V ol. 16　 N o .

1 　　　　　　　　　　　　 CHIN ESE JO U RN A L OF A ERO N A U TICS　　　　　　　　　　　　 February 2003

Hierarchical Evol utionary Al gorithms and Its

Appl ication in Transonic Airfoil Optimization in Aerodynamics
WANG Jiang -feng, WU Yi-zhao
( College of A erosp ace E ngineering , N anj ing University of A eronautics and A stronautics,
N anj ing 　210016, China)

Abstract: 　 Hierar chical evo lutionary algor ithms based o n g enetic algo rithms ( GA s) and N ash str at-
eg y o f game theory are proposed to accelerate the optimization process and im plemented in transonic
aerodynamic shape optimization problems . Inspir ed fr om the natural evolution hist ory that different
periods w ith cer tain env iro nments hav e differ ent criteria for the evaluations of individuals' fitness, a
hierarchical fidelity model is intro duced to reach high opt imization efficiency . T he shape of an
N ACA 0012 based airfoil is o ptimized in m ax im izing the lift coefficient under a g iven tr ansonic flow
condition . Optimized results are pr esented and co mpar ed with the sing le model r esults and tr aditional
GA .
Key words: 　tr ansonic flow ; aero dy namic optimization; finite element method; unstr uctured gr id;
hierarchical fidelity models; evolutionar y alg orithms
分级型进化算法及 其在 翼型 跨声 速气 动优 化中 的应 用 . 王 江峰, 伍 贻兆 . 中 国航 空学 报( 英文
版) , 2003, 16( 1) : 1- 6.
摘　要: 提出了变精度模型的分级型进化算法并对初始翼型为 NA CA 0012 的二维翼型进行了跨声
速流动下的形状增升优化。借 鉴自然进化中不同进 化阶段个体适应度 评估环境不同的机 理, 构造
了分级精度模型以加速优化 过程。对翼型进行了给 定气动条件下的形 状增升优化, 给出了优化结
果, 并与传统基因算法及单精度模型算法结果进行了对比。
关键词: 跨声速流; 气动优化; 有限元方法; 非结构网格; 分级精度模型; 进化 算法
文章编号: 1000-9361( 2003) 01-0001-06　　　中图分类号: V 211. 3　　　文献标识码: A

　 　 Genet ic algorit hms ( GAs ) have est ablished
them selves as a pow erf ul t ool for design opt imiza-
[ 1, 2]
tion in various fields . How ev er, one of t he
main draw backs of GAs is t hat they m ust require a
rat her larg e number of evaluations. T he problem is
particularly crucial in aerody namics, since CF D
codes are very time-consum ing. For inst ance,
codes t hat solve non-linear Navier -Stokes equations
take a considerable time t o perf orm a single sim ula-
tion of the flow past an airfoil. A t ypical opt imiza-
tion using GAs may involve several hundreds or
even a f ew thousand com put at ions, thus lim it ing
their applicat ion.
T he main focus of the studies has been to i-
dent if y st rat eg ies t o overcome t his draw back . One
such st rat eg y is hierarchical genet ic algorithms
w it h Nash equilibrium and game theory [ 3, 4] , w hich
off ers the possibilit y of using alternatively low cost
models for ex ploration and high f idelit y models for
ref inement . T his technique is used not as a st and
alone, but as a part of a larg er hierarchical st ruc-
ture.
Numerical result s are presented f or an
NACA0012 airf oil based shapes opt im ization prob-
lem, using respect ively a GA, a Nash GA
( N -GA ) [ 　
5]
and a hierarchical Nash GA ( HN -
GA ) . T he com parison in converg ence hist ory and
speedup rat io show s that t he hierarchical evolution-
ary algorithms have higher convergence ability and
could be used in diff erent research fields.

Received dat e: 2002-08-15; R evision received dat e: 2002-12-27

Foundation item: S tart -up f oundat ion it em of t he Educat ional D epart ment of China f or ret urnees
A rt icle U RL: ht t p: / / ww w . hkx b. n et . cn/ cja/ 2003/ 01/ 0001/
© 1994-2010 China Academic Journal Electronic Publishing House. Open access under CC BY-NC-ND license. http://www.cnki.net
 ・ 2・ W A N G Jian g -f eng , WU Y i-zhao CJA

