2008 International Conference on Intelligent Computation Technology and Automation

A Global Optimization Algorithm Based on Plant Growth Theory:

Plant Growth Optimization

Wei Cai, Weiwei Yang and Xiaoqian Chen

Multidisciplinary Aerospace Design Optimization Research Center, College of Aerospace and
Materials Engineering, National University of Defense Technology, Changsha, China. 410073
[email protected]

Abstract In the algorithm discussed in this paper we

A novel optimization algorithm, Plant Growth distribute the search activities over so-called “branches
Optimization (PGO), is proposed in this paper. and leaves”, that is, agents with very simple basic
According to the plant growth characteristics, an capabilities which, to some extent, mimic the behavior
artificial plant growth model is built including leaf of real plant. In fact, research on the virtual plant has
growth, branching, phototropism and spatial greatly inspired our work, so we introduce the basic
occupancy. Afterward, two mechanisms are introduced concepts of artificial plant growth in brief first.
and the basic process of PGO is presented in details.
Three classical test problems are adopted to test the 2.1. L-systems
performance of this new algorithm and the results
show that it is very effective to get global optimum L-systems are a mathematical formalism proposed
solution which benefits from its search mechanism. by the biologist Aristid Lindenmayer in 1968 [3] as a
Furthermore, the effects of population size on foundation for an axiomatic theory of biological
optimization performance are also discussed. development. Central to L-systems, is the notion of
rewriting, where the basic idea is to define complex
objects by successively replacing parts of a simple
1. Introduction object using a set of rewriting rules or productions. The
rewriting can be carried out recursively. Figure 1 is an
In recent years, great development has been example obtained by the L-system.
achieved in the application of artificial life especially
on optimization, such as Ant Colony Optimization [1],
Particle Swarm Optimization and Fish-swarm
Algorithm. These bionic optimization algorithms
simulate the way that nature adapting the environment
and build up a random, positive feedback and
distributed computation model. These algorithms are a
kind of uncertain algorithm and independent from the
mathematical characters of the problems. Figure 1. A L-system production
The existing bionic optimization algorithms usually
imitate animal behavior, but don’t pay much attention 2.2. Branch axes and branch point
to the plant. Compared to the animal, the movement of
the plant isn't so obvious, it always growth toward the The term axis denotes any linear stem structure
most beneficial direction. Inspired of this idea, some from its origin to its extremity. An axis may include
scholar proposed a plant growth simulation algorithm monopodial or sympodial branch points. A monopodial
for solving integer programming [2]. In this paper a axis is a sequence of branch segments, each of which
virtual plant growth model is built to simulate the plant extends its predecessor in the terminal position.
growth process from a new angle and a new algorithm, Sympodial branching indicates that the terminal axis
Plant Growth Optimization, is proposed. extends its predecessor and the child axis or axes
2. Basic concepts of artificial plant growth appear in the lateral position [4].

978-0-7695-3357-5/08 $25.00 © 2008 IEEE 1194

DOI 10.1109/ICICTA.2008.416

Authorized licensed use limited to: Indian Institute of TechnologyDharwad. Downloaded on February 18,2021 at 11:49:08 UTC from IEEE Xplore. Restrictions apply.
 2.3. Rashevsky model of phyllotaxis Where xi is the position of the point, N is the total
number of the points, f (.) is the fitness value of the
One tree has thousands branch points, which can
grow up into new stems is a complex process. The first point and f min is the minimum of the current points.
dynamic models of morphogenesis were suggested by
Rashevsky, Turing, and others [5]. Figure 2 shows an 3.2. Branching
idealized vine stalk. At the
tip of the growing stalk is a In artificial plant growth model the branch point is
growth bud containing a the point that can produce new points by branching,
mass of undifferentiated resembles the simple object in L-systems. The purpose
and totipotent cells. A cell of branching is to seek the optimal area. Based on the
is considered as a bag of Rashevsky model of phyllotaxis and inspired by the
fluid with homogeneous plant branching theory, we presented a random model
chemical composition. One Figure 2. Rashevsky to simulate the plant’s behavior of branching.
of the chemical constituents model of phyllotaxis
Two different critical values α and β , satisfied
is a growth hormone called
morphogen. The concentration of this morphogen is 0 ≤ α < β ≤ 1 , are selected randomly. They divide the
the observed parameter of the model. As the parameter state space of morphogen into three pieces. The branch
varies between 0 and 1, the state space of the model is point has three different branching modes accordingly.
a line segment. If the concentration of this morphogen ⎧Mode 1: Sympodial , Ai > β
exceeds a certain critical value, the growth function of ⎪ (2)
⎨Mode 2: Monopodial , α < Ai < β
the cell is turned on, the cell divides, and a branchlet ⎪Mode 3: No branching , A < α
comes into existence. ⎩ i

