Professional Documents
Culture Documents
Dr. Alan Brown, Virginia Tech and LCDR Mark Thomas, USN
2
Expert Opinion
Mission M&S
Mission Need Engineering Process /
Statement (MNS) Analysis LCC Model Schedule
Determine M&S Model
Mission Need Concept
Studies Zigzag down hierarchies
Approval
Constraints
R&D and
Exploratory
Design Define "possibilities" as function
of technology risk.
Proposed Design Process Framework threat and future mission scenarios consistent with
defense planning guidance. It determines joint force
It is helpful to think of the design process as a and naval force structure, identifies future mission
sequential mapping between four domains as shown deficiencies, and estimates the force level, mix and
in Figure 3: 1) the mission or customer domain; 2) mission performance required of future ships.
the functional domain; 3) the physical domain; and 4)
S h ip
the process domain (Suh, 1990). Decisions made in
each domain are mapped into the subsequent domain, M ission /
C u stom er
F u n ction al P hysical P roc ess/
B u ild S trateg y
moving from “what” to “how” in each mapping, and W arfare E n clo se an d p ro te ct S truc tu re M a x L ift W eig h t
M ob ility M a x D im en s ion s
then zigzagging down hierarchies in each domain as S u s tain ab ility
V u ln era bility
K eep ou t sea H u llform P late cu rvatu re
C atalog of S h ap es
C on trol floo din g S h ell
the design is defined in increasing detail. Customer S u s cep tab ility
R eliab ility
P artitio n volum e
P ro vid e d eck s h elter
W atertig h t Tbh d
D eck s
B lock size
N u clear/n on -n u clear
F lexibility
requirements (CRs) are the “what” for the functional P ro tect
P ro vid e m o bility
Tan ks
D eck h ou se
P rop ulsion
domain, functional requirements (FRs) are the P rovid e s tab le p latform
P rod u ce F & A Th ru s t M ain en g in es
“how”. FRs are the “what” for the physical domain, S u s tain
R ep len ish
Red u c tion g ear
S ha ftin g
M an eu ver P rop eller
design parameters (DPs) are the “how”. DPs are the P ro vid e elec tric p ower Con trol s ys tem
S teerin g
“what” for the process domain, process variables G en erate
Tra ns m it
E le ctric P ow er
3
threat, required operational capabilities (ROCs), In this research, a simple ship-synthesis model is
projected operational environment (POE), mission used to synthesize and balance designs in the physical
scenarios and MOEs. Also important are war- domain, and to calculate the first level of ship MOPs.
fighting model validation and sensitivity analysis in- Balance requires that physical and functional con-
cluding the identification of critical ship MOPs. straints are satisfied. A genetic algorithm (GA) is
Research, development, technology demonstra- used to search DP space in the physical domain, and
tion, and exploratory design identify, develop and to generate and identify concepts on the non-
evaluate new systems, technology, and ship concepts. dominated objective attribute frontier. More sophis-
Together with proven systems and concepts these ticated tools, models and simulations can be used
pieces must be assembled into balanced, feasible, and later in the design process on selected concepts to re-
cost-effective ship designs that must satisfy the fine the designs, demonstrate feasibility, and improve
evolving mission need in an evolving environment. MOP calculations. Analysis results are added to a
Products of R&D and Exploratory Design include design knowledge base and applied to update model
new technology and systems, ship concepts, and a parametric equations, MOP, cost and risk calcula-
preliminary definition of feasible ship and system tions. This provides a dynamic landscape or envi-
performance. Expected performance is described ronment for the genetic algorithm over the course of
using upper and lower bounds on MOPs. These the design process. The updated non-dominated
bounds are a function of technology risk that may be frontier is used to reevaluate and adjust earlier top
quantified and included as a system attribute. level DP decisions during the design process until
Mission Domain. Customer needs and require- further design changes are no longer cost-effective.
ments are determined, and effectiveness is assessed in Process Domain. Critical design parameters
the Mission Domain. Initial inputs to this domain must be related to process variables (PVs). These
are the MNS, POE, threat definition and mission sce- process variables may be synthesized in a build strat-
narios. Customer requirements (CRs) are the output. egy. The build process must be refined at each level
In this research, CRs are specified in three ways: 1) of the design hierarchy to insure feasibility and
required operational capabilities (ROCs); 2) per- maximize producibility. PVs effect cost and risk.
