See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/262390522

Ship design optimization considering uncertainty

Conference Paper · July 2010

CITATIONS READS
0 95

2 authors, including:

Wayne L. Neu
Virginia Polytechnic Institute and State University
41 PUBLICATIONS   339 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Wayne L. Neu on 18 August 2015.

The user has requested enhancement of the downloaded file.

 Ship Design Optimization Considering Uncertainty
Wayne L. Neu, Aerospace and Ocean Engineering, Virginia Tech, Blacksburg, VA/USA, [email protected]
Nathan Good, NAVSEA, Washington, DC/USA, [email protected]

Keywords: Multi-objective optimization, uncertainty, all criteria.

mean value method, confidence of success A traditional way of dealing with uncertainty is to in-
troduce margins or safety factors. This approach penalizes
Abstract all designs and does not allow the designer to identify
Multi-disciplinary analysis and design optimization regions of the design space that may produce more robust
allow designers to break away from the traditional design designs. It may even occur that the “penalty” being added
spiral approach and focus on searching the design space is used as a virtue by another aspect of the design. It is
for the best overall design rather than considering best suggested here that a more appropriate means of treating
discipline-specific designs. The effects of uncertainty in these uncertainties is to adopt a probabilistic approach
the disciplinary calculations of complex design simulation throughout the analysis. This approach, if implemented
models are considered. Uncertainty is quantified using a using a brute-force technique such as a Monte-Carlo si-
“confidence of success” calculation that is developed mulation would add unacceptable computational loads to
based on the Mean Value Method. The confidence of suc- an optimization. The key to this approach then, is to iden-
cess is the probability that a design will satisfy all con- tify a much simpler calculation technique which will yield
straints and meet performance objectives. Multiple sufficiently accurate results.
sources of uncertainty are quantified and represented in a In the following, we introduce a quantity we call
ship synthesis model. Optimizations using confidence of Confidence of Success (CoS), which is defined as the
success as a third objective, in addition to cost and effec- probability that a given design meets all objectives and
tiveness, in a multi-objective genetic algorithm are ex- constraints concurrently. It is insufficient to look at each
ecuted. It is demonstrated that designs lying along con- criterion (objective or constraint) and its distribution in-
straint boundaries, which are typically identified as opti- dependently, since all attribute values are generated by
mum designs in deterministic design optimizations, likely the same design choices and thus are interdependent. We
have a very low chance of being successful. accomplish this by first introducing uncertainties in a
number of disciplinary calculations in a multidisciplinary
1. INTRODUCTION ship synthesis model through the addition of uncertain
This paper presents an optimization tool that might variables with assumed probability distributions. We then
be used during the concept exploration phase of a ship examine the joint probability distribution, in all uncertain
design process. The concepts illustrated here, however, variables, of each appropriate ship characteristic deter-
apply to any optimization utilizing a process simulation mined by the model employing these, now random, dis-
that may contain uncertainties. In the case of ship design, ciplinary calculation results to determine what portion of
the simulation, or ship synthesis model, is the aggregate the uncertain variable space satisfies all criteria. This is
of several disciplinary calculations made to determine the done using the approximation of the Mean Value Method.
characteristics of a ship possessing a chosen set of The CoS is calculated as the integration of the joint prob-
attributes. During early stages of design, it is customary to ability of occurrence of designs over the portion of the
perform these disciplinary calculations using parametric, uncertain variable space satisfying all criteria. CoS
perhaps regression, analyses of varying accuracy. Deci- merges reliability and robust based design approaches
sion makers would benefit from having an understanding into an uncertainty calculation.
of how the uncertainties in these disciplinary calculations
compound in arriving at the ship design characteristics 2. MULTI-OBJECTIVE OPTIMIZATION AND
being calculated by their synthesis models. UNCERTAINTY
Further, it is important for them to consider the im- There is typically no single global optimum when
pact of these uncertainties on the results of optimization performing multi-objective optimization because of tra-
algorithms operating on these synthesis models. By the deoffs between associated objectives. Multi-objective
nature of the optimization process, the optimum solution, optimization results in a set of Pareto optimal design solu-
more often than not, resides along one or more constraint tions having the property that, within the feasible portion
boundaries. If the characteristics being calculated by a of the design space (defined by the constraints applied to
model (simulation) are uncertain, what are the chances the design), there exists no other design which can im-
that those characteristics will actually fall outside of the prove upon any single objective without decreasing the
constraint, in an infeasible region? In the work described desirability of at least one other objective, [Parsons and
below, we have shown that an optimum solution calcu- Scott 2004]. The set of these Pareto optimal solutions
lated without regard to model uncertainty often has a fair- forms what is called a Pareto front or non-dominated
ly small chance of being a successful design, i.e., meeting frontier. A typical Pareto front lies at the boundary be-
tween the feasible and infeasible regions in the objective sensitivity based approaches. Which is the most efficient
space. Every individual design is defined by a vector of approach is strongly problem dependent, as there is typi-
