Optimization of Process Synthesis and Design Problems - A Modified Differential Evolution Approach
Optimization of Process Synthesis and Design Problems - A Modified Differential Evolution Approach
www.elsevier.com/locate/ces
Received 21 November 2005; received in revised form 15 January 2006; accepted 4 March 2006
Available online 10 March 2006
Abstract
A large number of process synthesis and design problems in chemical engineering can be modeled as mixed integer nonlinear programming
(MINLP) problems. They involve continuous (floating point) and integer variables. A common feature of this class of mathematical problems
is the potential existence of non-convexities due to the particular form of the objective function and/or the set of constraints. Due to their
combinatorial nature, these problems are considered to be difficult. In recent years, evolutionary algorithms (EAs) are gaining popularity for
finding the optimal solution of nonlinear multimodal problems encountered in many engineering disciplines. In the present study, a novel
modified differential evolution [Angira, R., Babu, B.V., 2005a. Optimization of non-linear chemical processes using modified differential
evolution (MDE). Proceedings of the Second Indian International Conference on Artificial Intelligence (IICAI-05), Pune, India, December
20–22, pp. 911–923. Also available at http://discovery.bits-pilani.ac.in/discipline/chemical/bvb/publications.html], one of the evolutionary
algorithms, is used for solving process synthesis and design problems. To illustrate the applicability and efficiency of modified differential
evolution (MDE), seven test problems on process synthesis and design have been solved. These problems arise from the area of chemical
engineering, and represent difficult nonconvex optimization problems, with continuous and discrete variables. The performance of MDE is
compared with that of Genetic Algorithm, Evolution Strategy, and MINLP-Simplex Simulated Annealing (M-SIMPSA).
䉷 2006 Elsevier Ltd. All rights reserved.
Keywords: Optimization; Processes synthesis; Systems engineering; Design; Evolutionary algorithm; Mixed integer non-linear programming (MINLP)
problems; Differential evolution (DE); Modified differential evolution (MDE)
in methods and software for solving optimization problems, large scale or very ill conditioned and highly constrained
development of powerful modeling languages (General Alge- functions, these methods require the application of successive
braic Modeling System, GAMS), and technological advances relaxations which may substantially increase the effort in iden-
in computing. tifying feasible regions and attaining the global optimum and
In order to formulate the synthesis problem as a mathe- hence suited for small to medium scale problems. Cardoso
matical programming problem, a superstructure is postulated et al. (1997) compared the performance of the M-SIMPSA
which includes many alternate designs from which the optimal with MSGA (Salcedo, 1992). They concluded that for small-
process will be selected. Once the superstructure is specified, scale problems and with penalizing scheme its performance
the next task is to determine the optimal process flow sheet is comparable to MSGA algorithm, however, for large scale
through structural and parameter optimization of the superstruc- and/or ill conditioned problems, the M-SIMPSA algorithm
ture (which requires the solution of a mixed integer optimiza- performed better. Costa and Oliviera (2001) studied seven test
tion problem). In early 1980s, most of the process synthesis and problems using GA & Evolution Strategies (ESs) and com-
design problems have been formulated as mixed-integer linear pared the results with M-SIMPSA algorithm. They found that
programming (MILP) problems. Although these formulations the performance of M-SIMPSA with penalty is better than
(e.g. Papoulias and Grossmann, 1983) have proved to be quite that of M-SIMPSA without penalty. Also the performance
powerful, they have the limitation that nonlinearities in the pro- of M-SIMPSA is comparable to GA. Evolution Strategies
cess equations cannot be treated explicitly and approximated emerged as the best algorithms in most of the problems stud-
through the discretization. The need for the explicit handling of ied. However, ESs exhibited difficulties in highly constrained
the nonlinearities in the synthesis problem motivates the use of problems but in general, they are found most efficient in terms
mixed-integer nonlinear programming (MINLP). MINLP prob- of function evaluations. Also, all the algorithms (GA, ESs, and
lems, however, are much more difficult to solve than MILP M-SIMPSA) are found to have great difficulties with multi-
problems for which Branch and Bound methods perform rea- product batch plant problem (Grossmann and Sargent, 1979),
sonably well. which is highly constrained; the global optimum corresponds
A large number of process synthesis, design and control prob- to a point where a very small variation in any of the continuous
lems in chemical engineering can be modeled as mixed inte- variables produces infeasibility.