formance constraints, goals and thresholds; and 3) an
overall mission measure of effectiveness (OMOE) in- Building the OMOE Function
dex which defines mission effectiveness as a function
of ship MOPs. The definition of a quantitative rela- Early in the design process, designers and engi-
tionship between mission effectiveness and ship per- neers require a working model to quantify operators’
formance in the mission domain is an essential pre- and policy-makers’ definition of mission effective-
requisite to a disciplined search of design parameters ness, and define its functional relationship to ship
in the physical domain. Modeling, simulation and MOPs. This quantitative assessment of effective-
expert opinion are used to define the OMOE as a ness is fundamental to a structured optimization pro-
function of MOPs. Some MOPs are binary, the ship cess.
has a capability or it doesn't. Others are continuous, There are a number of inputs which must be in-
such as sustained speed or endurance. tegrated when determining overall mission effective-
Functional Domain. In the Functional Domain, ness: 1) defense policy and goals; 2) threat; 3) exist-
top level functional requirements might include en- ing force structure; 4) mission need; 5) mission sce-
close and protect, provide mobility, support and fight. narios; 6) modeling and simulation or war gaming re-
Consideration of functional requirements is most im- sults; and 7) expert opinion. Ideally, all knowledge
portant when new design solutions or technologies about the problem could be included in a master war-
are sought and when paradigms are reevaluated. gaming model to predict resulting measures of effec-
Physical Domain. Design parameters define the tiveness for a matrix of MOP inputs in a series of
ship in the Physical Domain. The selection, synthe- probabilistic scenarios. Regression analysis could be
sis and balance of DPs determine ship MOPs, and applied to the results to define a mathematical rela-
ultimately determine mission effectiveness. Cost and tionship between input ship MOPs and output MOEs.
risk are determined as a function of DPs and process The accuracy of such a simulation depends on mod-
variables (PVs). eling the detailed interactions of a complex human
4
and physical system and its response to a broad range hierarchy levels to an assessment of relative overall
of quantitative and qualitative variables and condi- system behavior or performance.
tions including ship MOPs. Many of the inputs and The first step in building an AHP hierarchy is to
responses are probabilistic so a statistically signifi- identify critical attributes affecting the decision or
cant number of full simulations must be made for system behavior. The level of detail of these attrib-
each set of discrete input variables. This extensive utes depends on the decision being made. These at-
modeling capability does not yet exist for practical tributes are then organized into a hierarchy structure
applications. that follows a logical breakdown or categorization as
An alternative to modeling and simulation is to shown in Figure 5. System options or alternatives
use expert opinion directly to integrate these diverse comprise the bottom hierarchy level.
inputs, and assess the value or utility of ship MOPs Next, the relative influence of each attribute on
in an OMOE function. This can be structured as a system performance and attribute values for each al-
multi-attribute decision problem. Two methods for ternative must be estimated. Saaty recommends a
structuring these problems dominate the literature: nine level dominance scale for the pair-wise compari-
Multi-Attribute Utility Theory (MAUT, Keeney and son of attribute influence on higher level attributes.
Raiffa, 1976) and the Analytical Hierarchy Process This results in a “ratio scale” comparison of attrib-
(AHP, Saaty, 1996). In the past, supporters of these utes. Pair-wise comparison or cardinal values may
theories have been critical of each other, but recently be used to assign attribute values for each alternative.
there have been efforts to identify similarities and Pair-wise comparison generates more information
blend the best of both for application in Multi- than is necessary with individual absolute measure-
Attribute Value (MAV) functions (Belton, 1986). ments or estimates. The AHP synthesizes and evalu-
This approach is adapted here for deriving an ates the consistency of this redundant information and
OMOE. calculates best-fit relative values.