Generally speaking, the current best branch points

3. Artificial plant growth modeling may not perform sympodial branching every
generation. In this paper, we dispose the critical values
A simple but effective algorithm is presented to as follows
simulate plant growth in a realistic way in this paper, α = α / 0.9 β = β / 0.9 (3)
taking into account branching, phototropism, leaf The new point produced by branching of the branch
growth and spatial occupancy. The main idea of the points is called growth point in this paper. In order to
model is to select the activated points by comparing abstain from getting stuck onto a local optimum,
the concentration of their morphogen to augment the artificial plant produce some new points called random
L-system. points in the areas with fewer axes randomly. So there
are four branching modes in all.
3.1. Concentration of the morphogen
3.2.1. Sympodial. Not only the axis elongates towards
It has already been proved by biological experiment the direction of the leaf but at the same time, child axes
that the morphogen is given by the environment of the appear in the lateral position. In this mode, the branch
cell’s position and the plant performs the characteristic point’s concentration of the morphogen is at a high
of phototropism obviously. In the model, a point level, that is, the fitness value is better than other
corresponds to a potential solution instance to the branch points. PGO search the optimum solution
problem and its activity, represented as the towards to the direction of the current best point and
concentration of the morphogen, is decided by the strengthen the searching around the branch point.
result of the objective function called fitness value.
In a minimization problem for example, the 3.2.2. Monopodial. The axis only elongates towards
the direction of the leaf. The concentration of the
concentration of the morphogen Ai is defined as
morphogen is at an average level in this mode and
⎧ f ( xi ) − f min PGO search the potential optimum solution towards to
⎪1 − N , f ( xi ) > fmin (1)
⎪ the direction of the current best point.
Ai = ⎨ ∑ ⎡⎣ f ( x j ) − fmin ⎤⎦ i = 1,2...N
⎪ j =1
⎪⎩1 , f ( xi ) = fmin 3.2.3. No branching. The concentration of the
morphogen is at a low level and the axis doesn’t grow
and only performs the leaf growth behavior.

1195

Authorized licensed use limited to: Indian Institute of TechnologyDharwad. Downloaded on February 18,2021 at 11:49:08 UTC from IEEE Xplore. Restrictions apply.
 3.2.4. Branching randomly. The points appear in the When a branch points have grown for a certain
sparse area randomly. This also matches an actual generations and still can’t find a point with a better
circumstance: The plants always produce smaller eligibility, it is considered to be mature. A mature
branches in the areas where the branches are sparse. branch point doesn’t take part in the growth. We record
The random points help to keep diversity of the points its information and compare it with the other branch
and may lead to a very good result. points in the next generations.

