You are on page 1of 5

ITAIC 2020 (ISSN 2693-2865)

Adaptive GASA algorithm for multidisciplinary


design optimization
2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC) | 978-1-7281-5244-8/20/$31.00 ©2020 IEEE | DOI: 10.1109/ITAIC49862.2020.9339141

Chao Fu1a, Yang Ou2b, Jihong Liu1c* , HongYan Yu3d, WenTing Xu2d
1. School of Mechanical Engineering and Automation, Beihang University, Beijing, China
2. AVIC Shenyang Aircraft Corporation, Shenyang, China
3. Beijing Institute of mechanical and electrical engineering, Beijing, China
a b d
superfurun@163.com, zaiwei123456@163.com, c*ryukeiko@buaa.edu.cn, cadsupport@163.com
*Corresponding author˖Jihong Liu Email: ryukeiko@buaa.edu.cn

Abstract—In the application of multidisciplinary design In this paper, an efficient GASA hybrid optimization
optimization, intelligent optimization algorithm can better algorithm is proposed by combining the parallel search
avoid the problems of "falling into local optimum" or structure of GA and the probabilistic jump characteristics
"unexpected termination" caused by conventional of SA. The GASA hybrid algorithm first uses the parallel
numerical optimization methods. In this paper, an adaptive search of genetic algorithm, optimizes the population
GASA algorithm is proposed, which combines the global genetic operation based on the idea of "survival of the
parallel search ability of genetic algorithm with the fittest" in the sense of probability, and then uses the serial
probability jump characteristic of simulated annealing structure of simulated annealing algorithm to search
algorithm, and uses adaptive mechanism to improve the
through probability jump, and finally tends to the global
population operation of intelligent optimization algorithm.
Finally, the proposed method is verified by using a
optimal solution. The crossover of genetic algorithm is
calculation example in combination with the collaborative conducive to inheriting the excellent information of the
optimization strategy. The results show that the proposed parent, mutation can strengthen the diversity of
method can effectively improve the optimization efficiency individuals in the population, the simulated annealing
and the quality of the optimization results. algorithm decreases from high temperature to low
temperature, so that the population evolves from a large-
Keywords—Multidisciplinary design optimization; scale global jump to a small-scale local movement; GASA
adaptive mechanism; genetic algorithm; simulated annealing integrates the optimization mechanism of the two
algorithm algorithms, complements the optimization behavior and
structure, reduces the adverse impact of parameter setting
I. INTRODUCTION on the optimization results, and enhances the search
ability of global and local variable space by enriching the
With the expansion of the scope of engineering search behavior in the population evolution process [4].
problems, the scale and complexity of MDO problems are
increasing [1]. Intelligent optimization algorithm has been
deeply studied and widely used in the field of MDO, but II. GASA HYBRID ALGORITHM
its own defects and deficiencies are gradually exposed, As shown in Figure 1, the GASA hybrid algorithm
and the optimization results of the algorithm are often not first uses the parallel search of genetic algorithm,
ideal. Take the genetic algorithm (GA) and simulated optimizes the population genetic operation based on the
annealing algorithm (SA) widely used in the field of idea of "survival of the fittest" in the sense of probability,
MDO as an example: the global search performance of and then uses the serial structure of simulated annealing
genetic algorithm is better, which is suitable for the algorithm to search through the probability jump, and
variable space of multi-dimensional CO method, but it is finally tends to the global optimal solution.
prone to premature convergence. To improve its
performance, it needs to change gene operation or adopt GASA hybrid algorithm has the above advantages, but
advanced evolution strategy [2]; The simulated annealing also inevitably introduces the shortcomings of the two
algorithm has strong local search ability when the initial algorithms: crossover, mutation and other operations of
temperature is higher, the number of sampling steps is genetic algorithm may lose the optimal individual, and
longer and the termination temperature is lower. However, simulated annealing algorithm has poor time performance
the overall analysis of the variable space of CO method is to achieve global convergence. The genetic algorithm and
insufficient, so it is difficult to enter the most reliable simulated annealing algorithm in GASA will affect the
variable space search, and the optimization efficiency is quality of the co optimization solution. Therefore, it is
low [3]. The parameter setting of the algorithm will affect necessary to optimize and improve GASA algorithm
the optimization performance of both methods. combined with the characteristics of multidisciplinary
design optimization.

