Professional Documents
Culture Documents
net/publication/312036821
On Global Smooth Path Planning for Mobile Robots using a Novel Multimodal
Delayed PSO Algorithm
CITATIONS READS
43 290
3 authors:
Lei Zou
Brunel University London
42 PUBLICATIONS 727 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Baoye Song on 30 November 2017.
Abstract The planning problem for smooth paths for mo- corresponding values in previous iterations) is added into
bile robots has attracted particular research attention, but the velocity updating model. A series of simulation exper-
the strategy combining the heuristic intelligent optimization iments are implemented on a standard collection of bench-
algorithm (e.g., particle swarm optimization) with smooth mark functions. The experiment results verify that the com-
parameter curve (e.g., Bezier curve) for global yet smooth prehensive performance of the developed MDPSO algo-
path planning for mobile robots has not been thoroughly rithm is superior to other well-known PSO algorithms.
discussed because of several difficulties such as the local Finally, the presented MDPSO algorithm is utilized in the
trapping phenomenon in the searching process. In this paper, global smooth path planning problem for mobile robots,
a novel multimodal delayed particle swarm optimization which further confirms the advantages of the MDPSO algo-
(MDPSO) algorithm is developed for the global smooth rithm over the traditional genetic algorithm (GA) investi-
path planning for mobile robots. By evaluating the evo- gated in previous studies. The multimodal delayed informa-
lutionary factor in each iteration, the evolutionary state tion in the MDPSO reduces the occurrence of local trapping
is classified by equal interval division for the swarm of phenomenon and the convergence rate is satisfied at the
the particles. Then, the velocity updating model would same time. Based on the testing results on a selection of
switch from one mode to another according to the evolu- benchmark functions, the MDPSO’s performance has been
tionary state. Furthermore, in order to reduce the occur- shown to be superior to other five well-known PSO algo-
rence of local trapping phenomenon and expand the search rithms. Successful application of the MDPSO for planning
space in the searching process, the so-called multimodal the global smooth path for mobile robots further confirms
delayed information (which is composed of the local and its excellent performance compared with the some typical
global delayed best particles selected randomly from the existing algorithms.
whose objective is to look for a feasible yet optimal path (e.g., stop, rotate, and restart) which leads to unnecessary
from the start point to the target point. Such an issue could waste on time and energy. Furthermore, the jerk resulting in
be regarded as an optimization issue on certain indices (e.g., the state switching is not permissible when the smoothness
shortest distance and minimum energy) with certain con- of the movement is required to ensure service quality [44].
straints (e.g., collision-free route), see, e.g., [4, 28, 29, 37]. Therefore, in addition to the path length, the path smooth-
Up to now, researchers have developed many heuristic algo- ness has been considered as another important criterion for
rithms to solve this problem (see, e.g., [10, 12, 32, 35, its close relation to other optimization criteria [5].
36], and the references cited therein), and particle swarm So far, the planning problem for smooth paths for mobile
optimization (PSO) algorithm is arguably the most studied robots has attracted particular research attention. For exam-
algorithms which have been widely employed in the path ple, Bezier curve has been applied to plan paths of a robot
planning for mobile robots [48]. soccer system modeled as multi-agents considering veloc-
In the past few years, a large number of PSO-based ity and acceleration limits [16]. A smooth path planning
approaches have been proposed for the path planning for approach has been proposed in [13] to generate a feasible
mobile robots. For example, a study on the comparison path composed of piecewise Bezier curve with curvature
of dynamic path planning has been presented for mobile constraint. A so-called A* algorithm has been introduced
robots in [11] where the PSO algorithm has shown the in [34] to find the minimum-cost path from the starting
better convergence performance compared with the genetic node to a given target node over a directed graph, and the
algorithm (GA). In [31], a new method called biogeogra- arc-line approach is utilized to smooth the generated path.
