ADM2006 - Paper No.309 Proceedings of the International Conference on Advanced Design and Manufacture 8-10 January, 2006, Harbin, China
A NOVEL EVOLUTIONARY ALGORITHMS BASED ON NUMBER THEORETIC NET FOR NONLINEAR OPTIMIZATION Fei Gao Dep. of Mathematics, Wuhan University of Technology, 430070, P. R .China E-mail:
[email protected]
Abstract: How to detect global optimums which reside on
optimize the problems such as genetic algorithms,
complex function is an important problem in diverse scientific
evolutionary algorithms, simulated annealing and etc.
fields. Deterministic optimization strategies, such as the
Stochastic methods have been proved to be capable
Newton–family algorithms, have been widely applied for the
of finding global optimum by an asymptotic convergence.
detection of global optimums. However, in the case of
These ideas require a primary stochastic evolving form
discontinuous /non-differentiable /non-convex complex functions,
and converge to one optimum. The samples generated by
this approach is not valid. In such cases, stochastic optimization
evolving operators such as selection, mutation and
strategies simulating evolution process have proved to be a
acceptance concentrate around from a local optimum to
valuable tool. In this paper, a novel evolutionary algorithm
global optimum in terms of computational efforts needs.
based on Number Theoretic Net for detecting global optimums of
This problem may be solved by improving the
Complex functions is introduced. It can be applied to any
distribution of points so as to mine the information of
functions with multiple local optimums. With some established
variable up mostly. This paper proposes a novel approach
techniques, such as the ideas of genetic algorithms and
Number theoretic evolutionary algorithm(NTEA) for
sequential number theoretic optimization to improve the property
global optimization of complex function by combining the
of convergence in large scale, the deflection and stretching of
idea
objective function to guarantee the detection of a different
(SNTO)[3~5] with the concepts from evolutionary
minimizer, it detects optimums of a function through adding
algorithms. This method adds the updating rules with the
genetic operations to the feasible points generated by number
ideas of mutation and other genetic operations to the
theoretic net sequentially. The experiments done indicate that the
feasible points generated by number theoretic net. The
proposed algorithm is robust and efficient.
experiments being done show that NTEA has the
Key words: Optimization, Evolutionary Algorithm, Number
advantages of high precision and high robustness on
Theoretic Net, Mutation, Deflection
multi-model function’s optimization problems to a certain
of
sequential
number
theoretic
optimization
extent. The rest of this paper is organized as follows. In 1 Introduction
Section 2, firstly we introduce the background of our
Detecting global optimums of complex functions is an
research, i.e. details of SNTO and the main idea of
important
fields.
evolutionary algorithms (EA), then the previous work on
Mathematically, complex functions are discontinuous/
extension of SNTO and EA is discussed. We propose a
nondifferentiable/ non-convex full of local optimums. An
novel approach NTEA with established techniques through
extensive literature review [1,2] indicates that most of the
function deflecting and adding evolutionary operations
existing deterministic optimization strategies for complex
and we examine the effectiveness of the novel method by
functions, such as the Newton–family algorithms, are
applying it into classical benchmark functions in Section4.
limited to the optimization of a local area with various
And Section 5 summarizes the paper.
problem
in
diverse
scientific
excess stories using stochastic approaches to globally
2 Backgrounds
Note P
2.1 Number-theoretic methods’ sequential realization Number theory method’s sequence realization SNTO is a Hua L. K. et al. In this method number theoretic net is
M ( t ) = f ( x (t ) ) ≥ f ( y ) , where ∀y ∈ Q ( t ) and
used to generate good points in terms of good lattice sequential contraction the searching field it find the global
x(t), M(t) are the current best value of x*, M*; (t)
(t)
(t)
0
(t)
enough small, c =( b -a )/2, if max c <δ, then
: Suppo ( n; h1 ," , hs )( s < n) is a integer
x(t), M(t) are accept, otherwise go to Step4;
[4]
vector s.t. 1 ≤ hi < n, common
δ >
Step3: Termination condition Judging. Given
optimum theoretically. Definition 1
∈ [a (t ) , b(t ) ] , x ( −1) = {φ} , take
x (t ) ∈ P ( t ) ∪ {x (t −1) } = Q (t ) s.t .
global optimization approach based on number theory by
points scattered in feasible domains, then through
(t )
hi ≠ h j (i ≠ j ) and greatest (n, hi ) = 1, i = 1," , s;
measure
⎧⎪ qki = khi (mod n), k = 1," , n , modify the where ⎨ ⎪⎩ xki = (2qki − 1) / 2n, i = 1," , s
Step4:
Region
contraction.