978-1-7281-5244-8/20/$31.00 ©2020 IEEE 725

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on March 24,2021 at 07:44:26 UTC from IEEE Xplore. Restrictions apply.
Population evolution space Variable space level GASA algorithm can adaptively adjust the crossover
Determine initial temperature Optimization and mutation operations during population evolution.
and initial solution objective function

° Pcini  Pc max  Pc min ˜ I c , fbig ! f avg


­
(2)
Evaluate each individual in the current Evaluate the best Pc ®
population fitness °̄ Pcini , fbig d f avg

Satisfy convergence YES ° Pmini  Pm max  Pm min ˜ I m , f w ! f avg


­
(3)
criteria? Pm ®
NO °̄ Pmini , f w d f avg

ln fbig  f avg ln f w  f avg


Crossover and mutation by GA

Among them, I c , Im .
Obtain initial population of SA, f max  f avg f max  f avg
SA search for individuals i = 1,Ă,pop

Where favg is the average fitness value of the current


Generate new individuals from SA state population, fmax is the maximum fitness value in the
generating function
current population, Pcini is the initial crossover probability,
Pmini is the initial mutation probability, Pcmax and Pcmin are
Accept new individuals
with probability the upper and lower limits of the crossover probability,
Pmmax and Pmmin are the upper and lower limits of the
YES NO
Cooling
Satisfies the metropolis mutation probability, fbig is the larger fitness of the two
criterion
individuals to be crossed, fw is the fitness of the individual
to be mutated.
Optimization System level
results optimization solution Crossover and mutation operations of genetic
Fig. 1 System level GASA hybrid algorithm flow chart of CO strategy algorithm are easy to generate new individuals and
increase population diversity. However, for the large-scale
variable space to be searched by MDO method at system
III. ADAPTIVE MECHANISM IN GASA ALGORITHM level, GASA algorithm should also have the ability to
In this paper, combined with the optimization process quickly jump from the parent population evolution space
of multidisciplinary design, the genetic algorithm and to the offspring population evolution space, so as to avoid
simulated annealing algorithm are deeply analyzed. the phenomenon that the GASA algorithm lingers on
Combined with the research results of genetic algorithm some old states for a long time, which leads to the
and simulated annealing algorithm, the adaptive premature convergence of MDO method.
mechanism is used to improve GASA algorithm. In
GASA algorithm, after the population is initialized, the Therefore, in GASA algorithm, as an effective
structure of genetic algorithm is used to optimize. In this supplement to the population operation of genetic
process, we also need to consider the effects of crossover algorithm, simulated annealing algorithm does the follow-
and mutation on population evolution. The essential up evolution operation on the population, and endows the
problem is that the crossover probability PC and mutation population with controllable probability jump
probability PM can be adjusted automatically with the characteristics at each temperature. At high temperature,
optimization. Firstly, information quantity I is constructed the migration step size of GASA algorithm is large, which
according to formula (1) to extract the change information can effectively avoid CO method falling into local
of population evolution process. minimum, and alleviate the situation that the algorithm
relies too much on the parameter setting due to genetic
ln f 0  f avg operation; When the temperature is low, the Markov chain
I (1) corresponding to simulated annealing makes GASA
f max  f avg converge to the global optimum with a probability close to
1. Metropolis sampling process will realize detailed local
The increase of I means that the individual fitness is search on the variable space of CO method, enhance the
close to the maximum fitness of the population. At this search ability of GASA algorithm and improve the
time, the crossover probability and mutation probability of optimization efficiency of MDO method.
the individual should be reduced to make the genetic
algorithm jump out of the local minimum. In I, fmax and From the above analysis, it can be seen that the
favg are the maximum and average fitness of the population operation design of simulated annealing algorithm has an
respectively, and f0 is the fitness of the individual to be important impact on the quality of the final optimization
evolved. Combining with the optimization process of CO individual. The state generating function for generating
method at the system level, the cross information quantity new individuals in simulated annealing algorithm is
Ic and the mutation information quantity Im are constructed shown in equation (4).
by formula (2) and (3), respectively, so that the system xk 1 xk  K[ (4)