phy PSO (BPSO) has been developed by combining the In [54], the Voronoi diagram and Dijkstra algorithm have
biogeography-based optimization (BBO) and PSO algo- been employed to plan a piecewise line path, where the end-
rithm to tackle the path planning problem in static environ- points of the lines in the path have been exploited as the
ments, where the BPSO algorithm is employed to optimize control points of the smooth Bezier curve path. A stochas-
the network of the paths through approximate voronoi tic PSO has been proposed in [7] to optimize certain cubic
boundary network (AVBN) modeling. An accelerated PSO spline described by polynomials whose coefficients are set
has been developed in [33] toward the global path plan- as variables/free parameters to form the swarm of parti-
ning for mobile robots, where only the global best particle cles. An approach based on the radial basis function (RBF)
of the PSO is utilized and the local best particle is dis- neural networks has been presented in [3] to deal with the
carded in the updating function. Furthermore, it has been smooth path planning of mobile robots where a Bezier curve
shown by simulation that such a simplified PSO gives the has been trained to realize the local path planning pro-
same order of convergence as the conventional PSO. In [14], cesses. Nevertheless, the strategy combining the heuristic
a hybrid heuristic GA-PSO scheme has been put forward intelligent optimization algorithm (e.g., PSO) with smooth
to plan the paths for mobile robots in a grid environment, parameter curve (e.g., Bezier curve) for global yet smooth
where crossover and mutation operators of GA are applied path planning for mobile robots has not been adequately
to the evolution of the particles of PSO, thereby avoiding discussed in the literature because of several difficulties
the premature convergence and time complexity in conven- such as the local trapping phenomenon in the searching
tional GA and PSO algorithms. A parallel metaheuristic process.
PSO algorithm has been developed in [15] for the global In this paper, we aim to develop a novel multimodal
path planning for mobile robots, in which three parallel PSO delayed PSO (MDPSO) algorithm combining with the
algorithms combined with a communication operator are Bezier curve to handle the global smooth planning for
utilized to generate the feasible line path, and the gener- mobile robot paths. The multimodal delayed information
ated feasible line path is then smoothed by a cubic B-spline adopted in the proposed MDPSO would decrease the trap-
smoother. However, the path planning algorithms proposed ping possibility to the local minimum and hence help
in the aforementioned papers have mainly been concerned explore the whole search space thoroughly. The main con-
with certain simple optimal criterion (e.g., the minimum tributions of this paper are outlined as threefold. (1) A
length of the path), while other important performances of novel multimodal delayed PSO (MDPSO) with adaptive
the path (e.g., the smoothness of the path) are seldom con- multimodal delayed information is proposed to overcome
sidered in the path planning [43]. Usually, a traditional path the local trapping phenomenon frequently appearing in the
planning algorithm could generate a path composed of sev- global smooth planning for mobile robot paths. (2) The
eral polygonal lines which involve inevitable sharp turns performance of the MDPSO is shown, via comprehensive
sometimes. Moving along such a path would cause fre- simulation experiments, to outperform the other well-known
quent switches for a mobile robot between different modes PSO algorithms on a standard collection of benchmark
Cogn Comput
functions. (3) The developed MDPSO is successfully where k is the number of the current iteration, w is the iner-
applied to globally smoothly planning paths for mobile tia weight, the acceleration coefficients c1 and c2 are called,
robots and the derived smooth path performs better than the respectively, cognitive and social parameters, and r1 and r2
one generated by GA in the previous studies. are two uniformly distributed random numbers on [0, 1].
The remainder of this paper is organized as follows. In
“Particle Swarm Optimization and its Developments” Developments of Traditional PSO
section, the particle swarm optimization and its develop-
ments are briefly introduced. In “A Novel Multimodal The traditional PSO scheme described above has widely
Delayed PSO Algorithm” section, the novel multimodal been used in various optimization problems for its sim-
delayed PSO is proposed and discussed in great detail. In ple concept and efficient implementation, and a variety of
“Simulation Experiments” section, we conduct the simu- approaches have been proposed to improve the capability of
lation experiments in order to compare and contrast the the traditional PSO [45].
performances of the MDPSO with several existing PSO PSO with linearly decreased inertia weight w on iteration
algorithms. In “Smooth Path Planning for Mobile Robots” generations (PSO-LDIW) has been introduced by Shi and
section, the strategy of MDPSO combining with Bezier Eberhart [39–41], the inertia weight of the current iteration
curve is exploited to the global smooth planning for mobile w is calculated as follows:
robot paths and the performance is then discussed. Finally,
itermax − iter
concluding remarks are given and future work is pointed out w = (w1 − w2 ) × + w2 , (2)
itermax
in the last section.