in the design variable space which maps to a point in the cally a trade-off between accuracy and computational
objective space through the simulation or synthesis model cost. For the multi-objective design optimization of a
which calculates values of the objectives from the design multi-disciplinary ship synthesis model, the approach of
variables. In an optimization with two objectives, the choice must have a low computational cost due the com-
Pareto front is a line in a two dimensional space. In a putational burden associated with a complex model and
problem with three objectives, the Pareto front would genetic optimization algorithm. Therefore, the CoS calcu-
form a three dimensional surface. lation in this work is based only on the first steps of the
The decision of which member of the Pareto optimal Mean Value Method (MV), i.e., a first order Taylor series
set to choose needs to be made by a decision maker. At- expansion, classifying it as a sensitivity based approach.
tractive designs include those at the extremes of the front This method is discussed in greater detail below.
and at “knees” in the curve or surface. Designs located at The Monte Carlo Simulation (MCS) or simple ran-
the “knees” are of interest because of their sharp increase dom sampling is a familiar method. MCS is implemented
in one objective without giving up too much in another by randomly simulating a process, given the stochastic
objective. properties of one or more random variables, with focus on
A deterministic optimization does not consider the characterizing the statistical nature of the responses of
possibility that the chosen design will actually perform as interest, [Handscomb and Hammersley 1964]. This simu-
predicted regarding each objective or any of a number of lation method has long been considered to be the most
constraints imposed on the design. Due to uncertainties exact method for calculation of probability distributions
associated with the problem definition or the synthesis of responses from systems dealing with uncertainty.
model, there is a risk that the chosen design may under- However, MCS typically calls for many thousands of sys-
perform and/or not meet the constraint requirements. As a tem evaluations and is impractical for a time consuming
result, there is a need to address uncertainty in any design analysis. MCS will be employed to verify the results ob-
optimization problem to produce both effective and relia- tained below from MV.
ble results.
Even if this risk can be evaluated, if the chosen de- 3. MEAN VALUE METHODS
sign is deemed too risky, it is not sufficient to simply se- Mean Value methods utilize most probable point
lect a more reliable design from a Pareto front determined (MPP) analysis, and, being based on first order Taylor
without consideration of this risk. If uncertainty has not series expansions, are appropriate for well-behaved re-
been included in the optimization process, there may not sponse functions that are computationally intensive,
even be a more reliable design on the Pareto front. More [Thacker et al. 2001]. Mean Value methods approximate a
reliable designs might be located well off the Pareto front cumulative distribution function (CDF) of the response
in the feasible region of the objective space. Therefore, function (here a model output), which can be differen-
the optimization process should be performed with a con- tiated to obtain a probability density function (PDF). In
straint on uncertainty, or uncertainty must be treated as a the analysis, a response, Z, is defined as a function of the
third objective. Confidence of Success (CoS) is defined uncertain random variables, X. Each point in the design
here as the probability that a specific design will satisfy space spanned by X has a specific probability density
all constraints and meet performance objectives, and it is according to the joint probability density function (JPDF)
treated as an additional objective in this work. With CoS of the uncertain random variables (RVs). Therefore every
included as an objective, the designer has the ability to response value Z(X) has an associated probability density,
make design decisions that are effective, efficient, and [Mavris et al. 1997].
reliable. The failure surface or the limit state, g, for an arbi-
Sources of uncertainty are abundant. They may arise trary response function Z is defined as:
from both qualitative and quantitative sources. This work
g = Z(X) – Zls = 0
focuses on the quantitative modeling and simulation un-
certainties that occur due to discrepancies in computa- where the limit state value Zls is a particular value of Z.
tional predictions. These are the uncertainties associated The limit state function (LSF) is defined such that g(X) =
with, perhaps parametric, disciplinary calculations that 0 is a boundary that divides the random variable space
may propagate through multi-disciplinary models due to into two regions, failure/infeasible and safe/feasible, see
the interdependence of modules. The computational out- Figure 1. We have now introduced a third space, the ran-
put of one discipline is typically the input of another dis- dom variable space, that spans the range of possible val-
ciplinary calculation, allowing the initial uncertainty to ues taken by the quantities we are considering to be un-
transform and interact with other uncertainties as it certain.
progresses through the model. The most probable point (MPP) is defined as the
There are a number of approaches to handling system point along the LSF where the uncertain variable combi-
uncertainty, [Haldar and Mahadevan 1999], [Koch and nation yields the highest probability of occurrence. For a
Cofer 2000]. Most of these can be classified into three JPDF of two standard normal and independent variables it
categories: random sampling, design of experiments, and is obvious that the point of minimum distance from the
 for each design criterion. Each resulting LSF may remove
another section of the RV space. The integration of the
f(x)
JPDF over the remaining RV space yields the probability
Approximate LSF
g̃ = ZMV – Z ls = 0 of simultaneously meeting all criteria. The result of this
integration is the confidence of success.