3.3. Leaf growth 4. The algorithms

After producing the new points, artificial plant As we are not interested in simulation of the natural
searches the optimum solution around these points plant growth process, but in the use of artificial plant
through the operation called leaf growth. This is used as an optimization tool, our system will have some
to ensure the accuracy of the solution. Two concepts major differences with a real one. The PGO take the
are introduced in this paper solution space of the problem as the growth area of the
z We call a point “leaf point”, around the branch artificial plant, one point of the plant as one potential
point, when it has a better fitness value; solution to the problem. The algorithm searches the
z The branch point is mature in the current optimal point in the solution space through two
generation when we can’t find a point with a better behaviors:
fitness value around the branch point; 1. Producing new points by branching to search the
If we can find the leaf point, record its information. optimal area where the optimum solution is;
Otherwise, the branch point is mature. By the way, the 2. Growing leaves around the branch points to find
area for the leaf growth shouldn't be too large. The way the accurate solution in the local area;
we search the leaf point, the local search strategy, can Given the definitions of the preceding section,
be any optimization algorithm. In this paper, we give formally the Plant Growth Optimization is:
priority to the pattern search method. We applied it to 0. Start
the test problems and found that this strategy can gain 1. Initialize:
higher accuracy with smaller calculation. Set NG=0 {NG is the generations counter}
Set NC=0 {NC is the convergence counter}
3.4. Selection mechanism Set NM=0 {NM is the Mature points counter}
Set the upper limit of the branch points N and
One tree only support a certain amount of branch initialize other parameters.
points but the current branch points produce a series of Select N0 branch points at random and perform
new points including growth points and random points. leaf growth.
It is necessary to establish a mechanism to select the 2. Assign morphogen
branch points for the next generation. Calculate the eligibility of the leaf point.
Most selection mechanisms choose the points with Assign the concentration of the morphogen of
better fitness values correspondingly. But under this each branch point by Eq.1.
mechanism, the convergence may occur very rapidly 3. Branching
so that it becomes impossible to reach to the optimum. Select two critical values between 0 and 1
This problem, which is called premature convergence, randomly and dispose by Eq.3.
is similar to the problem of getting stuck onto a local Produce new points by branching in four modes.
optimum encountered in local search methods. 4. Selection mechanism
The actual circumstance that the competition turns Perform leaf growth in all of the points.
more vehement where the points are too crowd and the Pick out the mature branch points, the number of
ones of stronger competition ability will restrain the which is k ( 0 ≤ k ≤ N ), by the maturity mechanism.
lower ones. In order to keep the variety of the branch Set NM = NM + k
points, we produce a new point in the center of the Produce a new point in the center of the crowded
crowded area and compare its fitness value with the area and select the best point to substitute the
closely points. The ones of lower competition ability crowded points.
will be eliminated and more chance is given to the Eliminate the lower competition ability branch
points in the scanty area. points and select N branch points for next generation.
5. Competition
3.5. Maturity mechanism Compare the current points with the mature
points and get the best fitness value f max

1196

Authorized licensed use limited to: Indian Institute of TechnologyDharwad. Downloaded on February 18,2021 at 11:49:08 UTC from IEEE Xplore. Restrictions apply.
 Set: NG = NG + 1 We set the initial point (-10, -10), which is on the
If ( f max < f max_ old ) Set: f max = f max_ old edge of the solution space. Figure 4 shows the
convergence history of the Rosenbrock function. The
If( | f max − f max_ old |< ε ) Set: NC = NC + 1 object fell below 1 after 21 cycles. In fact, the object
else was close to the global optimum after 47 cycles and
Set: NC = 0 satisfied the termination criteria after 54 cycles. PGO
got the optimal point (1.00000029, 1.00000055) at last.
else
The results indicate that the selection mechanism is
Set: NC = NC + 1 working and the PGO can overcome the problem of
6. Check the termination criteria: premature convergence.
If (NG < NGmax & & NC < NCmax & & NM < NM max )
Goto step2
else
Exit
7. Stop
One execution of the procedure from step2 through
step6 is called a generation or a cycle.

5. Experimental studyⅠ: Performance of Figure 4. Convergence history of Rosenbrock

the PGO function

We apply the proposed algorithm to three classical 5.2. Schaffer function

test problems to test its performance. The Rosenbrock
function tests the capability of avoiding premature The Schaffer function is defined as
convergence (Section 5.1). The Schaffer function tests (sin x12 + x22 ) 2 − 0.5
the performance of abstaining from get stuck onto a min f ( x1 , x2 ) = −0.5 +
(1 + 0.001( x12 + x22 )) 2
local optimum (Section 5.2). Finally, the Speed (5)
Reducer Problem tests the performance of solving s.t. x1 ∈ [ −10,10]
optimization problem with constraints (Section 5.3). x2 ∈ [ −10,10]
It is a continuous multimodal optimization problem
5.1. Rosenbrock Function with the global minimum -1 at the optimal point (0, 0).
There are countless
The Rosenbrock function is defined as local optimal points
min f ( x1 , x2 ) = 100( x12 − x2 )2 + (1 − x1 )2 around the global
(4) optimal point (Fig. 5)
s.t. x1 ∈ [ −10,10]
and algorithms usually
x2 ∈ [ −10,10] get stuck onto a local
It is a continuous unimodal optimization problem optimum.
with the global We set the initial
minimum 0 at the point (-10, -10), Figure 5. Schaffer function
optimal point (1, which is on the edge
1), which is on the of the solution space. Algorithm satisfied the
concave of a termination criteria after 17 cycles and finally got the
paraboloid (Fig. optimal point (-0.00000003, 0.00000000). We select 4
3). When region typical cycles to show how PGO found the optimal
of search come point in figure 6.
into the long and In cycle 1 the growth points were around initial
narrow place on Figure 3. Rosenbrock function point (-10, -10) and the random points guided the
the bottom, the direction of optimization. In cycle 2 and 6 the growth
variation of the and the random points distributed in the whole space.
object is very tiny. Commonly, algorithms repeat The growth points strengthened the search in the local
search in the concave and finally converge near the space and the random points helped to escape the local
optimal point. optimum. In cycle 17 the random points had little
contribution to the optimization, so we just show the

