Hindawi

Journal of Control Science and Engineering

Volume 2020, Article ID 3857894, 12 pages
https://doi.org/10.1155/2020/3857894

Research Article
An Improved Method of Particle Swarm Optimization for Path
Planning of Mobile Robot

Xun Li,1,2 Dandan Wu ,1 Jingjing He,1 Muhammad Bashir,1 and Ma Liping1

1
College of Electronics and Information, Xi’an Polytechnic University, Xi’an, China
2
Bernoulli Institute, University of Groningen, Groningen, Netherlands

Correspondence should be addressed to Dandan Wu; nakedbears@163.com

Received 6 February 2020; Revised 16 April 2020; Accepted 8 May 2020; Published 25 May 2020

Academic Editor: Carlos-Andrés Garcı́a

Copyright © 2020 Xun Li et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
The existing particle swarm optimization (PSO) algorithm has the disadvantages of application limitations and slow convergence
speed when solving the problem of mobile robot path planning. This paper proposes an improved PSO integration scheme based
on improved details, which integrates uniform distribution, exponential attenuation inertia weight, cubic spline interpolation
function, and learning factor of enhanced control. Compared with other standard functions, our improved PSO (IPSO) can
achieve better optimal results with less number of iteration steps than the different four path planning algorithms developed in the
existing literature. IPSO makes the optimal path length with less than 20 iteration steps and reduces the path length and simulation
time by 2.8% and 1.1 seconds, respectively.

1. Introduction and local search imbalance, low convergence precision, easy

falling into optimal local solution, and poor robustness.
Path planning is different from motion planning where To obtain better optimization results, a PSO optimiza-
dynamics must be considered. Its purpose is to find the tion technique is proposed in [12] to converge at the global
optimal path of motion in the least amount of time and to minimum, and a custom algorithm is used to generate the
model the environment completely [1]. For the path plan- coordinates of the search space. The coordinate values
ning problem of mobile agents, several researchers have generated by the custom algorithm are passed to the PSO
proposed many algorithms, which can be classified into two algorithm, which uses these coordinate values to determine
categories, i.e., traditional path planning methods and bionic the shortest path between two given end positions. Thus, it is
intelligent algorithm-based methods. The former method not limited to only finding the optimal value. Still, it can
includes the A∗ algorithm [2], Dijkstra [3], RTT [4], and improve the speed of the algorithm. However, the direct
artificial potential field method [5]. Bionic intelligent al- transfer of its two-point coordinates is easy to fall into the
gorithms include differential evolution algorithm [6], ge- local optimum, and the obtained optimal value is often
netic algorithm [7], ant colony algorithm [8], artificial fish considered as the global suboptimal value; the literature in
swarm algorithm [9], and PSO [10]. PSO algorithm is widely [13] proposed a random disturbance method.
used in the practical application and theoretical research of The adaptive PSO optimization method introduces a
mobile agent path planning due to its strong searchability, disturbing global update mechanism to the optimal global
fast convergence speed, and high efficiency [11]. The PSO is a position, which prevents the algorithm stalling. Besides, a
population-based randomization technique to find the op- new adaptive strategy is proposed to fine-tune the three
timal value for the foraging of the birds. PSO has the ad- control parameters in the algorithm. Moreover, the need for
vantages of fast search speed, memory, few parameters, and dynamic adjustment and exploration of the three parameters
simple structure and has easier implementation at the stage increases the computational complexity of the optimization
of validation. Besides, its shortcomings include global search algorithm as well as low execution efficiency. The work in
2 Journal of Control Science and Engineering

