Optimal Design of Viscous Damper Connectors for Adjacent

Structures using Genetic Algorithm and Nelder-Mead Algorithm

Kasra Bigdelia, Warren Hareb, Solomon Tesfamariama
a
School of Engineering, University of British Columbia, Kelowna, BC, Canada V1V 1V7
b
Mathematics, University of British Columbia, Kelowna, BC, Canada V1V 1V7

ABSTRACT

Passive dampers can be used to connect two adjacent structures in order to mitigate earthquakes induced pounding
damages. Theoretical and experimental studies have confirmed efficiency and applicability of various connecting
devices, such as viscous damper, MR damper, etc. However, few papers employed optimization methods to find the
optimal mechanical properties of the dampers, and in most papers, dampers are assumed to be uniform. In this study, we
optimized the optimal damping coefficients of viscous dampers considering a general case of non-uniform damping
coefficients. Since the derivatives of objective function to damping coefficients are not known, to optimize damping
coefficients, a heuristic search method, i.e. the genetic algorithm, is employed. Each structure is modeled as a multi
degree of freedom dynamic system consisting of lumped-masses, linear springs and dampers. In order to examine
dynamic behavior of the structures, simulations in frequency domain are carried out. A pseudo-excitation based on
Kanai-Tajimi spectrum is used as ground acceleration. The optimization results show that relaxing the uniform dampers
coefficient assumption generates significant improvement in coupling effectiveness. To investigate efficiency of genetic
algorithm, solution quality and solution time of genetic algorithm are compared with those of Nelder-Mead algorithm.
Keywords: Coupled structures, Earthquake damage control, viscous dampers, Derivative-free optimization, Genetic
algorithm.

1. INTRODUCTION
In most metropolitan areas, limited land availability and high demand for residency and office buildings, leads high-rise
buildings constructed in close proximity. During an earthquake the effect of pounding of two such adjacent buildings is
highly destructive. The pounding effect and frequency becomes even more pronounced in dense urban centers. For
example, severe damage has been observed in Mexico City earthquake (1985), Loma Prieta earthquake (1989), the Kobe
earthquake (1994), the Bhuj earthquake (2001), and the recent New Zealand earthquake (2011). Many studies have
introduced and investigated different dampers (visco-elastic, viscous fluid, MR dampers, etc.) in order to prevent
pounding induced damage, while increasing seismic resistance of the structure [1]. Experimental studies also confirmed
efficiency of damper connectors to mitigate vibrations of structures [2]. Disregarding the type of damping device, it is
common to assume that all adjacent floors of adjacent buildings are connected using identical dampers [3]. However, it
has been shown that the uniform distribution of dampers is not an optimal one [4].
Some preliminary studies investigated non-uniform damping coefficients of connectors for adjacent buildings [5].
However, they did not present an optimization method to find optimal damping coefficients. In [5], to find non-uniform
damping coefficients, the authors assume a simple linear function along the building. This assumption makes an intuitive
sense, as the force generated in viscous dampers is a linear function of relative velocity. However, (to the best of our
knowledge) few papers have examined the use of mathematical optimization on the optimal damping coefficients of
viscous dampers. The main objective of this paper is to present and compare methods to find the optimal distribution of
viscous dampers between two adjacent buildings. It should be noted that the objective value of the present problem,
representing performance of the system, is not an analytic function of the design variables. Instead, a simulation is
performed for each set of damping coefficients in order to determine the objective value. As such, conventional
optimization methods that are based on the derivatives of the objective function cannot be employed in this problem. To
solve an optimization problem of several design variables, where the derivatives are unknown, a derivative-free
optimization (DFO) method can be used.