ger nonlinear programming problems (Grossmann and Sargent, Previous studies (Storn, 1995; Storn and Price, 1997; Wang
1979; Kocis and Grossmann, 1987, 1988, 1989; Floudas et al., and Chiou, 1997; Babu and Sastry, 1999; Babu and Angira,
1989; Salcedo, 1992; Ciric and Gu, 1994 etc.). They involve 2002; Angira and Babu, 2003; Chakraborti and Kumar, 2003;
continuous (floating point) and integer variables. A common Chakraborti et al., 2004; Colaco et al., 2004, 2005; Babu
feature of this class of mathematical problem is the potential et al., 2005; Angira, 2005; Angira and Babu, 2005b etc. to
existence of nonconvexities due to the particular form of the name a few) have shown that differential evolution (DE) is
objective function and/or the set of constraints. Due to their an efficient, effective and robust evolutionary optimization
combinatorial nature, these problems are considered to be dif- method. The details on DE algorithm, various strategies of
ficult. DE and wide range of applications in various engineering,
The optimization of mixed integer nonlinear programming manufacturing and management areas are well documented
problems constitutes an active area of research. So far various in literature (Corne et al., 1999; Babu, 2004; Onwubolu and
methods such as branch and bound technique (Grossmann and Babu, 2004). In the present study, a novel modified differen-
Sargent, 1979), outer-approximation (OA)/equality-relaxation tial evolution (MDE) algorithm (Angira and Babu, 2005a) is
algorithm (Kocis and Grossmann, 1987, 1988, 1989), variant used for solving seven test problems on process synthesis and
of OA method (Diwekar et al., 1992), adaptive random-search design. These problems are difficult non-convex optimization
method (MSGA by Salcedo, 1992), branch-and-reduce algo- problems with continuous and discrete variables. Also a new
rithm (Ryoo and Sahinidis, 1995), MINLP Simplex Simulated approach for handling binary (discrete) variables is proposed
Annealing Algorithm (M-SIMPSA by Cardoso et al., 1997), and and compared with the nonlinear transformation for model-
genetic algorithm & evolution strategies (Costa and Oliviera, ing binary variables as continuous variables proposed by Li
2001) have been used for solving nonconvex MINLP problems. (1992). The performance of MDE is compared with that of
Gradient optimization techniques have only been able to DE, GA, ESs, and M-SIMPSA algorithms.
tackle special formulations, where continuity or convexity had
to be imposed, or by exploiting special mathematical struc- 2. Modified differential evolution (MDE)
tures. Stochastic algorithms, also known as adaptive random
search methods, have tackled MINLP problems, mostly in the The principle of modified DE is same as DE. The major
area of chemical engineering (Salcedo, 1992). These require difference between DE and MDE is that MDE maintains only
neither the prior step of identification or elimination of the one array. The array is updated as and when a better solution
sources of nonconvexities nor decomposition of the problem is found. Also, these newly found better solutions can take part
into sub problems which have to be iteratively solved. How- in mutation and crossover operation in the current generation
ever, various problem-independent heuristics related to search itself as opposed to DE (where another array is maintained
interval compression and expansion and to shifting strategies and these better solutions take part in mutation and crossover
are required for their effectiveness (Salcedo, 1992). Also, for operations in next generation). Updating the single array
R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721 4709
continuously enhances the convergence speed leading to less 2. Generate initial population in upper (UB) and lower (LB)
function evaluations as compared to DE (Angira, 2005; An- bounds of each decision variable for NP times in an array.
gira and Babu, 2005a; Babu and Angira, 2006). However, DE 3. Evaluate the cost (performance index) for each individual
maintains two arrays consuming extra memory and CPU-time including penalty terms.
(more function evaluations). This modification enables the 4. Perform mutation, crossover operation to obtain the trial
algorithm to get a better trade-off between the convergence vector for each target vector in the population.
rate and the robustness. By choosing the key parameters (NP, (a) For each vector Xt (target vector), select three distinct
CR, and F) wisely/appropriately, the problem of premature vectors Xa , Xb and Xc randomly from the population
convergence can be avoided to a large extent. array other than vector Xt. and create noisy random
Such an improvement can be advantageous in many real- vector as
world problems where the evaluation of a candidate solution
is a computationally expensive operation and consequently noisy random vector = Xc + F (Xa − Xb )
finding the global optimum or a good sub-optimal solution
with the original differential evolution algorithm is too time- (b) Perform crossover for each target vector with its noisy
consuming, or even impossible within the time available (Fan random vector to create a trial vector as below.
and Lampinen, 2003; Babu and Angira, 2006). This has been for k = 1 to D
found to be very true in examples such as optimization in If ((random no. (0, 1) < CR) or k = D)
the field of computational mechanics, computational magnet- trial vectork = noisy random vectork
ics, computational fluid dynamics and unsteady solidification else
(Rogalsky and Derksen, 2000; Stumberger et al., 2000; Colaco trial vectork = target vectork
et al., 2005). The pseudo code of MDE used in the present end for
study is given below: 5. Check whether the parameters of trial vector are within the
bounds. If an individual of this trial vector is outside the
Let P a population of size NP, bounds, then this parameter is assigned a value according to
and xj the jth individual of dimension D in population P, the bound handling method used (Section 6).