[14] proposed a modified particle swarm optimization population for the best companion to learn. That is, the next
(MPSO) with constraints to jointly obtain a smooth path, step of the particle is determined by the experience and the
but the method does not improve the efficiency of the best experience of the companion:
particle diversity, which makes the particles stagnate and fall
Vk+1 k k k k k
id � ωVid + c1 r1 􏼐Pid − Xid 􏼑 + c2 r2 􏼐Pqd − Xid 􏼑, (1)
into the optimal local solution. The studies in [15] proposed
a fusion of chaotic PSO and ant colony algorithm (ACO).
The algorithm effectively adjusts the particle swarm opti- Xk+1 k k+1
id � Xid + Vid , (2)
mization parameters. It reduces the number of iterations of
the ant colony algorithm, called the chaotic particle swarm where Vkid is the dth component of the flight velocity vector
algorithm, thereby effectively reduces the search time. of the particle i of the kth iteration; Xkid indicates the dth
However, due to the improved algorithm parameters, its element of the position vector of the particle i of the kth
speed and the position updates are proportional, and this iteration; c1 and c2 represent the learning factor, which is
can limit the particle global searchability. Moreover, it is a used to adjust the learned to the maximum step size; r1 and
solution that can fall into the optimal local solution. r2 are two random functions with a range of value from 0 to
The hybrid genetic particle swarm optimization algo- 1 and are used to increase the randomness searching; and ω
rithm (GA-PSO) is proposed in [16] that established a path is the inertia weight to adjust the searchability for the so-
planning. The mathematical model of the problem proposed lution space. Although the classical PSO is simple to im-
a time-first based particle-first iteration mechanism, which plement and has few adjustment parameters when it is used
makes the evolution process more directional and acceler- in the path planning method, it is prone to poor search-
ates the development of path planning problems. However, ability, falls into the optimal local solution, and has reduced
the hybrid algorithm has too many parameters that need particle diversity, low convergence precision, and low ac-
control. This increases the computational complexity and curacy of path planning. Therefore, this paper systematically
has reduced the execution efficiency of the algorithm. It improves the classical PSO algorithm by combining various
signifies that it is difficult to improve the diversity of particles improved methods.
due to the easy way of falling into the optimal local solution.
Therefore, to combine the advantages of the above- 2.2. Improvement of PSO Algorithm Integration
mentioned various improved algorithms and improve their
defects, this paper proposed a new method to improve the 2.2.1. Uniform Distribution. We aim to ensure that the PSO
PSO integration scheme based on improved details, which is algorithm does not lose the randomness of the particles
used to solve the global path planning of mobile robots in the when initializing the population and, at the same time, avoid
indoor environment. In the proposed algorithm, the navi- the problem of excessive concentration of the initial posi-
gation point model is selected as the working area model of tions of the particles, which is not conducive to global search
the mobile robot, and the uniform distribution, exponential and postprocessing, especially with the possibility that the
attenuation inertia weight, cubic spline interpolation search stagnation may occur in the late search. In this paper,
function, and learning factor of enhanced control are in- continuous uniform random distribution is used to initialize
troduced to PSO to improve its performance. Finally, the the particles so that the particles are relatively uniformly
advancement and effectiveness of the standard test function distributed in the search space to facilitate later search and
and obstacle environment are verified in the paper. The avoid particles falling into the local optimal solution. The
results show that, compared with other standard functions, formula for uniform position distribution is shown as
IPSO achieves better optimal results and more iteration follows:
steps. Compared with the other four path planning algo- x − xmin
rithms, IPSO reaches the optimal path length with less than x � max , (3)
n−1
20 iteration steps and reduces the path length and simulation
time by 2.8% and 1.1 seconds, respectively. ymax − ymin
y� , (4)
n−1
2. Improved PSO (IPSO) where xmax and ymax are the upper limits of the variables,
xmin and ymin are the lower limits of the variables, and n is
2.1. Classical PSO. The fundamental core of the PSO method the number of the decision variables.
is to share the information through the individuals in the
group so that the movement of the whole group can be
transformed from disorder to order in the problem of so- 2.2.2. Exponential Decay Inertia Weight. The change of the
lution space to obtain the optimal solution of the problem. inertia weight w affects the position of the particle. The larger
The answer to each optimization problem is the “particle” the value of w, the stronger the global searchability and the
speed and position update through (1) and (2). The first term weaker the local mining ability. Therefore, good results can
of (1) indicates that the next move of the particle is affected be obtained when the values of the w are dynamic (ad-
by the magnitude and direction of the last flight speed; the justable) compared to the fixed values. The value of w can
second term means that the following action of the particle vary linearly during the PSO search process, or dynamically
originates from its own experience; the third term indicates based on a measure function of PSO performance. By an-
that the next move of the particle originates from the alyzing and summarizing the linear decreasing inertia
Journal of Control Science and Engineering 3

weight proposed in [17], the improved particle swarm c1 � cos(ω), (7)

optimization algorithm based on the improved inertia
weight proposed in [18], and the enhanced method of c2 � sin(ω), (8)
fuzzy inertia weight strategy proposed in [19], this paper
presents an improvement in the above-mentioned
1 − ω > 0,
methods. The fixed, predictable, and iterative step adopted 􏼨 (9)
in the previous ways is a global transformation that is not 2ω + 2 > rand1 × c1 + rand2 ∗ c2 .
conducive to the particle, such as fixed transformation
The cosine function in (7) is a subtraction function in
that can narrow the search range of the particle and affect
the interval (0, π/2), which has a significant value at the
the diversity of the particle. This paper uses the study in
beginning of the search, and the value is decreasing in the
[20] to introduce the inertia weight of exponential decay
latter part of the search. The sine function in (8) is an
which is more in line with the characteristics of the ex-
increasing function in the interval (0, π/2) with a small
ponential function. By adopting the changes of the pre-
value before the quests, and its value increases after
steps and poststeps to the search area, its speed can be
searching. Not only does it satisfy the condition that the
successfully updated at different periods. This eliminates
PSO algorithm has better learning ability in the optimi-
the premature particles and improves the robustness of
zation process, but it also changes the inertia weight and the
the algorithm. Henceforth, the global search and local
learning factor into one variable, which is convenient for
mining ability are maintained, and this can solve the
practical application and strengthens the uniformity of the
problems mentioned above to some extent. The expres-
algorithm optimization process. Equation (9) is the guiding
sion is as follows:
condition for selecting and adjusting the PSO parameters
ln ωmax − ln ωmin given in [22].
A � it − ln ωmax , (5)
MaxIt

