Optimization for Engineering Design

(knowledge loop)
Design is an iterative proces Scientific Knowledge
+
Start: - poorly defined problem Existing Devices
knowledge Components
- refine it Materials
Manufacturing methods
- develope a model NEED (Society) Market
Economic conditions
- a solution Requirements Limitations
- More than one solution
CONCEPTUALIZATION
Decision making: selecting the best
choice from among a small number of
known design options
VERIFICATION
Feasibility analysis
Optimal Design means the best of all
feasible designs.
REVISION MODIFICATION
Optimization is the process of maximizing No No
Decision
a desired quantity or minimizing undesired (Convergence)
one. ?
Initial design
(size and shape parameters)
Yes
Numerical Analysis OPTIMIZATION
Performance measure
Sensitivity analysis

Modify the design to better achive PRODUCTION

critical performance citeria
İlhan Konukseven ME 407 Mechanical Engineering Design 1
Differential evolution. Evolutionary algorithms. Dynamic relaxation. Genetic algorithms.
Hill climbing. Particle swarm optimization.

Objective Function: Defines the value of the design in terms of the independent
variables (Design Variables) established.
F  F ( x1 , x 2 , . . . . . , x n )
Typical objective functions could be cost, weight, reliability, or producibility
Functional constraints: Constraints arise from physical laws and limitations or from
compatibility conditions on the individual variables.
Functional constraints g , also called equality constraints, specify relations that must
exist between the variables.
g1  g1 ( x1 , x 2 , . . . . . , x n )  0
g 2  g 2 ( x1 , x 2 , . . . . . , x n )  0
.
.
g 3  g 3 ( x1 , x 2 , . . . . . , x n )  0
Regional constraints: Reginal constraints h, also called inequality constraints, are imposed
by specific details of the problem. A type of regional Constraint that arises naturally in
design situations is based on specifications. Specifications are points of interaction with
other parts of the system
h1  h1 ( x1 , x 2 , . . . . . , x n )  L1
h2  h2 ( x1 , x 2 , . . . . . , x n )  L 2
.
.
hm  hm ( x1 , x 2 , . . . . . , x n )  L m
* * * * * *
Optimal Solution: F  F ( x1 , x 2 , . . . . . , x n ) F is min or max
İlhan Konukseven ME 407 Mechanical Engineering Design 2
 Example: Optimize the cost of a cylindrical tank to store a fixed volume of liquid V.
The tank will be constructed by forming and welding thin steel plate.

Design variables:
Tank diameter :D
Tank height :h

Objective Function : The cost depends directly on the area of the plate
(C is the cost per unit area)

D 2
F  C(  Dh)
2
Functional constraint : is introduced by the requirement that the tank must hold a
specified volume
 D 2h
V
4
Regional constraints : are introduced by the requirement for the tank to fit in a
specified location or to not have unusual dimensions

Dmin  D  Dmax ; hmin  h  hmax

İlhan Konukseven ME 407 Mechanical Engineering Design 3

 Optimization by Differential Calculus

A characteristic property of an extremum is that F is momentarily stationary at each point.

dF Stationary Point
0
dx1
d 2F
0 If the curvature is negative, SP indicates a local maximum
dx12

d 2F
0 If the curvature is positive, SP indicates a local minimum
dx12
Mathematical Maxima The tank problem:
Global maximum
Local maximum D 2
F  C(  Dh)
2
 D 2h 4V
V h
4  D2
F dF 4CV
 0  C D  2
dD D
Point of Inflection
1/ 3
 4V 
D   1.084 V 1 / 3
  

İlhan Konukseven ME 407 Mechanical Engineering Design 4

 Optimization by Lagrange Multipliers

The Lagrange multipliers provide a poweful method for finding optima in multi-variable
problems involving functional constraints.

Design variables:
x, y, z
Objective Function:

F  F1 ( x , y , z )
Subject to the functional constraints:

Functional constraint:
g1  g1 ( x , y , z )
g2  g2 ( x , y, z )
The Lagrange expression (LE):

LE  F1 ( x , y , z )  1 g1 ( x , y , z )   2 g 2 ( x , y , z )
1 and 2 are the Lagrange multipliers.
The following conditions must be satisfied at the optimum point
LE LE LE LE LE
0 0 0 0 0
x y z 1  2
İlhan Konukseven ME 407 Mechanical Engineering Design 5
 Optimization by Lagrange Multipliers (Example)

Problem: Select the dimensions of an open top rectangular storage container to

minimize its fabrication cost. The container must be able to store Vm3 of supplies,
and is to be fabricated from a stell bottom costing S $/m2. and wooden sides
costing W $/m2 As shown in Figure, we’ll let x and y be the dimensions of the base
and z be the height of the container.