where w1 and w2 are, respectively, the initial value and the
final value of the inertia weight; iter denotes the number
Particle Swarm Optimization of current iteration and itermax is the number of maximum
and its Developments iteration. Generally speaking, a larger inertia weight would
make the PSO tend to the global exploration and, on the
Traditional PSO Algorithm other hand, a smaller one could achieve the local exploita-
tion. Therefore, the initial and final values w1 and w2 are
As a heuristic intelligent optimization algorithm, PSO was customarily set as 0.9 and 0.4, respectively. Furthermore,
developed by Kennedy and Eberhart to simulate the swarm PSO with time-vary acceleration coefficients (PSO-TVAC)
behaviors of birds flocking or fish schooling, where each has been proposed in [38] as computed by the following
particle of the swarm acts as a potential solution of certain equations:
optimization problem [17].
For PSO, the swarm consisting of particles moves around itermax − iter
c1 = (c1i − c1f ) × + c1f , (3)
at certain velocity in the search space of D-dimension. itermax
At the kth iteration in the searching process, the posi-
tion of the ith particle (denoted by a vector xi (k) = itermax − iter
(xi1 (k), xi2 (k), · · · , xiD (k))) will be updated to reach the c2 = (c2i − c2f ) × + c2f , (4)
itermax
global optimum based on the corresponding velocity vector
(denoted by vi (k) = (vi1 (k), vi2 (k), · · · , viD (k))). More- where c1i (c2i ) is the initial value, and c1f (c2f ) is the final
over, the velocity vector of the ith particle at the kth iteration value of the acceleration coefficient c1 (c2 ). Usually, we set
will be updated according to (1) the historical best position c1i = 2.5 (c2i = 0.5) and c1f = 0.5 (c2f = 2.5) in this
of the ith particle, which is also called local best particle strategy. Additionally, PSO with constriction factor (PSO-
(pbest) as denoted by pi = (pi1 , pi2 , · · · , piD ), and (2) CK) has been proposed by Clerc and Kennedy to enhance
the historical best position of the entire swarm, which is the searching performance of PSO, where w = 0.729 and
also named as global best particle (gbest) as represented c1 = c2 = 1.49 are recommended [9]. The improve-
by pg = (pg1 , pg2 , · · · , pgD ). The details of the updating ment strategies of the traditional PSO mentioned above are
models for velocity and position of the ith particle at the mainly concerning with the parameter studies, while other
next iteration is given as follows: strategies including combination with auxiliary operations
and topological structures have also been constructed to
vi (k + 1) = wvi (k) + c1 r1 (pi (k) − xi (k)) + c2 r2 (pg (k) − xi (k)), enhance the capability of PSO [50]. One of the most remark-
xi (k + 1) = xi (k) + vi (k + 1), able research trends is the hybrid PSO, which combines
(1) some auxiliary operations with the traditional PSO, e.g.,
Cogn Comput
selection [2], crossover [8], mutation [1], local search [22] Framework of MDPSO
and differential evolution [53], etc. A fully informed particle
swarm optimization (FIPSO) scheme has been developed to The updating equations for the velocity and the position of
guide the particles of the swarm using the information of the novel MDPSO algorithm are presented by:
entire neighborhood [30]. A comprehensive-learning PSO
(CLPSO) introduced in [23] employs local best particle
from different neighbors to update the swarm flying at dif- vi (k + 1) = wvi (k) + c1 r1 (pi (k) − xi (k)) + c2 r2 (pg (k) − xi (k))
+si (k)c3 r3 (pi (k − τi (k)) − xi (k))
ferent dimensions so as to improve the performance in the +sg (k)c4 r4 (pg (k − τg (k)) − xi (k)),
case of multimodal optimizations. The analyses and exper- xi (k + 1) = xi (k) + vi (k + 1),
iments have shown that the performance of the traditional (5)
PSO has been promoted greatly by using these strategies.