Active and Passive Smart Structures and Integrated Systems 2012,

edited by Henry A. Sodano, Proc. of SPIE Vol. 8341, 83410M · © 2012 SPIE
CCC code: 0277-786X/12/$18 · doi: 10.1117/12.915471

Proc. of SPIE Vol. 8341 83410M-1

This paper uses a well-known evolutionary optimization technique, the genetic algorithm (GA), to solve the optimization
problem of non-uniform damping coefficients of dampers. GA is commonly regarded as to be an effective method to
solve problems where derivatives are not available. Results show that truly optimal damping coefficients is non-uniform
and assumption of uniformity of dampers imposes an additional constraint to the problem that prevents the finding of the
optimal damper coefficients [4]. The results obtained from GA are then compared with the results obtained from another
well-known optimization technique, the Nelder-Mead algorithm [4]. The comparison shows that not only Nelder-Mead
method results in much more accurate results; but also the solution time for Nelder-Mead is generally less than the
solution time for the GA.

Building1
mn+m,1

cn+m,1 kn+m,1

mn+m-1,1

cn+1,1
kn+1,1 Building 2
mn1 mn2

cn1 kn1 cn2 kn2

mn-1,1 cd,n-1 mn-1,2

cn-1,1 kn-1,1 cn-1,2 kn-1,2

mn-2,1 mn-2,2

cn-2,1 kn-2,1 cn-2,2

kn-2,2

mn-3,1 cd,n-3 mn-3,2

c31 k31 c32

k32

m21 cd2 m22

c21 k21 c22

k22

m11 m12

c11 k11 c12 k12

Base Excitation

Proc. of SPIE Vol. 8341 83410M-2

 2. MODELLING
2.1 Assumptions and limitations
The ground motion and dynamic response of the buildings are assumed to be unidirectional. In order to prevent the effect
of torsional vibrations of the buildings, both buildings are assumed to be symmetric and their centers of mass are located
in the same plane. Fig 1 shows that each building is modeled as a multi-degree of freedom system consisting of lumped
masses, representing mass for each floor, linear spring, representing stiffness of columns, and, linear viscous damper.
In order to connect adjacent buildings, both buildings are assumed to have the same floor elevations. However, the
height of each building does not need to match. Damper connectors are modeled as viscous dampers with a damping
coefficient variable (the damping coefficient is uniform across all dampers). Many mechanical models are available to
represent dynamic behavior of the fluid dampers. In this study, fluid dampers are modeled as a set of parallel linear
springs and dampers. Based on results presented in [1,6], it is further assumed that the stiffness of fluid damper can be
neglected without any significant effects on the model and performance.

2.2 Modeling and Analysis

As shown in Fig 1, buildings 1 and 2 have n + m and n storeys, respectively. The mass, shear stiffness, and damping
coefficients for the ith story are mi1, ki1, and ci1 for building 1 and mi2, ki2, and ci2 for building 2. The damping coefficient
and stiffness coefficient of the damper at the ith floor are cdi and kdi, respectively. The dynamic model for both structures
is taken to be a 2n + m degree of freedom system.
Let xi1 (t ) and xi 2 (t ) be the displacement of the ith floor of buildings 1 and 2 in time domain, respectively. Consequently,
th
x&i1 (t ) and x&i 2 (t ) represent the velocity, and &x&i1 (t ) and &x&i 2 (t ) represent the acceleration of the i floor of buildings 1 and 2,
respectively. To find a general and reliable objective function, the model is analyzed in pseudo-excitation frequency
domain rather than the time domain, which is usually associated with a specific real earthquake record. Therefore,
displacement, velocity and acceleration values are transformed from functions of time into functions of frequency ω , as
xi1(ω) , xi 2 (ω) , x&i1(ω) , x&i 2 (ω) , &x&i1(ω) , and &x&i 2 (ω) . As before, the first index corresponds to the floor number and the second
index corresponds to the building number. Since calculated parameters in frequency domain are complex values, the
squared magnitude, also known as the auto-spectral density, is used for optimization. For example, auto-spectral density
of displacement for the ith floor of the building b is written as:

Sxib (ω ) = x ib (ω ) • conj (x ib (ω )). Eq. 1

Subsequently, applying integration over the frequency domain will result in statistical parameters, which represent
standard deviations of displacement, velocity and acceleration of each floor. For instance, for the ith floor of building b,
its standard deviation of displacement response is
1/ 2
σ xib = ⎡∫ S xib (ω )dω ⎤ .
+∞
Eq. 2
⎢⎣ −∞ ⎥⎦

2.3 Simulation

In order to evaluate σ xib it is necessary to solve the governing equation of the system:

MX&& (t ) + (C + Cd ) X& (t ) + ( K + K d ) X (t ) = M E g (t ) Eq. 3

where
M = diag [ m1,1 ,..., mn+m ,1 , m1, 2 ,..., mn , 2 ] , Eq. 4
⎡K 0⎤
K =⎢ 1 ⎥, Eq. 5
⎣0 K2 ⎦

Proc. of SPIE Vol. 8341 83410M-3

 ⎡C 0⎤
C=⎢ 1 ⎥, Eq. 6
⎣ 0 C2 ⎦
⎡k11 + k 21 − k 21 ⎤
⎢ −k k 21 + k31 − k31 ⎥
⎢ 21 ⎥
⎢ O ⎥
K1 = ⎢ ⎥, Eq. 7
⎢ ⎥
⎢ − k n+ m−1,1 k n+ m−1,1 + k n+ m ,1 − k n+ m ,1 ⎥
⎢ ⎥
⎢⎣ − k n+ m ,1 k n+ m ,1 ⎥⎦
⎡ k12 + k 22 − k 22 ⎤
⎢ −k k 22 + k32 − k32 ⎥
⎢ 22 ⎥
⎢ O ⎥
K2 = ⎢ ⎥, Eq. 8
⎢ ⎥
⎢ − k n−1, 2 k n−1, 2 + k n , 2 − k n , 2 ⎥
⎢ ⎥
⎣⎢ − kn,2 k n , 2 ⎦⎥
⎡c11 + c21 − c21 ⎤
⎢ −c c21 + c31 − c31 ⎥
⎢ 21 ⎥
⎢ O ⎥
C1 = ⎢ ⎥, Eq. 9
⎢ ⎥
⎢ − cn+ m−1,1 cn+ m−1,1 + cn+ m ,1 − cn+ m ,1 ⎥
⎢ ⎥
⎣⎢ − cn+ m ,1 cn+ m ,1 ⎦⎥
⎡c12 + c22 − c22 ⎤
⎢ −c c22 + c32 − c32 ⎥
⎢ 22 ⎥
⎢ O ⎥
C2 = ⎢ ⎥, Eq. 10
⎢ ⎥
⎢ − cn−1, 2 cn−1, 2 + cn , 2 − cn , 2 ⎥
⎢ ⎥
⎢⎣ − cn , 2 cn , 2 ⎥⎦
⎡ Kd 0 − Kd ⎤
⎢ ⎥
Kd = ⎢ 0 0 0 ⎥, Eq. 11
⎢⎣ − K d 0 K d ⎥⎦
⎡ Cd 0 − Cd ⎤
⎢ ⎥
Cd = ⎢ 0 0 0 ⎥, Eq. 12
⎢− Cd 0 Cd ⎥
⎣ ⎦
C d = diag [cd 1 ,..., cdn ] , Eq. 13
K d = diag [ k d 1 ,..., k dn ] , Eq. 14