als are always t hose w ho are most suitable to t he

1　Com bination w ith GAs and
Nash Equilibrium
Due t o t he large dimension of global opt imiza-
tion problems and access t o low cost dist ributed
parallel environment s such as PC cluster it is a
quite nat ural idea replacing a global opt im ization by
decentralized local sub-optimizat ions using Nash
[ 4]
game int roduced by J . F. Nash in 1951 . Since it
appeared f irst in econom ics , t he not ion of play er is
oft en used in t he sequel. F or a mult i crit eria opt i- F ig. 1　 T r aditional evo lut ionary algo rithms vs
hierar chical ev olutionary alg orit hm s
mizat ion w it h N object ives, a Nash strategy con-
sist s of N players, each in charg e of one object ive, present environment . Based on t his case, a t ech-

and each has it s ow n st rat egy set and its ow n crit e- nique named the hierarchical f idelit y model, w hich

rion . During t he game , each play er looks for t he has diff erent mechanism f rom variable f idelit y
best strategy in its search space in order t o improve models described in Ref . [ 7] , is int roduced int o al-
it s own crit erion w hile crit eria of ot her players are gorit hms combined wit h GA and Nash g ames.

fixed. When no player can f urther im prove it s cri- T his t echnique consist s in dividing a global opt i-

terion , t he system has reached a st at e of equilibri- mizat ion process into sub processes, w it h each one
associated a num erical model diff erent in CP U cost
um named Nash equilibrium .
Using GAs t o build the Nash equilibrium ( comput e accuracy ) to evaluate individuals' fit-

makes t he not ion of Nash GAs ( N -GAs) . During ness. T he mig rat ion of best individuals occurs a-

a Nash gam e, each play er uses a GA to improve mong diff erent models, w hich t ransport s the best
inform at ion and realizes the transf ormation f rom
his ow n criterion along g enerat ions constrained by
ex plorat ion in searching space t o refinement of so-
strat egies of t he ot her player . In applications, de-
sign variables are geometrically split bet ween play- lutions, as show n by t he low er -half in F ig . 1.
T he main objects of t he low -fidelity model
ers w ho ex change sy mmet rically t heir best st rat e-
( L F m odel ) are: dif fuse t he individuals in
gies ( best chrom osom es ) during t he evolut ion ;
searching space; maintain t he variat ion of t he
more det ails about Nash GAs could be f ound in
Ref. [ 4～6] . population ; ex plore in searching space; ! avoid
t he local converg ence. And the m ain object s of t he
2　Hierarchical Fidelity M odels and high-f idelit y model ( HF model ) are: reserve
Hierarchical Algorit hms best individuals ( best pot ent ial solut ions ) ; re-
fine global opt imized solut ions; enforce t he co-
2. 1　Hierarchical fidel ity model s
operation bet w een t he exploration and the ref ine-
For traditional evolutionary algorit hm s, a sin-
ment .
gle model is of ten used for fitness com put at ion for
2. 2　Hierarchical algorithms
diff erent individuals during t he w hole evolutionary
Wit h the technique of a hierarchical f idelit y
hist ory , as show n by the upper-half in F ig . 1.
model, the w hole optimizat ion process is devised
Observing the biological nat ural evolution his-
int o diff erent sub processes w hich are associat ed by
tory , it is alw ays t he case that , during t he proce-
a certain model to evaluat e the fitness of a popula-
dure of t he w hole evolution, diff erent periods have
tion. Diff erent algorit hm s associat ed construct a
diff erent environment s, which makes individuals [ 6]
family of hierarchical alg orit hms , for ex am ple,
have diff erent fitness values, but the best individu-
© 1994-2010 China Academic Journal Electronic Publishing House. Open access under CC BY-NC-ND license. http://www.cnki.net
 Hierarchical E volut ionary A lgorith ms and it s
February 2003
A pplicat ion in Tran sonic A irf oil Opt imizat ion……
hierarchical g enetic algorit hms ( HGAs) and hier-
archical Nash g enetic alg orit hms ( HN -GAs) w hich
coupled GA and Nash game.
Fig. 2 shows the t opology of an HN-GA used
in t he present st udy w ith a t w o -f idelit y model
( w it h hierarchy on CF D solver ) , w here the w hole
opt imization process is devised into t w o sub pro-
cesses. S is a chain composed by variables X and
Y . Each m odel ( sub -process ) could have an algo-
rit hm wit h dif ferent control parameters, such as
F ig. 2　HN -G A w ith a 2-F i model and tw o-play er
population size, et c. F or model L F, a relat ive big
population is used for ex plorat ion in search space ,
and a relative small population for model HF since
it is expensive in CPU time. T his diff erence, w ith
diff erent definitions of models, const ruct s t he hier-
[ 6]
archical topology for hierarchical algorithms .
T he convergence criterion is defined on t he
HF model. T he arrow s in Fig . 2 m ean t he mig ra-
tion direct ion of individuals betw een tw o f idelit y
models. Define t he ev olut ion st art ing from L F
model and ending at HF m odel as one cycle; t hus,
the convergence of algorit hm s in one cycle corre-
sponds t o a mono-direction evolut ion; otherw ise
backward transports of individuals f rom HF model
to L F model make alg orit hms looped evolut ion ,
w hich is quit e diff erent from Ref . [ 7] . Since dif-
ferent models could hav e diff erent populat ions and
they are independent except the exchange of indi-
viduals, they could inherit naturally the parallelism
of GAs. T he hierarchical evolutionary algorithms
proposed here are hig h eff icient memory distribut-
ed parallel opt imization met hods.
Diff erent models could be def ined in hierarchi-
cal algorit hms, such as meshes w it h dif ferent num-
bers of nodes ( similarly t o multigrids met hods ) ,
© 1994-2010 China Academic Journal Electronic Publishing House. Open access under CC BY-NC-ND license. http://www.cnki.net
・ 3・
diff erent CFD solvers, etc. In t he present study, a
t wo -level CFD solver on the same unst ruct ured
grid is used as diff erent f idelit y m odels:
L F m odel: low accuracy CFD Euler equations
solver w hich uses less CPU t ime , serves t o ex plore
in search space and transport the best individuals
( quasi-solut ions) .
HF model: high accuracy CF D Euler equa-
tions solv er which uses more CPU t ime , serves to
ref ine t he solut ions—t he f inal t arget of optimiza-
tion.
3　CFD Flow Solver
In the present st udy , the f it ness comput ation
( lif t coef ficient) for individuals is realized using 2D
Euler equat ions finit e volume met hod with
+
AUSM f lux splitt ing scheme on an unst ruct ured
grid. Det ails about t his CF D solver could be found
in Ref . [ 8] .
During t he evolut ion , t he unstruct ured grid
must be adjust ed f or every new individual. T o save
t he CPU t ime spent on m esh re-g enerat ion, a
t echnique named dynamic grid method, used of ten
in unst eady aerody namics, is im plement ed . With
t his technique, it is not necessary anym ore t o re-
generat e a g rid f or individuals, but adapt t he orig i-
nal grid with new surface configurat ion . T his pro-
cess saves CPU time for each individual and con-
tributes benef it t o the global convergence of algo-
rit hm s.
4　P roblem Definition and Parameters
for A lgorit hms
4. 1　Parameterization of airfoil shape and
fitness definition
T he t arg et of the opt im ization is to maximize
t he lif t coef ficient of an airfoil w it h shape variables
under t ransonic flow conditions. T he shape of t he
airfoil is represent ed using 11t h order Bezier
[ 9]
spline , and t he tot al number of variables t o be
opt imized is reduced to t went y ( t en for w indw ard
side and t en for leew ard side of airfoil) . See Ref .
[ 6] f or more details.
T he transonic f low conditions are
 ・ 4・ W A N G Jian g -f eng , WU Y i-zhao CJA