4. SHIP SYNTHESIS MODEL AND GENETIC

ALGORITHM OPTIMIZER
Exact LSF The multi-disciplinary ship synthesis model used in
g = Z – Z ls = 0
this project has been under development for several years.
x1 The history of its development, additions, and modifica-
x2 MPP
tions can be found in [Brown and Thomas 1998] and
[Stepanchick and Brown 2006]. This particular ship syn-
Figure 1: JPDF f(x) of two standard normal variables. thesis model was constructed to represent modern United
The MV approximate LSF is used to define the infeasi- States Navy (USN) surface combatants such as Ticonde-
ble region of the variable space, from which the JPDF roga Class Cruisers, Arleigh Burke Class large multi-role
has been cut off. Also plotted are the exact LSF and the Destroyers, and Oliver Hazard Perry Class Frigates. The
MPP, after [Thacker et al. 2001]. DDG 51 Arleigh Burke was selected as a model evalua-
tion design point (DP) and was chosen as the initial de-
origin to the LSF represents the most probable combina- sign in all optimization efforts. It appears multiple times
tion of the random variables, and is therefore named the below.
most probable point, as seen in Figure 1, [Thacker et al. ModelCenter 6.1, a commercial process integration
2001]. For nonlinear limit states, identifying the MPP and design optimization software package from Phoenix
becomes an optimization problem in itself. Non-normally Integration, Inc., was used to build and execute this com-
distributed random variables may be transformed into plex ship synthesis model. This work uses Phoenix Inte-
normal form with some distortion of the LSFs. Once the gration’s multi-objective genetic optimization plug-in,
MPP has been identified, it can be used as a basis to de- Darwin version 1.1.3. The results presented below make
velop approximate polynomial LSFs. use of Darwin’s Multiple Elitist selection scheme with a
Assuming the response function, Z, is “smooth” or population size of 300 and 100 designs preserved between
can be “smoothed” and that a first order Taylor series generations for crossover and mutation. Since genetic
expansion of Z exists at the mean, μx , of the uncertain algorithms do not need to calculate gradients, they have
variables X, the response function Z can be expressed as: the ability to treat discrete choice design variables.
The complete model is comprised of an input mod-
n
 Z 
Z ( X)  Z (μ x )      ( X i   Xi )  H ( X) ule, primary ship synthesis model modules, a constraint
i 1  X i   module, three objective modules, a numerical differentia-
n tion calculation, and the genetic algorithm optimization
 a0   ai  ( X i   Xi )  H ( X) module. The primary ship synthesis model is comprised
i 1 of several interlinked disciplinary modules, each contain-
 Z MV ( X)  H ( X) ing considerable detail that will not be discussed here.
The primary ship synthesis modules include combat sys-
where the derivatives are evaluated at the mean of the tems, propulsion, hull form, electric power, resistance,
uncertain variables, μXi is the mean of uncertain variable i, weight and stability, space required, space available, tan-
and n is the number of uncertain variables. ZMV (X) kage, feasibility, cost, and effectiveness.
represents the MV method function and is the sum of the Each ship design is defined by 26 design variables
first order terms only; H(X) represents the remaining (DV) and 21 fixed input parameters, which are set as ei-
higher order terms. The coefficient a0 is simply the ex- ther fixed constants or fixed constraints in accordance
pected value of Z at the mean values, and the coefficients with the DDG 51. The nine constraints defined in this
ai are generally computed using numerical differentiation ship synthesis model are listed in Table 1 since they are
with a minimum of (n+1) Z-function evaluations, [Wu et referred to below. There are four static constraints, which
al. 1990]. Since the function ZMV (X) is linear and explicit, are constant for every ship design, and five dynamic con-
a CDF can be calculated and differentiated to obtain a straints, which change based on other DP attributes.
PDF. Figure 1 shows the JPDF of two independent stan-
dard normal variables. The MPP, the exact LSF of a typi- 4.1 Objectives
cal response function Z, and the MV approximated LSF of The two objectives originally considered in this ship
Z are plotted. The approximate LSF is used to define the design optimization problem are to minimize total owner-
infeasible region of the design space, which has been cut ship cost (CTOC) and to maximize effectiveness. To
away from the JPDF, leaving the feasible region and ap- these, maximization of COS is added as a third objective.
propriate probability density only. The CTOC is weight-based with consideration of inflation
In the calculation of CoS, this procedure is followed and other financial effects, [Good and Brown 2006].
 Table 1: Surface Combatant Design Constraints reflect realistic design choices.
Parameter Constraint Type Boundary The CoS computation is divided into two distinct
Ata Total Available Minimum Dynamic steps, a numerical differentiation calculation to obtain the
Area required criteria responses and a probability calculation
Ada Available Minimum Dynamic that trims away the infeasible portion of the JPDF. The
Deckhouse Area optimization process is accelerated by setting CoS equal
Vs Sustained Speed Minimum 28 knots to zero for all deterministically infeasible DPs, eliminat-
Pbpengend Brake Propulsion Minimum Dynamic ing the need for excessive CoS calculations. Although it is
Power at Endurance possible that a deterministically infeasible DP has a rea-
Speed sonable CoS, it is unlikely that any decision maker would
choose a deterministically infeasible design. All nine con-
KWg Generator Electric Minimum Dynamic
straints and both the cost and overall effectiveness objec-
Power
tives are used as criteria for the CoS evaluation. The cost
Cgmbmin Minimum GM to Minimum 0.05
and effectiveness limit state functions are defined such
Beam Ratio
that designs with costs exceeding a specified percentage
Cgmbmax Maximum GM to Maximum 0.15 of the deterministic CTOC are excluded as are designs
Beam Ratio with effectiveness less than another specified percentage
D10 Hull Depth at Minimum Dynamic of the deterministic OMOE.
Station 10
E Endurance Range Minimum 3500 nm 5. INTRODUCING UNCERTAINTY
Randomness is assumed to arise only from the em-
Effectiveness is quantified by a parameter called the bedded uncertainties in the analysis; all input variables are
Overall Measure of Effectiveness (OMOE). OMOE is a presumed deterministic for each DP. Uncertainties within
quantitative measure of a ship’s ability to perform over a the analysis arise mainly from the parametric equations
range of mission types and mission scenarios, [Good and located in the primary ship synthesis model modules. Un-