and CR denotes the crossover probability 6. Evaluate the cost for trial vector as in step 3.
input D, NP 4; F ∈ (0, 1+); CR ∈ [0, 1], and initial 7. If the performance index of trial vector is better than perfor-
bounds: lower(xi ); upper(xi ); i = 1, . . . ..D mance index of target vector then replace the target vector
initialize P = { x 1 , . . . .., xNP } as with trial vector in the array in current generation itself else
For each individual j ∈ P target vector will be retained.
j
xi = lower(xi ) + randi [0, 1] × (upper(xi ) − 8. Repeat steps 4–7 until the termination criterion is satisfied.
lower(xi )); i = 1, . . . . . . D
end For each 3. Problem formulation
Evaluate P
while the stopping criterion is not satisfied do The chemical process synthesis problem involves selecting
forall j NP the optimal flowsheet structure as well as the parameters which
Randomly select r1 , r2 , r3 ∈ (1, . . . . . . .NP), describe the operation of a desired process. In order to define
j = r1 = r2 = r3 the search space of candidate flowsheet alternatives, a super-
randomly select irand ∈ (1, . . . . . . ..D) structure is to be postulated based on preliminary screening.
forall i D, This superstructure can then be modeled as a MINLP problem
⎧ x r3 + F × (x r1 − x r2 )
⎪ i
⎨ i i of the form (Kocis and Grossmann, 1987):
if (random[0, 1) < CRi = irand
xi = j
⎪
⎩ xi F = min cT y + f(x),
x,y
otherwise
end forall s.t. h(x) = 0,
if f ( x )f ( xj )
g(x) 0,
Then, xj = x ; f ( x );
x j ) = f (
end forall Ax = a,
end while By + Cx d,
Print the results.
x ∈ X = {x | x ∈ Rn , xL x xU },
The basic version of MDE/rand/1/bin (Angira and Babu, 2005a; y ∈ Y = {y | y ∈ {0, 1}m , Ey e},
Babu and Angira, 2006) consist of following steps:
where x is the vector of continuous variables specified in the
1. Input: dimensions of problem (D) and key parameters (the compact set X, and y is the vector of 0–1 variables which
population size NP, the scaling factor F, and the crossover must satisfy linear integer constraints Ey e. f(x), h(x)=0, and
constant CR). g(x) 0 represent nonlinear functions involved in the objective
4710 R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721
y2 v2 1 − y1
+ 50
x2 z2 0.8[1 − exp(−0.4v2 )]
y1
+ 50
0.9[1 − exp(−0.5v1 )]
Fig. 1. Superstructure for two-reactor problem.
s.t. 0.9[1 − exp(−0.5v1 )] − 2y1 0,
0.8[1 − exp(−0.4v2 )] − 2(1 − y1 ) 0,
one between two candidate reactors (as shown in Fig. 1) in
order to minimize the production cost. Also, it has been solved v1 10y1 ,
by Diwekar et al. (1992), Diwekar and Rubin (1993), Cardoso
v2 10(1 − y1 ),
et al. (1997), and Costa and Oliviera (2001)
v1 , v2 0,
Min f (x, y1 , y2 , v1 , v2 ) = 7.5y1 + 5.5y2 + 7v1 + 6v2 + 5x
y1 ∈ {0, 1}.
s.t. y1 + y2 = 1,
z2 = 0.8[1 − exp(−0.4v2 )]x2 , Problem 5 (Process synthesis problem). This problem was
studied by Floudas et al. (1989), Salcedo (1992), Ryoo and
z1 + z2 = 10, Sahinidis (1995), Cardoso et al. (1997), and Costa and Oliviera
x1 + x2 = x, (2001). This problem features nonlinearities in both continuous
and binary variables and has seven degrees of freedom.
z1 y1 + z2 y2 = 10,
Min f (x1 , x2 , x3 , y1 , y2 , y3 , y4 )
v1 10y1 ,
= (y1 − 1)2 + (y2 − 1)2 + (y3 − 1)2
v2 10y2 ,
− ln(y4 +1) + (x1 −1)2 + (x2 −2)2 + (x3 −3)2
x1 20y1 ,
s.t. y1 + y2 + y3 + x1 + x2 + x3 5,
x2 20y2 ,
y32 + x12 + x22 + x32 5.5,
x1 , x2 , z1 , z2 , v1 , v2 0,
y1 + x1 1.2,
y1 , y2 ∈ {0, 1}.
