm[C].

6th International

conference on machinery, materials, environment, biotechnology and computer

(MMEBC 2016), Tianjin, China. 2016: 11-12.

(21) Bosona T, Gebresenbet G, Nordmark I, et al.

Integrated logistics network for

the supply chain of locally produced food, Part I: Location and route

optimization analyses[J]. Journal of service science and management

, 2011,

4(02): 174.

(22) Lozano Murciego Á, Jiménez-Bravo D M, Pato Martínez D, et al.

Voice

assistant and route optimization system for logistics companies in

depopulated rural areas[J]. Sustainability, 2020, 12(13): 5377.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(23) He R, Ma C, Ma C, et al.

Optimisation algorithm for logistics distribution

route based on Prufer codes[J].

International Journal of Wireless and Mobile

Computing, 2015, 9(2): 205-210.

(24) Konstantakopoulos G D, Gayialis S P, Kechagias E P.

Vehicle routing problem

and related algorithms for logistics distribution: A literature review and

classification[J]. Operational research, 2020: 1-30.

(25) Bin Z, Xiao-Jun L.

Study on logistics distribution route optimization based

on clustering algorithm and ant colony algorithm[J].

The Open Cybernetics

& Systemics Journal, 2015, 9(1).

(26) Sun Y, Geng N, Gong S, et al.

Research on improved genetic algorithm in

path optimization of aviation logistics distribution center[J].

Journal of

Intelligent & Fuzzy Systems, 2020, 38(1): 29-37.

(27) Fan Q, Nie X X, Yu K, et al.

Optimization of logistics distribution route based

on the save mileage method and the ant colony algorithm[C].

Applied

Mechanics and Materials. Trans Tech Publications Ltd, 2014, 448: 3683-3687.

(28) Damayanti T R, Kusumaningrum A L, Susanty Y D, et al.

Route optimization

using saving matrix method–A case study at public logistics company in

Indonesia[C].

International Conference on Industrial Engineering and

Operations Management. 2020: 1583-1591.

(29) Wang C, Zhou J, Xiao B, et al.

Uncertainty Estimation for Stereo Matching

Based on Evidential Deep Learning. Pattern Recognition, 2021. https://doi.org/

10.1016/j.patcog.2021.108498

(30) C. Yan, G. Pang, X. Bai, et al.,

Beyond triplet loss: person re-identification

with finegrained difference-aware pairwise loss,

IEEE Trans. Multimedia

(2021), https://doi.org/10.1109/TMM.2021.3069562.

(31) Zhang K, Qiu B, Mu D.

Low-carbon logistics distribution route planning with

improved particle swarm optimization algorithm[C].

2016 International

Conference on Logistics, Informatics and Service Sciences (LISS). IEEE

, 2016:

1-4.

(32) Zheng C, Gu Y, Shen J, et al.

Urban Logistics Delivery Route Planning

Based on a Single Metro Line[J]. IEEE Access, 2021, 9: 50819-50830.

(33) Yao C.

Research on logistics distribution path analysis based on artificial

intelligence algorithms[J]. International Journal of Biometrics

, 2020, 12(1):

100-108.

(34) Bai X, Zhou J, Ning X, et al. 3D data computation and visualization[J]

Displays, 2022: 102169.

(35) Zhang Wenguang.

Research on terminal distribution mode and route

optimization of e-commerce logistics based on ant algorithm [D].

Guizhou

Normal University, 2016.

(36) W Cai, B Zhai, Y Liu, et al.,

Quadratic polynomial guided fuzzy C-means and

dual attention mechanism for medical image segmentation. Displays

, vol.

70, no. 102106, 2021. https://doi.org/10.1016/j.displa.2021.102106.

(37) X. Ning, P. Duan, W. Li, and S. Zhang,

Real-time 3D face alignment using an

encoder-decoder network with an efficient deconvolution layer,

IEEE Signal

Processing Letters, vol. 27, pp. 1944–1948, 2020. https://doi.org/10.1109/

LSP.2020.3032277.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023

(38) Jiao Yunlong.

Research on logistics transportation vehicle routing

optimization based on travel time prediction［D]. Dalian Maritime University,

2015.

(39) Martins L B, Iori M, Moreira, Mayron César O, et al.

On Solving the Time

Window Assignment Vehicle Routing Problem via Iterated Local Search.

2021.

(40) J. Miao, Z Wang, X. Ning, X. Nan, W. Cai, R. Liu.

Practical and Secure

Multifactor Authentication Protocol for Autonomous Vehicles in 5G.

Software:Practice and Experience, 2022. https://doi.org/10.1002/SPE.3087

(41) X. Ning, K. Gong, W. Li, and L. Zhang,

JWSAA: Joint Weak Saliency and

Attention Aware for person re-identification, Neurocomputing, 2021, vol. 453,

pp. 801-811. https://doi.org/10.1016/j.neucom.2020.05.106.

(42) Shan W.

Digital streaming media distribution and transmission process

optimisation based on adaptive recurrent neural network[J]. Connection

Science, 2022, 34(1): 1169-1180.

(43) Ying L, Qian Nan Z, Fu Ping W, et al.

Adaptive weights learning in CNN

feature fusion for crime scene investigation image classification[J].

Connection Science, 2021, 33(3): 719-734.

(44) Medina, R., Breña, J. L., y Esenarro, D. (2021).

Efficient and sustainable

improvement of a system of production and commercialization of

Essential Molle Oil (Schinus Molle). 3C Empresa. Investigación y pensamiento

crítico, 10(4), 43-75. https://doi.org/10.17993/3cemp.2021.100448.43-75

(45) Ma Yanran,Chen Nan & Lv Han.(2021).

Back propagation mathematical

model for stock price prediction.

Applied Mathematics and Nonlinear

Sciences(1). https://doi.org/10.2478/AMNS.2021.2.00144.

https://doi.org/10.17993/3cemp.2023.120151.68-85

3C Empresa. Investigación y pensamiento crítico. ISSN: 2254-3376

Ed. 51 Iss.12 N.1 January - March, 2023