The analytical hierarchy process (AHP) is a tool Although the AHP was developed primarily for
developed by Saaty (1996) for solving multi-attribute comparison of management alternatives, it has also
decision problems. It uses a hierarchical structure to proven to be a robust method for application in
abstract, decompose, organize and control the com- MAVT (Belton, 1986). The AHP provides a struc-
plexity of decisions involving many attributes, and it tured method for deriving an additive weighted value
uses informed judgment or expert opinion to measure function, and by careful application can also be used
the relative value or contribution of these attributes to derive non-linear attribute value or utility without
and synthesize a solution. Pair-wise comparison and the more cumbersome lottery comparison approach.
a maximum eigenvalue approach extract and quantify The OMOE function must include all important
this relative value. The method allows and measures effectiveness/performance attributes, both discrete
inconsistency in value measurement, and is able to and continuous, and ultimately be used to assess an
consider quantitative and qualitative attributes. unlimited number of ship alternatives. Successful
A hierarchy is a simplified abstraction of the application AHP/MAVT to this problem requires a
structure of a system used to study and capture the very structured and disciplined process as follows:
functional interactions of its attributes, and their im- 1. Identify, define and bound decision attrib-
pact on total system behavior or performance. It is utes. Identify critical mission scenarios. Identify
based on the assumption that important system enti- MOE(s) for each mission scenario. Establish goals
ties or attributes, which must first be identified, can and thresholds for all MOEs. Identify ship MOPs
be grouped into sets, with the entities of one group or critical to mission scenario MOE assessment and
level influencing the entities of the neighboring group consistent with the current design hierarchy level.
or level. One can conceptualize a hierarchy as a Set goals and thresholds for these MOPs.
bottoms-up synthesis of influence on the top level be- 2. Build OMOE/MOP hierarchy. Organize
havior of a system, or as the top down distribution of MOEs and MOPs into a hierarchy as shown in Fig-
influence of top level behavior to low level attributes. ure 5, with specific ship MOPs at the lowest level.
Alternatives are compared in terms of the lowest level Association with the performance of a discrete sys-
attributes and this comparison is rolled up through tem may define some MOPs. Others are continuous
performance variables such as sustained speed.
5
3. Determine MOP value and hierarchy the ship design, primarily the hull form and installed
weighting factors. Use expert opinion and pair-wise power. The threshold for this MOP is 26 knots, and
comparison to determine MOP value and the quanti- the goal is 32 knots. Pair-wise comparison is ac-
tative relationship between the OMOE and MOPs. complished for discrete values of speed at one knot
increments, and a value function is fit to these results
O verall M ission
E ffectiven ess to calculate intermediate values. Again, the pair-wise
O MOE
comparison is structured to compare the relative
SA G
SA G OMOE 1
A R G E sc ort
ARG OMO E2
M C M E sc ort
MCM OMO E3
value of MOP options to achieve a particular MOE
O rd nan ce on O rdn ance on # S h ips Left # L and ing C raft
(Ordnance on target, etc.) in a specific scenario.
In lan d Targets S hore T arg ets w /M in D eg radation R eac hed B each
MOE1 M OE2 MO E3 MOE4
Value
M ission S u stainab ilit y M ob ilit y V uln erab ility S u scept ab ility 1.000
A A W -M O P 1 W eap on s-M O P 9 S p eed -M O P 12 P S ink /H it - M O P 1 5 I R S ig nitu re - M O P 17 0.900
A S U W -M O P 2 R an ge-M O P 10 S eakeepin g-M O P 1 3 R edu nd ancy - M O P 1 6 A coust ic S ig natu re
A S W -M O P 3 D u rat ion -M O P 11 R eliab ility - M O P 14 - MO P1 8 0.800
C 4 I-M O P 4 H u ll R C S - M O P 1 9
M C M -M O P 5 T opsid e R C S - M O P 20 0.700
N S F S -M O P 6
S E W -M O P 7
S TK -M O P 8
0.600
1.000 0.200
0.900 0.100
0.800 0.000
0.700 26 27 28 29 30 31 32
0.600
Value
6
students and faculty, and specifically for use with a the basic process shown in Figure 9.
genetic algorithm (GA) by Shahak (1998). It follows
MOP
MOP17-IR Signature
MOP18-Acoustic Signature
MOP16-Redundancy
MOP19-Hull RCS
MOP12-Sustained Speed
MOP11-Stores Duration
MOP8-STK
MOP20-Topside RCS
MOP15-CBR
MOP5-MC M
MOP13-Seakeeping
MOP10-Range
MOP7-SEW
MOP14-Reliability
MOP3-ASW
MOP9-W eapons
MOP4-C4I
MOP2-ASUW
MOP1-AAW
MOP6-NSFS
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Weight
Figure 8. Measure of Performance (MOP) Weights
Input Design
Estimate Full
Calculate (LCC) are calculated. The GA uses these results to
Start Load Weight Resistance
Principle
Parameters and Deckhouse
Volume
Characteristics
and Power assess fitness and breed the next generation of ship
No variants.