ω � exp(−A), (6) 2.2.4. Improvement Based on Robot Dynamics Requirements

where ωmin is the minimum weight, ωmax is the maximum (1) Cubic Spline Interpolation. The experimental results show
weight, as well as the current number of iterations, MaxIt is that (see Section 4.2 for details) the path planning of the
the maximum number of iterations, and it is the current improved PSO has more turning points with a rough path,
number of iterations. which will affect the dynamic characteristics of the robot
when it is moved. Therefore, it is necessary to improve the
aforementioned improved PSO algorithm further to have a
2.2.3. Improved Learning Factors. The standard PSO smooth path with the improved robot dynamic adaptability
expressed in (1) represents the acceleration weight of each of the algorithm requirements. The cubic spline curve is
particle moving to the optimal local value and the optimal fitted by several interpolation intervals based on cubic
global value. Lower values allow the particles to linger polynomials to provide a smooth curve, defined as follows:
outside the target area, while the higher values are causing On the range [a, b], n + 1 takes a node, and the node
the particles to suddenly rush toward or cover the target coordinates are (x1 , y1 ), (x2 , y2 ), ..., (xn , yn ). If the follow-
area. Several methods have been effectively improved by the
ing conditions are met, it is called a cubic spline function.
learning factor, although the learning factor and inertia
weight weaken the uniformity of the optimization algorithm (i) Within each cell (xi , xi+1 ), where i � 0, 1, . . ., n, there
to a certain extent. Hence, it is not conducive to the global is a cubic polynomial:
and local search algorithms. 2
zi � ai + bi x − xi 􏼁 + ci x − xi 􏼁 + di x − xi 􏼁 .
3
(10)
The work in [21] analyzes the PSO with a compression
factor that considered the learning factor as a function of
inertia weight. This kind of method has the advantage of (ii) The function and its first and second derivatives are
transforming the three variables involving inertia weight continuous at the interpolation point.
and learning factor into one variable, which not only (iii) zi (x) commonly uses endpoint conditions that can
facilitates practical application but also enhances the satisfy the following three requirements:
uniformity of the algorithm optimization process.
(a) Free boundary: the second derivative at the
However, the function used is complicated, costly, and
endpoint is zero.
time-consuming. Besides, the possibility of particle
(b) Fixed limitation: the range value of the differ-
stagnation in the later search is high, increasing the risk of
ential function from the beginning to the end is
particles falling into optimal local values. Therefore,
specified.
according to the characteristics of high self-learning
(c) Nonnode boundary: the third derivative at the
ability in the early search period and high demand for
2nd to the last node is continuous.
“social learning ability,” this paper adopts the dynamic
method to improve the learning factor; its expressions are
as shown in (7) and (8). It is shown that this expression is (2) Particle Coding. Particle coding is the assignment of
constrained by (9): coordinate positions to several path nodes. The path node is
4 Journal of Control Science and Engineering

the intersection of any two cubic spline intervals. Path nodes Step 4: calculate the fitness value of the particle
can be selected arbitrarily or according to the environment. according to (11).
Suppose that the coordinates of n path nodes are Step 5: update the velocity and position of the particle
(xn1 , yn1 ), (xn2 , yn2 ), ..., (xnn , ynn ). The start and end coor- according to (1), (2), and (5) to (9), respectively. Update
dinate of the trail are (xs , ys ), (xt , yt ). Interpolating the the local optimal value Pbest and the global optimal
coordinates of the interpolation points with cubic spline value Gbest.
interpolation is on the interval of
Step 6: determine whether the updated particle in-
(x1 , y1 ), (x2 , y2 ), . . . , (xm , ym ). The path of the particle code
tersects the obstacle according to (13), and obtain a path
planning is the connection of the interpolation points. This
composed of the path node coordinates after updating.
paper uses the path node, the interpolation point, and the
The number of iterations is increased by one.
link of the start and endpoints of the path as the running
trajectory of the robot. Step 7: if the termination condition is met (the
threshold error is good enough to be negligible.), then
(3) Evaluation Function. This paper provides the shortest the algorithm ends, and the optimal path is output.
way that does not intersect with the obstacle. The design Otherwise, it goes to Step 3 and repeats the procedure.
evaluation function of our proposed algorithm is as The flowchart of the algorithm is shown in Figure 1.
shown in (11), where L is the path length of the mobile
agent from the ith path point to the i + 1th path point, 4. Simulation Experiment and Result Analysis
which is often used as an index in the path planning, and
its mathematical expression is given in (12); xi and yi are To verify the effectiveness and feasibility of the improved
the coordinates of the ith path point; xi+1 and yi+1 are the algorithm, the improved algorithm of this paper (denoted as
coordinates of the i + 1th path point; α is the weight IPSO), classical PSO algorithm (referred to as PSO), liter-
coefficient and is set to 100 here, which is used to exclude ature [23] stochastic inertia weight PSO algorithm (indicated
the illegal path, i.e., the way through the obstacle; and P is as Rand PSO), literature [24] trigonometric learning factor
a penalty function of obstacle avoidance constraint in the PSO algorithm (denoted as TFPSO), and research [25] PSO
selected indoor environment model, which is used to set with contraction factor algorithm (recorded as NCFPSO),
up a safe distance. The calculation formula is as shown in the optimal values of the typical functions are compared and
(13), where Rm is the radius of the mth obstacle, m is the analyzed through path planning experiment. The simulation
number of obstacles, and c and d are the center coor- experiment environment is carried out on Windows 10, core
dinates of the obstacle; the smaller the value of P, the i7, CPU (2.2 GHZ), memory 8 GB, Matlab (2018a).
higher the safety coefficient of the final path. If the trail
passes the mth obstacle, it is a number greater than 0. If 4.1. Standard Test Function Optimization. At present, several
the obstacle is avoided, then the value is 0: researchers work on intelligent bionic algorithms to evaluate
and compare the performance of algorithms by finding the
F � L ×(1 + α × P), (11) optimal values for the five typical functions. The Ackley function
is generally used to detect the global convergence rate of the
n 􏽱���������������������
2 2 algorithm, the Rastrigin function is used to find the optimal
L�􏽘 xi+1 − xi 􏼁 + yi+1 − yi 􏼁 , (12) global value of the algorithm, and the Griewank function detects
i�1
the ability of the algorithm to jump out of the optimal local
􏽱����������������� value. The Sphere and Rosenbrock functions are unimodal
m 2 2
⎜ ⎜ xi − c􏼁 + yi − d􏼁 ⎟ ⎟ functions that are used to solve optimal global solutions.
P � 􏽘⎛
⎝MAX⎛
⎝1 − ⎠⎞
, 0⎞ ⎠. (13)
m�1 Rm In this paper, the above five functions are used as objects
to compare and evaluate the effectiveness of our proposed
algorithm. The basic mathematical of the basic functions
properties are shown in Table 1.
3. Improved Algorithm Flow
This section provides the detailed step by step explanation of 4.1.1. Standard Test Function and Parameter Setting. This
how to improve the PSO algorithm, as stated in Section 2. section compares the performance of our proposed algo-
Step 1: the number of path nodes, together with the rithm with some of the developed algorithms in the current
number of interpolation, is determined according to literature. The maximum number of iterations of each al-
the actual environment. gorithm is set to MaxIt � 1000, the population size is set to
Npop � 50, the dimension is set to D � 30, and the lower and
Step 2: set the parameters of the particle uniformly and upper limit of the dynamic inertia weight are set as wmin �
initialize the particle position using (3) and (4), re- 0.4 and wmax � 0.7, respectively. The five algorithms were
spectively. Then, the particle population and velocity run 20 times, and the optimal values, mean values, standard
are initialized. deviations, and average simulation run times are computed
Step 3: compute the coordinates of m interpolation based on the experimental results for each algorithm, re-
points in x and y directions of each particle. spectively. The performance of our proposed (IPSO)
Journal of Control Science and Engineering 5