1197

Authorized licensed use limited to: Indian Institute of TechnologyDharwad. Downloaded on February 18,2021 at 11:49:08 UTC from IEEE Xplore. Restrictions apply.
 branch points and the growth points. The growth points ⎡⎛ 745.0x ⎞2 ⎤
0.5

helped to search the accurate solution and finally got A1 = ⎢⎜ 4

⎟ +16.9 ×10 ⎥
6
B1 = 0.1x63
x x
⎣⎢⎝ 2 3 ⎠ ⎦⎥
the global optimum. The results of this test show that 0.5
PGO can abstain from getting stuck on a local ⎡⎛ 745.0x ⎞2 ⎤
A2 = ⎢⎜ 5
⎟ +157.5×10 ⎥ B2 = 0.1x7
6 3

optimum and find the global optimum. ⎢⎣⎝ x2 x3 ⎠ ⎥⎦

2.6 ≤ x1 ≤ 3.6 , 0.7 ≤ x2 ≤ 0.8 , 17 ≤ x3 ≤ 28, 7.3 ≤ x4 ≤ 8.3
7.3 ≤ x5 ≤ 8.3, 2.9 ≤ x6 ≤ 3.9 , 5.0 ≤ x7 ≤ 5.5
We set the initial point (3.6, 0.8, 28, 8.3, 8.3, 3.9,
5.5), which is breach of constraints. Figure 8 shows the
convergence history of the object and constraint of
SRP. PGO converged after 74 cycles and found the
optimum solution 2994.35502622 at the point (3.5, 0.7,
17, 7.3, 7.71531991, 3.35021467, 5.28665447),
(a) Cycle 1 (b) Cycle 2 satisfied the constraints. The results show that the PGO
can solve the optimization problem with constraints.

(c) Cycle 6 (c) Cycle 17

Figure 6. Typical cycles to solve Schaffer function
Figure 8. Convergence history of SRP
5.3. Speed reducer problem
6. Experimental studyⅡ: Population size
The Speed Reducer Problem (SRP) represents the
design of a simple gear box shown in figure 7. The In this section we discuss experiments which have
objective is to minimize the speed reducer weight deepened our understanding of the Plant Growth
while satisfying a Optimization. PGO requires parameter tunings of the
number of constraints population size in order to achieve desirable solutions
imposed by gear and and performance for an arbitrary complex problem.
shaft design practices. Its The population size discussed here is:
global optimum is on the z N0: the size of the initial points;
edge of the solution z m: the size of the random points;
space. In this paper we The tests were carried out on the Rosenbrock
function. Based on these tests we’ll give some advices
used it to test the for the tunings of these parameters.
Figure 7. Speed
performance of solving Reducer Problem
optimization problem 6.1. Size of initial points
with constraints. SRP is defined as
min f ( x) = 0.7854 x1 x22 (3.3333 x32 + 14.9334 x3 − 43.0934) We compared the performance of PGO when the
2 2 3 3
(6 initial points are positioned on a unique starting point
− 1.5079 x1 ( x + x ) + 7.477(x + x )
6 7 6 7
with the performance obtained when the initial points
+ 0.7854( x4 x62 + x5 x72 ) are positioned on the whole space evenly. In the first
) case, we set the initial point at (-10, -10), which was on
s.t. g1: 27.0 /( x1 x22 x3 ) − 1 ≤ 0 g 2 : 397.5 /( x1 x22 x32 ) − 1 ≤ 0 the edge of the solution space (N0=1). In the second
g3 : 1.93x43 /( x2 x3 x64 ) − 1 ≤ 0 g 4 : 1.93x53 /( x2 x3 x74 ) − 1 ≤ 0 case, we use Latin Hypercube Sampling (LHS) to
g5 : A1 / B1 − 1100 ≤ 0 g 6 : A2 / B2 − 850 ≤ 0
produce initial branch points (N0=16). The size of
random points m=10 in both cases.
g 7 : x2 x3 − 40.0 ≤ 0 g8 : 5.0 ≤ x1 / x2 ≤ 12.0
The result is shown in Figure 9. PGO found the
g 9 : (1.5 x6 + 1.9) / x4 − 1 ≤ 0 g10 : (1.1x7 + 1.9) / x5 − 1 ≤ 0 optimal points in both cases. Convergence occurred
after 54 cycles in case 1 and 36 cycles in case 2. With

