Beruflich Dokumente
Kultur Dokumente
Ivan Zelinka
University of Vaasa
Department of Information Technology
and Production Economics
P. O. Box 700, FIN-65101 Vaasa, Finland
phone: +358-6-3248 111, fax: +358-6-3248 467
E-mail: Jouni.Lampinen@UWasa.fi
Abstract: This article discusses solving non-linear programming problems containing integer, discrete and
continuous variables. The Part 1 of the article describes a novel optimization method based on Differential
Evolution algorithm. The required handling techniques for integer, discrete and continuous variables are described including the techniques needed to handle boundary constraints as well as those needed to simultaneously deal with several non-linear and non-trivial constraint functions. In Part 2 of the article a mechanical
engineering design related numerical example, design of a coil spring, is given to illustrate the capabilities and
the practical use of the method. It is demonstrated that the described approach is capable of obtaining high
quality solutions. The novel method is relatively easy to implement and use, effective, efficient and robust, which
makes it as an attractive and widely applicable approach for solving practical engineering design problems.
Keywords: evolutionary algorithms, differential evolution, non-linear optimization, engineering design
1 Introduction
In order to illustrate capabilities of the solution techniques introduced in Part 1 of this article, a practically oriented numerical example is given here in Part 2. The design of a coil spring (Figure 1) was optimized by using
Differential Evolution (DE) algorithm and soft-constraint approach as described in Part 1 of the article. This nonlinear engineering design optimization problem, containing one discrete, one integer and one continuous design
variable, was first investigated by Eric Sandgren [San90] and subsequently by other researchers [CT93, WC95]
by applying various optimization techniques. The example problem is based on a real-world mechanical engineering design optimization. Thus, it demonstrates optimization situations with discrete, integer and continuous
variables occurring in practice. The particular methods used for solving the example are discussed in detail and
the results obtained are compared with the earlier results, found in literature, that are obtained by using other
optimization methods.
0.009
0.014
0.023
0.047
0.105
0.192
0.307
0.0132
0.020
0.041
0.092
0.177
0.283
0.500
2 x2 x32 (x1 + 2)
4
subject to
8C f Fmax x2
S 0
g1 ( X ) =
x33
g2 ( X ) = l f lmax 0
f (X ) =
g3 ( X ) = d min x3 0
g4 ( X ) = x2 Dmax 0
x
g5 ( X ) = 3.0 2 0
x3
g6 ( X ) = p pm 0
g7 ( X ) = p +
g8 ( X ) = w
Fmax Fp
K
Fmax Fp
K
+ 1.05( x1 + 2) x3 l f 0
0
where
Cf =
4(x2 / x3 ) 1 0.615 x3
+
0
4(x2 / x3 ) 4
x2
Gx34
8x1 x23
Fp
p =
K
F
l f = max + 1.05(x1 + 2 )x3
K
K=
The objective function, f(X), computes the volume of spring steel wire as a function of the design variables. The
design limitations are specified as follows:
3dmin x2 Dmax
constraint g3
dmin x3
constraint g4
Dmax
3
i =1
i=6
gi ( X ) > 0
Item
Optimum solution
Chen and Tsao Wu and Chow
[CT93]
[WC95]
9
9
x1 (N) [1]
Sandgren
[San90]
10
x2 (D) [inch]
1.180701
1.2287
1.227411
1.2230410
x3 (d) [inch]
0.283
0.283
0.283
0.283
g1(X)
54309
415.969
550.993
1008.8114
g2(X)
8.8187
8.9207
8.9264
8.94564
g3(X)
0.08298
0.08300
0.08300
0.083000
g4(X)
1.8193
1.7713
1.7726
1.77696
g5(X)
1.1723
1.3417
1.3371
1.32170
g6(X)
5.4643
5.4568
5.4585
5.46429
g7(X)
0.0
0.0
0.0
2.6758110-16
g8(X)
0.0
0.0174
0.0134
5.0751510-16
2.7995
2.6709
2.6681
2.65856
100.0%
95.4%
95.3%
95.0%
f(X) [inch3]
This article
9
Type of
variable
integer
continuous
discrete
Generally, a robust search and high convergence velocity are conflicting objectives. An experiment was performed to illustrate this by using alternative values for DEs control parameters, namely F=0.70 and Cr=0.80,
which values were selected to favor a high convergence velocity rather than a robust search. In this case 200
generations, corresponding 8000 function evaluations, were computed. With respect to the original settings,
F=0.90 and Cr=0.90, the convergence velocity increased by 325%. Only 5 out of 100 independent runs ended up
with a result that was worse than that reported in Table 3. Thus, a significantly faster convergence was achieved
by sacrificing some robustness.