Start

Determine the number of path nodes and

interpolation

Set and initialize the various parameters, and uniformly initialize

the position of the particles according to equations (3) and (4)

Number of iterations plus 1 Number of iterations plus 1

Calculate the coordinates of the interpolation point in the x and y
directions

Calculate the individual fitness value according to equation (11)

Update the speed and position of Pbest and Gbest according to

equations (1) and (2) and (5) ~ (9), respectively

i ≥ NP?

Met the end conditions

Output optimal path

End

Figure 1: Flowchart of improved PSO algorithm.

Table 1: Standard test function.

Optimal Range of independent
Function name Expression
solution variables
􏽱�����������
Ackley f(x) � −20 exp(−0.2 (1/D) 􏽐D 2 D
i�1 xi ) − exp((1/D) 􏽐i�1 cos 2 πxi ) + 20 + e 0 (−32, 32)
√�
Griewank f(x) � (1/4000) 􏽐D 2 D
i�1 xi − 􏽑i�1 cos(xi / i ) + 1 0 (−600, 600)
Rastrigin f(x) � 􏽐D 2
i�1 (xi − 10 cos 2 πxi + 10) 0 (−5.12, 5.12)
Rosenbrock f(x) � 􏽐D−1
i�1 100(xi+1 − x2i )2 + (1 − xi ) 2
0 (−2.048, 2.048)
D
Sphere f(x) � 􏽐i�1 x2i 0 (−100, 100)

algorithm was evaluated using the experimental results of but it can be seen from the test curve in the subgraph of
these four aspects and the iterative curve. Figure 2(a) that our proposed IPSO algorithm has the
fastest convergence and highest precision among all the
four algorithms developed in the literature. The subgraph
4.1.2. Comparative Analysis of Algorithm Simulation Results, depicted in Figure 2(b) is the iterative curve of the
Standard Function Test, and Result Analysis. The results for Griewank function. Griewank function is a typical non-
the five standard functions test curves are shown in linear multimode stating function with a large search
Figure 2. The test curve for the Ackley function is shown in space and has many local advantages. As can be seen from
the subgraph of Figure 2(a). The Ackley function is an n- the graph shown in Figure 2(b), our proposed IPSO al-
dimensional function with several local optimal values. gorithm can obtain the optimal value better than the other
Therefore, it is difficult to find the optimal global value, developed algorithms in the existing literature, and its
6 Journal of Control Science and Engineering

Ackley iteration curve Griewank iterative curve

12 700

10 600

500
8
Fitness value

Fitness value
400
6
300
4
200

2 100

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Number of iterations Number of iterations

PSO NCFPSO PSO NCFPSO

RandWPSO Improved RandWPSO IPSO
TFPSO TFPSO

(a) (b)
Rastrigin iterative curve Rosenbrock iterative curve
300 9000

8000
250
7000

200 6000
Fitness value

Fitness value

5000
150
4000

100 3000

2000
50
1000

0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Number of iterations Number of iterations

PSO NCFPSO PSO NCFPSO

RandWPSO IPSO RandWPSO IPSO
TFPSO TFPSO

(c) (d)
Sphere iterative curve
30

25

20
Fitness value

15

10

0
0 100 200 300 400 500 600 700 800 900 1000
Number of iterations

PSO NCFPSO
RandWPSO IPSO
TFPSO

(e)