Brown 2006]. OMOE ranges from 0.0 to 1.0, representing certainties were considered in the resistance, weight, cen-
the least effective ship design possible in the given design ter of gravity, endurance range, required area and electric-
space and the most effective ship design possible. In prac- al load calculations. Each uncertain variable Y is modeled
tice, however, there may not be a feasible design with an through the use of an extra random variable X:
OMOE close to 1.0. Definition of the OMOE is structured Y   y  (1   x  X )
as a multi-attribute decision problem, and a Multi-
Attribute Value function approach is used to develop a where X Є N(0,1) and is independent of any other RV, μy
measure spanning several mission types. Specific meas- is the mean value (i.e., the deterministically calculated
ures of performance (MOPs) are developed and values of value) of the uncertain variable, and δx is the coefficient
performance (VOPs) are set to specify the value of a par- of variation of the uncertain variable.
ticular MOP to a specific mission area for a particular The resistance calculation is based on the Holtrop-
mission type. USN defense policy and goals, projected Mennen regression equation [Holtrop 1984] with a wave
operational environment and threat, mission need, model- resistance correction [Good 2006]. This calculation was
ing and simulation, war gaming results, and expert opi- verified by comparison to match runs of the USN’s tool
nion are only some of the inputs that must be considered for early-stage ship design, ASSET. Match run outputs
when developing an OMOE metric. MOPs are determined have been previously adjusted to accurately represent the
based on DVs and required operational capabilities. Goal actual characteristics of particular ships. Match runs were
and threshold values or options are identified for each obtained for Arleigh Burke Class Flight I and Flight IIa
MOP, and MOP weights are calculated using pair-wise Destroyers, the Oliver Hazard Perry Class Frigate and a
comparison and expert opinion. The dot product of the Ticonderoga Class Cruiser. The standard deviation of
MOP weight vector and a specific VOP vector generates total resistance was found to be 7.98% and thus δ for un-
an OMOE value for a particular ship design, [Brown and certain resistance was set to 0.0798.
Thomas 1998]; The Weight and Stability Module evaluates Ship
Work Breakdown Structure (SWBS) weight groups to
OMOE   wiVOPi ( MOPi ) calculate ship weight. Each SWBS weight group calcula-
i
tion is the sum of smaller subsystem weights calculated
It should also be noted that 9 of the 20 MOPs used in using parametric equations, or a parametric equation it-
this work originate directly from combat system selection, self. This fact combined with weight’s significant impact
and they represent almost 77% of the OMOE metric. A on the weight-based cost algorithm makes weight a criti-
weight-based cost algorithm does not adequately reflect cal uncertainty to analyze. Three RVs are used to model
the cost variation of designs with various combat system weight uncertainties as shown in Table 2. These three
combinations but is the only cost algorithm available to coefficients of variation were calculated from the same set
this study. Consequently, the results shown below should of ASSET match runs. Also, it should be noted that in the
be considered illustrative only. They may not accurately first model execution involving only two RVs,
 Table 2: SWBS Weight Group RVs Table 4: CoS Calculation compared to MCS for DDG 51
Random δ Value Description Criterion MCS (%) CoS (%) Error (%)
Variable CTOC 84.25 84.40 0.15
W16 9.94% SWBS 100 & 600 OMOE 88.67 84.05 -4.62
weight groups Ata 77.83 78.30 0.47
W235 5.74% SWBS 200, 300, & 500 Ada 100.00 99.99 -0.01
weight groups Vs 100.00 99.99 -0.01
W47 9.65% SWBS 400 & 700 Pbpengend 100.00 99.99 -0.01
weight groups KWg 100.00 99.99 -0.01
Table 3: Other Sources of Uncertainty Cgmbmin 99.89 99.88 -0.01
Random δ Value Description Cgmbmax 100.00 99.99 -0.01
Variable D10 100.00 99.99 -0.01
Atr 10.00% total required area E 75.73 75.60 -0.13
E 10.00% endurance range CoS 53.56 53.89 0.33
KG 3.65% center of gravity
KWmflm 13.20% maximum functional deviation departure from the deterministic value
electric load represented a 2.74% departure from the deterministic val-
ue while a 1.58% departure from the deterministic value
of the OMOE represented two standard deviations. These
total lightship weight is summarized in one RV with a
percentage departures are used to set the LSFs for CTOC
coefficient of variation of 10%.
and OMOE. A full calculation of CoS was performed for
Although the two most significant sources of uncer-
the DDG 51 DP and the results are compared to a 100,000
tainty are bare hull resistance and weight, other critical
sample Monte Carlo Simulation (MCS) in Table 4. Per-
sources of uncertainty are embedded within the ship syn-
centages represent the probability of feasibility for each
thesis model. The most important are those that have a
specific criterion. As expected, the OMOE error is the
direct impact on the CTOC and OMOE objectives as well
largest but the error in the overall CoS calculation is only
as the nine design constraints. A few of the calculated
0.33%.
parameters that directly affect the nine design constraints
This agreement can be visualized by overlaying the
are identified in Table 3. Coefficients of variation were
MCS distribution with the MV approximated LSFs as in
calculated based on ASSET match runs where data was
Figure 2. Figure 2 shows the two-dimensional RV sub-
available or assumed to be 10%.
space spanned by the standard normal RVs. The bare hull
resistance and total weight means are located at zero, and
6. COS VALIDATION
each unit away from the mean represents one standard
Properties of the CoS calculation are examined con-
deviation. This design subspace is scattered with the solu-
sidering only the two most important sources of uncer-
tions produced from the MCS; the feasible solutions are
tainty, resistance and weight. For these calculations, the
marked with a black ‘x’ and the infeasible solutions are
weight uncertainties are combined into a single RV
marked with a gray ‘+.’ Those approximated constraint
representing total lightship weight. The MV method used
and objective LSFs from the CoS calculation that lie in
to calculate CoS will work best for responses that vary
slowly with the uncertain variations of the RVs locally 4
within the design space. Criteria response behavior was x - feasible
investigated at a selection of feasible DPs and was found 3 + - infeasible
to be nearly linear over the range of interest for most of
2
the criteria that depend on the RVs. Exceptions were
OMOE, which was discontinuous due to its stepwise de- 1 CTOC
finition and two continuous but non-linear responses; sus-
Weight