Define
a new
region
D(t+1)=[a(t+1), b(t+1)] as
⎧ai(t +1) = max( xi(t ) − γ ci( t ) , ai( t ) ), (i = 1, 2," , s ) ⎨ (t +1) (t ) (t ) (t ) ⎩ bi = min( xi + γ ci , bi ), where γ∈(0,1) is pre-given contraction ratio, generally γ=0.5, t =t+1, go to Step1; 2.2 Evolutionary algorithm
≤ n,
Evolution has created wonders like us. There is much to
then the lattice point set of generic vectors ( n; h1 ," , hs )
study of computational systems which use ideas and get
common congruence method to make qki s.t . 1 ≤ ki
defined by Pn
be learned from Nature. Evolutionary computation is the inspirations from natural evolution. The Evolutionary algorithm (EA) [6~9] has been
= {xk = ( xk1 ," , xks ), k = 1," , n}.
widely used for solving optimization problems. In EA
Suppose f(x) is continuous function on bound and
initial population generated by random points in feasible
closed set D , take a NT-net P = { xk , k = 1, 2, , n }
region of problem which is likely to select the fittest
xn∗ ∈ {xk }
the environment survive at higher probabilities in the next
…
and
let
stand
∗
point s.t . M n = f ( xn ) = max f ( xk ) n
1≤ k ≤ n
for
, we expect
by
uniformly
scattered
points
generation is formed through evolutionary operations such as crossover, mutation and saving the best for respective series of these operations. We expect that the number of
The essence of Number theory method is to replace points
generation among those that forms a generation. The next
individuals. Solution searching is pursued by repeating a
∃xn∗ ∈ { xk } s.t . f ( xn∗ ) → M , xn∗ → x∗ (n → ∞) random
individuals so that those proven to have greater fitness for
individual with higher fitness (that is, those closer to
on
optimal solutions) increases as the search make progress.
s-dimensional unit cube. When SNTO is used to solve the
Thereby an optimal solution can be achieved. The above
multi-modal function optimization problem, it requires not
describes the basic concept of EA.
computing f(x)’s derivative at point P but f(x)’s value at P,
2.3 Previous studies
and it’s independent of derivation and the choice of initial
But as the other global optimization method, SNTO is
points.
exhaustive and deterministic, and it fails to jump off local [4]
Now we give the main progress of SNTO as below: Algorithm 1. SNTO :
points, but it doesn’t mine more information from these (0)
(0)
(0)
Step0: Initialization. t=0, D =D, a =a, b =b; (t)
Step1: Generating a NT-net P uniformly distributed on (t)
optima. In beginning of the method, it has a big sample
(t)
[a , b ] by using a number theory method; Step2: Computing the new fitness.
points. So mining these points may generate more effective points globally. Evolutionary algorithms is a kind of self-adaptive heuristic method simulating evolution process effective on
complex functions’ optimization problems, and it’s similar
s.t . M ( t ) = f ( x (t ) ) ≥ f ( y ) , where ∀y ∈ Q ( t )
to SNTO on "Multi-points", thus this gives the idea of
and x (t), M (t) are the current best value of x*, M*;
combining the idea of SNTO with the concepts from
If the objective function f(x) is full of local optimums
evolutionary algorithms. of
and more than one minimizer is needed, we choose
evolutionary calculation technique with number theoretic
another established techniques to guarantee the detection
method have been done in two ways mainly: First, using
of a different minimizer, such as deflection and stretching
number theoretic net to generate initial feasible points for
are introduced. Suppose objective function is f(x), we use
EA [10]; second, parallel strategies [11] for SNTO using
deflection technique as below to generate the new
small initial feasible population and using evolutionary
objective function F(x)[8]:
Some
Previous
studies
of
combination
operator crossover to these point.
k
F ( x) = ∏ ⎡⎣ tanh ( x − xi∗ ) ⎤⎦
3.1 Proposed Technique NTEA Based on the above mentioned studies, We propose a novel number-theoretic evolutionary algorithm (NTEA) by doing evolutionary operations (mainly mutation) on equi-distribution set Pn generated by number theoretic net in SNTO while keeping SNTO’s advantages in our technique, just modifying the step2 of SNTO, the NTEA is given as below in details: Algorithm 2. NTEA Step2: Computing the new fitness.