Proc. of SPIE Vol. 8341 83410M-4

X = [ x11 , x21 ,..., xn+ m ,1 , x12 ,..., xn , 2 ]T , Eq. 15

M, C and K are the given mass, damping and stiffness matrices of the buildings, respectively; Cd and Kd represent
damping and stiffness matrices of the connectors, respectively; and X is the displacement vector of the system. Also, E is
a vector with all elements equal to one, and g(t) is the ground acceleration during the earthquake. Eq 3 describes the
equation of the motion in time domain. Assuming that the ground excitation is a stationary random process, the ground
acceleration, g(t), can be written as:

g(t) = S g (ω )e iωt
Eq. 16

where Sg(t) is the spectral density function of ground acceleration for a defined frequency of ω. Considering ground
excitation as a series of harmonic loads, one can rewrite the displacement vector in frequency domain as
X (t ) = X (ω ) e iωt . Eq. 17
Substituting Eqs. 16 and 17 into Eq. 3, the governing equation in frequency domain is obtained,
− M ω 2 X (ω ) e iωt + (C + C d )iω X (ω ) e iωt + ( K + K d ) X (ω ) e iω t = − ME S g (ω ) e iωt Eq. 18
and,
[
X (ω ) = [−Mω 2 + (C + C d )iω + (K + K d )] × −ME S g (ω ) ].
−1

Eq. 19
In this study, a Kanai-Tajimi filtered white noise function is used for spectral density function of ground acceleration:

2
⎛ω ⎞
1 + 4ζ ⎜ ⎟ 2
⎜ω ⎟
g

S g (ω ) = ⎝ g⎠ S0
Eq. 20
2
⎛ ⎛ω ⎞ ⎞ 2 2
⎜1 − ⎜ ⎟ ⎟ + 4ζ 2 ⎛⎜ ω ⎞
⎟
⎜ ⎜ω ⎟ ⎟ g⎜ ⎟
⎝ ⎝ g⎠ ⎠ ⎝ ωg ⎠

where ω g , ζ g and S0 represent dynamics characteristics and intensity of earthquake and are chosen based on geological
characteristics of a specific zone.
Having developed a formula for S g , it is now possible to numerically approximate σ xib . To do this, we impose upper
and lower limits of ± 20rad / s and apply a trapezoid rule approximation with a step size of 0.02 to the integral in Eq. 2. It
is worth noting that previous studies show that the effect of the frequencies greater than 20 rad / s on the response of the
structure is negligible [1].

2.4 Dampers' Mechanical Properties

It is clear that the standard deviation of displacement ( σ xib ) for a set of dampers is highly dependent on the mechanical
properties of the damper connectors since damping and stiffness matrices of the connectors in Eq. 3, Cd and Kd, are
dependent on cdi and kdi (Eqs. 3 and 11-14). Results from [1,6] show that the stiffness of the connectors does not change
the objective function significantly as long as cdi’s are optimal and kdi’s are small. Moreover, the value of σ xib will
increase if kdi has a very large value, i.e., rigid connectors. Thus, we assume that kdi=0 for all d and i. This simplifies the
problem, but the final result will remain the same. Thus the objective is to find an optimal set of cdi’s for two adjacent
buildings.

2.5 Objective function

The standard deviations of displacement, σ xib , represent the inter-story drift for each floor and provide our performance
measure for the effectiveness of a collection of dampers. In particular, we seek to minimize the inter-story drift. We

Proc. of SPIE Vol. 8341 83410M-5

therefore consider the objective function as the maximum value of squared inter-story drift. The mathematical
formulation of each objective function is:
{
F2 = max max{σ d i1 : i = 1..n + m}, max{σ d i 2 : i = 1..n} ,
2 2
} Eq. 21
where σ dib is the standard deviation of inter-story drifts, for ith floor of building b computed using a specific set of
damper coefficients.

3. OPTIMIZATION METHOD
In this study, we employ a derivative-free method, namely genetic algorithm (GA), to find the optimal damping
coefficients of damper connectors. GA is a well-known evolutionary search method imitating natural evolution process.
Many successful applications of GA to different engineering problems, particularly structural dynamics, are found in
literature (e.g. see: [7,8]). The GA used in this study follows standard GA steps. It starts with a random initial population
which is followed by fitness evaluation, selection, and reproduction steps. This iterative procedure continues until a
stopping condition is satisfied. Numerical examples are presented in order to examine efficiency of the presented
method. Efficiency and solution time of GA are compared with those of Nelder-Mead algorithm, another DFO method,
from a previous study [4].