1198

Authorized licensed use limited to: Indian Institute of TechnologyDharwad. Downloaded on February 18,2021 at 11:49:08 UTC from IEEE Xplore. Restrictions apply.
 uniform distributing initial points PGO had a better that a moderate size (about half of the branch points) of
performance. The formal Design of Experiments the random points could improve the PGO’s
(DOE) techniques may aid in forming the dispersal performance.
pattern to achieve a reasonable coverage of the domain.
Usually there are too many factors in a complex 7. Conclusion
problem and LHS method will be feasible just need
less sampling points. A global optimization algorithm, Plant Growth
Optimization, based on the plant growth theory is
proposed. The main idea of PGO is that each branch
point searches optimum in the local space and the plant
guides the search direction by refreshing the branch
points to enhance the search around the good points
which is closer to optimum. It is applied to three
classical test problems and the results show that PGO
Figure 9. Convergence history of Rosenbrock has the desirable characteristics as following:
Problem with different N0 z It is a global optimization algorithm. The search
mechanism of the PGO can help to overcome the
6.2. Size of random points problems of getting stuck on a local optimum and
premature convergence.
The purpose of producing random points is to z It is adaptable which can be applied to the
abstain from getting stuck on a local optimum and we optimization problems with constraints or not.
had already proved it in Section 5.2. This experiment z It has a good openness. The local search strategy of
was run in order to study whether the size of the the PGO can be chosen as any optimization
random points influences the PGO’s performance. We algorithm. It means that PGO can be improved to
tested whether there is any difference among the cases be a hybrid algorithm easily with other excellent
in which m=2, 10, 50. In all of these three cases the algorithm.
initial point was set (-10, -10) and the number of the Further research is recommended to the study of the
branch points N=20. practical applications for complex systems.

8. References
[1] M.Dorigo, V.Maniezzo, and A.Colorni “The Ant System:
Optimization by a colony of cooperating agents”,IEEE
Trans. On Systems, Man, and Cybernetics, Vol.26, 1996, pp.
1-13.

Figure 10. Convergence history of Rosenbrock [2] LI Tong, WANG Chun-feng, WANG Wen-bo, and SU
Wei-ling, “A Global Optimization Bionics Algorithm for
Problem with different m
Solving Integer Programming - Plant Growth Simulation
The convergence histories are shown in Fig. 10. Algorithm”, SYSTEMS ENGINEERING - THEORY &
PGO found the optimal points in all of the three cases. PRACTICE, No.1, 2005, pp.77-85.
Convergence occurred
after 95 cycles in case [3] A.Lindenmayer, “Mathematical models of cellular
1, 54 cycles in case 2 interaction in development, Parts I and II”, Journal of
and 40 cycles in case Theoretical Biology, 1968, 18: 280–315.
3. The results indicated
that the more the [4] P. Prusinkiewicz, W. Remphrey, “Characterization of
architectural tree models using L-systems and Petri nets”,
random points is the
The Tree 2000: Papers presented at the 4th International
faster PGO escaped the Symposium on the Tree, pp. 177-186.
local optimum.
Figure 11 shows the Figure 11. Number of the [5] M.Klaus, Thinking in Complexity, Springer Berlin
total number of all points with different m Heidelberg New York, 2007.
points (random
points, growth points and branch points) in the
solution, which can be used to evaluate the elapsed
task time. It was least when m=10, the results show

1199

Authorized licensed use limited to: Indian Institute of TechnologyDharwad. Downloaded on February 18,2021 at 11:49:08 UTC from IEEE Xplore. Restrictions apply.