y2 + x2 1.8,
The binary variables y1 and y2 denote the existence (nonex-
istence) of reactor 1 and 2 when their value is 1 (0). In the y3 + x3 2.5,
objective function, there are fixed charges for purchasing reac- y4 + x1 1.2,
tor 1 (7.5) or reactor 2 (5.5), linear terms in v1 and v2 (reactor
volumes) and the purchase price for raw material x. The two y22 + x22 1.64,
nonlinear equations are the input–output relations for the re-
actors which define the output flows (z1 and z2 ) in terms of y32 + x32 4.25,
the input flows (x1 and x2 ) and the volumes. The raw material
x is split into the reactor input flows x1 and x2 ; a total de- y22 + x32 4.64,
mand of 10 units must be met by the output flows z1 , z2 . The x1 , x2 , x3 0,
next four inequalities are logical constraints which insure that
if a given reactor does not exist (e.g. y1 = 0), then the corre- y1 , y2 , y3 , y4 ∈ {0, 1}.
sponding volume and feed stream are zero. The last constraint
requires that either reactor 1 or 2 be selected. The suboptimal The global optimum is (x1 , x2 , x3 , y1 , y2 , y3 , y4 ; f ) =
solution corresponding to (y1 , y2 )=(0, 1) has an objective func- (0.2, 1.28062, 1.95448, 1, 0, 0, 1; 3.557473).
tion value of 107.376 at (x1 , x2 ) = (0.0, 15.0) and (v1 , v2 ) =
(0.0, 4.479). The global optimum is: (x, y1 , y2 , v1 , v2 ; f ) = Problem 6 (Process design problem). It is a maximi-
(13.36227, 1, 0, 3.514237, 0; 99.245209). zation problem studied by Cardoso et al. (1997), and
4712 R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721
Table 2 Table 3
Constants for Problem 6 Values of Sij and tij of Problem 7
− 37.29329y1 + 40792.141 The values of Sij and tij [i = 1.2 (rows); and j = 1–3
(columns)] are given in Table 3. The global optimum is:
s.t. a1 + a2 y2 x3 + a3 y1 x2 − a4 x1 x3 92,
(N1 , N2 , N3 , V1 , V2 , V3 , B1 , B2 , T1 , T2 ; f )=(1, 1, 1, 480, 720,
a5 + a6 y2 x3 + a7 y1 y2 + a8 x12 − 90 20, 960, 240, 120, 20, 16; 38499.8).
a9 + a10 x1 x3 + a11 y1 x1 + a12 x1 x2 − 20 5,
5. Handling of integer and binary variables
27x1 , x2 , x3 45,
y1 ∈ {78, . . . , 102}, integer, 5.1. Integer variables
Vj Sij Bi ,
5.2. Binary or discrete variables
NJ TLi tij ,
1 Nj Nju , In the present study, integer variables are handled in the
same way as described above however, a new procedure is
Vjl Vj Vju , proposed and evaluated to handle binary (discrete) variables.
This procedure is called Approach 1 and is compared with
l
TLi TLi TLi
u
, another procedure (Li, 1992), called Approach 2 in the present
Bjl Bj Bju , study. These are described below:
where, for the specific problem considered, M = 3, N = 2, H = Approach 1. In this procedure, for the function evaluation,
6000, j = 250, j = 0.6, Nju = 3, Vjl = 250 and Vju = 2500. binary variables are handled as follows:
The values of TLil , T u , B l and B u are given by
Li j j 0 if xi 0.5,
yi =
1 otherwise,
l
TLi = max tij /Nju ,
where xi is a continuous variable 0 xi 1. However bound-
u
TLi = max tij , ary constraint is handled in the same way as for continuous
R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721 4713
variables. The only difference is that lower and upper bound of the following methods: (1) If there is bound violation for a
are set to zero and one respectively. parameter, then assign the upper or lower bound value if upper
or lower bound is violated (forced bound), (2) if there is bound
Approach 2. It is a nonlinear transformation for modeling bi- violation for a parameter, then that parameter is again gener-
nary variables as continuous variables proposed by Li (1992). ated randomly between given lower and upper bound (without
A binary variable y ∈ {0, 1} can be modeled as a continuous forcing) using the following equation:
variable x ∈ [0, 1], simply by the addition of the following
j
constraint in the problem: xi = lower(xi ) + randi [0, 1] × (upper(xi ) − lower(xi ));
i = 1 . . . D,
x(1 − x) = 0, 0 x 1
where D is the number of decision variables. In the present
which forces x to take either 0 or 1. Hence with this trans- paper, the first method is called forced bound method (FBM)
formation any MINLP model can be converted into an equiv- while the second method is called method without forcing the
alent NLP model. The function x(1 − x) is a non-convex bound (MWFB).