= P (t ) ∈ [a ( t ) , b(t ) ] , do mutation to
(t )
f ( x)
i =1
3 NTEA and Numerical Experiments
Note P0
−1
where
xi∗ (i=1,2,…,k) are k minimizers founded,
λ∈(0,1). We also introduce stretching technique[8] to generate the new objective functions G(x) and H(x) as new objective functions:
G ( x) = f ( x) + β1 x − xi∗ ⎡⎣1 + sgn( f ( x) − f ( xi∗ )) ⎤⎦ 1 + sgn( f ( x) − f ( xi∗ )) H ( x) = G ( x) + β 2 tanh ⎡⎣δ (G ( x) − G ( xi∗ )) ⎤⎦ where β1, β2, δ > 0. Fig.1. shows deflection and stretching effects on f(x)= cos x at x=π
P0(t ) in probability p1 , for xi ∈ P0(t ) , ⎧ xi ,k + Δ(t , y ), rand = 0 , ⎩ xi ,k − Δ(t , y ), rand = 1
Let xi , k = ⎨
, Δ (t , y ) = y[1 − r
y = bk(t ) − ak(t )
(1− Tt )b
] , T
denotes the algorithm’s largest iteration number, t is the iteration number now, b is coefficient, r=rand [0,1]; (t )
(t )
Then we get P1
, do crossover to P1
as
Fig.1.Deflection and stretching effects on f(x)
below: Suppose x1 , x2 ∈ P1
(t )
[0,
1],
k
then x1, k = 1
P1(t )
;
is
, β is a random number on
random
integer
β x1,k + (1 − β ) x2,k Note
on
[1,
s],
, and now we get
x ( −1) = {φ}
,
x (t ) ∈ P0( t ) ∪ {x (t −1) } ∪ P1(t ) ∪ P2(t ) = Q ( t )
take
In this way, we see that the searching algorithms will not locate x=π. So we can combine this technique with NTEA to find more minimizers needed more efficiently. 3.2 Benchmark functions To check the effective of these algorithms, we performed a series experiments. First we define standard functions [1] introduced mainly by De Jones and often used as quantitative evaluation means for the benchmark test of
EA and other technique. They are defined as below:
point is (0,0), and they are classic benchmark functions to
eg.1. f1(x,y)=21.5+xsin(4πx)+ ysin(20πy), -3.0 ≤
evaluate EA’s global convergence and local exploring
x ≤ 12.1,
4.1 ≤ y ≤ 5.8, its global maximum value is known as [1]
performance.
38.827553 . This is a classical multimodal function whose global
3.3 Comparisons of results among EA, SNTO and
optimum is surrounded with many local points.
For all the testified functions, we take the same generic
NETA
f2 (x,y)=-[20+xsin (9πy) +ycos (25πx)], where (x,y)
vector for number net to generate good points as these:
∈ D= {( x, y ) x + y ≤ 9 } , its minimum value
first cycle we take (987; 1,610), then we take (233; 1,144)
eg.2.
2
2
2
-32.71788780688353, correspond global minim point is [9]
in the latter cycles. To
avoid
non-convergence
for
long-playing
(-6.44002582194051, -6.27797204163553) . Fig.2 is f2
circulation we replace max c(t)<δ by min c(t)<δ; and
on D, where its value out of D is put as -40.
generally take δ=10-15, r=0.6. And we add mutation operator in contraction region. Especially, for f1 we choose mutation ratio 0.05, require the newer optimal not smaller than old one. For f2 we take δ=10-8, r =0.86, mutation ratio=0.85 and we make the function value 3 at points out of defined region and we take a new mutation method as this: Generate a random number α on [0, 1], if α>0.95 or α<0.05, then mutation in defined region; otherwise mutation in correspondent contraction region. For f6, we require the newer optimal not smaller than
Fig.2. f2 on D f2 is a multimodal function difficult to be optimized. Its defini region D is big, xsin (9πy) and ycos (25πx) oscillate in different directions in their ways and it has
old one not strictly. For f7, we take δ=10-30 and require the newer optimal bigger than old one not strictly. Now we use Matlab6.5 to program NTEA and its
deep valley clatters near to 4 points.
performance in contrasted to SNTO, simple evolutionary
eg.3. Easom function f3 (x1, x2) =-cos x1 cos x2
Algorithm SEA. The global search capability, convergence
2
2
exp[-(x1-π) -(x2-π) ], its global minimum is known as -1,
speed and robustness of the algorithms proposed is
correspond global minim point is (-π,π).