Recently, an adaptive PSO (APSO) has been put forward where w is the inertia weight determined by Eq. 2; c1
by Zhan et al., which introduces an evolutionary factor to and c2 are the coefficients for acceleration updated by
quantify the mean distance (between the global best particle Eqs. 3 and 4, and c3 and c4 are equal to c1 and c2
and other particles). By a series of fuzzy membership func- without loss of generality, i.e., c1 = c3 and c2 = c4 ,
tions according to the evolutionary factor, four states have respectively; ri (i = 1, 2, 3, 4) are the random uniformly
been defined, which are the exploration state, the exploita- distributed numbers in [0, 1]; τi (k) and τg (k) are the ran-
tion state, the convergence state, and the jumping out state. dom delays uniformly distributed in [0, k] for the local
These four states have been used to adaptively control the and the global delayed best particle, respectively; si (k) and
inertia weight and the acceleration coefficients in each iter- sg (k) are the intensity factor of the newly added terms in
ation [50]. A switching PSO (SPSO) has been proposed by the velocity updating model depending on the evolutionary
Tang et al. to overcome the shortcomings of Zhan’s algo- factor.
rithm. A Markov chain is used to predict the next state In the novel MDPSO algorithm, the newly added terms
according to the current state as decided by the evolu- in the velocity updating model is closely related to the evo-
tionary factor, and the velocity updating rule is switched lutionary factor (EF) defined in [50] to describe the swarm
from one mode to another depending on the evolutionary distribution properties. According to the EF in the search-
state [45]. Moreover, a switching delayed PSO (SDPSO) ing process, the four states (i.e., the exploration state, the
has been proposed more recently, which could switch the exploitation state, the convergence state, and the jumping
velocity updating model according to the evolutionary state out state) are denoted by ξ(k) = 1, ξ(k) = 2, ξ(k) = 3 and
predicted by a Markov chain. In the SDPSO, in addition ξ(k) = 4, respectively. The mean distance between the ith
to the inertia weight and acceleration coefficients that are particle and the other particles in the swarm denoted as di
adaptively adopted based on the evolutionary factor and could be calculated by
the evolutionary state, the local/global best particles are
randomly chosen based on the corresponding values from D
1
S
previous iterations [49]. di = (x − x )2 , (6)
ik jk
S−1
j =1,j =i k=1
A Novel Multimodal Delayed PSO Algorithm where S and D denote the swarm size and the parti-
cle dimension, respectively. Accordingly, the evolutionary
The main purpose of this section is to develop a novel factor denoted as Ef could be calculated by
multimodal delayed PSO (MDPSO) so as to improve the
searching performance further. The main novelty of such a dg − dmin
Ef = , (7)
new PSO algorithm is to add two delayed terms in the tra- dmax − dmin
ditional velocity updating model of the PSO algorithm. The
new terms, composed of both the local and global delayed where dg is the mean distance between the global best par-
best particles selected from the corresponding values in the ticle and the other particles in the swarm. dmax and dmin
previous iterations stochastically, are added into the veloc- are the maximum and minimum of di in the swarm, respec-
ity updating model according to the evaluated evolutionary tively. The evolutionary state are classified according to the
state. This improvement strategy aims to reduce the conver- evolutionary factor by a series of fuzzy functions in [50],
gence speed of the traditional PSO and thereby decreasing equal division strategy is used for the evolutionary state
the likelihood of converging to local minimum. As such, the classification in [45, 49] for the state prediction depend-
entire search space could be explored more thoroughly. ing on a Markov chain. In this paper, the formulation of
Cogn Comput
evolutionary state classification in [45, 49] has been adopted Table 1 Strategies for multimodal delayed information
and expressed as follows:
State Mode si (k) sg (k) τi (k) τg (k)
⎧
⎪
⎪ 1, 0 ≤ Ef < 0.25, Convergence ξ(k) = 1 0 0 – –
⎨
2, 0.25 ≤ Ef < 0.5, Exploitation ξ(k) = 2 Ef (k) 0 k · randτi –
ξ(k) = (8)
⎪
⎪ 3, 0.5 ≤ Ef < 0.75, Exploration ξ(k) = 3 0 Ef (k) – k · randτg
⎩
4, 0.75 ≤ Ef ≤ 1. Jumping-out ξ(k) = 4 Ef (k) Ef (k) k · randτi k · randτg
D−1
Rosenbrock: f2 (x) = (100(xi+1 −xi )2 +(xi −1)2 ). (10)
i=1
D
1 D
−0.2 2
i=1 xi
1
Ackley: f3 (x) = −20e D −eD i=1 cos 2πxi
+ 20 + e.