726

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on March 24,2021 at 07:44:26 UTC from IEEE Xplore. Restrictions apply.
Where η is the disturbance amplitude parameter, ξ is control the optimization. Therefore, the definition of
the random disturbance variable, and xk and xk+1 are the equation (7) can be adopted, and c is the estimated value
current solution and the new solution respectively. of f.
Random disturbance variables can obey different
1
distribution functions, such as Cauchy distribution, fc f c d 0
Gaussian distribution, uniform distribution, etc. When ξ 1 f  c (7)
obeys Cauchy distribution, the scale parameter α ­ 1
fc ® c t 0, f  c t 0
determines the distribution of probability density function. ¯1  f  c
In order to improve the optimization performance of
simulated annealing algorithm, the adaptive concept is On the basis of adaptive state generating function,
introduced, and the scale parameter α is constructed with GASA affects the search behavior of MDO by controlling
the maximum inconsistent information, as shown in the the population evolution operation of genetic algorithm
following formula (5). and simulated annealing algorithm. The multi-point
parallel probability search weakens the dependence of
1 a MDO method on initial value, realizes global optimization,
f [ u ,  f  [  f,
S a [2
2
(5) enriches the domain search structure in the optimization
process, enhances the ability of GASA to search variable
a max J i z j  xij i 1, 2, n space, and improves the ability of GASA to search
variable space and the convergence speed of MDO
Equation (5) indicates that the adaptive scale method.
parameter α is constructed by the largest term among the
inconsistent information of the system. The geometric IV. EXAMPLE VERIFICATION
meaning of α is shown in Figure 2. The system level
optimization coordinates the optimization points of three The oil-gas buffer is a common component in most
subsystems, and continuously approaches the best Z1, In modern aircraft landing gears. It deforms and does work
the search space of Z1, X2 is the farthest from Z1, and the during landing and absorbs the impact energy of the
inconsistent information between the two points is the airframe. The design requirements of the buffer are as
largest. Taking Z1 as the center of the circle and the follows: (1) it can absorb the landing work determined by
distance between two points as the radius, a circle is the strength specification; (2) it does not produce an
constructed, which includes all subsystem optimization impact load exceeding the allowable strength value; (3)
points. The radius of the circle represents the scale the compression stroke of the buffer is within the
parameter α. The Cauchy distribution based on adaptive allowable range during landing. The reasonable
scale parameter α produces large step length disturbance, optimization of the parameters of the oil-gas buffer plays
which is beneficial to improve the global optimization an important role in the buffer performance and impact
performance of CO method. load. The optimization of oil-gas buffer parameters is a
typical multidisciplinary design optimization problem,
Zy which involves many disciplines such as materials science,
X2 heat, fluid mechanics and so on. Reference [5] is available
X3
Z1
for the complete information of the example
In the process of buffer design optimization, the
X1 influence of multiple disciplines should be considered
comprehensively. During landing, the strength of the
shock absorber strut should withstand the impact of F, and
O Zx the oil should be compressed to absorb the impact energy
under the condition of adiabatic compression. The
Fig. 2 Structure chart of adaptive scale parameters optimization problem is divided into three subsystems,
The objective function optimized by MDO method including structure, heat and fluid mechanics. The
generally belongs to the category of seeking extremum. If optimization structure of buffer design problem
the objective function is to find the maximum or constructed by collaborative optimization strategy is
shown in Figure 3. Z is the system level design variable,
minimum problem, it is commonly defined by formula (6)
Xi is the design variable of subsystem i, and Ji is the
fc f fc f (6) constraint of subsystem O.

The above definition is simple and easy to implement, GA algorithm, SA algorithm and adaptive GASA
but because the population of GASA algorithm has to algorithm are used as system level optimization
undergo the evolutionary operation of genetic algorithm algorithms to solve the example. In the original design,
and simulated annealing algorithm, when the optimization the damping efficiency was 59%. After optimization of
results are scattered in the variable space, the average GA algorithm, SA algorithm and adaptive GASA
fitness of the population may not be conducive to reflect algorithm, the efficiency is improved to 70.5%, 69.1%
the distribution of the population, and it is difficult to and 77.4% respectively. Figure figure 4 to 7 are the

727

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on March 24,2021 at 07:44:26 UTC from IEEE Xplore. Restrictions apply.
impact load diagram and work diagram of landing gear force of the buffer strut changes with time and
buffer strut optimized by four algorithms. In the impact displacement respectively. The following conclusions can
load diagram and the work quantity diagram, the impact be drawn by comparing and analyzing each diagram.
System level optimization
³ F  Foil  Fm ˜ v ya dt
t

0 air
min 1
F max u S max
X ª
¬ Aair , P, V , Ah , Aoil , d oil , d h , v ya º
¼
s.t J1 d H , J 2 d H , J 3 d H
J1ˈX1 J3ˈX3
Z1 Z3
J2ˈX2 Z2
Structural analysis and optimization Fluid mechanics analysis Thermal analysis optimization
Aair  Aair 0  P  P0 
2 2
min J1 optimization P  P0  V  V0
2 2
min J 3
Aair  Aair 0  Ah  Ah 0
2 2
V  V0  Ah  Ah 0  ( Aoil  Aoil 0 )2
2 2
min J 2
 Aair  Aair 0
2

s.t F Aair  Aoil d >V @  P  P0  V  V0


2 2
s.t P ˜ V J const
s.t 7  d oil d h d 0

Fig. 3 Optimization structure of oil-air buffer for shock absorption efficiency

As shown in Figure 4, after the landing gear buffer cause body turbulence and bring adverse impact on the
strut is optimized by the adaptive intelligent CO method riding comfort of the crew; in the work diagram, the work
based on GA algorithm, the fluctuation degree of impact process of impact force with displacement is also slightly
force curve is large in the impact load diagram, which steep, and the landing process is not as smooth as the
indicates that in the later stage of compression, the optimized design scheme based on other algorithms.
amplitude of reverse impact is large, which is easy to