3. Conclusions
As demonstrated, the DE appears to be promisingly efficient, effective and robust optimization algorithm. It is
both easy to implement and easy to use. It was capable of optimizing all integer, discrete and continuous variables, and capable of handling non-linear objective functions with multiple non-trivial constraints. Furthermore,
a high quality solution was obtained for the discussed practically oriented test problem. The solution found by
DE was better than the best solutions found in literature. Generally, the used test problem, that was earlier considered to be rather difficult to solve, was found to be not enough difficult for finding the limits of the described
approach.
In order to demonstrate the robustness of the optimum-seeking algorithm, 100 independent trials were performed. All runs of the optimization program yielded the reported value of f(X) or better. All of the obtained
solutions were within the feasible region of the design space.
DE is not remarkably sensitive to the values of its control parameters npop , F and Cr. This characteristic is
demonstrated here by selecting the values for the control parameters F and Cr among the multiples of 0.1 and
selecting the population size npop among the multiples of 10. Similarly, the number of generations to be evaluated
was selected to be a multiple of 50.
The values of npop , F and Cr , however, still affect on the convergence velocity and robustness of the algorithm. Here the values were selected to favor robustness rather than convergence velocity. Because computationally inexpensive objective function was used, the highest possible convergence velocity was not important. Despite this, convergence was still relatively fast. However, as demonstrated, it is possible to increase the convergence velocity even more than by a factor of 3 by favoring fast convergence over robustness.
It is sometimes argued that the evolutionary optimization requires a large number of cost-function evaluations.
Thus, the computational cost would be high. Before deciding whether or not this is true in case of DE, some
further facts should be considered. The real-world optimization problem discussed here can be solved on an
ordinary PC using less than 1 second of computation time. Observe that the test problem is relatively difficult to
solve, too difficult for most existing optimization methods. Unfortunately insufficient data was available to accurately compare the number of objective function evaluations required by the other methods to those required by
DE. For some unknown reason the articles, which were referred for comparisons, do not report this information
properly. Anyway, the number of objective function evaluations would be important only if the compared algorithm is, in fact, capable of finding the optimal solution.
Despite the method was found promisingly efficient, effective and robust, the current test problem failed to uncover any major limitation of the discussed DE algorithm. In future, a larger problem set, preferably containing
even more difficult problems than the one discussed here, should be a subject of investigations. Because the
target of this article was mainly to introduce and demonstrate the novel approach, the convergence properties of
DE were neither studied, nor discussed comprehensively, which should also be included to the future studies in
order to create a more complete picture about the properties of the novel optimization technique.
Acknowledgements
The authors would like to thank Kenneth Price and Rainer Storn for their constructive comments, advises, suggestions and discussions concerning this investigation. The authors would like to thank especially for sharing
their latest knowledge and experiences on Differential Evolution algorithm.
References
[CT93] Chen, J. L. and Tsao, Y. C. (1993). Optimal design of machine elements using genetic algorithms. Journal of the
Chinese Society of Mechanical Engineers, 14(2):193199, 1993.
[Sid82] Siddall, James N. (1982). Optimal engineering design: principles and applications. Mechanical engineering series /
14. Marcel Dekker Inc. ISBN 0-8247-1633-7
[SPEC95] SPECint95 and SPECfp95 computer benchmarks (1998). The Standard Performance Evaluation Corporation.
http://www.specbench.org/
[San90] Sandgren, E. (1990). Nonlinear integer and discrete programming in mechanical design optimization. Transactions
of the ASME, Journal of Mechanical Design, 112(2):223229, June 1990. ISSN 0738-0666
[WC95] Wu, S.-J. and Chow, P.-T. (1995). Genetic algorithms for nonlinear mixed discrete-integer optimization problems
via meta-genetic parameter optimization. Engineering Optimization, 24(2):137159, 1995. ISSN 0305-215X