Balance requires that physical and functional
Feasible? Weight and Area and
Converge? Stability Volume
Tankage
constraints are satisfied. The ship must float. It
Yes must have adequate stability, volume, area, electric
power, etc. It must provide required capabilities and
Risk Cost MOP's OMOE
satisfy minimum thresholds for performance. The
Figure 9. Ship Synthesis Model Process ship synthesis model uses regression-based equations
for weight, volume, area and electric power. Resis-
Most recently modules have been added to inter- tance is calculated using Gertler/Taylor Standard Se-
face with a payload database, and calculate acquisi- ries (Gertler, 1954). Cost is calculated using a modi-
tion cost, seakeeping index and effectiveness. It has fied weight-based algorithm. LCC as defined for this
also been updated to be more consistent with regres- analysis includes only follow-ship acquisition cost,
sion-based equations for weight, area and electric life cycle fuel cost and life cycle manning cost.
power as they have evolved in ASSET (DTMB, Seakeeping is assessed using the McCreight Index.
1990). The current version is self-balancing, and
written in FORTRAN 90. Cp Cx C∆L CBT CD10
In the GA application of this synthesis model, in- 0.61 0.82 80 2.9 11.1
put design parameters (genes) are specified in a ship CRD Cmanning AAW ASUW ASW
design matrix (chromosome). An example is shown 0.2 0.5 1 2 1
in Figure 10. Design parameter descriptions are C4I MCM NSFS SEW Weapons
1 4 1 1 1
listed in Table 1. Specific payload systems with Range Stores Shafts CPS ICR/GT
weight, area and power requirements are associated 3 2 2 1 1
with each payload description. The ship is balanced Figure 10. Design Parameter Chromosome
and resulting MOPs, OMOE, and Life Cycle Cost
7
Table 1. Design Parameter Descriptions
In the simplified design problem presented in this
paper, each design variant requires 12.5 seconds on a
Design Parameter Description 200 MHz PC to balance and evaluate. An exhaus-
1 - Prismatic Coefficient (Cp) 0.5-0.7; 20 increments
2 - Maximum Section Coefficient 0.7-0.9; 20 increments tive search would assess over ten trillion variants re-
(Cx) quiring over 4 million years on this machine! Ran-
3 - Displacement to Length Ratio 60.0-90.0; 15increments
(C)L)
dom search does not require a closed-form solution
4 - Beam to Draft Ratio (CBT) 2.8-3.7; 9 increments and has advantages of simplicity and insensitivity to
5 - Length to Depth Ratio (CD10) 10.0-15.0; 10 increments discontinuities, but it still requires many iterations,
6 - Raised Deck Ratio (CRD) 0.0-0.4; 4 increments
7 - Manning Factor (CManning) 0.5-1.0; 5 increments
and is computationally impractical for a large design
8 - AAW Payload 1 - Theater TBMD problem. Exponential random search improves the
2 - Area TBMD efficiency of random search, but also requires many
3 - Area Defense
4 - Limited Area Defense concept iterations. Genetic algorithms (GA) offer
5 - Self Defense great promise to tackle this difficult problem.
9 - ASUW Payload 1 - Long Range
2 - Medium Range
Twenty-five years ago John Holland developed a
3 - Short Range genetic algorithm to abstract and explain the adaptive
4 - Self Defense processes of natural systems for use in the design of
10 - ASW Payload 1 - Area Domonance
2 - Adverse ASW Environment artificial systems (Goldberg, 1989). Since that time,
3 - Good ASW Environment this algorithm has been applied with success to a
4 - Torpedo Defense
11 - C4I Payload 1 - Advanced wide range of problems. Genetic algorithms use
2 - Current models of natural selection, reproduction and muta-
12 - MCM Payload 1 - Limited Clearance tion to improve a population of individuals or vari-
2 - Mine Recon
3 - Mine Avoidance ants based on the “survival of the fittest”, or in the
4 - Limited Mine Advoidance case of Pareto Genetic Algorithms (PGAs), based on
13 - NSFS Payload 1 - Advanced (VGAS,
NATACMS, ATWCS) the dominance and distribution of variants (Thomas,
2 - Full 1998). GAs are ideally suited to optimizing discon-
3 - Medium
4 - Minimum tinuous and disjointed functions, and to optimization
14 - SEW Payload 1 - Advanced where no closed-form function exists (or no mathe-
2 - Current matical function at all, as with experimental data).