Figure 2: (a) Ackley iteration curve. (b) Griewank iteration curve. (c) Rastrigin iteration curve. (d) Rosenbrock iteration curve. (e) Sphere
iteration curve.
Journal of Control Science and Engineering 7

convergence speed is fast. Its optimal value is received at convergence speed at the beginning, as the optimization
the iteration step of less than 120. This shows that our process continues, the IPSO algorithm in this paper per-
proposed algorithm has better local optimal value bounce forms better than the other four algorithms at the middle
ability. The subgraph shown in Figure 2(c) represents the and later stages.
iterative curve of the Rastrigin function. Rastrigin func-
tion, which is similar to Griewank function, is a multipeak
function. Its algorithm is easy to fall into the local 4.2. Path Planning
maximum with an excellent solution. From the subgraph
of Figure 2(c), our proposed IPSO algorithm approaches 4.2.1. Experimental Environment and Parameter Setting.
the optimal global value in a shorter iterative step when The IPSO algorithm of this paper uses the navigation point
compared with the other four algorithms existing in the model of [26] to construct the experimental environment
literature. Our proposed algorithm converges at 110 it- model and the obstacle expansion treatment. The experi-
eration steps, which verifies its ability to jump out from mental environment is divided into a simple environment
the optimal local value. and a complex environment. The number of simple envi-
The subgraph in Figure 2(d) is the iterative curve of the ronmental obstacles is set to 5, the number of path nodes is 3,
Rosenbrock function. The global advantage of the Rose- and the number of interpolation points is set to 100. Sim-
nbrock function lies in a smooth and narrow parabolic ilarly, the number of complex environmental obstacles is set
valley. The function optimization algorithm provides limited to 11, the number of path nodes is 5, and the number of
information, and it is challenging to distinguish the search interpolation points is set to 100. The boundary is a nonnode
direction. This implies that it is challenging to solve optimal boundary.
global value. Nevertheless, it can be seen from the iteration Simulation parameter selection: the population size and
curve that our proposed IPSO algorithm can find the op- the maximum number of iterations of the five algorithms are
timal global value in less than 60 iterations, which is the consistent: Npop � 150, MaxIt � 100; wmin � 0.4, and
fastest among the compared four existing algorithms in the wmax � 0.9.
literature. This indicates that our proposed algorithm has
better global searchability.
The subgraph shown in Figure 2(e) is the iteration curve of 4.2.2. Path Planning and Algorithm Performance Analysis
the Sphere function. From Figure 2(e), our proposed algorithm
has shown a certain advantage in terms of convergence speed (1) Simple Environment. In a simple environment, the
and convergence precision that are not obvious. Still, the starting point (■) is the map coordinate system of (0, 0), and
optimal value obtained by our proposed algorithm is the the endpoint (★) coordinates are (9, 9), as shown in
smallest, because the Sphere function is a unimodal function Figure 3(a).
with a unique global minimum value, which does not make it It can be seen from Figure 3(a) that the path of the IPSO
difficult to find the optimal value. Hence, the degree of dis- algorithm is smoother and the dynamics of the robot motion
crimination of the algorithm is not obvious. is improved; the path of the classical PSO planning does not
Table 2 presents the performance comparison for the test find the optimal solution at the beginning of optimization
result data obtained by the standard functions Ackley, and encounters stagnation along its path. It should be easy
Griewank, Rastrigin, Rosenbrock, Sphere, and our proposed for the algorithm to fall into the optimal local solution. The
algorithm for the verification. It can be seen from Table 2 ability to find the optimal global solution is weak due to poor
that, for the five functions, except for the Ackley function, particle diversity with small disturbance. The TFPSO al-
the best result is the same as the other four comparison gorithm starts to search for a short stagnation phenomenon.
algorithms. However, the other four functions are optimized Finally, the localized optimal solution is obtained by the
by our proposed algorithm, and the results are improved by disturbance of the algorithm, which signifies that the al-
at least 45%. After the optimization of the five functions by gorithm has a weak ability to jump out from the optimal
the proposed algorithm in this paper, the average value is local solution.
increased by at least 0.1%, and the standard deviation is Figure 3(b) shows that our proposed IPSO algorithm has
increased by at least 0.2%. The experimental data shows that the fastest convergence rate and converges at 13 iteration
the IPSO algorithm is superior to the other four comparison steps; the RandWPSO algorithm converges at around 20
algorithms in terms of optimal value, average value, and iterative steps; the TFPSO converges at the 48th iteration; the
standard deviation. However, the average running time is NCFPSO algorithm converges at the 14th iterative step; the
more than that of the classical PSO. This is because the classical PSO converges at the 34th iteration.
function of the parameters of our proposed algorithm is Table 3 provides the performance comparison based
improved, while the classical PSO improved the parameters average path length and the average simulation run time for
by constant values. Moreover, our proposed algorithm will the five algorithms. As compared with the other four al-
increase the function to some extent. The limitation of the gorithms, the longest path length of our proposed algorithm
IPSO algorithm in this paper is being time-consuming. Our is reduced by at least 3.3%, the shortest path length is re-
future studies would further consider the issue of achieving a duced by at least 2.9%, the average path length is reduced by
reduced time. Although the IPSO algorithm of this paper at least 5.2%, and the average simulation time is reduced by
optimizes the Rastrigin and Sphere functions with the slower 1.2 s.
8 Journal of Control Science and Engineering

Table 2: Optimization results of five functions.