Fig. 4 Impact load and work diagram of buffer strut optimized by GA based adaptive intelligent CO method

Fig. 5 Impact load and work diagram of buffer strut optimized by SA based adaptive intelligent CO method

728

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on March 24,2021 at 07:44:26 UTC from IEEE Xplore. Restrictions apply.
Fig. 6 Impact load and work diagram of buffer strut optimized by GASA based adaptive intelligent CO method

As shown in Figure 6, the buffer performance has simulated annealing algorithm, an improved hybrid
been significantly improved after using GASA hybrid genetic simulated annealing algorithm (GASA) is
algorithm. In the impact load diagram, the rebound proposed. The optimization results of an example verify
amplitude of impact force is the smallest at the end of the the effectiveness and efficiency of the multidisciplinary
period; in the power diagram, the curve of impact force optimization method based on intelligent optimization
work is also more round and full, which shows that the algorithm. It is worth noting that although intelligent
optimized landing gear buffer strut can ensure the whole optimization algorithm has high optimization performance
landing process is relatively stable during compression, at the system level or discipline level, it can play a greater
and in the later stage of landing, the buffer strut can role when combined with efficient optimization strategies.
absorb the impact energy and avoid violent rebound. Therefore, the research of multidisciplinary optimization
Based on the GASA hybrid algorithm, in the variable strategy for adaptive intelligent optimization algorithm is
space, the CO method tends to be the global optimal the next research focus of our research team.
through the global parallel search of GA, so it can
efficiently converge to the optimal point.
By comparing the time axes of the three power charts, ACKNOWLEDGMENT
it can be seen that the work time of the buffer strut The authors gratefully acknowledge the fund support
optimized by GASA hybrid algorithm is the shortest; the from the National Key Research and Development
work time of the buffer strut optimized by GA or SA Program of China (2018YFB1700800) and funds of
based adaptive intelligent co method is in the middle. The (41401030301).
optimized buffer can control the slow growth of load in
the initial stage of landing, avoid the sharp increase of
REFERENCES
load in a very short period of time, and reduce the impact
on the body and passengers; with the compression of the [1] AIAA Multidisciplinary Design Optimization Technical
Committee. Current State of the Art on Multidisciplinary Design
buffer strut, in the later stage of landing, the descending Optimization[A]. AIAA Washington DC, 1991: 1-63
speed decreases, and the load increases slowly, forming
[2] Jun S., Jeon Y.H., Rho J., et al. Application of Collaborative
the maximum load at the back end of the journey, so that Optimization Using Genetic Algorithm and Response Surface
the impact energy can be effectively absorbed and the Method to an Aircraft Wing Design[J]. Journal of Mechanical
buffer can avoid working It appears to be too rigid and Science and Technology, 2006, 20(1): 133-146
improves the damping efficiency. The optimized buffer [3] N. Qiang, Y. Zhao. Improvement Multidisciplinary Collaborate
can absorb additional work and has strong adaptability. Optimization based on Simulated Annealing and Artificial Neural
Networks[J]. Open Cybernetics & Systemics Journal, 2015,
9(1):2306-2311.
V. CONCLUSION [4] Jodei J., Ebrahimi M., Roshanian J.. Multidisciplinary design
In this paper, the computational and convergence optimization of a small solid propellant launch vehicle using
difficulties in multidisciplinary design optimization system sensitivity analysis[J]. Structural and Multidisciplinary
Optimization, 2009, 38(1): 93-100
methods are discussed. An adaptive GASA algorithm is
[5] Wong J , Ryan L , Kim I Y . Design optimization of aircraft
proposed to replace the traditional numerical optimization landing gear assembly under dynamic loading[J]. Structural &
algorithm to solve the problem of computational difficulty. Multidiplinary Optimization, 2017.
Combining the advantages of genetic algorithm and

729

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on March 24,2021 at 07:44:26 UTC from IEEE Xplore. Restrictions apply.

You might also like