nonlinear function. Li (1992) referred to this as the “binary The penalty function methods are one of the most popular
condition” to model binary variables and found this procedure techniques in EAs to handle constraints. The techniques trans-
to be more convenient than current approaches as branch-and- form the constrained problem into an unconstrained problem
bound method and implicit enumeration method. The resulting by penalizing unfeasible solutions. In addition, the penalty
NLP was solved using a modified penalty function method, function methods are easy to implement and considered effi-
however only local optimal solutions were found due to the ad- cient. In the present study, an absolute value of constraint vi-
dition of non-convexities. Also, the use of standard NLP solver olation is multiplied with a high penalty and added/subtracted
like SQP (sequential quadratic programming) and GRG (gen- to objective function depending upon the type of problem,
eralized reduced gradient) is ruled out as they could be sensi- i.e., minimization/maximization. In case of more than one
tive to initial guess and hence get stuck at local optima. Ryoo constraint, all such absolute violations are first multiplied
and Sahinidis (1995) proposed a specialized branch and reduce with high penalty and then added or subtracted from objective
algorithm for the global optimization of NLPs and MINLPs function value (for minimization or maximization problem,
wherein they refer to the usage of such procedure for handling respectively).
binary and discrete variables. The present study evaluates the
application of DE and MDE algorithms for solving the result- 7. Results and discussion
ing nonconvex NLP problems to global optimality.
7.1. Effect of key parameters (CR, F, and NP)
6. Handling of bound violations and constraints
The performance of DE and MDE algorithms depends on
Through manipulating its own scaling factor, one can control key parameters, namely, CR, F, and NP. By choosing the key
the scaling of difference of two solutions/vectors. However, parameters (NP, CR, and F) wisely/appropriately, the problem
it does not guarantee that the random noisy vector/solution of premature convergence can be avoided to a large extent. To
generated after adding this weighted difference to the third study the effect of key parameters, two highly multimodal mul-
vector/solution would be lying within the given bounds. For tidimensional test functions, discussed in Fan and Lampinen
unconstrained optimization problems, when the variables are (2003), are used. Also, a basic key parameter set (CR = 0.85,
out of bounds, their objective function values are also worse, F = 99, and NP = 30) is first chosen as mentioned in Fan
and hence they are eliminated automatically via selection by and Lampinen (2003). Then, all the numerical simulations are
comparison of the function values in subsequent generations, carried out around this basic set taking maximum number of
but definitely not by controlling the scaling factor. But, in the generation to be 250. A systematic numerical simulation is
case of constrained optimization problems, the points outside performed keeping two key parameters constant while vary-
the bounds may have a better function value but may not come ing the third at a time. The two test functions are defined
under the feasible region. In such cases, a method for bringing below:
the solutions to within the bounds of feasible region is required. Ackley’s function: This is a continuous, highly nonlinear func-
Most of the process synthesis and design problems are tion that causes the search with moderate complications.
constrained. The difficulty of using EAs in the constrained op-
timization is that the evolutionary operators used to manipulate ⎛ ⎞ n
n
the individuals of the population often produce solutions which 1 1
⎝
f1 = −20 exp −0.2
⎠
xi − exp
2 cos(2xi )
are unfeasible. There may be many ways of doing it. Two such n n
i=1 i=1
methods (prescribing externally) are discussed, applied and
compared in the present study, which are discussed below. + 20 + e, −20 xi 30, n = 30.
Bound violations (whether upper or lower) may occur after
mutation step of DE and MDE. This can be repaired by one The global minimum is: f1 = 0 with xi = 0, i = 1, 2, . . . ., n.
4714 R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721
Function Value
60
50
50
40
40
30 30
20 20
10 10
0 0
0 50 100 150 200 250 300
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
NP
CR
Fig. 4. Effect of Population Size (NP).
Fig. 2. Effect of Crossover constant (CR).
Table 4
Results of DE using Approach 1 DE MDE
100000
Problem NFE/NRC/CPU-time Key parameters
90000
No. (NP/CR/F)
FBM MWFB 80000
70000
1 802/90/0.022 812/100/0.0 20/0.8/0.5
2∗ 610/100/0.011 626/100/0.011 20/0.8/0.5 60000
NFE
3 801/10/0.011 2727/100/0.071 30/0.8/0.7 50000
4∗ 1080/100/0.040 2196/100/0.083 30/0.8/0.5
40000
5 11739/100/0.610 13265/90/0.696 30/0.9/0.6
6 1045/100/0.039 6430/100/0.280 20/0.8/0.5 30000
7 46090/100/3.225 89490/0∗∗ /6.291 100/0.8/0.5 20000
∗∗ Converged to a non-optimal solution. 10000
0
1 2* 3 4* 5 6 7
Table 5 Problem No.
Results of MDE using Approach 1
Fig. 5. NFE variation using DE and MDE (MWFB, Approach 1).
Problem NFE/NRC/CPU-time Key parameters
No. (NP/CR/F)
FBM MWFB
92
90
100% using method without forcing the bound. Similarly, for 88
forced bound method NRC is almost 100% excepting Problem 86
3. It is because when upper limit of bound is violated, the
value of variable is forced to the upper limit that resulted in 84
1 2* 3 4* 5 6 7
convergence to non-optimal solution. However, for Problem 7,
the NRC is 100% for forced bound method and 0.0% for method Problem No.
without forcing the bound, as global optimum is located on the Fig. 6. NRC variation using DE and MDE (MWFB, Approach 1).
bound of decision variables. CPU-time is found to be more in
case of method without forcing the bound.