(11)
D
Rastrigin: f4 (x) = (xi2 − 10 cos 2π xi + 10). (12)
i=1
D
D
Schwefel 2.22: f5 (x) = | xi | + | xi | . (13)
i=1 i=1
D i
Schwefel 1.2: f6 (x) = ( xj )2 . (14)
i=1 j =1
Table 2 Configuration of
benchmark functions Functions Name Dimension Search space Threshold Minimum
convergence performances of the algorithms are quite dif- from the traditional interpolation-based curves such as poly-
ferent from each other. Obviously, the convergence rate nomials and cubic splines, the Bezier curve consists of a
of the proposed MDPSO is slightly slower than some of number of control points, which are not in the curve except
the other PSO algorithms, e.g., PSO-CK and SPSO on the the start and end points. Let a set of control point vec-
Sphere function, yet better solutions could be achieved for tors P0 , P1 , · · · , Pn be given. In this case, the Bezier curve
all the other benchmark functions. It is worthwhile to note denoted as P(t) is defined as
that the multimodal delayed information in the MDPSO
make it not easy to converge to the local minima, and thus
n
P(t) = Bin (t)Pi , 0 ≤ t ≤ 1, (15)
the whole search space could be explored more thoroughly
i=0
as a result of the super capability of escaping from the
local optima. Hence, the proposed MDPSO algorithm out-
where t is the normalized time variable; Pi = (xi , yi )T
performs the other PSO algorithms for both unimodel and
stands for the coordinate vector of the ith control point with
multimodel benchmark functions in a series of criteria such
xi and yi as X and Y coordinate components, respectively;
as success ratio and mean fitness value.
Bin (t) is the Bernstein polynomial expressed as:
n i
Smooth Path Planning for Mobile Robots Bin (t) = t (1 − t)n−i , i = 0, 1, . . . , n. (16)
i
Preliminary of Bezier Curve
The smoothness of a Bezier curve based path is closely
related to the curvature function of the path. In the
As a parametric curve, the Bezier curve has been success-
fully used in practice such as computer graphics. Different
Rosenbrock
Sphere 9
50 PSO−LDIW
8 PSO−TVAC
PSO−CK
0 SPSO
7 SDPSO
−50 MDPSO
6
Mean fitness: log(f(x))
Mean fitness: log(f(x))
−100
5
−150
4
−200 3
PSO−LDIW
−250 PSO−TVAC
2
PSO−CK
SPSO
−300 1
SDPSO
MDPSO
−350 0
0 5000 10000 15000 20000 0 5000 10000 15000 20000
Number of generations: (Population=20, Dimension=20) Number of generations: (Population=20, Dimension=20)
Fig. 2 Performance test for sphere function f1 (x) Fig. 3 Performance test for Rosenbrock function f2 (x)
Cogn Comput
0 5
−2
0
−4
Mean fitness: log(f(x))
Fig. 4 Performance test for Ackley function f3 (x) Fig. 6 Performance test for Schwefel 2.22 function f5 (x)
n−2
where R(t) is the curvature’s radius; Ṗx (t), Ṗy (t), P̈x (t), P̈(t) = n(n − 1) Bin−2 (t)(Ṗi+2 − 2Ṗi+1 + Ṗi ). (19)
and P̈y (t) are the coordinate components on X and Y of the i=0
−5
1.8
1.6
−10
1.4 PSO−LDIW
PSO−TVAC
1.2 PSO−CK
−15
SPSO
1 SDPSO
MDPSO
0.8 −20
0 5000 10000 15000 20000 0 5000 10000 15000 20000
Number of generations: (Population=20, Dimension=20) Number of generations: (Population=20, Dimension=20)
Fig. 5 Performance test for Rastrigin function f4 (x) Fig. 7 Performance test for Schwefel 1.2 function f6 (x)
Cogn Comput
f1 (x) Minimum 2.53 × 10−145 2.49 × 10−201 0.0000 0.0000 1.23 × 10−12 2.72 × 10−94
Mean 3.03 × 10−67 3.07 × 10−191 0.0000 0.0000 3.09 × 10−9 4.50 × 10−64
Std. Dev. 2.14 × 10−66 0.0000 0.0000 0.0000 1.00 × 10−8 3.10 × 10−63
Ratio 100 % 100 % 100 % 100 % 100 % 100 %
f2 (x) Minimum 2.21 × 10−4 1.85 × 10−2 4.45 × 10−12 2.23 × 10−3 3.5339 2.70 × 10−5
Mean 1.17 × 101 7.41 × 103 5.47 × 103 3.79 × 103 1.67 × 101 7.7743
Std. Dev. 1.48 × 101 2.46 × 104 2.15 × 104 1.77 × 104 1.52 × 101 5.1952
Ratio 100 % 80 % 90 % 88 % 100 % 100 %
f3 (x) Minimum 2.66 × 10−15 2.66 × 10−15 6.21 × 10−15 2.66 × 10−15 1.99 × 10−7 2.66 × 10−15
Mean 5.64 × 10−15 4.72 × 10−15 2.0694 7.91 × 10−1 2.14 × 10−5 6.64 × 10−15
Std. Dev. 1.81 × 10−15 1.77 × 10−15 2.2971 2.3036 5.03 × 10−5 2.22 × 10−15