Function Performance index PSO RandWPSO TFPSO NCFPSO IPSO
Average value 0.34 0.15 0.35 0.17 0.10
Best result 0 0 0 0 0
Ackley
Standard deviation 1.45 1.14 1.52 1.22 0.97
Running time (s) 0.31 0.46 0.43 0.51 0.31
Average value 130.56 19.33 45.80 98.19 8.00
Best result 36.73 12.71 23.62 15.93 1.10
Griewank
Standard deviation 9.95 × 103 1.10 × 103 7.35 × 103 1.40 × 104 1.01 × 103
Running time (s) 0.33 0.46 0.50 0.61 0.40
Average value 179.70 79.53 175.79 52.4226 47.54
Best result 113.62 61.48 118.97 47.55159 34.09
Rastrigin
Standard deviation 3.64 × 103 6.52 × 102 3.62 × 103 1.10 × 103 2.21 × 102
Running time (s) 0.25 0.34 0.37 0.26 0.24
Average value 2.77 × 103 4.38 × 102 4.11 × 103 3.37 × 102 8.91 × 10−1
Best result 5.67 × 102 7.95 × 102 6.00 × 102 2.22 × 102 2.86 × 101
Rosenbrock
Standard deviation 4.51 × 106 2.52 × 106 9.64 × 106 4.73 × 105 2.38 × 105
Running time (s) 0.21 2.40 3.98 0.24 0.23
Average value 2.70 × 10−1 2.10 × 10−2 1.20 × 10−2 2.65 × 10−2 5.8 × 10−3
Best result 2.00 × 10−1 1.51 × 10−9 2.88 × 10−2 1.87 × 10−2 7.97 × 10−192
Sphere
Standard deviation 1.00 × 10−1 1.71 × 10−1 8.88 × 10−2 7.60 × 10−1 5.3 × 10−2
Running time (s) 0.88 3.36 5.33 0.95 1.09

Path planning of PSO in simple environment Simple environment iterative curve

9 32

8 30

7 28
26
6
Fitness value

24
5
y 22
4
20
3
18
2
16
1
14
0 12
0 2 4 6 8 10 0 10 20 30 40 50 60 70 80 90 100
x Number of iterations
PSO IPSO PSO NCFPSO
RandWPSO Start RandWPSO IPSO
TFPSO Destination TFPSO
NCFPSO
(a) (b)

Figure 3: Experimental comparison of five algorithms in a simple environment. (a) Simple environment path plan. (b) Simple environment
iteration graph.

Table 3: Comparison of path length of five algorithms in simple environment.

Algorithm Longest path Shortest path Average path Average time (s)
PSO 28.52 15.23 15.67 9.91
RandWPSO 22.89 13.51 13.90 14.10
TFPSO 27.06 13.29 14.36 15.16
NCFPSO 23.55 15.13 15.41 9.81
IPSO 22.14 12.91 13.18 8.71
Journal of Control Science and Engineering 9

Path planning of PSO in complex environment Complex environment iterative curve

9 30

8 28

7 26

6 24

Fitness value
5 22
y
4 20

3 18

2 16

1 14

0 12
0 2 4 6 8 10 0 10 20 30 40 50 60 70 80 90 100
x Number of iterations

PSO IPSO PSO NCFPSO

RandWPSO Start RandWPSO IPSO
TFPSO Destination TFPSO
NCFPSO
(a) (b)

Figure 4: The first experimental comparison of five algorithms in the complex environment. (a) The first type of complex environment path
planning diagram. (b) The first type of complex environment iteration graph.

(2) Complex Environment. To verify the universality of the algorithms, the average path is reduced by at least 2.1%, and
experiment, two types of experiments were carried out. The the average simulation running time is reduced by 1.1 s.
first experiment is carried out from the starting point to the Furthermore, path planning has higher accuracy.
endpoint, while the second experiment is conducted from The second type of experiment: the starting point (■) is the
the endpoint to the starting point. origin of the map with the coordinate system of (9, 9), and the
The first type of experiment: the starting point (■) is the endpoint (★) coordinates are (0, 0), as shown in Figure 5(a).
map coordinate system (0, 0), and the endpoint (★) coor- By considering Table 5, the performance of the path
dinates are (9, 9), as shown in Figure 4(a). planning of our proposed algorithm in this paper is as
It can be seen from Figure 4(a) that the path planning of follow: the average path length is reduced by at least 5.2%,
our proposed algorithm is not only the shortest but the the shortest path length is reduced by at least 1.5%, the
smoothest. The path planning of a complex environment is longest path length is reduced by 7.2%, and the average
relatively concentrated. This is because there are many ob- simulation running time is reduced by 1.2 s.
stacles in the complex environment with fewer paths to choose In summary, it is verified that our proposed IPSO al-
from. The classical PSO and TFPSO appear to be stagnant at gorithm has the shortest path, the highest precision, and
the beginning of the algorithm, and the IPSO algorithm is more least processing time when compared with the four algo-
prominent. An excellent solution is achieved after the algo- rithms existing in the literature.
rithm is optimized for a while. The optimal local solution is
jumped out after its disturbance is settled. It shows that two
algorithms (i.e., classical PSO and TFPSO) have small dis- 4.2.3. Summary of Path Planning. In the simple environ-
turbances, few particle diversity, and weak ability to jump out ment, the path planning by the five algorithms is relatively
from the local optimal solutions. Figure 4(b) presents the scattered. This is because there are fewer obstacles and ample
results obtained from the complex environment; from the space with many paths available for the algorithm to select.
iterative curves shown in Figure 4(b), the IPSO algorithm has When comparing the first type of experiment with the
the fastest convergence rate and the highest convergence second type experimental path planning diagram in a
precision with 22 iteration steps. The NCFPSO algorithm has complex environment, the five types of experiments in the
27 iterative steps. The RandWPSO algorithm has 28 iterations first type of experiment are more concentrated when the
before it starts to converge. TFPSO iterates 39 times before the robot starts running, and appear to be more dispersed in the
left and right algorithms start converging; the PSO algorithm later stages. This is because the first type of experiment is
starts to converge at around 93 iterative steps. Table 4 compares near the starting point. In the presence of obstacles, the
the path lengths of five algorithms in the first type of complex algorithms avoid the obstacles according to their optimi-
heterogeneous environment. zation ability. At the later stages of the optimization, the
Table 4 shows that the shortest path of the IPSO algo- obstacles at the endpoint are sparse, thereby making the path
rithm is reduced by at least 3% compared with the other four more concentrated.
10 Journal of Control Science and Engineering