Design Parameters: x, y, z

Objective function to be mimimized: C  Sxy  2Wz ( x  y )

1st Constraint: xyz  V

One approach to deal with the constraint on the volume is to solve constraint
aquation for one of the variables, then substitude back to eliminate one variable:
C C
 0, 0
x y
In practice it may not be possible to solve the constraint equation. Suppose the
constrait equation was of the form:
Cos( xy )  z 2  e yz

İlhan Konukseven ME 407 Mechanical Engineering Design 6

 Optimization by Lagrange Multipliers (Example)

The Lagrange multiplier technique for optimizing an objective function subject to a

constraint avoids substituting the constraint into the objective function.

Rewrite the constraint: xyz V  0

Construct a Lagrangian function L: L  C   ( xyz  V )

L  Sxy  2Wz( x  y )   ( xyz  V )
Optimizing L is the same as optimizing C. L is a function of four
variables: ( x , y , z ,  )
The original constrained optimization problem is transferred to an
unconstrained optimization problem.
L
1   0  Sy  2Wz  yz
x
L
2   0  Sx  2Wz  xz
y
L
3   0  2W  x  y   xy
z
L
4   0  xyz  V

İlhan Konukseven ME 407 Mechanical Engineering Design 7
 Optimization by Lagrange Multipliers (Example)
2W ( x  y )
Solve for  in terms of x, y, and z as:  3
xy
2Wz ( x  y )
5  4
V
2Wyz 2 ( x  y )
Sy  2Wz  0  1 5
V
2Wxz 2 ( x  y )
Sx  2Wz  0  2 5
V

2Wz 2 ( x  y )( x  y )
S( x  y)  0
V

 2Wz 2 ( x  y ) 
( x  y ) S  0
 V 

 2Wz 2 ( x  y ) 
( x  y)  0 S  0
 V 

This tells us that the optimum design for the container is one in which the bottom
plate is square

İlhan Konukseven ME 407 Mechanical Engineering Design 8

 Search Methods

1. Deterministic search: is assumed to be free from appreciable experimental error

2. Stochastic search: the existence of random errors must be considered.

All engineering problems cannot be expressed by analytical functional relations

Instead of differentiating an analytical expression to find the optimum conditions,
a search strategy can be devised for finding the optimum in the fewest possible
experiments

Uniform Search

In the uniform search method the trial points are spaced equally
over the allowable range of values

x lies over the range x = 0 to x = 10

four experiments available (n)

Each interval width L/n+1

The optimum will lie between 2 < x < 6 (interval of uncertainty)

with four experiments we have narrowed the range of values that require
further search to 40 percent of the total range.

İlhan Konukseven ME 407 Mechanical Engineering Design 9

 Uniform Dichotomous Search

In the uniform dichotomous search method, experiments are performed in pairs to

establish whether the funtion is increasing or decreasing.

the experiments are spaced evenly over the entire range.

x lies over the range x = 0 to x = 10

four experiments available (n)

Pair # n/2
Each interval width L/[(n/2)+1]

Two experiments in each pair are seperated in x

by an amount slightly greater than the experimental
error in x.
x=3.30 and 3.36 and x=6.64 and 6.70

ya < yb the region 0 < x < 3.33 is excluded

yc > yd, the region 6.67 < x < 10 is excluded
(interval of uncertainty)
with four experiments we have narrowed the range of values that require further
search to 33 percent of the total range.

İlhan Konukseven ME 407 Mechanical Engineering Design 10

 Sequential Dichotomous Search
In sequential search the experiments are done in sequence, each taking advantage
of the information gained from the preceding one.

x lies over the range x = 0 to x = 10

four experiments available (n)

We first run a pair of experiments near

the center of the range, x=5.0

Each pair are seperated in x by an amount slightly

greater than the experimental error in x.
x=5.00 and 5.06
Since yb > ya we eliminate x > 5.0

The second pair of experiments are run at the center

of the remaining interval,
x=2.5 and 2.56
Since yd > yc, we eliminate the region 0 < x < 2.5

(interval of uncertainty) 2.56 < x < 5.00

with this search technique four experiments have narrowed the range of values that
require further search to 25 percent of the total range.