4. RESULTS
For all numerical examples, ground acceleration parameters are considered as ωg = 15 rad / s , ζ g = 0.6 , ωk = 1.5 rad / s ,
ζ k = 0.6 and S0 = 4.65 × 10−4 m 2 / rad .s 3 , same as [1]. Three different sets of mechanical properties for the adjacent
buildings are defined in Table 1.
Table 1 – Mechanical properties of buildings
Building (a) Building (b)
ma (kg) ka (N/m) ca (N.s/m) mb (kg) kb (N/m) cb (N.s/m)
Set I 1.29E+06 4.00E+09 1.00E+05 1.29E+06 2.00E+09 1.00E+05
Set II 2.60E+06 1.20E+10 2.40E+06 1.60E+06 1.20E+10 2.40E+06
Set III 4.80E+06 1.60E+10 1.20E+06 4.00E+06 2.30E+10 1.20E+06

In Table 2 we tabulate the 8 different building height relations considered in this research. For each case the different
building height are given in fa and fb (denoting number of floors for buildings “a” and “b”).
Table 2 - Different sets for numerical tests
fa fb Case
10 10 Case 1
10 20 Case 2
20 10 Case 3
10 40 Case 4
40 10 Case 5
20 20 Case 6
20 40 Case 7
40 20 Case 8

Numerical results of the GA and Nelder-Mead algorithm are shown in Tables 3-5. It is seen that the Nelder-Mead
algorithm results in smaller objective values in a shorter time. In order to show efficiency and applicability of each
method, performance profiles [9] are computed and plotted in Figs. 2-4. Recall that performance profiles are designed to
graphically compare both speed and robustness of algorithms across a test set. This is done by plotting, for each

Proc. of SPIE Vol. 8341 83410M-6

algorithm, the percentage of problems that are solved within a factor of the best solve time. Mathematically, we first
compute the performance ratio by
t p ,a Eq. 22
rp,a = , for p ∈ Ρ
min{t p,a , a ∈ Α}

where Ρ is the set of all problems, Α is the set of algorithms, and t p ,a is the solving time of algorithm a for problem p.
The percentage of problems that are solved within a factor τ ∈ ℜ of the faster algorithm is next computed by
| { p ∈ Ρ : rp,a ≤ τ } |
ρa (τ ) = ×100% Eq. 23
|Ρ|
The percentage ρ a of all problems solved by algorithm a at τ gives us an overall assessment of the performance of
algorithm a. High values of ρ a near τ = 1 represents fast solve times. High values of ρ a , when τ is large, represent
high success rates. For a good algorithm, the plot of ρ a (τ ) should therefore be found above the other algorithms. For a
more detailed description of performance profiles we refer the reader to [9]. To calculate the performance profile, one
needs to define when a method “solves” a specific problem. In this paper, a method is considered as a “failed method” if
the difference between the objective value obtained using that specific method and the best objective value obtained for
that problem exceeds the defined allowable tolerance. The performance profiles show that Nelder-Mead is much better
than GA, except for few cases where the solution quality is not important (allowable tolerance of 10%). For those special
cases, GA is able to solve the problem while Nelder-Mead is not. Results suggest that if GA is more likely to be able to
find global optimum; however, it is not strong to find an exact optimal value. On the other hand, if NM converges to a
local minimum, it is very unlikely to be able to find global minimum. Results show that NM generally results in more
accurate optimal values; however, if it converges to a local minimum, the optimal value obtained using GA is much
better than the optimal value obtained using NM.