Table 4: The first comparison of path length of five algorithms in the complex environment.
Algorithm Longest path Shortest path Average path Average time (s)
PSO 31.21 13.73 15.84 10.89
RandWPSO 22.49 13.07 13.67 15.10
TFPSO 26.50 13.00 14.00 16.22
NCFPSO 26.40 13.07 13.92 10.83
IPSO 22.44 12.84 13.38 9.70

Path planning of PSO in complex environment Complex environment iterative curve

9 30

8 28

7 26

6 24

Fitness value
5 22
y
4 20
3 18
2 16
1 14
0 12
0 2 4 6 8 10 0 10 20 30 40 50 60 70 80 90 100
x Number of iterations
PSO IPSO PSO NCFPSO
RandWPSO Start RandWPSO IPSO
TFPSO Destination TFPSO
NCFPSO
(a) (b)

Figure 5: The second experimental comparison of five algorithms in the complex environment. (a) The second type of complex envi-
ronment path planning diagram. (b) The second type of complex environment iteration graph.

Table 5: The second comparison of path length of five algorithms in the complex environment.
Algorithm Longest path Shortest path Average path Average time (s)
PSO 25.66 13.59 15.30 11.14
RandWPSO 28.86 13.07 13.76 15.07
TFPSO 25.62 13.12 14.41 16.42
NCFPSO 23.41 13.04 13.67 11.07
IPSO 21.72 12.84 12.96 10.00

The second type of experiment is the opposite of the first algorithm. The introduction of the local optimization and
type. The three-path planning experiment is carried out to exponential decay for inertia weight improves the search-
verify the effectiveness of our proposed IPSO algorithm ability of our proposed algorithm and increases the dis-
among the algorithms existing in the literature. The per- turbance of our algorithm, which subsequently improves the
formance comparison for the algorithms is based on the diversity of the particles.
longest path, the shortest path, the average path, and the
average simulation running time. In all the comparisons 5. Conclusion
carried out, our proposed IPSO algorithm has performed
better than the other four algorithms in the literature. This is In this paper, particle swarm optimization (PSO) and cubic
because the algorithm proposed in this paper introduces a spline interpolation are combined to solve the minimum of
uniform distribution initialization strategy. Furthermore, five test functions and robot path planning problems. In
the inertia weight and the learning factor interact in the view of the problems of the classical particle swarm opti-
optimization algorithm, which reduces the parameters of the mization algorithm, such as poor searchability, low con-
proposed algorithm. Increasing the uniformity of the im- vergence accuracy, easy falling into local optimal solution,
proved algorithm can better balance the global search of the poor robustness, and poor path smoothness, the classical
Journal of Control Science and Engineering 11