M a = 0. 8, = 1. 25° ( 1) w ard shape w ith t he w indw ard shape fix ed by

w here M a is the M ach number and t he angle of Play er 1.
at t ack ( deg) . F or symmetric airf oil NACA0012, For HN-GA w it h L F -HF model, each model
the m ax imum thickness d = 0. 06. T he search do- uses a Nash GA w it h the same f it ness definition as
main for alg orit hms in the present st udy is: for Eq. ( 3) and Eq. ( 4) , t he diff erence being the ac-
w indw ard side, d ∈[ 0. 05　 0. 08] ; for leew ard curacy of CF D solver , as defined in Sect ion 2. 2.
side, d∈[ - 0. 08　- 0. 05] . 4. 2　Control parameters for al gorithms
For GA, the fitness is def ined as T he cont rol paramet ers for alg orithms ( GA,
M ax im ize f = C L ( V U , V L ) ( 2) Nash GA - N -GA , Hierarchical Nash GA - HN -
w here V U is t he variable set f or windw ard side, GA ) are grouped in T able 1, w here P m and P c de-
and similarly V L for leeward side , C L t he lift coeff i- not e respectively t he rates of crossover and of mu-
cient . tation ; t he default param et ers are: tournament se-
For Nash GA, these tw o set s are dist ributed lect ion , elitist scheme , binary code. For HN -GA ,
to tw o players, w ith f it ness defined as follow s. diff erent P m and P c are used for L F and HF models
*
Play er 1: Max imize f 1 = C L ( V U , V L ) ( 3) t o realize dif ferent opt imization funct ions: L F
*
Play er 2: Max imize f 2 = C L ( V , V L )
U ( 4) model for ex ploration and HF model for ref ine-
w here* denot es that the variables are fixed ment ; the exchange frequency bet ween t w o models
to the best ones ( the best st rat egy ) obtained by is five generat ions, w hich means t he transm ission
diff erent players. Player 1 opt im izes the w indw ard of best individuals occurs at t he end of each five
side shape w ith t he leew ard side shape fixed by generat ions .
Play er2, and similarly , Play er2 opt imizes t he lee-
Table 1　 Control parameters f or al gorithms
Algorit hm GA N ash G A Hierarchical N ash G A
LF -HF model
M odel Single Single
LF model HF model
N b. of players 2 2 2
Population size 40 20×2 20×2 10×2
N b. of variables( t w o Bezier splines) 20 10×2 10×2 10×2
Crossover an d P c U niform 0. 85 U nif orm 0. 85 U nif orm 0. 9 U niform 0. 85
M ut at ion P m 0. 01 0. 01 0. 05 0. 01
N b. of generat ions 100 40 20 10

