Bhaskar 2000
Bhaskar 2000
ABSTRACT
1. INTRODUCTION
2
On leave from: Department of Chemical Engineering, Indian Institute of
Technology, Kanpur 208016, India. E-mail: [email protected]
To whom correspondence should be addressed; E-mail: [email protected]
apart from the major goal of achieving economic efficiency. Until a few years
ago, these several objective functions were combined into a single scalar
objective function, using arbitrary weight factors, so that the problem could
become computationally tractable. This 'scalarization' of what is really a
vector objective function suffers from several drawbacks. One is that the
results are sensitive to the values of the weighting factors used, which are
difficult to assign on an α-priori basis. What is even more important is the
less-recognized fact that there is a risk of losing some optimal solutions
(Chankong and Haimes, 1983; Haimes, 1977). This happens if the
non-convexity of the objective function (a set is said to be convex if the line
joining two points of the set lies within that set) gives rise to a duality gap.
[Numerically, a duality gap is the difference between the primal and the dual
objective values (Farber, 1986; Goicoechea et al., 1982). Dual is another
mathematical program with the property that its objective is always a bound
on the original mathematical program, called the primal. If the primal is
feasible, the dual cannot be unbounded, and vice versa; if the dual is feasible,
the primal cannot be unbounded. A dual provides a sufficiency test for
optimality, for if feasible χ and y can be found such that f(x) = F(y), it
follows that χ is optimal in the primal and y is optimal in the dual].
Therefore, it is necessary that a vector form of all the objective functions be
used in formulating and solving real-life optimization problems. Such
multiobjective optimization problems form the subject of this review.
The concept of multiobjective optimization is attributed to the economist,
Pareto (1896). After several decades, this concept was recognized in
operations research and has recently become popular in engineering.
Extensive literature is now available on multiobjective optimization but we
limit ourselves to problems of Chemical Engineering interest only.
We start with a discussion of a simple example of a multiobjective
optimization problem to illustrate some elementary concepts. Most
techniques of solving such problems comprise of two phases - an objective
phase, which is more precise and mathematical, followed by a subjective
phase, which is statistical in nature and involves intuition and interaction with
human beings (called decision-makers, DMs) who are familiar with the
industrial operation being studied. In the first phase, a set of several optimal
solutions is generated. This is referred to as the Pareto set. For example, a
typical two-objective function minimization problem can be represented
mathematically as
I,
Pareto-optimal set
I,' I,
possible functions could be used (which do not change the location of the
optima), and a simple and popular transformation is F; s 1/(1 +1,).
A classic example of a 2-dimensional Pareto set (two-objective function
problem) in Chemical Engineering is provided by an isothermal batch reactor
in which the following series reaction is taking place:
(2)
We are interested in maximizing both the yield and the selectivity of the
desired product, B, simultaneously. These are defined by
0
0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.05
Selectivity
Fig. 2: Pareto optimal solutions (QR) for the yield-selectivity maximization
problem, for the reaction, A— l —>B 2
—»C, in an isothermal
batch reactor. Feed is pure A, k( = 0.1, k2/ki = 0.1. PQ represents the
inferior set. Corresponding plot for time vs. selectivity also shown.
The generation of the Pareto set (QR in Figure 2) was quite easy, and was
done using qualitative arguments once the yield vs. selectivity plot (PQR in
8
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et αϊ. Reviews in Chemical Engineering
Table 1
Multiobjective Optimization in Chemical Engineering:
Based on Areas in Chemical Engineering
Design of chemical plants with robust dynamic operability Palazoglu and Arkun, 1986
characteristics
process
Table 1 continued
in a batch reactor
reactors
Dynamic optimization of a Nonvaporizing nylon 6 batch Wajge and Gupta, 1994
reactor
Optimization of an industrial semibatch nylon 6 batch reactor Sareen and Gupta, 199S
Free radical bulk polymerization reactor using genetic Garg and Gupta, 1999
algorithm
Industrial nylon 6 semibatch reactor system using genetic Gupta and Gupta, 1999
algorithm
10
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
K Bhaskar et al. Reviews in Chemical Engineering
Table 1 continued
in a bioreactor
Waste treatment costs in discrete process synthesis and Ciricand Huchette, 1993
optimization problems
MIXING
1991a
CATALYTIC REACTORS
ELECTROCHEMICAL PROCESS
acid
GENERAL
11
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
cess.