It is observed from Table 5 that NRC is 100% for all the
problems excepting Problem 7 using method without forcing compared to DE in all the seven problems. NFE is about
the bound while NRC is zero and 80% for Problems 3 and 5, 13.17%, 21.73%, 27.61%, 18.17%, 10.18%, 14.54%, and
respectively, using forced bound method. NFE and hece CPU- 11.3% more, respectively, for Problems 1, 2, 3, 4, 5, 6, and 7 in
time is found to be more for method without forcing the bound case of DE as compared to MDE. NRC for MDE, as shown in
as compared to forced bound method. In all the seven problems, Fig. 6, is 100% in all the problems except Problem 7 (for which
NFE for forced bound is less than that for without forcing NRC is zero). For DE too, it is 100% for all the problems
but the difference is quite significant for Problems 3, 4∗ , 5, 6, except Problem 5 (for which NRC is 90).
and 7 than for Problems 1 and 2∗ . NFE for without forcing Figs. 7 and 8 show the comparison of DE and MDE in
is about 1.86 times for Problem 3, 1.52 times for Problem 4∗ , terms of NFE and NRC, respectively, using forced bound
1.17 times for Problem 5, 6.35 times for Problem 6, and 1.96 method. As can be seen from Fig. 7. NFE using MDE is less
times for Problem 7, more than that of forced bound method. as compared to DE for Problems 1, 2∗ , 5, 6, and 7. However,
This observation is similar to that seen for results reported in NFE using DE is less for Problems 3 and 4∗ . NFE is about
Table 4 using DE. 13.97%, 19.67%, 13.71%, 17.22%, and 12.02% more, respec-
tively, for Problems 1, 2, 5, 6 and 7 in case of DE as compared
7.2.1.1. Comparison of DE and MDE. Figs. 5 and 6 show to MDE. As shown in Fig. 8, NRC for MDE is 100% in all the
the comparison of DE and MDE in terms of NFE and problems except Problem 5 (for which NRC is 90). For DE
NRC, respectively, using method without forcing the bound. too, it is 100% for all the problems except Problems 1 and 3
As can be seen from Fig. 5. NFE using MDE is less as (for which NRC is 90 and 10, respectively).
4716 R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721
Table 7
DE MDE
Results of MDE using Approach 2
50000
45000 Problem NFE/NRC/CPU-time Key parameters
40000 No. (NP/CR/F)
35000 FBM MWFB
30000
1 742/60/0.011 1408/90/0.022 20/0.8/0.5
NFE
25000
2∗ 560/100/0.016 1218/100/0.055 20/0.8/0.5
20000
3 882/20/0.030 3297/100/0.110 30/0.8/0.5
15000
4∗ 927/100/0.027 4602/100/0.176 30/0.8/0.5
10000 5 13566/60/0.703 45801/90/2.511 30/0.9/0.6
5000 7 52350/100/3.621 110970/0∗∗ /7.786 100/0.8/0.5
0
1 2* 3 4* 5 6 7 ∗∗ Converged to a non-optimal solution.
Problem No.
120000
DE MDE
100 100000
90
80000
80 NFE
60000
70
60 40000
NRC
50 20000
40
0
30 1 2* 3 4* 5 7
20 Problem No.
10
Fig. 9. NFE variation using DE and MDE (MWFB, Approach 2).
0
1 2* 3 4* 5 6 7
Problem No.
lems using either forced bound method or without forcing the
Fig. 8. NRC variation using DE and MDE (FBM, Approach 1). method. It is important to note that for Problem 7, the NRC
is 100% using forced bound method and 0.0% using without
forcing method as found for Approach 1. CPU-time is found
Table 6 to be more in case of without forcing the bound method.
Results of DE using Approach 2
It is observed from Table 7 that NRC is not 100% for all the
Problem NFE/NRC/CPU-time Key parameters problems using either forced bound method or method without
No. (NP/CR/F) forcing the bound but certainly it is better or equal in case of
FBM MWFB
method without forcing the bound for all the problems except-
1 964/90/0.011 1910/60/0.055 20/0.8/0.5 ing Problem 7 (where NRC is zero). NFE and hence CPU-time
2∗ 654/100/0.011 1332/100/0.011 20/0.8/0.5 is found to be twice or more for without forcing the bound
3 801/10/0.027 2754/100/0.083 30/0.8/0.7
method as compared to force bound method. To be precise, NFE
4∗ 1071/100/0.055 5328/100/0.187 30/0.8/0.5
5 16114/80/0.852 53158/90/3.016 30/0.9/0.6 for without forcing is about 47.3%, 54.0%, 73.25%, 79.9%,
7 57330/100/4.022 122170/0∗∗ /8.681 100/0.8/0.5 70.4%, and 52.8%, respectively, for Problems 1, 2∗ , 3, 4∗ , 5,
∗∗ Converged to a non-optimal solution.
and 7, more than that of forced bound method. This observation
is similar to that seen for results reported in Table 6 using DE.