Proc. of SPIE Vol. 8341 83410M-7

Table 3- Comparison between GA and NM for building set 1
solution time final objective value
fa fb GA NM GA NM
10 10 175 157 8.47E-07 1.07E-06
10 20 211 180 1.82E-06 1.80E-06
20 10 212 182 1.58E-06 1.53E-06
10 40 310 235 5.49E-06 5.52E-06
40 10 312 236 6.38E-06 5.88E-06
10 10 192 164 4.28E-07 4.02E-07
10 20 226 186 2.40E-07 2.39E-07
20 10 228 186 3.31E-07 3.23E-07
10 40 326 242 2.89E-06 2.70E-06
40 10 326 242 1.29E-06 1.22E-06
10 10 209 171 4.80E-07 4.61E-07
10 20 243 191 4.44E-07 4.36E-07
20 10 247 193 5.34E-07 5.58E-07
10 40 492 245 6.13E-06 9.16E-06
40 10 346 247 2.48E-06 2.36E-06

Maximum tolerance of 1%
100

90

80

70

60
P(τ)

50

40

30

20

10 Genetic Algorithm
Nelder-Mead Algorithm
0
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
τ

Figure 2- Performance profile for maximum tolerance of 1%

Proc. of SPIE Vol. 8341 83410M-8

Table 4- Comparison between GA and NM for building set 2
solution time final objective value
fa fb GA NM GA NM
10 10 75 128 8.53E-07 1.07E-06
10 20 110 148 1.96E-06 1.80E-06
20 10 117 149 1.57E-06 1.53E-06
10 40 208 203 5.66E-06 5.52E-06
40 10 210 203 6.40E-06 5.88E-06
10 10 92 132 4.26E-07 4.02E-07
10 20 129 152 2.42E-07 2.39E-07
20 10 131 154 3.54E-07 3.23E-07
10 40 227 209 2.78E-06 2.70E-06
40 10 230 209 1.26E-06 1.22E-06
10 10 111 137 4.87E-07 4.61E-07
10 20 146 157 4.50E-07 4.36E-07
20 10 149 160 5.78E-07 5.58E-07
10 40 350 214 6.19E-06 9.16E-06
40 10 250 214 2.66E-06 2.36E-06

Maximum tolerance of 5%
100

90

80

70

60
P(τ)

50

40

30

20

10 Genetic Algorithm
Nelder-Mead Algorithm
0
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
τ

Figure 3- Performance profile for maximum tolerance of 5%

Proc. of SPIE Vol. 8341 83410M-9

Table 5- Comparison between GA and NM for building set 3
solution time final objective value
fa fb GA NM GA NM
10 10 130 143 8.76E-07 1.07E-06
10 20 162 162 1.96E-06 1.80E-06
20 10 167 164 1.58E-06 1.53E-06
10 40 261 220 5.59E-06 5.52E-06
40 10 263 220 6.39E-06 5.88E-06
10 10 142 149 4.23E-07 4.02E-07
10 20 180 169 2.39E-07 2.39E-07
20 10 182 170 3.48E-07 3.23E-07
10 40 279 225 2.73E-06 2.70E-06
40 10 280 226 1.22E-06 1.22E-06
10 10 158 155 4.87E-07 4.61E-07
10 20 194 174 4.37E-07 4.36E-07
20 10 198 176 5.00E-07 5.58E-07
10 40 293 230 8.17E-06 9.16E-06
40 10 296 233 2.48E-06 2.36E-06

Maximum tolerance of 10%

100

90

80

70

60
P(τ)

50

40

30

20

10 Genetic Algorithm
Nelder-Mead Algorithm
0
1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2
τ

Figure 4- Performance profile for maximum tolerance of 10%

Proc. of SPIE Vol. 8341 83410M-10

Fig 5 shows optimal uniform and optimal non-uniform distributions of damping coefficient along the building for
different problems (for the different sets for numerical tests shown in Table 2). For all cases, as expected, the Cd value is
decreasing floor heights. It is interesting to note that, the non-uniform and uniform damping coefficients intercept around
floor 6. The value is approximately the average of the linearly increasing damping coefficient.