thod
lied
is
w
M
Technique used fo
decision making
ll
||||J 1
l Hi
•2 e ζ- αe·
•S §8
ο _ I «
Ά
Is
0 3
II
ls
il
5s
ε s
I|
; ·-
Z t,
12
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Λ
JO
«s
υ
o
T3
υ
V. Bhaskar et al.
Mi
_ 'S
S ts
° 's
we
S.-Π Μ
«—
™
is
la l
1. Maximization of co
o: > t a
SIJi
g^ii
2
2 S SS
J£S:S
si-s
ί Ss
i Mi
-2 2 ε s
molecular weight
Iι_ Iυ S«ο ρl
|
w
O
1. Maximization of co
molecular weight
Styrene-acrylonitrile syst
1. Maximization of co
molecular weight
Opt design o the Discretization techniques empli
Authenticated | 134.99.128.41
withrobustdynamic op circumvent [he problem of inflr
13
Brought to you by | Heinrich Heine Universität Düsseldorf
Vol. 16, No. 1, 2000 Applications ofMultiobjective Optimization
In Chemical Engineering
g 11 i ·=
3! S-i |:
!S|&f-g~
sii*u*
ΐΙ|ϊ·5ΐ|
li|5ll|
Iff§!£!
gJ g« 3 3g
I!
*·£
iε
1-2
Ξ «η
.3·
.Ξ
*j
c l
.2
JU i
IS
l
!|
li
i!
13
.ε <t
o
III
II
11!
II
14
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et αϊ. Reviews in Chemical Engineering
24*
a 3ο -β
Ο ·* u
<= €
5o —P.a
« ο
i
oill*
· S.J H o S
o.
J
f
•σ
\Λϊ Ι
S o ** ,S β
ω
*·«ΐ
Έ|ϊί £§·!
ug Σ Is .Ε 5 « !
ee
H
l
li
15
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
16
ι2
,C
Ξ
υ
T3
0
υ
3
"
.C
β>
0
-—
cu
(Λ
.C
ΐέ
^
ο
ώ
Ι
S
Ο
Equality constraint on the desired
Ν
Ι
I
1
if
ο
Ο
number average molecular weight ι
1
II
1ft
J-jj.2
lit
«|J
Stopping condition on the desired
•-ε
i -i -s
ι||
ϊ|!
conversion imposed. Smoothening
'S
•nil
1 »1 8
Tg
1=
the Pareto sets to yield suboptimal
Ko/. 7r5, No. 1. 2000
| Sι
_
.2 .a E .g >. § Ε
•e
.£
w
Ο
??
E
CO
C
as
ON
Ϊ
•S
ε
-5
I
eft
Equality constraint on the desired
η
^
i
1
-i S
ο—
ll
number average molecular weight ι
Η Isis t ll
£·!
Έ
.S 15 α
.S "5 £
**· c c
8«
Stopping condition on the desired
S S-8
8 -| §
ll!
-8a
.a
^^
.C
•β
M
E
O
n
J
a
j^i
Endpoint constraints on the numb«
i
1
1
Έ
If
If
?g
"So
0 <
ll
II SS
average chain length and the mono
Ί ^-,
-il
8.
,1
|ί
11
!
conversion incorporated. Unique
optimal solution obtained.
Ε
V
-H
^c
ο
CO
*i
a
j;
S
Equality constraint on the desired
SP
I
1
fa
ll
O2
number average molecular weight ι
Jif
Authenticated | 134.99.128.41
.Is«
if"
's § §
.lit
11.
Stopping condition on the desired
111
1 e1
|||
generated.
es
«S
.c
υ
T3
2
υ
f
.a
E
C
"c
Q
c
v
•s
'S
Ο
r
00
S
Mapping of the decision varial
J,
S
imization ofthe methane feed rate
S
1
1
I«
Ό Ο
3 ?
S ΐ
C ^j
done inside the NSGA code tc
<"
imization ofthe flowrate of CO in
zS.
[as for a fixed production rate of the bounds chromosome .§
c
Ogen in an existing side-fired steai speciflc(dcpending on the cho
.ε 9 ρ -5 £
rmer other decision variables).
ο
.a
Si
ε
C
<S
'c
m
c
I
Γ*
•3
NSGA adapted to use inequali
0
0
00
£·
s
°
imization ofthe residence time of
5
1
Ή
1
M
„
constraints on certain product
·§ α<
mer melt and the concentrations ο
!