DE MDE DE MDE
100 100
90
90
80
80
70
70
60
NRC
50 60
NRC
40 50
30 40
20
30
10
20
0
1 2* 3 4* 5 7 10
Problem No. 0
1 2* 3 4* 5 7
Fig. 10. NFE variation using DE and MDE (MWFB, Approach 2).
Problem No.
Fig. 12. NRC variation using DE and MDE (FBM, Approach 2).
DE MDE
70000
DE (Approach-1) DE (Approach-2)
60000
140000
50000
120000
40000
NFE
100000
30000
80000
NFE
20000
60000
10000
0 40000
1 2* 3 4* 5 7
20000
Problem No.
0
Fig. 11. NFE variation using DE and MDE (FBM, Approach 2).
1 2* 3 4* 5 7
Problem No.
DE is same for all the problems except Problem 1 where NRC Fig. 13. NFE variation for DE using Approaches 1 and 2 (MWFB).
for DE is 60% as compared to 90% for MDE. It is to be noted
that NRC is zero for Problem 7 using both DE and MDE.
Figs. 11 and 12 show the comparison of DE and MDE in 7.2.3. Comparison of Approaches 1 and 2
terms of NFE and NRC respectively using forced bound method Figs. 13 and 14 show the comparison of Approaches 1 and 2
and Approach 2. As can be seen from Fig. 11, NFE using for DE in terms of NFE and NRC, respectively, using method
MDE is less as compared to DE for all the problems except for without forcing the bound. As can be seen from Fig. 13, NFE
Problem 3 where NFE using DE is slightly less (about 9.0%) using Approach 2 is significantly more as compared to Ap-
than that of MDE. NFE is about 23.03%, 14.37%, 13.44%, proach 1 for all the problems except for Problem 3 where NFE
15.81%, and 8.68% more, respectively, for Problems 1, 2∗ , using is almost same (2727 using Approach 1 and 2754 using
4∗ , 5, and 7 in case of DE as compared to MDE. As shown Approach 2, respectively) in both the approaches. NFE is about
in Fig. 12, NRC for MDE and DE is 100% for Problems 2, 57.49%, 53.00%, 58.78%, 75.04%, and 26.75% more, respec-
4∗ , and 7. For Problems 1 and 5, NRC is higher for DE as tively, for Problems 1, 2∗ , 4∗ , 5 and 7 in case of Approach 2 as
compared to MDE but for Problem 3, NRC for DE is only compared to Approach 1. This is because of addition of con-
10% as compared to 20% for MDE. It is to be noted that straints (as binary or discrete variable is modeled as continuous
NRC is not 100% for Problems 1,3 and 5 using both DE and variable) and addition of nonconvexities to the problem in case
MDE. As compared to without forcing, the NRC is less for all of Approach 2.
the problems using forced bound method. The savings in NFE As shown in Fig. 14, NRC for Approaches 1 and 2 is same
using MDE is almost same in both the methods (forced bound i.e. 100%, 100%, 100%, 90%, and zero, respectively, for the
method and method without forcing the bound). This indicates Problems 2∗ , 3, 4∗ , 5, and 7. For Problem 1, NRC is 100% for
that the method without forcing the bound is better than the Approach 1 as compared to 60% for Approach 2. It is to be
forced bound method. noted that NRC is zero for Problem 7 using both the approaches.
4718 R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721
100 100
90 90
80
80
70
70
60
60
NRC
50
NRC
50
40
40 30
30 20
20 10
10 0
0
1 2* 3 4* 5 7
1 2* 3 4* 5 7 Problem No.
Problem No.
Fig. 16. NRC variation for MDE using Approaches 1 and 2 (MWFB).
Fig. 14. NRC variation for DE using Approaches 1 and 2 (MWFB).
DE (Approach-1) DE (Approach-2)
70000
MDE (Approach-1) MDE (Approach-2)
120000
60000
100000
50000
80000
40000
NFE
NFE
60000
30000
40000
20000
20000
10000
0
1 2* 3 4* 5 7 0
Problem No. 1 2* 3 4* 5 7
Problem No.
Fig. 15. NFE variation for MDE using Approaches 1 and 2 (MWFB).
Fig. 17. NFE variation for DE using Approaches 1 and 2 (FBM).