Ratio 100 % 100 % 18 % 72 % 100 % 100 %
f4 (x) Minimum 5.9697 3.9798 2.78 × 101 2.18 × 101 2.9852 1.9899
Mean 1.12 × 101 1.46 × 101 5.86 × 101 5.16 × 101 2.08 × 101 9.1337
Std. Dev. 3.4876 1.26 × 101 1.95 × 101 1.90 × 101 1.26 × 101 2.8119
Ratio 100 % 100 % 32 % 48 % 98 % 100 %
f5 (x) Minimum 3.86 × 10−43 7.84 × 10−120 5.68 × 10−31 1.63 × 10−162 1.04 × 10−7 6.50 × 10−40
Mean 6.00 × 10−1 1.26 × 101 8.40 1.05 × 108 2.4000 3.10 × 10−29
Std. Dev. 2.3989 1.21 × 101 9.33 1.67 × 108 4.7638 1.14 × 10−28
Ratio 94 % 32 % 46 % 2% 78 % 100 %
f6 (x) Minimum 1.31 × 10−29 2.63 × 10−28 1.05 × 10−106 2.83 × 104 2.64 × 10−1 1.41 × 10−30
Mean 1.00 × 102 2.43 × 103 2.70 × 103 7.26 × 104 6.04 × 102 3.77 × 10−18
Std. Dev. 7.07 × 102 4.01 × 103 3.57 × 103 2.46 × 104 2.39 × 103 1.67 × 10−17
Ratio 98 % 66 % 58 % 0% 0% 100 %
A two-dimensional workspace with several obstacles is sup- 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
posed to be the working environment of a mobile robot, 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
and the purpose of planning the smooth path is to seek a 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207
feasible and optimal Bezier curve path from the start to the
176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
target points. The whole workspace is divided into many
160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
square grids according to the requirement of the smooth
path planning, e.g., Fig. 8 shows the workspace of a mobile 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
robot with numbered grids. For each grid, it is defined to be 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
either empty (denoted as the white square grid) or occupied 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127
(denoted as the black square grid), which is determined by 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
whether the boundary of the obstacles is within the square
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
grid. A mobile robot in the workspace could be treated as
a point. In terms of the size of a mobile robot, the bound- 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
ary of the obstacles has been constructed from (1) the actual 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63
boundary and (2) the minimum safety distance. 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
In this paper, the proposed MDPSO is combined with 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
the Bezier curve for planning the smooth path of a mobile
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
robot by considering feasibility, smoothness, and distance
of the path at the same time. Generally, a Bezier curve Fig. 8 Workspace of a mobile robot with numbered grids
Cogn Comput
160 0.1
MDPSO
140 0.09 GA
120 0.08
0.06
80
0.05
60
0.04
40
0.03
20
0.02
0
0 50 100 150
0.01
Fig. 9 Smooth path planning from (5, 5) to (155, 155)
0
0 50 100 150 200
Path points of the smooth path
path comprises several segments denoted as a sequence of
control points. Hence, the smooth path planning is to find a Fig. 11 Path curvature in the case of Fig. 9
sequence of control points that define a feasible and short-
est Bezier curve path by using the MDPSO algorithm. The
objetive function is given as follows: workspace in Fig. 8 in order to show the usefulness of the
proposed approach. The parameters of the experiments are
n
given as follows: the population of the swarm is set as 100,
min J = Pi (t) + No × Pr , (20)
the maximum of the iteration generation is set as 100, the
i=1
dimension of the particle that denotes the amount of control
where Pi (t) denotes the length of the ith segment of a points of the Bezier curve is taken as 7, and the experimental
Bezier curve path; No and Pr denote the amount of the occu- penalty ratio is taken as 50 for each occupied obstacle.