«
£C„
isirable side products formed in th properties by "artificial killing
«trial continuous wiped film PET chromosomes in the process ο
?w^
.2
'S
tor.Equality constraint on the desil optimization. Unique optimal
£ £-3 -g a a
•J
.Ξ
.0
E
Γ»
c
'S
·§
0
S
c
&
c
S
<a
Additional bounds and over-ri
is
Limization ofthe overall collection
1
1
JS_ ">Ο
^
constraints needed to be impo;
Sz,
;iency and minimization ofthe pre
If
ΣΪ Ι
* IP e
sg
ι / overall cost ofthe cyclone sepai Bounds were made chromosol
specific.
.a
ο
1"»
(Λ
^
W
c
g
.··
•i
I
ee
limization ofthe percentage remo\ Procedure to develop useful d
8
a
1
1
Έ
•g Ο
charts representing trade-offs ι
1<
llcohol from beer and minimizatio
ζ ε.
J •5
!·-
«
U s f II
ϊ2*
and few other optimization pn
•s^-S
•emoval ofthe 'extract' (taste chem |
studied.
—
_c
ε
TS
-S
.2
g
e
g
s
3
00
s
s
limization ofthe monomer conven Constraints on the number avi
1
1
1
oOT
·§1<
o
g
fi it
a S=
imization ofthe length ofthe film imposed
S
£
tor in the continuous casting proce
E S£
5
vIA
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Reviews in Chemical Engineering
The utility function is optimized to obtain the preferred solution. The utility
function is an abstract variable indicating goal attainment. For a chemical
engineer, the objectives in a typical industrial setting could either be
quantifiable quantities, as, for example, minimization of the cost of
production while meeting specifications on the products and effluents, etc.,
or, a decision maker (DM) could define the utility to include intuitive
quantities, e.g., operability and controllability of the plant, reliability of the
process, etc. In this interpretation, the possibility of measuring the utility does
exist, though formal techniques are not available. Thus, the utility is a way of
allowing the decision maker to describe his goals in an abstract way. The
drawback of this technique is that the utility function, usually, cannot be
determined systematically (Haimes et al., 1975). No Chemical Engineering
examples have been solved using this method, to the best of our knowledge.
18
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
19
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16. No. I, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
(5)
20
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
21
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. I, 2000 Applications ofMulliobjective Optimization
In Chemical Engineering
6 batch reactor. They minimized the (i) reaction time, tf, and (ii)
concentration, [C2]f, of the undesired side product, the cyclic dimer, in the
product, while requiring the monomer conversion and the number average
chain length in the product to lie at design or desired values (xm.d and μη,α,
respectively). The temperature history, T(t), was used as the decision
variable. Pontryagin's minimum principle was used with the first-order
control vector iteration method (Ray and Szekely, 1973) to solve the problem
with the ε-equality constraint incorporated. These workers reported some
numerical oscillations in the Pareto sets. They also encountered convergence
problems when using larger values of ε. Sareen and Gupta, 1995, used the
same objective functions and end-point constraints (on the product) to
optimize an industrial semi-batch nylon 6 reactor. In this reactor, monomer
and water vaporize and build up the pressure in the vapor space above the
liquid reaction mass, at a rate depending on the release of the vapor mixture
through a control valve. They used the jacket-fluid temperature, Tj (a scalar
value), as well as the pressure history, p(t) (a continuous function of time, t),
in the reactor as the decision variables. The pressure history was
parametrized (shape fixed) and was represented in terms of five parameters
(constants, rather than a function). Sequential quadratic programming (SQP),
developed by Gill et αϊ., 1981 (which has a first-order convergence), was
used for solving the single objective function problem obtained with the ε-
equality constraint method. Some amount of scatter was obtained in the
Pareto set. A 'smoothening procedure' was used to get sub-optimal but
smoother Pareto sets. Substantial improvements in the operation of the
industrial reactor were indicated when operating the reactor at any of the
points on the Pareto set.
The inequality form in Eq. 6 is very useful when the DM wishes to solve
the problem interactively. The Kuhn-Tucker (positive) multipliers associated
with the optimization problem reflect the sensitivity of the primary objective,
Ii, to the active constrained objectives, Ij, and so give an indication of the
trade-offs among the objectives. The values of the Kuhn-Tucker multipliers
indicate the price that one needs to pay in terms of one objective, if one
wishes to improve another, while still remaining on the non-inferior set. The
drawback of this technique is that the test of non-inferiority needs to be
performed. This leads to an increase in the computational time. This method
also locates some of the points that are inferior (Chankong and Haimes,
1983; Clark and Westerberg, 1983). This method has been used extensively
(Sophos et σ/., 1980; Tsoukas et al., 1982; Fan et al., 1984; Farber, 1986;
22
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
Palazoglu and Arkun, 1986; Videla et al., 1990; Luyben and Floudas, 1994 a,
b) in Chemical Engineering.