İlhan Konukseven ME 407 Mechanical Engineering Design 11

 Fibonacci Search
Sequential technique which is based on the use of the Fibonacci number series.

Fibonacci series is given by Fn=Fn-2+Fn-1, where F0=1 and F1=1.

n 0 1 2 3 4 5 6 7 8 9 ..
Fn 1 1 2 3 5 8 13 21 34 55

d1 = (Fn-2/Fn)L from each end of the range values

For n=4
d1 = (2/5)10=4
Since y4 > y6, the interval 6 < x < 10 is eliminated

d2 is obtained by letting n2 = n – 1
d2 = (Fn-3/Fn-1)L
d2 = 1/3(6)=2
Since y4 > y2, the interval 0 < x < 2 is eliminated

Last experiment:
x=4.06 to determine whether the optimum is 2 < x < 4 or 4 < x < 6
(interval of uncertainty)
with this search technique four experiments have narrowed the range of values to
4 < x < 6 that require further search to 20 percent of the total range.

İlhan Konukseven ME 407 Mechanical Engineering Design 12

 Golden Search

Fibonacci search has the disadvantage that it requires an advance decision on the number
of experiments before we have any information about the behavior of the function near
the maximum.

Golden section search does not require an advance decision on the number of trials.
n 0 1 2 3 4 5 6 7 8 9 ..
Golden mean: Fn-1/Fn=0.618 where n > 8 Fn 1 1 2 3 5 8 13 21 34 55

F(x) F(x)

0.618L1 0.618L2
0.618L1 0.618L2

a b x c a b x
0 1 0 1
0.382 0.618 0.236 0.618
0.382

F(b) > F(c) [b, 1 ] eliminated F(a) > F(c) [0, c ] eliminated

İlhan Konukseven ME 407 Mechanical Engineering Design 13

 Golden Search

Golden mean (0.618): a length divided into two unequal segments such that the ratio of the
length of the whole to the larger segment is equal to the ratio of the length of the larger
segment to the smaller segment.

F(x) F(x)
L3=0.382

0.618L2 0.618L3
0.618L2 0.618L3

x a x
0 c a b 1 0 0.236c 0.382 d b
0.618 1
0.236 0.618 0.472
0.382

F(a) > F(c) [0, c ] eliminated F(d) > F(a) [0, a ] eliminated

(interval of uncertainty)
with this search technique four experiments have narrowed the range of values to
0.382 < x < 0.618 that require further search to 23 percent of the total range.

İlhan Konukseven ME 407 Mechanical Engineering Design 14

 Comparison of Univariable Search Methods

Reduction ratio: The ratio of the original interval of uncertainty to the interval remaining after
n trials.

Reduction ratio
Search Method For n=4 For n=13
Uniform (n+1)/2 2,5 7
Uniform dichotomous (n+2)/2 3 7,5
Sequential dichotomous 2^n/2 4 90
Fibonacci Fn 5 377
Golden search 4,3 250

İlhan Konukseven ME 407 Mechanical Engineering Design 15

 Multivariable Search Methods
Lattice Search:
Two-dimensional grid lattice is super-imposed over the contour plot.

The starting point is selected near the center of the region, at point 1. The objective function is
evaluated for points one through nine. If point five turns out to be the largest value, it becomes
the central point of the next search. The procedure continues until the location reached is one at
which the central point is greater than any of the other eight points. Frequently, a coarse grid is
used initially and a finer grid is used after the maximum is approached.

İlhan Konukseven ME 407 Mechanical Engineering Design 16

 Univariate Search:
Univariate search is a one-at-a-time method. All of the variables are kept constant
except one, and it is varied to obtain an optimum in the objective function. That optimal
value is than substituted into the function and the function is optimized with respect to
another variable.
Starting at point 0 we move along x2 = const to a maximum by using any one of the single
variable search techniques. Then we move along x1 = const to a maximum at point 2 and
along x2 = const to a maximum at 3.

If the responce surface contains a ridge, then the univariate search can fail to find an
optimum.

İlhan Konukseven ME 407 Mechanical Engineering Design 17

 Steepest Ascent:
The path of steepest ascent (max) or descent (min) is the gradient method.
The direction of the search is changed in the direction of the maximum slope.

Steepest ascent method is applicable to many variables.

F F
F  i1  i2
x1 x 2
To move in the direction of the gradient vector, take the
step lengths x1 and x1 in proportion to the
components of the gradient vector.

 F 
For the general case of n independent variables, F = F(x1, x2, . . . . . . .xn).  x 
 1
 F 
 x 
F   2 
.
 