since t hree algorit hms obt ained the same opt im ized
airfoil, just t he result of HN-GA is present ed.
5　Results and Analysis
Fig. 5 present s the comparison of surficic pressure
T he init ial airf oil is the Bezier spline represen- distribution for t hree algorit hms. T he convergence
tation of NACA0012 wit h 150 surf icic points. T he hist ory of diff erent algorit hms is show n in F ig . 6.
CFD Euler solver is on an unst ruct ured grid w ith T able 2 grouped t he optimized result s for t hree al-
about 5000 nodes and 10000 t riangles. T he low gorit hms and t he CPU t ime.
accuracy Euler solver is 1st order AU SM + schem e Tabl e 2　 Comparison of resul ts
- 3
w it h stop crit erion 1×10 ; the high accuracy Eu- I nitial Optimized
+
ler solver is 2nd order AUSM scheme w ith st op GA N -GA HN -G A
- 5 CL 0. 330 0. 861 0. 864 0. 862
criterion 1 × 10 . Fig. 3 show s t he isopressure
CD 0. 029 0. 064 0. 065 0. 065
distribution around the init ial airf oil ( NACA 0012)
CL / CD 11. 46 13. 41 13. 31 13. 30
and the opt imized airf oil. CP U 170 h 60 h 36 h
T he com parison of shapes betw een t he initial Speedup 1. 0 2. 8 4. 7
airfoil and t he opt im ized airfoil is show n in Fig. 4;
© 1994-2010 China Academic Journal Electronic Publishing House. Open access under CC BY-NC-ND license. http://www.cnki.net
 Hierarchical E volut ionary A lgorith ms and it s
February 2003 ・ 5・
A pplicat ion in Tran sonic A irf oil Opt imizat ion……