Sophos et al., 1980, optimized a petrochemical industry using multiple
objectives. These were to optimize (i) the energy utilization (which included
maximization of the change in the thermodynamic availability and
minimization of lost work) and (ii) carbon utilization efficiency (which is
equivalent to minimization of the feed stock consumption). Compromise
'bounding structures' (optimal solutions of each objective) were generated
for the system studied. The two objectives given above were considered first
and a Pareto set was obtained between the feed consumption and the lost
work. Pareto surfaces were also generated for a three objective function
problem using linear programming. Two preferred solutions were identified
from the Pareto surface. In one case, the method used was goal programming
(discussed in Section 2.6) while in another, commensuration of the objective
functions for a specific case was used before solving a single objective
function optimization problem.
Tsoukas et al., 1982, were the first to introduce the concept of multi-
objective optimization in polymerization reactors, where the presence of
conflicting objectives is common. Pareto optimal solutions were determined
for the dual objectives of (i) narrowing the copolymer composition and (ii)
narrowing the molecular weight distribution for styrene-acrylonitrile (SAN)
copolymerization in a semi-batch reactor. Temperature, initiator con-
centration and monomer addition histories (functions of time) were used as
the decision variables. Farber, 1986, extended this work to continuous
copolymerization reactors. They considered both methyl methacrylate-vinyl
acetate (MMA-VA) as well as SAN systems. Pareto sets were obtained
[composition vs. molecular weight; conversion vs. molecular weight] for the
MMA-VA system using the temperature history of the reactor as the decision
variable. In the SAN system, a Pareto optimal set was obtained for
conversion vs. molecular weight by using both temperature history and
residence time as decision variables. Fan et al., 1984, also used this technique
for chain polymerizations in a batch reactor (the same as discussed in the
previous study) and compared the results with those obtained using the
parametric approach.
Palazoglu and Arkun (1986) studied the problems of operability
associated with improper design, modeling and control system design in the
presence of uncertainties. They proposed a process design methodology to
find optimal operating regions for chemical units. This approach guides the
23
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
24
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
The hybrid form does not require any non-inferiority tests and is, therefore,
more efficient when the user is interested in generating only the Pareto
optimal solutions numerically. Interactive decision making is not possible
with this method.
Though the ε-constraint technique can also be used to solve problems
which are non-convex, the difficulty with this method is that an a-priori
knowledge of the appropriate range of values of εί is required.
25
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
n
Minl = (8)
26
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
27
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
28
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
1 0 0 1 1 1 0 0 0 1 1 1 1 0 0 1 1 1 0 1 1 1 0 0
(9)
0 0 1 1 0 1 0 1 1 1 0 0 0 0 1 1 0 1 0 0 0 1 1 1
parent chromosomes daughter chromosomes
In Eq. 9, the crossover site for this pair of parent chromosomes is taken just
after the fifth binary digit. It is hoped that the daughter strings are superior. If
they are worse than the parent chromosomes, they will slowly die a natural
death over the next few generations (the Darwinian principle at work).
The mutation operator is required for the following reason. In Eq. 9, let us
assume that all the chromosomes in the gene pool have a 0 at the second
position. There is a finite probability of this happening, since the generation
of the binary numbers is done probabilistically. It is obvious that cross-over
will never be able to generate chromosomes with a 1 at this position, and if it
so happens that the optimum is, indeed, located at a point described by such a
chromosome, GA will be unable to reach this solution. The mutation operator
looks at each binary digit in every daughter chromosome in the gene pool,
and transforms a 0 into a 1 (or vice versa) with a small probability. In effect,
it moves the chromosome locally in the x-space, to create a better
chromosome. The entire process is repeated till some termination criterion is
met (the specified maximum number of generations is attained, or the
improvements in the values of the objective functions become lower than a
specified tolerance). A more elaborate description of SGA is available in
Holland, 1975, Goldberg, 1989, and Deb, 1995.