15 - Weapons Capacity (VLS) 1 - 128 cells
2 - 64 cells The robustness of a particular GA depends on its ex-
3 - 32 cells ploration and efficiency qualities. Exploration refers
16 – Range or fuel capacity 1 - 10000 nm
2 - 7000 nm to its ability to master the design space and consis-
3 - 5000 nm tently identify the global optima. Efficiency refers to
4 - 4000 nm
the effort required to identify the global optima. Ro-
17 - Stores Duration 1 - 60 days
2 - 45 days bustness implies an effective balance between these
18 - Shafts 1 or 2 qualities. Genetic algorithms are very robust relative
19 - CPS 0 (none) or 1 (full)
20 - ICR or GT 0 (ICR) or 1 (LM2500)
to other methods.
Figure 11 illustrates the PGA process used in this
application. An initial population of 200 variants is
Design Optimization created by random selection of design variables
within the design space specified in Table 1. A
Ship design optimization is not a new concept, chromosome with 20 design parameters as in Figure
but it poses difficult problems (Leopold, 1965, Man- 10 represents each variant. The ship variants defined
del, 1966 and Mandel, 1972). As discussed previ- by these chromosomes are balanced, and evaluated
ously, ship design space is non-linear, very discon- using the ship synthesis model resulting in an assess-
tinuous, and bounded by a variety of constraints and ment of feasibility and objective values (OMOE and
thresholds. These attributes inhibit effective applica- LCC) for each variant. Fitness indicates a variant’s
tion of mature gradient-based optimization techniques relative dominance in the population based on OMOE
including Lagrange multipliers, steepest ascent meth- and LCC. A Goldberg (1989) ranking scheme is
ods, linear programming, non-linear programming used. Variants are sorted into layers of Pareto-
and dynamic programming. dominance. Each layer contains all variants that are
dominant to subsequent layers. Variants are sorted
8
with the best variant in the highest layer getting a Once a surviving population is selected, 25 per-
rank of one and the poorest in the lowest layer a rank cent of these are chosen in pairs at random for cross-
of 200 (population size). A geometrically decreas- over. A cut is made at a random location in the
ing probability of selection is assigned to each variant chromosomes of each pair. Design parameters below
based on its rank. Equivalent variants (same domi- the cut are swapped between the parents producing
nance layer) are ordered randomly within their layer, new variants. A small percentage of individual de-
probabilities are averaged for variants in the same sign parameters (genes) in the selected variants are
layer, and the same average value is ultimately as- chosen randomly to mutate. In mutation, the value of
signed to each. Variants are penalized for infeasibil- a single design parameter is replaced with a new
ity and for similarity to other variants. This mini- value chosen at random. After these operations are
mizes niches and duplicate variants, and forces the completed, new variants in the new population are
selection to spread out over the objective frontier. evaluated and the process cycles until convergence.
Solution space
Each cycle defines a new generation.
boundaries Compared to other optimization techniques, ge-
netic algorithms are particularly well suited for gen-
Other input erating a Pareto-optimal frontier because they im-
No Forced
prove the fitness of a population of concepts simulta-
Initialization
Method neously. By penalizing for niching, this population
par.in data file Yes
DDG_forced
data file
spreads out over non-dominated values of the objec-
Random
initialization
tive attributes, and ultimately defines the Pareto-
optimal frontier.
gen=1,gen_max
Ship Synthesis
Surface Combatant Design
Self-balanced Model
A simple surface-combatant concept design is
Yes No
Penalty? Feasible? used to evaluate and demonstrate this design method-
Feasibility Penalty No Yes ology. A Mission Need Statement, Required Opera-
tional Capabilities (ROCs) and mission scenarios
Sort into Dominance Layers / Niche were developed for a notional Land Attack Surface
Penalty
Combatant. A group of warfare experts consisting of
Selection MIT naval officer students and faculty were briefed
on this mission need, and an effectiveness hierarchy
Crossover
similar to Figure 5 was developed. Measures of Ef-
fectiveness were identified for each of three critical
Mutation
mission scenarios: 1) Surface Action Group (SAG);
No 2) Amphibious Readiness Group (ARG) escort; and
Convergence?