Fig . 6　Co mpar ison of converg ences

　 　 With t hese comparisons, one can conclude
( a ) Initial airfoil
t hat :
1) T he f inal opt im ized airf oil is asy mmet ric,
t he thickness of w indw ard side increases and that
value of leeward side decreases. T he final shapes
t ake on som e propert ies of high-lift airfoil.
2) T he w indward side shock is stronger and
moves t o the t railing edge ; t he leew ard side weak
shock disappeared.
3) T he C L increases 160% ( f rom 0. 33 to
0. 86) .
4) T he drag coeff icient C D and C L / C D are not
( b) O ptim ized airfoil
F ig . 3　Isopressure comparison
considered as optimizat ion const raints in t he pre-
sent st udy, but one can observe that t he CD in-
creases 120% ( from 0. 0288 to 0. 0642) and t he
rat io of CL / C D increases 17% ( from 11. 46 to
13. 31) .
5 ) Suppose that t he speedup ratio of GA is
1. 0. T hen the N -GA is 2. 8 and HN -GA is 4. 7
for t he present case. Note t hat this speedup value
could be still improved in a parallel environment ,
such as PCs-cluster.
F ig . 4　Com parison of air foil shapes T he alg orit hm s w it h the single model ( GA
and N -GA ) and t he hierarchical algorit hm ( HN -
GA ) obtained t he same opt imal result s, but t he
lat er is more eff icient in converg ence hist ory and
CPU time. In hierarchical alg orit hms, t he ex-
change of t he best individuals among dif ferent
models is construct ed: from t he view point of biolo-
gy evolution, this migrat ion of best individuals
keeps bett er variat ion of populat ions; f rom t he
view point of evolut ionary alg orithms, t his mig ra-
Fig. 5　Compar iso n o f pressure distribution
tion conserves Building Blocks and utm ost ly makes
t hem be represent ed in nex t g enerat ions. Dif ferent
© 1994-2010 China Academic Journal Electronic Publishing House. Open access under CC BY-NC-ND license. http://www.cnki.net
 ・ 6・ W A N G Jian g -f eng , WU Y i-zhao
models in hierarchical algorit hm s have dif ferent
particular properties, such as ex plorat ion in search
space for LF model and ref inement of solutions for
HF m odels; hereby small populat ion size could be
used and t his will not cause premat urity .
6　Conclusions and Future Work
An NACA0012 based airfoil is opt imized in
shapes for maximizat ion of t he lif t coeff icient under
the t ransonic f low condit ion. T he new evolution-
ary algorit hms—hierarchical Nash GA, are devel-
oped and implemented; numerical results are ob-
tained and compared w ith t radit ional GA and Nash
GA . T he comparisons show t hat HN-GAs have
high eff iciency and robust ness in m ultidisciplinary
opt imization problems . With the hierarchical algo-
rit hm s developed in this paper , some ext ended pa-
rameters, such as Mach num ber, ang le of at tack,
C D, C L / C D, etc. , could be easily added in and t he
game could be constructed as ‘ real game’for con-
[ 6]
flict optimizat ion t arget s .
Besides t he im plicit parallelism in g ame st rat e-
gy and genet ic algorit hm s, the opt im izat ion process
of diff erent models in hierarchical algorit hms has
im plicit synchronizat ion ; hereby mem ory distribut-
ed parallel hierarchical alg orit hms on P Cs-Clust er
syst em w ill im prove great ly t he opt imizat ion eff i-
ciency , and af ford robust application t ools f or eng i-
neering design .
Acknowl edgements
T his research is support ed by the start -up
foundat ion it em of the Educat ional Depart ment of
China f or ret urnees.
© 1994-2010 China Academic Journal Electronic Publishing House. Open access under CC BY-NC-ND license. http://www.cnki.net
CJA
References
[ 1] 　 Holland J H . A daptation in nat ural and art if icial system [ M ]
. A nn A rbor : U niversity of M ichigan Press, 1975.
[ 2] 　 Goldberg D E . G enet ic algorit hms in search , opt imization,
and machine learning [ M ] . Reading : Addison-W esley ,
1989.
[ 3] 　 von N eumann J . Zur T heorie der Gasellschaft sspiele [ J ] .
M at h A nn, 1928, 100: 295- 320.
[ 4] 　 N ash J F. N oncooperat ive games[ J] . A nnals of M athemat -
ics, 1951: 154- 289.
[ 5] 　 Wang J F, W u Y Z, Periaux J . T ransonic Euler equation
airf oil aerodynamic opt imization based on genet ic algorit hms
and game th eory[ J ] . Journal of N anjing U niversity of A ero-
naut ics and A stronaut ics, 2002, 33( 6) : 531- 536
[ 6] 　 Wang J F. Opt imisat ion distribu e mult icrit re par algo-
rithm es g n tiques et t h orie des jeux & applicat ion la sim-
ulat ion n um rique de probl mes d' hypersust ent at ion en
a rodynamique [ D ] . Paris: U niversit Pierre et M arie
Curie, 2001.
[ 7] 　 A lexan dr ov N M , Lew is R M , Cumbert C R , et al. Opt i-
mizat ion w it h variable-f idelit y models applied to w ing design
[ R] . A IA A 2000-0841, 2000.
[ 8] 　 Wang J F, W u Y Z. A new flux split t in g scheme on un-
structured grid: AU SM + [ J] . Journal of U niversit y of Sci-
ence and T echnology of China, 1999, 29( 1) : 122- 126.
[ 9] 　 B zier P E. Essai de d finit ion num rique des courbes et des
surfaces exp riment ales [ D ] . Paris: U niversit Pierre et
M arie Curie, 1977.
Biography:
WANG Jiang -feng 　 Born in 1970, he receiv ed B . S . ,
M . S . and doctoral deg ree fr om Nanjing U niversity of
A eronautics and A str onautics ( N U AA ) in 1993, 1996 and
2000 respectively , and then became a teacher t her e. A fter
three years Ph . D. studies and cooperation at the U niversi-
ty P aris 6 and Dassault Av iatio n in Fr ance, he got the sec-
ond doctoral degr ee. He has published several scientific pa-
pers on CFD , evo lutionary alg orithm s and par allel co mpu-
tation in v arious per iodicals. T el : ( 025 ) 4892735,
4891595, E -mail : w ang jf @ nuaa. edu . cn