exhibited in Fig3~6: Fig.3 is NTEA contrast simulation
eg.4. Bohachevsk1 function
progress of the best, average and worst function value in
f4 (x1, x2) =
x12 + 2 x22 − 0.3cos(3π x1 ) − 0.4 cos(4π x2 ) + 0.7 .
iteration for f1. Fig.4 shows the local convergence of
eg.5. Bohachevsk2 function
contour of Easom. Fig.6 is contrast simulation progress of
f5 (x1, x2) =
x12 + 2 x22 − 0.3cos(3π x1 ) ⋅ cos(4π x2 ) + 0.3 . eg.6. Schaffer function
0.5 +
sin
2
(1.0 + 0.001( x
2 2
40
2 1
+ x )) 2 2
35 Best function avalue 30 Average function avalue
2
. And it is called
Multimodal Sine Envelope Sine Wave Function [2]. eg.7.
the three algorithms–SEA, SNTO and NTEA for f6.
f6 (x1, x2) =
x + x − 0.5 2 1
NTEA for f2. Fig.5 is the progress of optimization in
f7 (x1, x2) =
25 Worst function avalue f1 20
15
10
5
( x12 + x22 )0.25 ⎡⎣1.0 + sin 2 ( x12 + x22 )0.1 ⎤⎦ .
0
0
10
20
30
40
50
60
70
Iteration number
eg.3.~eg.7. are defined in -100≤x1, x2≤100, their global minimum is known as 0, correspond global minim
Fig.3. Comparison of Best, Average, Worst fitness for f1
No.2 [-6.44002582235322, -6.27797201356924], No.3 [-6.4400258222786, -6.2779720141167] No.4 [-6.44002582226788, -6.27797201447922], No.5 [-6.44002582235673, -6.27797201347218], No.6 [-6.44002582213847, -6.27797201162225], No.7 [-6.44002582211448, -6.27797201373397], No.8 [-6.44002582217536, -6.27797201223602], No.9 [-6.44002582208887, -6.27797201222687]
−6.276
−6.2765
−6.277
Y
−6.2775
−6.278
−6.2785
−6.279
−6.2795
From the Table1 and Table2, We could draw a
−6.28
conclusion that the performance of NTEA is much better
−6.2805
−6.442
−6.4415
−6.441
−6.4405
−6.44
−6.4395
−6.439
−6.4385
−6.438
than SNTO and SEA, especially for f1, f2, F6, F7. And for
X
Fig.4. Local convergence of NTEA for f2 in contour map
f4, f5 NTEA does as well as SNTO but superior to SEA increasingly.
10
9
8
1.48
1.4803e−016
Y
016
Cases studies illustrate that NTEA With established
1.4803e−016
7
6
4 Conclusions.
03e−
1.4803e−016
techniques in Section 3. for complex function optimization 1.4
1.4803e−016
−0
16
1.4803e−016
−0.8
1.4803e−016
1.4803e−016
4
. −0
−0
.6 −0.4 −0.2
2
1.4803e−016 1
3e
1.4803e−016
−0.2
4
3
1
can improve the global convergence speed and has the
80
5
2
3
advantages of high precision and robustness to such a certain extent. In this paper we propose a technique in which SNTO
1.4803e−016 4
5
6
7
8
9
10
X
Fig.5. Finding optimal for Easom in contour map
is combined with the concept of evolutionary calculation technique, and verified its effective using standard
0.35
functions. The improved-version NETA has inherited the
0.3
features of both: simple SNTO is strong at raw exploration and EA is robust on multimodal functions. At the end,
0.25
through experiments, NETA gave better performance than
SEA 0.2
either SNTO or EA.