10

6
Floor

5
Set I (uniform)
4 Set I (non-uniform)
Set II (uniform)
3 Set II (non-uniform)
Set III (uniform)
2 Set III (non-uniform)

1
5 5.5 6 6.5 7 7.5 8
Log(Cd)

Figure 5- Optimal damping coefficients

Fig 6 compares the dynamic behavior of the buildings with uniform dampers and those with non-uniform dampers
(obtained using NM). It is seen that the maximum drift (objective function) is decreased by almost 10% when non-
uniform dampers are employed. This confirms that assumption of uniform dampers is not a reliable assumption. Thus it
is worth to employ a derivative free method to find the optimal damping coefficients.

10

6
Floor

Building 1 (uniform dampers)

3
Building 2 (uniform dampers)
Building 1 (non-uniform dampers)
2
Building 2 (non-uniform dampers)

1
0 0.2 0.4 0.6 0.8 1 1.2
Inter story drift -3
x 10

Figure 6- Inter-story drift

Proc. of SPIE Vol. 8341 83410M-11

 5. CONCLUSION
In this study, optimal damping coefficient of dampers between two adjacent structures is found. Since no derivative is
available for the present problem, conventional optimization techniques which are based on derivatives cannot be
employed here. Also, it should be noted that the number of design variables is equal to the number of dampers which is
definitely greater than one. In other words, this is an n-dim optimization problem where derivatives are not available.
Therefore, we consider GA and NM to solve this problem. The results show that the conventionally-accepted assumption
of uniformly distributed dampers is not a valid assumption and a globally optimal solution to the dampers coefficient is
not uniform. The new solution can reduce the objective function up to 10% which is a significant improvement. Finally,
a comparison between GA and NM shows that NM results in more accurate results in less solution time. Therefore, NM
is a more reliable choice to solve optimal damping coefficient of dampers between two adjacent structures.

REFERENCES

[1] Y. Xu, Q. He, and J. Ko, “Dynamic response of damper-connected adjacent buildings under earthquake
excitation,” Engineering Structures 21, 135–148, Elsevier (1999).

[2] Z. Yang and X. Lu, “Experimental seismic study of adjacent buildings with fluid dampers,” Journal of
Structural Engineering 129, 197 (2003).

[3] S. Bharti, S. Dumne, and M. Shrimali, “Seismic response analysis of adjacent buildings connected with MR
dampers,” Engineering Structures 32, 2122–2133, Elsevier (2010).

[4] K. Bigdeli, W. Hare, and S. Tesfamariam, “Determining optimal non-uniform damping coefficients for adjacent
buildings via a Nelder-Mead approach,” in 23rd Canadian Congress of Applied Mechanics, 2011, Vancouver,
BC, Canada (2011).

[5] C. Patel and R. Jangid, “Seismic response of dynamically similar adjacent structures connected with viscous
dampers,” The IES Journal Part A: Civil & Structural Engineering 3, 1–13, Taylor & Francis (2010).

[6] W. Zhang and Y. Xu, “Dynamic characteristics and seismic response of adjacent buildings linked by discrete
dampers,” Earthquake Engineering & Structural Dynamics 28, 1163–1185, John Wiley & Sons (1999).

[7] R. Alkhatib, G. Nakhaie Jazar, and M. Golnaraghi, “Optimal design of passive linear suspension using genetic
algorithm,” Journal of Sound and Vibration 275, 665–691, Elsevier (2004).

[8] S. Ok, J. Song, and K. Park, “Optimal design of hysteretic dampers connecting adjacent structures using multi-
objective genetic algorithm and stochastic linearization method,” Engineering Structures 30, 1240–1249,
Elsevier (2008).

[9] E. D. Dolan and J. J. Moré, “Benchmarking optimization software with performance profiles,” Mathematical
Programming 91, 201–213 (2002).

Proc. of SPIE Vol. 8341 83410M-12