pied grids in the path and the penalty ratio for each occupied Two results of numerous simulation experiments have
grid, respectively. been illustrated in Fig. 9, where the start position and tar-
get position are (5, 5) to (155, 155) and (155, 5) to (5, 155),
Results and Analyses
160 0.08
Curvature of the smooth path
140 0.07
120 0.06
100
0.05
0.04
80
0.03
60
0.02
40
0.01
20
0
0 50 100 150 200
0
0 50 100 150 Path points of the smooth path
Fig. 10 Smooth path planning from (155, 5) to (5, 155) Fig. 12 Path curvature in the case of Fig. 10
Cogn Comput
200 Conclusions
40. Shi Y, Eberhart R. Parameter selection in particle swarm opti- predictive control based on neuro-dynamics optimization. IEEE
mization. Proceedings of the 7th International Conference on Trans Ind Electron. in press, doi:10.1109/TIE.2016.2606358.
Evolutionary Programming; 1998. p. 591–600. 48. Zeng N, Zhang H, Chen Y, Chen B, Liu Y. Path planning
41. Shi Y, Eberhart R. Empirical study of particle swarm optimiza- for intelligent robot based on switching local evolutonary PSO.
tion. Proceedings of IEEE Congress on Evolutionary Computa- Assem Autom. 2016;36(2):120–126.
tion; 1999. p. 1945–1959. 49. Zeng N, Wang Z, Zhang H, Alsaadi FE. A novel switching
42. Shu H, Zhang S, Shen B, Liu Y. Unknown input and state estima- delayed PSO algorithm for estimating unknown parameters of lat-
tion for linear discrete-time systems with missing measurements eral flow immunoassay. Cognitive Computation. 2016;8:143–152.
and correlated noises. Int J Gen Syst. 2016;45(5):648–661. 50. Zhan Z, Zhang J, Li Y, Chung H. Adaptive particle swarm
43. Song B, Wang Z, Sheng L. A new genetic algorithm approach optimization. IEEE Trans Syst Man Cybern Part B Cybern.
to smooth path planning for mobile robots. Assem Autom. 2009;39(6):1362–1381.
2016;36(2):138–145. 51. Zhang J, Ma L, Liu Y. Passivity analysis for discrete-time
44. Song B, Tain G, Zhou F. A comparison study on path smoothing neural networks with mixed time-delays and randomly occurring
algorithms for laser robot navigatioed mobile robot path planning quantization effects. Neurocomputing. 2016;216:657–665.
in intelligent space. Journal of Information and Computational 52. Zhang W, Wang Z, Liu Y, Ding D, Alsaadi FE. Event-
Science. 2010;7(1):2943–2950. based state estimation for a class of complex networks with
45. Tang Y, Wang Z, Fang J. Parameters identification of unknown time-varying delays: a comparison principle approach. Phys Lett
delayed genetic regulatory networks by a switching particle A. 2017;381(1):10–18.
swarm optimization algorithm. Expert Syst Appl. 2011;38:2523– 53. Zhang WJ, Xie XF. DEPSO: Hybrid particle swarm with dif-
2535. ferential evolution operator. Proceedings of IEEE International
46. Wen C, Cai Y, Liu Y, Wen C. A reduced-order approach to filter- Conference on Systtems, Man, and Cybernetics; 2004. p. 997–
ing for systems with linear equality constraints. Neurocomputing. 1006.
2016;193:219–226. 54. Zhou F, Song B, Tian G. Bezier curve based smooth path plan-
47. Xiao H, Li Z, Yang C, Zhang L, Yuan P, Ding L, Wang ning for mobile robot. Journal of Information and Computational
T. Robust stabilization of a wheeled mobile robot using model Science. 2011;8(1):2441–2450.