SNTO
f
6
NTEA
In the future works, we will do more researches on
0.15
adjusting the parameters of genetic operations in NTEA
0.1
and their relations to the algorithm’s stability to improve
0.05
algorithm’s robustness and precision much more. And we
0 0
10
20
30
40
50
60
70
80
90
100
Iteration number
Fig.6. Comparison of Best fitness for f6
plan to apply this technique to more complicated and difficult problems. We would like to develop other algorithms that perform well by combining EA’s concepts
Table1 and Table2 are the results of contrasted
with the proposed algorithms.
optimal searched by NTEA, SNTO and SEA and correspondent optimal points to the optimal values in
ACKNOWLEDGEMENT
Table1 searched by NTEA. With deflection technique, we can detect 6 minimizers of f1 with the same value: No.1 [11.6255447026864, 5.72504424431332], No.2 [11.6255447027255, 5.7250442445237], No.3 [11.6255447026949, 5.72504424490065], No.4 [11.6255447027843, 5.72504424471521], No.5 [11.6255447027431, 5.72504424413162], No 6 [11.6255447027118, 5.72504424463833] And 9 minimizers of f2 with the same value are also founded: No.1 [-6.44002582207099, -6.2779720144943],
We thank the anonymous reviewers for their helpful remarks and comments. F. Gao was supported by the Foundation (Grant No.XJJ2004113), Project of educational research, and the UIRT Project (Grant No. A156, A157) granted by Wuhan University of Technology in People's Republic of China.
References [1]
Michalewicz
Z.
Genetic
Algorithms
+Data
Structures=Evolution Programs (3rd. edition).Berlin: [2]
[8]
orbits of nonlinear mappings through particle swarm
Srinivas M., Patnail L. M., Adaptive Probabilities of
optimization, in Proceedings of the 4th GRACM
Crossover and Mutation in Genetic algorithms. IEEE
Congress on Computational Mechanics, Patras,
Trans. on Sys. Man & Cybernetics, Vol., 24, No.4,
Greece, 2002. [9]
Apr.1994, pp. 656~666. [3]
[4]
Parallel
solving a system of nonlinear equations. J. Comp.
Optimization
Math, 1991, 9: pp.9~16.
2001(38): pp.1381~1386(In Chinese).
LIU B. D., ZHAO R. Q. Statistical Programming Fuzzy
Programming.
Beijing:
Algorithms
for
.Theoretical
Computer
Function Science,
[10] ZHANG Ling, ZHANG Bo, Good Point Set Based
Tsinghua
Genetic Algorithm. Journal of Computer, 2001(24):
University Press, 1998.
[6]
KANG L.S., LIU P., CHEN Y.P., Two Asynchronous
Fang K.T. and Wang Y. A Sequential Algorithm for
and [5]
Parsopoulos K., Vrahatis M., Computing periodic
Springer -Verlag, 1996.
pp.917-922(In Chinese).
KANG L.S., XIE Y., YOU S.Y. etc. Non-numerical
[11] LIU H.Q., YUAN X.G., FANG K.T. Application of
parallel algorithms: Simulating annealing algorithms.
evolution
Science Press, Beijing, 1997(In Chinese).
Optimization. Trans. of Tianjin University, 2001(8):
PAN Z.J., KANG L.S., CHEN Y. P. Evolutionary
pp.221-225(In Chinese).
Sequential
Number
Theoretic
Computation. Beijing: Tsinghua University Press, 1998(In Chinese). [7]
Hua L. K. and Wang Y. Applications of Number theory to Numerical analysis. Springer -Verlag & Science Press, Berlin & Beijing, 1981.
Table1 Optimum searched by NTEA, SNTO and SEA and known Value Test Function
NTEA
SNTO
SEA
Known Value
f1
38.8502944794474
38.8502944794474
38.827553
38.827553
f2
-32.7178878068835
-32.4245444521219
-32.4396133897864
-32.71788780688353
f3
-1
-1
-0.99988661032628
-1
f4
0
0
0.00204985947697
0
f5
0
0
0.00596033477647
0
f6
0
/
0
0.07035405596053
0
f7
1.69641423986802×10
0.00971590987751 -16
5.91645474391892×10
-14
Table2 Correspond Optimal point for the optimum in Table1 searched by NTEA Test Function
Correspondent optimal points in Table1 searched by NTEA
f1
(11.6255447027118, 5.72504424463833)
f2
(-6.44002582208887, -6.27797201222687)
f3
(3.14159264305311, 3.14159265381822)
f4 f5 f6 f7
Unknown (-6.44002582194051, -6.27797201463553) (-π,π)
-9
-9
(-3.00171076217092×10 ,-6.14027470452886×10 ) -10
-11
(-1.5805059471004×10 ,1.65931843809982×10 (-2.87741108834877×10
-32
(0,0)
-9
(0,0)
)
(0,0)
-34
(0,0)
(-8.27493200622298×10 ,-2.02504829295479×10 ) -8
Correspondent Known point
,-4.85869247033871×10 )