We now turn our attention to extensions of SGA to solve problems
involving multiobjective optimization. Since GA uses a population of points,
it seems very natural to use GA for such problems to capture a number of
solutions simultaneously. Schaffer (1984) was the first to apply an adapted
vector evaluated genetic algorithm (VEGA) to solve multiobjective
optimization problems. Though it was simple to implement, the method has a
bias towards some Pareto-optimal solutions (Goldberg, 1989; Srinivas and
Deb, 1995). In order to overcome this problem of bias with some of the
optimal solutions, Goldberg proposed a non-dominated sorting procedure.
29
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
The idea was implemented in different ways by Fonseca and Fleming (1993),
Horn et al. (1994) and Srinivas and Deb (1995). The algorithm implemented
by Srinivas and Deb (1995) is called the non-dominated sorting genetic
algorithm (NSGA). It is to be noted that NSGA overcomes the pitfalls of the
previous two techniques (Fonseca and Fleming, 1993, Horn et al. 1994).
Fonseca and Fleming (1998 a, b) have recently extended their algorithm,
applying the principles of niche and sharing. The concept of non-dominated
sorting is first discussed and then we turn our attention to the applications of
this technique in Chemical Engineering. It is interesting to note that most of
the recent applications in Chemical Engineering have used NSGA (adapted
as required by the problem).
NSGA uses a ranking selection method to emphasize the good points and
a niche method to create diversity in the population without losing a stable
sub-population of good points. Principally, NSGA differs from SGA only in
the way of selection. A check for non-dominance is first carried out among
all the chromosomes in the gene pool before reproduction is performed. All
the non-dominated chromosomes from the entire population are first
identified and assigned a front number (Front No. = 1). These non-dominated
chromosomes are then assigned a dummy fitness value (which is usually the
number of chromosomes, Np, but could be any other arbitrarily selected, large
value instead). The dummy fitness value of any chromosome in this front is
then modified according to a sharing procedure (Goldberg and Richardson,
1987; Deb, 1989; Deb and Goldberg, 1991) by dividing it by the niche count
of the chromosome. The niche count of a chromosome represents the number
of neighbors around it, with distant neighbors contributing less than those
nearby. The niche count, thus, gives an idea of how crowded the
chromosomes are in the x-space. This is obtained, say, for the ith
chromosome, by computing its distance, dy, from another, say, jth
chromosome in the x-space, and using a sharing function, Sh, as given below
d
Sh(djj) = l - (— —ü T , if dij<a , ;0, otherwise
$h re (10)
l CT shareJ
30
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
31
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
front =front + 1
Fig. 3: A flowchart describing NSGA (Mitra et al., 1998; Garg and Gupta,
1999).
32
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
values, xm.d and μ^, are attained for these. The first constraint was taken care
of by adding it in the form of a penalty function (Deb, 1995) to both the
objective functions. The objective functions were, thus, modified to
The second equality constraint, μη?Γ = μη><ι, was used as a stopping condition,
i.e., integration of the model equations, a set of ordinary differential
equations (ODEs), was stopped when this constraint was met. The decision
variables used in this study were (i) the rate of release, V-Kt), of the vapor
from the semi-batch reactor (a function of time, t) which influenced the
pressure in the reactor, and (ii) the jacket fluid temperature, Tj (a scalar).
NSGA had to be adapted to apply to decision variables that were continuous
functions of time. This was achieved by discretizing the continuous function
into several, equi-spaced (in time) scalar values, VTii; i = 1, 2, ..., Q, and
constraining the value, VTi, to lie within a small range of the previous value,
VTin. This ensured that the final continuous function, Vj(t) [obtained by
curve-fitting the digitized values] was implementable. Pareto-optimal
solutions were obtained for a specified value of the feed water concentration
(see Figure 4). These results were compared with the solutions obtained
earlier by Sareen and Gupta, 1995, who parameterized the pressure history,
and so introduced some artificial constraints into the problem (since the
shape of the pressure history was assumed to be fixed). Mitra et al. (1998)
found that the solutions obtained by NSGA were superior. Interestingly,
considerable improvement in the operation of the industrial reactor was
indicated by this study, and we understand that these results were
implemented on the industrial reactor.