particle swarm optimization algorithm is improved from the environments,” IEEE Transactions on Industrial Electronics,
following aspects: (1) The uniform initialization strategy is vol. 66, no. 8, pp. 6117–6127, 2019.
adopted for the population to improve the later searchability [3] A. Sedeño-noda and M. Colebrook, “A biobjective Dijkstra
of our IPSO algorithm. This prevents the algorithm from algorithm,” European Journal of Operational Research,
falling into the local optimal solutions, because the random vol. 276, no. 1, pp. 106–118, 2019.
initialization of the particles is not evenly distributed, which [4] X. Cao, X. Zou, C. Jia, M. Chen, and Z. Zeng, “RRT-based path
planning for an intelligent litchi-picking manipulator,”
is not conducive to the searchability at the later stages. (2)
Computers and Electronics in Agriculture, vol. 156, pp. 105–
The introduction of the exponential decay for inertia weight
118, 2019.
makes the particles grow up in the early stage of the search [5] A. Azzabi and K. Nouri, “An advanced potential field method
and is beneficial to the global search. The particle step size in proposed for mobile robot path planning,” Transactions of the
the later stages of the search is small in the local develop- Institute of Measurement and Control, vol. 41, no. 11,
ment, and the optimization accuracy is high. Experimental pp. 3132–3144, 2019.
results show that the method increases the disturbance and [6] X. Li and S. R. Qu, “An effective differential evolution al-
diversity of particles to a certain extent. (3) The use of sine gorithm for collision avoidance of vehicles under IOT (In-
and cosine function can control the independent variable as ternet of things),” Journal of Northwestern Polytechnical
the inertia weight. The learning factor makes the three University, vol. 30, no. 5, pp. 729–733, 2012.
variables become one variable; this reduces the parameters of [7] J. W. Wu, D. Bin, X. B. Feng et al., “GA based adaptive
the improved algorithm and thus reduces the complexity of singularity-robust path planning of space robot for on-orbit
our algorithm. The interaction of inertia weight and learning detection,” Complexity, vol. 2018, Article ID 3702916,
factor is increased, the unity of algorithm optimization is 11 pages, 2018.
[8] C. Xiong, D. Chen, D. Lu, Z. Zeng, and L. Lian, “Path planning
improved, and the performance of the algorithm is im-
of multiple autonomous marine vehicles for adaptive sam-
proved to a certain extent. (4) The evaluation function is
pling using Voronoi-based ant colony optimization,” Robotics
constructed, and a smooth path is planned with cubic spline
and Autonomous Systems, vol. 115, pp. 90–103, 2019.
interpolation, which improves the accuracy and dynamic [9] Y. Q. Huang, Z. K. Li, Y. Jiang et al., “Cooperative path planning
characteristics of the path. However, when compared with for multiple mobile robots via HAFSA and an expansion logic
the classical PSO algorithm, the time advantage of the strategy,” Applied Sciences-Basel, vol. 9, no. 4, p. 10, 2019.
proposed algorithm is weak, which will be a crucial issue to [10] Li Xun, D. Wu, Z. Zhao et al., “Path planning method for
be further studied in future research. Nevertheless, this does indoor robot based on improved PSO,” Computer Measure-
not affect the competitiveness of the improved algorithm. ment &Control, vol. 28, no. 3, pp. 206–211, 2020.
Besides, in the follow-up study, virtual obstacles will be used [11] T. T. Gao, L. Zhang, B. D. Li et al., “Study on path planning of
for the simulation experiment of multirobot path planning, soccer robot based on improved particle swarm algorithm,”
and it will be used in the real environment of an operating Journal of Xi’an Polytechnic University, vol. 30, no. 5,
system with hardware platform as turnabout and software pp. 609–615, 2016.
platform as ROS (robot operation system), to improve the [12] P. I. Adamu, H. I. Okagbue, and P. E. Oguntunde, “Fast and
practicability of the algorithm. optimal path planning algorithm (FAOPPA) for a mobile
robot,” Wireless Personal Communications, vol. 106, no. 2,
pp. 577–592, 2019.
Conflicts of Interest [13] B. Tang, Z. Zhanxia, and J. Luo, “A convergence-guaranteed
particle swarm optimization method for mobile robot global
The authors declare that they do not have any conflicts of path planning,” Assembly Automation, vol. 37, no. 1,
interest. pp. 114–129, 2017.
[14] B. Song, Z. Wang, L. Zou, L. Xu, and F. E. Alsaadi, “A new
approach to smooth global path planning of mobile robots
Acknowledgments with kinematic constraints,” International Journal of Machine
This work was supported by the National Natural Science Learning and Cybernetics, vol. 10, no. 1, pp. 107–119, 2019.
[15] P. Chen, Q. Li, C. Zhang et al., “Hybrid chaos-based particle
Foundation of China (No. 51607133); Shaanxi Natural
swarm optimization-ant colony optimization algorithm with
Science Basic Research Project (No. 2019JM567); China
asynchronous pheromone updating strategy for path plan-
Textile Industry Federation Science and Technology Guid- ning of landfill inspection robots,” International Journal of
ance Project (No. 2018094); and College Student Innovation Advanced Robotic Systems, vol. 16, no. 4, p. 11, 2019.
and Entrepreneurship Training Program (No. [16] L.-Z. Du, Z. Wang, S. Ke et al., “Research on multi-load AGV
201910709019). path planning of weaving workshop based on time priority,”
Mathematical Biosciences and Engineering, vol. 16, no. 4,
References pp. 2277–2292, 2019.
[17] Y. Shi and R. C. Eberhart, “A modified particle swarm op-
[1] L. Yang, J. Qi, D. Song, J. Xiao, J. Han, and Y. Xia, “Survey of timizer,” in Proceedings of IEEE Congress on Evolutionary
robot 3D path planning algorithms,” Journal of Control Sci- Computation, IEEE Service Center, Piscataway, NJ, USA,
ence and Engineering, vol. 2016, Article ID 7426913, 22 pages, pp. 69–73, May 1998.
2016. [18] J. Q. Wang and X. D. Wang, “Particle swarm optimization
[2] J. Votion and Y. Cao, “Diversity-based cooperative multi- algorithm with improved inertia weight,” Journal of Xi’an
vehicle path planning for risk management in costmap Polytechnic University, vol. 31, no. 6, pp. 835–840, 2017.
12 Journal of Control Science and Engineering

[19] Y. Shi and R. Eberhart, “Fuzzy adaptive particle swarm opti-

mization,” in Proceedings of the IEEE Congress on Evolutionary
Computation, IEEE, San Francisco, USA, pp. 101–106, May 2001.
[20] Y. G. Wang, T. T. Qu, and S. Li, “Disruption particle swarm
optimization algorithm based on exponential decay weight,”
Application Research of Computers, vol. 37, no. 4, pp. 1020–1024,
2020.
[21] Y. Zhao and Z. Fang, “Particle swarm optimization algorithm
with weight function’s learning factor,” Journal of Computer
Applications, vol. 33, no. 8, pp. 2265–2268, 2013.
[22] M. G. C. Resende, C. C. Ribeiro, F. Glover et al., “Scatter
search and path-relinking: fundamentals, advances, and ap-
plications,” Handbook of Metaheuristics, vol. 146, pp. 87–107,
Springer, Berlin, Germany, 2010.
[23] Z. G. Zhao, S. Y. Huang, and W. Q. Wang, “Simplified particle
swarm optimization algorithm based on stochastic inertia
weight,” Application Research of Computers, vol. 31, no. 2,
pp. 361–391, 2014.
[24] A. Khare and S. Rangnekar, “A review of particle swarm
optimization and its applications in Solar Photovoltaic sys-
tem,” Applied Soft Computing, vol. 13, no. 5, pp. 2997–3006,
2012.
[25] S. Zhao, J. H. Zhang, and Y. F. Zhu, “An adaptive constriction
factor particle swarm optimization algorithm using fuzzy
inference engine,” Journal of Qingdao University, vol. 33,
no. 3, pp. 9–37, 2018.
[26] J. Yang and L. Li, “Improved biogeography-based optimi-
zation algorithm for mobile robot path planning,” in Pro-
ceedings of 2017 Chinese Intelligent Systems Conference,
Singapore, 2018.