Yes
3) Mine Counter Measures (MCM) Group escort.
Report Ship MOPs were identified for each MOE. Goals
and end
and thresholds were set for all MOEs and MOPs.
Figure 11. Pareto Genetic Algorithm (PGA) Process The AHP was applied using questionnaires and group
discussion. AHP results were used to build an
Once selection probabilities have been assigned, OMOE function. Typical MOP value functions and
selection can proceed. A roulette wheel is con- MOP weights are shown in Figures 6 through 8.
structed with 200 segments representing each variant. Once the OMOE function was determined, criti-
The area of each segment is equal to the variant se- cal design parameters to be varied were selected, and
lection probability. Baker’s (1987) selection method structured as a GA chromosome. These DPs in-
is used. This method “spins” 200 (population size) cluded hull form parameters, payload packages to
equally space markers once (vice spinning one achieve various levels of warfare area performance,
marker 200 times) to select 200 variants (some mul- ship endurance, number of shafts, use of a Collective
tiple times) for survival and reproduction. Protection System (CPS), propulsion gas turbine se-
lection, and a manpower reduction factor.
9
Table 2. Select Non-dominated Feasible Variants
Ship A Ship B Ship C Ship D
LBP (ft) 422.8 440.4 520.9 588.5
Beam (ft) 52.5 59.9 55.5 75.0
D10 (ft) 36.8 40.0 35.9 52.6
Draft (ft) 17.5 21.2 19.8 24.2
Displacement (lton) 5524.2 6713.8 8899.2 16612
Shafts 1 1 2 2
Range (nm) 4000 5000 7000 10000
Sustained Speed (knt) 27.1 27.0 31.0 29.2
GM/B .108 .129 .101 .111
Generators/kW 3/3000 3/3000 3/3000 7/3000
CPS no yes full full
AAW RAM,SPS- ESSM,SM,SPS-49,X-Band ra- ESSM,SM,SPY-1D, X- ESSM,SM,SPY-1D, X and S Band
49,CIWS,SDS dar, Mk92 MFCS Band radar,Mk99 GMFCS radar, GMFCS
ASUW 5”/54 w/ERGM,GFCS Harpoon, 5”/54 Harpoon,AN/SWG- TASM/TMMM,
w/ERGM,GFCS 1,VGAS,GFCS ATWCS,VGAS,GFCS
ASW 1.5m sonar, SSTD, helo 5m sonar(passive), SSTD, 5m sonar, SSTD, LAMPS 5m sonar, SSTD, LAMPS MK3,
haven, SVTT,NIXIE NIXIE, SVTT,helo haven,VLA MKIII,NIXIE,SVTT,VLA NIXIE,SVTT,VLA,SQQ-89,
LBVDS
C4I Baseline Baseline CEC,JTIDS, digital comm, CEC, JTIDS, digital comm, TA-
TADIX/TACINTEL DIX/TACINTEL
MCM Degaussing Degaussing Mine avoidance sonar, de- Mine avoidance sonar, Remote
gaussing Minehunting System, degaussing
NSFS N-ATACMS, 5”/54 N-ATACMS, 5”/54 VGAS, N-ATACMS, VGAS, N-ATACMS, ATWCS
w/ERGM,GFCS w/ERGM,GFCS ATWCS
STK TWCS,TLAMs,UAVs TWCS,TLAMs,UAVs ATWCS,TLAMs,UAVs ATWCS,TLAMs,UAVs
SEW SLQ-32V2, DLS SLQ32V2,DLS AIEWS,DLS AIEWS,DLS
VLS cells 32 64 64 128
Hello hangar / helos 0 0 Yes/2 Yes/2
Crew 108 120 184 266
Follow ship Acquisi- 547.6 656.6 888.6 1242.3
tion Cost ($M)
LCC ($M) 628.3 943.5 1319.7 1867.5
OMOE 0.2296 .4627 .7523 .9562
Ship manning was calculated based on historical objective attributes. The optimization was run for 50
data and reduced by a manpower reduction factor. generations, taking 26 hours on a 200 MHz PC.