Gupta and Gupta (1999) extended this work on the industrial nylon-6
reactor system to consider the multiobjective optimization of the
reactor-cum-control valve assembly. They considered the fractional-opening
of the control valve as one of the decision variables (again, a function of
time), instead of the rate of release of vapor from the reactor. The second
decision variable was the temperature of the jacket fluid, a scalar value. The
Pareto optimal solutions obtained for this system were found to be worse
when compared to the solutions obtained by Mitra et al., who had studied the
reactor alone (see Figure 4). This was because the operation of the control
valve, which released the vapors, excluded certain sets of values of V-r(t),
33
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. I, 2000 Applications ofMultiobjective Optimization
In Chemical Engineering
0.55
0.45
n
0.35
3.45
iNO.25
U
•—•0.15
CM
Ο
'3.45
0.05
0.40 0.50 0.60 0.70 0.80 0.90
VW
Fig. 4: Comparison of the Pareto solutions for an industrial semi-batch
nylon 6 reactor for two different choices of the decision variables
(Mitra et al., 1998; Gupta and Gupta, 1999). Subscript, ref, indicates
values being used in industry before changeover to near-optimal
conditions.
which were permitted in the study of Mitra et al. (1998). It is clear that for
industrial systems, the optimization of the entire system is more valuable than
that of its major parts (Aatmeeyata and Gupta, 1998).
Garg and Gupta, 1999, applied NSGA to the multiobjective optimization
of free radical bulk polymerization reactors, wherein diffusional effects (the
Trommsdorff, cage and glass effects) are manifested. The two objective
functions used were the minimization of (i) the total reaction time, tf, and (ii)
the polydispersity index, Qf, of the product. The manufacture of polymethyl
methacrylate in a batch reactor was chosen as the example system. Equality
constraints were used on the value of the number average chain length, μΠιη
and the monomer conversion, xm>r, in the final product, similar to that for the
34
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
35
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
26.0
- - FH2. WM
i-- i>«. «m-3450 Igftr
25.5
m
Ό
16.0
12.0
8.0
10 11 12 13 14 15
w~3\,
Fig. 5: Pareto sets for three different values of FH2,unit- Ii (minimized) =
Few.™, t (minimized) = 108/FCOiOUt. Point Q represents the current
operating conditions of the industrial reformer studied, with FH2,unit=
3450 kg/hr. P is an arbitrarily selected point on the Pareto for the
same value of FH2,unit, which is better than Q in terms of both the
objective functions (Rajesh et al., 2000).
36
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
K Bhaskar et al. Reviews in Chemical Engineering
37
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
cyclones, and seven geometric ratios (shape) of the cyclones. The bounds on
two of the geometric ratios were adapted (depending on the values selected
for some of the other decision variables, chromosome-specific bounds) to
ensure that the cyclone shape was physically meaningful. The technique used
was similar to what was used in the optimization of the steam reformer. An
upper limit was imposed on the inlet velocity of gases (an inequality
constraint). This constraint was implemented using penalty killing, as was
done for the PET wiped film reactor optimization problem. Δρ vs. η0 Pareto
sets were obtained (see Figure 6). It was found that the optimal values of the
decision variables could not, at first sight, be explained physically. This
problem is quite common in multiobjective optimization problems associated
with a considerable amount of 'freedom'. One then needs to solve several
ι ι ι
6500 -
5500
/
4500 ί9
03 i
^3500 ί
2500 f
1500 S
_
500
I I I
0.6 0.7 0.8 0.9 1.0
Fig. 6: Pareto set (Δρ vs. η0) for a train of identical cyclone separators in
parallel (Ravi el al., 2000a). 165mVs of dust-laden gases from a
paper mill are processed.
38
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
simpler multiobjective optimization problems, each with only one, or, at most
two decision variables, to find out which of the several decision variables in
the original problem are the most sensitive and important in deciding the
Pareto set. Such a study leads to considerable physical insight and, at times,
even helps in developing appropriate bounds of the decision variables, as
well as in selecting appropriate constraints to be used in solving the
multiobjective optimization problem, which may not be known too precisely
at the beginning. Such detailed studies also help in explaining the scatter that
may be present in the final results - both in the Pareto set obtained as well as
in the plots of the individual decision variables over the range of objective
functions covered by the Pareto set. A similar study is being carried out (Ravi
et al., 2000b) on the multiobjective optimization of venturi scrubbers. Pareto
optimal solutions are being obtained in the preliminary studies completed till
now.