E
OMOE
tained speed and endurance range. While the first order 0
series expansion of the MV is expected to have significant -1 Ata
problems approximating a response surface to the discon-
tinuous OMOE, it was found that the OMOE objective -2
has a very low standard deviation and that, within one
-3
standard deviation of the deterministic mean, the OMOE Cgmbmin
response is well behaved for most DPs. It was also found -4
that the LSF for OMOE is rarely in effect, that is, the -4 -2 0 2 4
Resistance
OMOE LSF rarely eliminates any designs. As a result, the
OMOE approximation error was judged to be acceptable. Figure 2: Monte Carlo simulation determination of fea-
An evaluation of the variability of CTOC and OMOE sibility vs. Mean Value approximation of limit state
at the DDG 51 DP showed that, for CTOC, a one standard functions. Also shown are the approximations of limit
state functions for CTOC and OMOE.
the design subspace domain have been superimposed on
the scatter plot. The approximated constraint LSFs follow
the edges of the infeasible regions very well, and only the
E and Ata LSFs are active. The two approximated objec-
tive LSFs are also superimposed on the scatter plot to
show the region used for the actual CoS percentage calcu-
lation. The OMOE LSF is slightly active in this regard.
The CTOC LSF is not a factor in the calculation. These
approximated LSFs are determined from only three model
evaluations.