Weight and cost for automation was increased con- LCC was defined to include acquisition cost, dis-
sistent with the manpower reduction. Weight, area counted life cycle fuel cost and discounted life cycle
and power were specified for each payload package. manpower cost. Results of this search are presented
Chromosome DPs are described in Table 1. Other in Figure 12. The data points represent LCC and
DP’s were held constant for all designs. OMOE values for a sample of feasible variants in
generations 1, 40 and 50. Generation 1 is a random
OMOE selection of design parameters. Convergence to a
1 non-dominated frontier can be seen in the evolution
Ship C Ship D
0.9 from Generation 1 to Generation 40 and finally to
0.8 Generation 50. Generation 50 results approximate
0.7 Ship B
the design frontier. Many of these points are non-
0.6
Ship A dominated.
0.5
Ships A, B, C and D are non-dominated and rep-
0.4
0.3
resent “knees in the curve”. Characteristics for these
Gen 50
0.2 ships are listed in Table 2. Variants between Ships C
Gen 40
0.1 Gen 1 and D on the frontier have 2 shafts. Variants below
0 Ship B on the frontier have one shaft. Between Ships
0.5 1 1.5 2 2.5 3 B and C, there is a mix of one and two shaft ships.
Life Cycle Cost ($B) Ship D is the feasible ship with the highest OMOE.
Although non-dominated, none of these ships can
Figure 12. OMOE/LCC Frontier
be identified as “the best”. Selection of the preferred
A PGA search was completed for the design
design is up to the customer, but Figure 12 provides
space specified in Table 1 using OMOE and LCC as
10
the customer with important information to make this • It provides a practical method for the ship
selection: 1) the engineer can assure the customer designer to calculate an Overall Measure
with confidence that non-dominated variants have of Effectiveness (OMOE) which represents
been identified; 2) the non-dominated frontier pro- customer requirements and relates ship
vides a perspective on the design space; and 3) some measures of performance (MOPs) to mis-
variants stand out as providing good value given a sion effectiveness. This is an essential pre-
range of acceptable cost. In this example, Ships A, requisite to a disciplined search of design
B, C and D are noteworthy. parameters.
A discussion with the customer might consider • It includes an efficient method to search
the following: design space for non-dominated concepts.
• Ship A represents a low-end alternative. It has • It provides a consistent format for pre-
good performance in Naval Surface Fire Support senting and trading off a manageable set of
(NSFS) with other MOPs at threshold values. It dissimilar objective attributes (effective-
represents the best alternative if acquisition cost is ness, cost, and risk).
limited to $500-600M. Most war fighters would not
be impressed, but it is the best for the money. Essential future work includes:
• Ship D is the high-end variant. It achieves • Integrate with a design database and prod-
goal performance in most MOPs. The big cost driver uct model.
for ships to the right of Ship C is the addition of an • Extend to later stages of design using a dy-
S-band radar. This radar has a large ship impact due namic landscape PGA.
to its power (seven 3000 kW generators) and cooling • Validate the AHP/MAVT approach to de-
requirements. Ship D would be an excellent candi- fine and calculate an OMOE.
date for an Integrated Power System (IPS), an option • Integrate with a process/shipbuilding
which is not included in the present model. As a me- model.
chanical drive ship, it is only worth the money if full • Extend to include risk as a third objective
TBMD capability is essential. attribute.
• Ship B is the most effective of the single shaft • Evaluate existing ship designs using this
ships at a reasonable price. Its AAW and ASW sys- methodology.
tems are much more capable than Ship A and the ac- • Compare the GA performance to other op-
quisition cost is still low. It is an excellent choice for timization techniques.
a reasonable low-end capability.
• Ship C stands out as a particularly sharp knee The methodology described in this paper does not
in the curve. It has excellent capability in all areas replace imagination and experience. It provides a
and the price is comparable to a number of variants practical tool to manage a complex total-system
with much less effectiveness. It is a “Best Buy”, and problem that cannot be managed by experience and
if the acquisition cost is manageable, this would be intuition alone. It represents essential change in how
an excellent choice. we do naval ship concept design.