Multiobjective optimization of hollow fiber membrane modules was
carried out by Chan et al. (2000), using the dialysis of beer to produce
low-alcohol beer, as an example. The two objective functions considered
were (i) the maximization of the removal of alcohol from the beer, and (ii)
the minimization of the removal of 'taste chemicals (called extract)' by the
module. Inequality constraints were imposed on the pressure drops on both
the shell and the tube sides, as well as on the inner diameter of the shell. The
decision variables considered were the flow rate of water on the shell side,
the length of the fiber, the fractional free area in the shell (related to the tube
pitch), the internal radius of the individual hollow fibers and the thickness of
the hollow fiber membranes. Again, this problem is associated with an
immense amount of freedom, and several simpler one- and two- decision
variable problems have been studied to develop physical insight. It was found
that the internal radius of the membrane was the most important decision
variable influencing the Pareto set.
Another application of considerable industrial importance is the
optimization of the continuous casting of polymethyl methacrylate (PMMA)
films. In this process, a prepolymer is first produced in an isothermal plug
flow tubular reactor (PFTR). The product from this reactor flows as a thin
film through a furnace. The temperature, Tw(z), of the upper and lower
surfaces of the polymer film varies with the axial location, z, in the furnace.
The two objective functions (Zhou et al., 2000) are (i) the maximization of
the cross section-average value of the monomer conversion at the end of the
furnace, xm.av.f, and (ii) the minimization of the length, Zf, of the furnace. The
39
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications ofMultiobjective Optimization
In Chemical Engineering
end-point constraint used was that the section-average value of the number
average chain length in the product, μη,»ν,Γ, should be equal to a desired value,
μη>(1. In addition, a 'local' constraint is to be satisfied in this problem. This
takes the form that the temperature at any point in the film must be below an
upper safe value, to prevent degradation (discoloration) of the polymer film.
The decision variables used were the temperature of the isothermal PFTR,
concentration of the initiator in the feed to the PFTR, monomer conversion at
the end of PFTR, film thickness (all scalars), and the temperature
programming, Tw(z), in the furnace (a continuous function). In order to
simplify the problem, the temperature of the surface of the film, Tw(z), was
assumed to be a cubic function of z. Thus, the continuous function was
replaced by four constants. The local constraint was taken care of by using
the penalty killing procedure. In this study, the requirement, μη,,ν,Γ ~ un,d> was
taken care of by using it both as a penalty function in the two objective
functions, and also by using it as a stopping condition. This leads to faster
convergence to the Pareto solution.
40
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
41
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
changing the steady state design on the performance indices was studied
using the bounding technique. A trade-off was found to exist between them,
and three non-inferior designs were determined.
Clark and Westerberg (1983) reviewed the mathematical tools used for
multi-criteria decision making while addressing problems in design
optimization where an Outer' optimization problem is constrained by an
'inner' one (embedded optimization). They proposed a single level solution
to locate the local optimum quickly. Two approaches to solve this problem
were suggested. The first was based on the active set strategy for solving the
inner problem, while the second involved the relaxation of complimentary
conditions. The concept of embedded optimization was illustrated on the
calculation of the equilibrium composition for multiphase, reacting systems.
An interactive CAD tool, CONSOLE, was used by Butala et al. (1988)
for the multiobjective optimization of a semi-batch reactor for the
manufacture of styrene-acrylonitrile (SAN) copolymer. Dynamic
optimization was used to obtain the two decision variables, the temperature
and the monomer feed rate histories, to produce copolymer having specified
values of the average copolymer composition and the average molecular
weight. A third-order polynomial was assumed for the two decision variables
and their coefficients were obtained using the feasible direction algorithm of
Nye and Tits (1986). The problem specifications (both functions and
numbers) were classified into objectives, as well as hard and soft constraints.
The initial solution had to satisfy the hard constraints first. The soft
constraints and the objectives compete simultaneously in the second phase of
optimization. In the final phase, the objectives were improved till the DM is
satisfied. This process was achieved by scaling of the specification space
(GOOD/BAD) based on the DM's knowledge of the process. It is to be
mentioned here that this process is highly intuitive and the DM should have a
very good knowledge about the problem, before he starts to solve it. Choi and
Butala (1991) extended this work to develop open loop optimal control for a
semi-batch methyl methacrylate-vinyl acetate (MMA-VA) co-polymerization
reactor. The feed rates of the monomers and the temperature of the reactor
(both functions of time) were used as decision variables. Copolymer having
specified values of the average composition and molecular weight was to be
produced. Hie optimal control policies were tested on an experimental stirred
tank polymerization reactor. These workers experienced some plant-model
mismatch. The need for improving the optimization method to minimize the
plant-model mismatch was indicated.