7. MULTI-OBJECTIVE GENETIC OPTIMIZA-

TION (MOGO) CALCULATION
The first example calculation done with the ship syn-
thesis model and optimizer described above uses only the
resistance and the single lightship weight uncertainties.
The analysis was performed on a PC with a 2.66 GHz
CPU and 1.0 GB of RAM. The optimizer terminated after Figure 3: Three-dimensional scatter plot of Pareto op-
46.5 hours or 166 generations due to lack of improve- timal design solutions. CoS obtained assuming uncertain
ment. A total of 49,800 individual design variable combi- bare hull resistance and lightship weight. Choice of
nations were evaluated of which, 43,037 were determinis- “Best DP” is explained in the text. Points lying closer to
tically infeasible and 6,763 were deterministically feasi- the foreground are shown larger.
ble. With three model evaluations being required for the
feasible DPs, for which the CoS calculation was made, 51 as well as an increase in CoS.
the total number of design evaluations was 63,326. This An important feature of Figure 3 are the points lying
represents a 27% increase in the number of model evalua- along the CoS = 0 plane. These are DPs that are slightly
tions over an optimization considering the same number deterministically infeasible, i.e., the deterministic value of
of DPs without the CoS calculation, however, due to one or more of the constraint criteria values falls slightly
overhead savings in the extra CoS model evaluations, the outside of the constraint boundary. However, because the
time increase was only 18%. optimizer adds a 5% tolerance to these boundaries and
The calculation resulted in a three-dimensional non- these DPs do not exceed the constraint values by an
dominated frontier made up of 225 Pareto optimal design amount that is greater than that tolerance, they are treated
solutions. While it is difficult to visualize the front in a as feasible by the optimizer and retained in the optimiza-
static two dimensional plot, it is useful to present such a tion. On the other hand, the CoS calculation procedure
figure here. Figure 3 shows the Pareto optimal designs in does not employ such a tolerance, identifies the DP as
the three-dimensional objective space along with a desig- deterministically infeasible, and does not perform the CoS
nation of a “Best DP”. Of course, the designer’s choice of calculation but sets CoS = 0 to accelerate the optimization
design from among the Pareto optimal DPs is highly sub- process. These DPs do not actually have a zero chance of
jective and may be based on a number of unquantifiable success but the actual CoS values are small. Attention is
preferences. Our designation of a “Best DP” means only drawn to these points here but they will be discussed in
that this is the point that yields the highest objective ratios detail in the next section.
for all possible objective combinations concurrently; Figure 3 is very hard to interpret in printed form, and
OMOE/CTOC, CoS/CTOC, and OMOE·CoS. Combining a surface fitted to the Pareto front may be misleading and
these yields the definition of a quantity we are calling hard to translate into specific design solutions. The set of
Fitness; designs on the Pareto front does not form a continuous
 OMOE  CoS  Table 5: Objective values of selected design solutions
Fitness     1000
 CTOC ($ M )  Solution CTOC($M) OMOE CoS
Lowest 720.7 0.2420 0.3052
Our “Best DP” is thus the point among the Pareto op- CTOC DP
timal set that has the greatest Fitness. Table 5 gives the Highest 1364.9 0.7921 0.2331
objective values of the best individual objective points, OMOE
the “Best DP”, and the DDG 51 DP. The DDG 51 DP has DP
a relatively high OMOE due to its emphasis on combat Highest 1001.5 0.3525 0.8246
systems. While the DDG 51 DP lies on this particular CoS DP
Pareto front, subsequent optimization runs did find other
“Best” DP 1081.2 0.6928 0.6833
combinations of DVs that gave performance increases in
(see text)
all three objectives over the DDG 51. Many other discov-
DDG 51 1455.8 0.7685 0.4276
ered designs offer significant cost savings over the DDG
DP
 0.8
"Best DP"
0.7 0.9
"Knee" 4
0.6 0.8
"Knee" 1
0.5
"Knee" 2 0.7
"Knee" 3
OMOE .

