Professional Documents
Culture Documents
KEYWORDS: Optimization, Pareto front, TOPSIS, Desirability function, Design of experiments, Tactical missile design
Concept design requiring complicated feed-forward and feed-back processes makes it difficult for engineers to determine the global
behaviors of design variables and objective functions in a design space. Although design of experiments and response surface models
have been applied to overcome these problems, the design variables satisfying the objective functions can't be found due to violations
of given constraints. In this study, a new optimization process, i.e., the GEO (Generate, Explore and Optimize) process for the concept
design of a tactical missile, based on a MOGA (Multi-Objective Genetic Algorithm) was proposed, which was first adapted to
generate a Pareto Front in order to simultaneously satisfy the constraints and the objective functions. In the first step, the weights
between the objective functions were determined by using an AHP (Analytic Hierarchy Process). Then, the design space exploration
followed, and was implemented by surrogate models constructed from the Pareto Front with a neural network. In the last step, a
TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) and a desirability function were applied together to
determine the optimal solution. The TOPSIS merged the multi-objective design problem to a single entity, and the desirability function
normalized each objective function.
Manuscript received: July 10, 2013 / Revised: March 5, 2014 / Accepted: March 19, 2014
user baseline systems that are used to perform trade studies. The rocket Fig. 4. Pareto alternatives presented by a Parallel Plot of six design
is based on the sparrow MRAAM (Medium-Range Air-to-Air Missile) variables and six responses
while the air-breathing baseline is based on the ramjet powered
ASALM (Advanced Strategic Air-Launched Missile). In the case of the
MRAAM, six design variables including three geometric and three
performance design variables and six responses are considered, as
shown in Tables 1 and 2.10
3.1 Generation
3.1.1 Formulation of multi-objective genetic algorithm
Before executing a MOGA, an AHP (Analytic Hierarchy Process)5
was used to define the importance of the objective functions as that has
been defined by expert knowledge. Rather than prescribing a “correct”
decision, the AHP helps decision makers find a choice that best suits
the goal and understanding of the problem. The AHP provides a
comprehensive and rational framework for structuring an approach to
Fig. 5 Distribution and correlation analysis of six responses
solving a problem, for representing and quantifying the related elements,
for relating those elements to overall goals, and for evaluating alternative
solutions. help them decide if a selected alternative is satisfied within a given
MOGA formulation has four objective functions, which are design range.
maximization of the final range (FinalRange), maximization of the Based on the distribution and correlation analysis of the 1,280
maximum velocity (Max_Velocity), minimization of the final time Pareto alternatives, as shown in Fig. 5, information from the descriptive
(FinalTime), and minimization of the horizontal turning radius statistics for six responses was gathered and response space was
(HorizTurnRadius); and MOGA has three constraints, which are a explored for the optimization formulation.
lower specification of the FinalRange, a lower specification of the Fig. 5 also shows the trade-offs between the responses based on the
Max_Velocity, and an upper specification of the HorizTurnRadius, as correlations found among the 1,280 Pareto alternatives. There were
shown in Table 3. very strong trade-offs found between the FinalRange and the
FinalTime. Since the FinalRange always degenerated when the
3.1.2 Pareto front FinalTime was minimized, an optimal solution would set the FinalTime
With the MOGA formulation shown in Table 3, a total of 1,280 to a specification that would be higher than the lower value of the
Pareto alternatives were generated. With the help of the Parallel Plot FinalRange. The Pareto Front that fell between that of the FinalTime
shown in Fig. 4, all of the Pareto alternatives can be shown in one and the Max_Velocity is shown in Fig. 6, which accounts for the trade-
figure. Among them, the thick line shows one alternative, which is the off between the FinalRange and the FinalTime. By considering three
maximum FinalTime. Because the plot shows the position for six constraints of the MRAAM, 46 optimal candidates were selected and
design variables and six responses, engineers have graphic evidence to displayed. The final optimal solution should be among these 46
1374 / JULY 2014 INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 15, No. 7
3.2 Exploration
3.2.1 Neural network modeling
An artificial neural network (NN) is a computational model for
storing and retrieving acquired knowledge. The artificial NN has been
studied for many years in the hope of imitating the human brain's ability
to solve problems that are ambiguous and that require a large amount
of processing. Human brains accomplish this data processing by utilizing
massive parallelism between millions of neurons working together to
solve complicated problems.11 Similarly, the NN models consist of many
computational elements, called “neurons” that correspond to their
biological counterparts operating in parallel and connected by links with
variable weights. The NNs consist of dense interconnected computing
units that are simple models of the complex neurons in biological
systems. Knowledge is acquired during the learning process and is
stored in the synaptic weights of the inter-nodal connections. These
weights adapt during the training process, most commonly through the
back-propagation algorithm, by presenting the neural network with
examples of input-output pairs exhibiting the relationship that the
network is attempting to learn. The main advantage of neural networks
Fig. 8 Profiler plot of NN models for a MRAAM between six design
is their ability to represent complex input and output relationships.12
variables and six responses
In the present study, two hidden layers and three nodes were
determined for the modeling of a MRAAM from 1,280 Pareto
alternatives, as shown in Fig. 7.13 Among 1,280 Pareto ones, 853 data Wing_Area affects only the HorizTurnRadius, and the Sustain_Pc acts
was used for training the neural network models and 427 data was used on only the Sustainlsp, but the Booster_Pc and the Expand_Ratio
for validating the neural network ones. The accuracy of the modeling influence both the Boostlsp and the Max_Velocity. The profiler plot
is summarized in Table 4. In order to validate the NN models of a shows that the Diameter and the Nose_Length are the most important
MRAAM, the deterministic coefficients, R2adj, for training and factors among the four objective functions and three constraints. The
validating were calculated. The Max_Velocity shows the worst value of Diameter and the Nose_Length also demonstrate the trade-off between
R2adj, which was 0.9987124, which equates to 99.87124% and exceeds the FinalRange and the FinalTime. Those facts show that the optimal
99%. Because the worst value of R2adj exceeded 99%, all the NN solution can be determined by the proper adjustment of the Diameter
models were appropriate for use in the optimization process. and the Nose_Length.