Conclusions References
It is estimated that more than 80 percent of a na- Baker, J.E. (1987), “Reducing Bias and Inefficiency
val ship’s ultimate acquisition cost is locked in during in the Selection Algorithm”, pp. 14-21, Proceed-
concept design. For a class of ships, this means tens ings of the Second International Conference on
of billions of dollars. An “ad hoc” process for mak- Genetic Algorithms, Hillsdale, NJ.
ing these critical design decisions is not adequate. Belton, V. (1986), “A comparison of the analytic hi-
Figure 11 appears to be a simple and somewhat in- erarchy process and a simple multi-attribute value
tuitive result, but it is not, and its implications are function”, European Journal of Operational Re-
revolutionary. Without this kind of information, we search.
cannot make responsible decisions. DTMB (1990), ASSET/MONOSC User Manual,
Key elements addressed by this methodology are: Version 3.3, DTMB, Carderock Division, Naval
Surface Warfare Center.
11
Gertler, Morton (1954), “A Reanalysis of the Origi- tions at the International Maritime Organization
nal Test Data for the Taylor Series”, Report 806, (IMO) in tanker design, oil outflow, intact stability,
Navy Department, DTMB. damaged stability and tanker risk. He is chairman
Goldberg, D.E. (1989), Genetic Algorithms in of the SNAME Ad Hoc Panel on Structural Design
Search, Optimization and Machine Learning, and Response in Collision and Grounding, a mem-
Reading: Addison-Wesley Publishing Company, ber of the SNAME Ship Design Committee and
Inc.. SNAME Panel O-44, Marine Safety and Pollution
Keeney, R.L. and Raiffa, H. (1976), Decisions with Prevention. He is Northeast Regional Vice Presi-
Multiple Objectives: Preferences and Value dent of SNAME, a member of the ASNE Council
Tradeoffs, John Wiley and Sons, New York. and Past Chairman of the New England Section of
Kramer, R. et. al. (1996), "Virtual Notational Ship SNAME. He received a PhD in Marine Engineer-
Simulation Supporting SC-21 Design", Modeling, ing from MIT in 1986.
Simulation and Virtual Prototyping Conference,
ASNE. LCDR Mark Thomas, USN graduated from Okla-
Hockberger, W.A. (1996), "Total System Ship De- homa State University with a bachelor's degree in
sign in a Supersystem Framework", Naval Engi- electrical engineering. His first Navy assignment
neers Journal. was aboard USS David R. Ray (DD-971) where he
Leopold, R. (1965), Mathematical Optimization served as First Lieutenant, Gunnery Officer, and
Methods Applied to Ship Design, MIT Department Auxiliaries Officer. After four years, he returned to
of Naval Architecture and Marine Engineering, Surface Warfare Officers School (SWOS)
Report 65-8. Coronado as an instructor and course director for
Mandel, P., and Leopold, R. (1966), "Optimization the DD-963 Engineering Officer of the Watch
Methods Applied to Ship Design", SNAME Trans- (EOOW) School. He transferred to the ED com-
actions, Vol. 74. munity in 1991, and served two years as a main
Mandel, P., and Chryssostomidis, C. (1972), "A De- propulsion inspector at the Pacific Board of In-
sign Methodology for Ships and Other Complex spection and Survey. He reported to MIT in May
Systems", Philosphical Transactions, Royal Soci- 1993, and received the degrees of Naval Engineer
ety of London, A273. and Master of Science in Electrical Engineering
Saaty, T.L. (1996), The Analytic Hierarchy Process, and Computer Science in June 1996. He then began
RWS Publications, Pittsburgh. a two-year period of doctoral research and received
Shahak, S. (1998), “Naval Ship Concept Design: an the degree of Doctor of Philosophy in hydrody-
Evolutionary Approach”, Master of Science The- namics in June 1998. LCDR Thomas' next perma-
sis, MIT Department of Ocean Engineering. nent duty station will be at the Supervisor of Ship-
Suh, N.P. (1990), The Principles of Design, Oxford building, Pascagoula MS.
University Press, New York.
Thomas, Mark (1998), “A Pareto Frontier for Full
Stern Submarines via Genetic Algorithm” PhD
Thesis, MIT Department of Ocean Engineering.
Tibbitts, B. and Keane, R.G. (1995), "Making De-
sign Everybody's Job", Naval Engineers Journal.
12