0.4 0.6

OMOE
0.3 CoS DP
High (CoS > 0.70) 0.5

0.2
Medium High (0.60 < CoS < 0.70) 0.4
0.1 0.3 2 Objective Deterministic Pareto Front
Medium (0.50 < CoS < 0.60)

0
0.2 3 Objective Deterministically Infeasible
700 800 900 1000 1100 1200 1300 1400
CTOC ($M)
(CoS = 0) Front
0.1

Figure 4: Pareto optimal designs in two-dimensional 0

CTOC and OMOE space with deterministically infeasi- 600 700 800 900 1000 1100 1200 1300
ble and low CoS DPs removed. Possibly attractive CTOC ($M)
points at knees are indicated.
Figure 5: 2-D objective space comparison of the two-
surface since many of the DVs are discrete. Points along a objective deterministic Pareto front with the three-
surface fitted to the Pareto front do not always correspond objective deterministically infeasible (CoS = 0) front.
to a feasible design solution; only the actual design solu-
tions in the objective space have a feasible set of DVs in dimensional Pareto front with the deterministically infeas-
the design space. Figure 4, a scatter plot of the Pareto ible (CoS = 0) DPs of Figure 3.
optimal design solutions in two-dimensional CTOC and The difference in the fronts reflects a real difference
OMOE space, with CoS indicated by marker, may be eas- in the numerical details of the two optimizations. In the
ier to interpret. All design solutions with a CoS less than two-objective deterministic optimization, the entire popu-
50% were removed from this plot because it is unlikely lation is pushed to the constraint boundaries as they
that any designer would be interested in a solution that represent the most desirable designs. In the three-
would most likely not meet requirements and/or perfor- objective optimization considering uncertainty through
mance objectives. (Note that this includes the DDG 51 DP the CoS calculation, those deterministically infeasible
in this model.) It is also unlikely that any designer would designs found which lie close to constraint boundaries are
be interested in a deterministically infeasible solution and flagged by having CoS set to zero but they are not excep-
these have been removed as well. The advantage of this tionally desirable to the optimization process since it is
two-dimensional scatter plot is that the Pareto fronts at seeking designs of high CoS as well as desirable values of
different CoS levels can be visualized. This method of CTOC and OMOE. The optimizer thus does not probe the
examining alternative design choices presents a number constraint boundaries to the extent done in the two-
of attractive designs to the designer for further analysis. dimensional deterministic optimization but rather encou-
rages the population to spread out into regions of higher
8. THE COS = 0 DESIGNS CoS. Thus the deterministic optimization finds better de-
Perhaps the most important point of this paper is to signs in terms of CTOC and OMOE but it turns out that
be made by considering the DPs lying along the CoS = 0 all but 2 of the 41 Pareto optimal DPs found by the de-
plane in Figure 3. The process of setting CoS = 0 for de- terministic optimization are actually deterministically
terministically infeasible DPs (as discussed in the pre- infeasible in the sense of the CoS calculation. An evalua-
vious section) in conjunction with the optimization, effec- tion of the CoS of these 41 DPs was carried out and the
tively sorts out those DPs lying along constraint bounda- results are shown in the histogram of Figure 6.
ries and having desirable levels of CTOC and OMOE. From these results, it can be seen that an optimization
Because objective values usually improve as constraint performed without consideration of the inherent uncer-
boundaries are approached, this is precisely the type of tainty in the associated simulation pushes the resulting
points sought by a deterministic two-objective optimiza- optimal design(s) to the extremes of constraint boundaries
tion considering only CTOC and OMOE. This line of where there may actually be little chance of successfully
points in the CoS = 0 plane of Figure 3 closely approx- meeting the design criteria due to analysis uncertanty.
imates the Pareto front that would be generated by a de-
terministic optimization process with only CTOC and 9. INCREASING THE NUMBER OF UNCERTAIN
OMOE as objectives. Such a deterministic optimization QUANTITIES
process was created for this work by simply eliminating The CoS computation described above was extended
the CoS objective from the three-objective optimization. to accommodate three and four uncertain quantities at a
The remainder of the optimization problem formulation time. Runs were made considering those sources of uncer-
was left unchanged. Figure 5 compares the resulting two- tainty previously identified in various combinations. The
 than wise choices if those uncertainties are not quantified
18 and accurately represented in the analysis.
16
14 References
12 Brown, A.; and M. Thomas, 1998, “Reengineering the
Frequency