+ -
Sj = Max ( wj yij ), Sj = Min ( wj yij ),
(2)
i = 1, 2, …, M, j = 1, 2, …, N
where w means the weight of the criterion, x means the score of
alternatives, y means the normalized score of alternatives, M is the
number of alternatives, and N is the number of the criterion.
Second step: For the first alternative, find the difference between its
score in the first criterion and the best value, S+, for that criterion and Fig. 10 Optimal solution of MRAAM from TOPSIS between six
square this difference. Next, find the difference between the score of design variables and six responses
this first alternative and the second criterion and its best S+ and square
the difference. Repeat the process for all criteria. When completed, add
all these squared values and extract the square root of this sum, this is of C was sufficient for use in the optimization process. The profile plot
the Euclidean Metric, and it is the solution for the first alternative. of C as shown in Fig. 9 shows that the Diameter and the Nose_Length
Repeat the procedure for all alternatives. Then, repeat the procedure but were the most effective design variables, which was similar to the
now consider the worst values, S-. As a final result there will be a tendency of the response for the FinalTime. This was because the
matrix with columns that are alternatives and rows that are the best, D+, weight of the response for the FinalTime in Table 3 was larger than for
and worst distances, D-. any of the other three responses. In order to maximize C, the Diameter
should be maximized, and then the Nose_Length should be minimized.
N 2
+ +
Di = ∑ (Si – wj yij) , j = 1, 2 , …, M (3)
j=1
3.3.3 TOPSIS optimization
With the help of the desirability function, the optimal solution was
N
- - 2
Di = ∑ (wj yij – Si ) , j = 1, 2, …, M (4) achieved from maximizing a closeness index C and satisfying
j=1
constraints, as shown in Fig. 10.15 The optimal solution satisfied whole
Third step: Compute a closeness index, C, with a range that is three constraints, where the HorizTurnRadius should be less than
between 0 and 1, by using the following formula for each alternative. 4,000, the Max_Velocity should be larger than 2,400, and the
The alternative with the largest C is the first to be calculated, followed FinalRange should exceed 10. In order to verify the optimal solution
by the others in decreasing C order. from the NN models, the TMD was simulated using the same design
- variables as the NN models shown in Table 5. The largest difference
Di
Ci = ----------------
+
-- , i = 1, 2, …, M (5) between the NN models and the TMD occurred in the FinalRange with
Di + Di a difference of only 0.083%, so that the use of the approximate NN
models was verified. The optimal solutions were compared with the
3.3.2 Neural network model initial values. The initial solution of the Max_Velocity was violated
The deterministic coefficients, R2adj, of the closeness index, C, were because it was below 2,400, but the optimal solution satisfied the
0.9998177 and 0.9997673, respectively, for training and validating. constraint because it exceeded 2,400. Because the main objective
Because the worst value for R2adj of C exceeded 99%, the NN model function was to minimize the FinalTime, it should have improved the
1376 / JULY 2014 INTERNATIONAL JOURNAL OF PRECISION ENGINEERING AND MANUFACTURING Vol. 15, No. 7
7. Lee, K. K., Lee, K. H., and Han, S. H., “Use of an Orthogonal Array
4. Conclusions based on the Kriging Model to Maximize the Fatigue Life of a
Turbine Blade,” International Journal of Structural Integrity, Vol. 2,
A process for the optimization of concept design was proposed by No. 3, pp. 303-312, 2011.
using a Pareto Front and TOPSIS together with a neural network
8. Eugene, L. F., “Tactical Missile Design,” AIAA Education Series,
algorithm and desirability function. The obtained results are as follows:
AIAA, pp. 1-18, 2001.
(1) Neural network models of a MRAAM, generated from Pareto
sets, were verified as accurate for exploring and optimizing the design 9. Lee, K. K., Han, J. W., and Han, S. H., “TOPSIS-Based Multi-
space of a MRAAM, because the worst deterministic coefficient was Objective Shape Optimization for a CRT Funnel,” Transactions of
only 0.9987, which meant that neural network models represent the Korean Society of Mechanical Engineers A, Vol. 35, No. 7, pp.
99.87% of the TMD simulation. 729-736, 2011.
(2) By using the suggested GEO (Generate, Explore and Optimize)
10. Frits, A. P., Fleeman, E. L., and Mavris, D. N., “Use of a Conceptual
process, optimal solutions were calculated, where the Pareto sets were
Sizing Tool for Conceptual Design of Tactical Missiles (U),” Proc.
generated via the MOGA. The obtained optimal solutions were verified
of the Missile Sciences Conference, 2002.
via the simulation results from the TMD analysis. The optimal
solutions provided good agreement with the simulated results from FE 11. Ramu, M., Raja, V. P., Thyla, P., and Gunaseelan, M., “Design
analysis to within an error range of 0.083%. Optimization of Complex Structures Using Metamodels,” Jordan
(3) With maximizing the closeness index C which translates multi- Journal of Mechanical & Industrial Engineering, Vol. 4, No. 5, pp.
objectives into single objective, the objective function FinalTime was 653-663, 2010.
improved 23.19% and whole constraints were satisfied within the lower
12. Hamza, R. M. A., “Multi-Objective Neural Network Modeling for
and upper specifications.
Improving Stud Arc Welding Process Joining,” Journal of Engineering
Science and Technology, Vol. 6, No. 3, pp. 382-391, 2011.