42
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
Ciric and Huchette (1993) tried to minimize the waste treatment costs and
maximize the profits of a chemical process, by employing the basic
relationship between the sensitivity problem and a multiobjective
optimization problem. The proposed approach uses a modified form of the
outer approximation method to identify discretely different regions of the
Pareto set. The complete solution set is generated using a sequential
approximation algorithm (Ciric and Jia, 1992). The proposed algorithm was
applied to the production of ethylene glycol from ethylene oxide and water.
This approach is better than the conventional sensitivity analysis techniques
such as linear sensitivity analysis and exhaustive search.
Liu et al. (1997) applied the concepts of multiobjective optimization to
the multimodel-based minimum bias control of a benchmark paper machine
process. They used modified goal programming for solving the problem.
Köllner et al. (1989 a, b) determined the optimal process flow structure
for the chlorination of benzene. They used the modified simplex method
(Neider and Mead; Box complex method; Beveridge and Schechter, 1970;
Edgar and Himmelblau, 1988) incorporating the inequality constraints on the
decision variables, with a stochastic global-search method (Schammler,
1988). They obtained compromise solutions between the total investment
cost and variable costs per annum for the optimal process flow structure.
They could eliminate four heat exchangers in the process using their study
involving multiobjective optimization.
Dimkou and Papalexandri (1998) proposed a decomposition-based
algorithm for solving convex MINLP (mixed integer non-linear
programming) problems and identifying the Pareto sets. Parametric nonlinear
optimization is used to identify the upper bounds of the optimal parametric
solution. Lagrangian information from the deterministic non-linear
programming (NLP) problems is used to determine the lower bounds to the
solution, as is done in Benders' decomposition method (Geoffrion, 1972).
The algorithm is illustrated using two simple examples. The assumption of
convexity limits the applicability of the proposed algorithm for real life
problems.
Tijsen et al. (1999) generated Pareto sets for the factors affecting the
degree of substitution of the hydroxyl groups in starch by ether groups, using
experimental data. They optimized the process conditions for the
modification of starch. They found the important factors by a backward
elimination strategy (Kleinbaum, 1994).
Methods used for generating non-inferior solutions and their applications
43
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
in Chemical Engineering have been discussed. This forms the objective phase
of the decision-making process. A few methods are available to select the
best (or preferred) solution from among the Pareto optimal points, using
additional, and often non-quantifiable, information. This forms the second,
subjective phase. In addition to the technique presented earlier, involving
consultation with several DMs to obtain their rankings of the Pareto
solutions, the surrogate worth trade-off method (Haimes and Hall, 1974)
seems to be popular in Chemical Engineering (Nishitani et al., 1980; Wajge
and Gupta, 1994a; Sareen and Gupta, 1995). This method uses the
Lagrangian multipliers obtained while generating the Pareto sets, to analyze
the trade-offs between the non-commensurate objectives. The preferred
solution is usually the one at which the improvement in one of the objective
functions is equivalent to the degradation that results in the other objectives.
3. FUTURE DIRECTIONS
44
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
ACKNOWLEDGEMENTS
For their contributions to the literature on the multiobjective optimization
applications using NSGA, the works of K. Deb, K. Mitra, S. Garg, R.R.
Gupta, J.K. Rajesh, Dr. G.P. Rangaiah, C.Y. Chan, Aatmeeyata, R. Gopalan,
Dr. M.B. Ray and F.B. Zhou are gratefully acknowledged. We thank Dr.
M.V. Garland for helping us with the translation of the article from German
to English.
45
NOMENCLATURE
Symbol Description
Subscripts / Superscripts
av Cross-section average value
d Desired or design value
f, out Final, outlet value
46
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
J Jacket-fluid value
ref Reference value (currentlyused in industrial reactor)
Τ Transpose
w Wall or surface value
Greek
REFERENCES
47
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications ofMultiobjective Optimization
In Chemical Engineering
49
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
50
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
V. Bhaskar et al. Reviews in Chemical Engineering
51
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. 1, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
52
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
K Bhaskar et al. Reviews in Chemical Engineering
53
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM
Vol. 16, No. I, 2000 Applications of Multiobjective Optimization
In Chemical Engineering
54
Brought to you by | Heinrich Heine Universität Düsseldorf
Authenticated | 134.99.128.41
Download Date | 9/24/13 12:39 PM