10 Naval Ship Concept Design Process,” From Research to

8 Reality in Ship Systems Engineering Symp., ASNE, Ty-
6 sons Corner, VA.
4
2 Good, N. 2006, “Multi-Objective Design Optimization
0 Considering Uncertainty in a Multi-Disciplinary Ship
0 0.05 0.1 0.15 0.2 0.25 Synthesis Model,” M.S. Thesis, Virginia Polytechnic In-
stitute and State University, Blacksburg, VA.
CoS
Good, N. and A. Brown, 2006 “Multi-Objective Concept
Figure 6: CoS values of the two-objective deterministic Design of an Advanced Logistics Delivery System Ship
Pareto optimal DPs in Figure 5 (ALDV),” ASNE Joint Sea Basing Symposium, Arling-
ton, VA.
extension of the Taylor series expansion to n RVs re-
quires n+1 model evaluations at each DP. The extension Haldar, A. and S. Mahadevan, 1999, Probability, Reli-
of the numerical integration of the JPDF over n dimen- ability and Statistical Methods in Engineering Design,
sions quickly becomes the limiting factor as n increases. ISBN 0471331198, John Wiley & Sons, New York.
The limitation of the 1 GB RAM capacity of the PC being Handscomb, D.C. and J.M. Hammersley, 1964, Monte
used for this effort precluded calculations of more than 4 Carlo Methods, Chapman and Hall, London.
uncertain quantities at a time. The resolution of this inte-
gration needed to be reduced in order to reach four uncer- Holtrop, J. 1984, “A Statistical Re-analysis of Resistance
tain quantities. The computational times required for the and Propulsion Data,” Int. Shipbuilding Progress Vol. 31,
three and four uncertain quantity cases were 31% and No. 363, pp.272-276.
100% greater than that required for a deterministic calcu-
lation evaluating the same number of DPs, respectively. Koch, P.N. and J.I. Cofer, 2000, “Simulation-Based De-
The exponential increase in both computational time and sign in the Face of Uncertainty,” Joint Army-Navy-
memory required means that further extensions of n will NASA-Air Force (JANNAF) Interagency Propulsion
require significant computational resources. Committee, 2nd Modeling and Simulation Subcommittee
Mtg., Destin, FL.
10. CONCLUSIONS Mavris, D.N., G.C. Mantis, and M.R. Kirby, 1997,
A method of treating uncertainty in a simulation “Demonstration of a Probabilistic Technique for the De-
model and its implementation in a multi-disciplinary, termination of Aircraft Economic Viability,” AIAA-1997-
multi-objective ship design optimization has been demon- 5585, AIAA and SAE World Aviation Congress, Ana-
strated. The technique, which is based on the Mean Value heim, CA.
Method is shown to be well suited to this application,
enabling an accurate evaluation of the confidence of Parsons, M.G. and R.L. Scott, 2004, “Formulation of mul-
meeting all design criteria, for a given design solution, ticriterion design optimization problems for solution with
using only n+1 model evaluations where n is the number scalar numerical optimization methods,” J.Ship Research
of uncertain quantities being considered. Treating the 48/1, pp.61-76.
confidence of success as an additional objective in a de- Stepanchick, J. and A. Brown, 2006, “Revisiting
sign optimization can give a designer significant insight DDGX/DDG-51 Concept Exploration,” ASNE Day, Ar-
not afforded by deterministic analyses. Further, it was lington, VA.
demonstrated that deterministic optimizations drive opti-
mal solutions to the extremes of constraint boundaries Thacker, B.H., D.P. Nicolella, S. Kumaresan, N.
where uncertainties in the model calculations will likely Yoganandan, and F.A. Pintar, 2001, “Probabilistic Finite
mean that these solutions have a small chance of actually Element Analysis of the Human Lower Cervical Spine,”
meeting the imposed design criteria. While extension of J. Mathematical Modeling and Scientific Computing
this technique to consider more than five or six uncertain 13/1-2, pp.12-21.
quantities simultaneously will require significant compu-
tational resources, those resources are readily available Wu, Y.T., H.R. Millwater, and T.A. Cruse, 1990 “Ad-
and the additional information that would be provided to vanced probabilistic structural analysis method for im-
decision makers is well worth the investment for design plicit performance functions,” AIAA Journal 28/9,
projects of any significant magnitude. It is the opinion of pp.1663-1669.
the authors that optimization of simulations involving
uncertainties may actually lead the designer to make less

View publication stats