Starting at the initial point, we take a small finite difference x i in each  . 
variable and evaluate the function at each x i  x i in turn, holding all  F 
other x i at their initial values. Fi  Fi  F0  x 
 n

İlhan Konukseven ME 407 Mechanical Engineering Design 18

 Steepest Descent: F ( x1 , x 2 )  x1  x 2  2 x12  2 x1 x 2  x 22
 0
1. Start with arbitrary xi i=1 X1   
 0
2. Find Search direction  F 
x1   1  4 x1  2 x 2 
S i   Fi   F ( x i ) F  
 F

   1  2 x1  2 x 2 
*  x 2 
3. Determine the optimal step i and X 1  0 0
T

*
  1
xi 1  xi   i S i S1   F   
1
*
4. Test xi+1   0.1
if xi+1 is optimum stop search  0   1   0.1
X 2     0.1    
 0  1   0.1 
5. i=i+1 goto step 2   0.8
S 2   F   
 0.6 
Convergence Criteria:   0.1   0.8   0.18
X3     0.1  
 0.1   0.6   0.16 
F ( xi 1 )  F ( xi )
  1 Change in function value is small S 3   F  
  0.6

F ( xi )  1.04 
  0.18   0.6   0.24
F Components of Grad. is small
X3     0.1  
 2  0.16   1.04   0.26 
x i .
.
xi 1  x i   3 Change in design vector is small   1
Xn   
1.5

İlhan Konukseven ME 407 Mechanical Engineering Design 19

 Linear Programming

Design variables (x1, x2, , , ,xn)

Objective function F (x1, x2, , , ,xn) without violating the constraints.

The most widely studied class of optimization problems involving inequality constraints
consists of both an objective function F and a set of constraints that are linear functions
of the design variables. This class of problems is called linear programming.
(mathematical term)

Maximize the objective function F

F ( x1, x 2 , . . . ., x n )   k i x i
n

Where xi and n design variables, subject to m constraints on the design variables of the
form
ai
ij xi  rj j  1,2, . . . , m

In the above expressions, aij and ki are constants presumed to have known values in any
particular problem. The standart form of linear programming problems also requires that
the design variables be non-negative, that is
xi  0

İlhan Konukseven ME 407 Mechanical Engineering Design 20

 Linear Programming (Example)

We wish to add two additives to the parent metal iron to design a speciality steel with
certain enhanced properties.

Percent composition of the first additive x1

Percent composition of the second additive x2

Design variables : x1 and x2

Objective : is to increase in crack resistance
Every percent of additive 1 in the alloy increases crack resistance by 2%
Every percent of additive 2 in the alloy increases crack resistance by 1%

F=2x1+x2

First constraint: cost constraint

Each percent increase in additive 1 increases the cost by 3%
Each percent increase in additive 2 increases the cost by 1%
The overall cost increase to no more than 9%.
3x  x  9 ( 1 )
1 2

Second constraint: undesirable effects of the alloy additives on the metal’s corrosion
resistance
Additive 1 increases the corrosion resistance by 1% for every 1% incereases
Additive 2 decreases the corrosion resistance by 2% for every 1% increases
The maximum acceptable decreases in corrosion resistance is %4

 2 x2  x1  4 ( 2 )
İlhan Konukseven ME 407 Mechanical Engineering Design 21
 Linear Programming (Example)

Third constraint: reduction in melting temperature

Each percent increase in additive 1 decreases the melting temperature by 1%
Each percent increase in additive 2 decreases the melting temperature by 1%
We want to limit the total decrease in melting temperature to no more than 4%.
x1  x2  4 ( 3 )
Fourth constraint: stoct to provide
Supplier only has enough in stock to provide additive 2 at most a %5 level in the alloy
x2  5 ( 4 )
Generic Linear programming constraints: ( 1 )

x1  0 ( 5 )
( 4 ) redundant
x2  0

( 2 )

( 3 )

İlhan Konukseven ME 407 Mechanical Engineering Design 22

 Linear Programming (Example)

Rewrite objective function

x2=F-2x1

All lines have a slope of-2; the position of each line is determined by the value of F.

As we continue moving in the direction of

the lines of increasing F, a limiting case
is reached for which a single solution
accurs at the corner point A.

U=2(2.5) + 1.5 = 6.5

For this problem, the maximum value is

F=6.5

F = 6.5
F=0 F=2 F=4 F=6

İlhan Konukseven ME 407 Mechanical Engineering Design 23