100 100
90 90
80 80
70 70
60 60
NRC
NRC
50 50
40 40
30 30
20 20
10 10
0 0
1 2* 3 4* 5 7
1 2* 3 4* 5 7
Problem No.
Problem No.
Fig. 20. NRC variation for MDE using Approaches 1 and 2 (FBM).
Fig. 18. NRC variation for DE using Approaches 1 and 2 (FBM).
50000
7.2.4. Comparison of MDE, GA, M-SIMPSA, and ( + )-ES
In the present study, DE and MDE are used to solve process
40000
synthesis and design problems and their performance is com-
pared. Further, the performance of MDE is evaluated and com-
pared with that of GA, M-SIMPSA, and (+)-ES algorithms.
NFE
30000
Table 8 shows the comparison of MDE with GA, M-SIMPSA,
20000
M-SIMPSA-pen, and ( + )-ES. The NFE in MDE is about
82% less than of that in GA for all the problems (The range is
10000
82.0–98.16% to be precise).
Similarly, the NFE in MDE is about 81–96.0% less than
0
that of M-SIMPSA for various test problems. While comparing
1 2* 3 4* 5 7 ( + )-ES and MDE it is found that for Problems 1 and 2
Problem No. NFE for MDE is about 53.56% and 78.27% less than that of
( + )-ES but for Problems 3, 5, and 6, NFE for ( + )-ES
Fig. 19. NFE variation for MDE using Approaches 1 and 2 (FBM). is, respectively, about 11.48%, 43.68%, and 53.85% less than
that of MDE. It is important to note that ( + )-ES is found
to converge to a non-optimal solution for Problems 4∗ and 7,
bound method. As can be seen from Fig. 19, NFE using Ap- whereas MDE is able to locate the global optimum for all the
proach 2 is more as compared to Approach 1 for all the problems problems solved in the present study.
except for Problems 3 and 4∗ (where NFE is about 16.95% and It may be noted that GA could not converge to global opti-
21.37% more using Approach 1 than that of Approach 2, re- mum, while execution was reported to be halted in M-SIMPSA
spectively, for Problems 3 and 4∗ ). NFE is about 7.0%, 12.50%, for Problem 7. However, for Problem 7, NRC is 100% only
25.33%, and 22.54% more, respectively, for Problems 1, 2∗ , 5 for MDE, and 92% for MIMPSA-pen, while GA, M-SIMPSA
and 7 in case of Approach 2 as compared to Approach 1. and ( + )-ES were not able to converge to global optimum.
As shown in Fig. 20, NRC for Approaches 1 and 2 is same The performance of ( + )-ES and MDE is found to be com-
i.e. 100%, 100%, and 100%, respectively, for the Problems parable in terms of function evaluations but MDE has the ad-
2∗ , 4∗ , and 7. For Problems 1, 3, and 5, NRC for Approach 1 vantage over ES in terms of convergence to optimal solution.
is higher as compared to Approach 2. Therefore, using MDE, The performance of MDE is found to be better than that of
the Approach 1 is found to be better than Approach 2 in GA & M-SIMPSA in optimizing the mixed integer nonlinear
terms of both NFE and NRC. This indicates that nonconvexity programming problems considered in the present study. Also,
(induced due to modeling of binary variable as continuous results clearly show the potential of ( + )-ES and MDE for
variable) affects the NRC (Figs. 14, 16, 18, and 20). Also, it solving such process synthesis problems. We have already seen
is important to note that NRC using MDE is higher than that that MDE is better than DE in terms of functions evaluations
of using DE (Figs. 18 and 20). Therefore, MDE using Ap- (NFE) and convergence to optimal solution (NRC).
4720 R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721
Table 8
Comparison of MDE, GA, ES, M-SIMPSA & M-SIMPSA-pen
8. Conclusions Angira, R., Babu, B.V., 2005a. Optimization of non-linear chemical processes
using modified differential evolution (MDE). Proceedings of the second
Seven test problems on process synthesis and design from Indian International Conference on Artificial Intelligence (IICAI-05), Pune,
India, December 20–22, pp. 911–923. Also available at http://discovery.
chemical engineering have been solved using DE and MDE in bits-pilani.ac.in/discipline/chemical/bvb/publications.html.
the present work. Performance of various algorithms (e.g. GA, Angira, R., Babu, B.V., 2005b. Process synthesis and design using modified
M-SIMPSA, M-SIMPSA-pen, ( + )-ES, DE, and MDE) is differential evolution (MDE). Proceedings of the International Symposium
compared in terms of NFE and NRC. Also, a new approach for & 58th Annual Session of IIChE (CHEMCON-2005), December 14–17,
handling binary variables is studied and compared with the ap- New Delhi, Also available at http://discovery.bits-pilani.ac.in/discipline/
chemical/bvb/publications.html.
proach of Li (1992) along with two methods for handling bound Babu, B.V., 2004. Process Plant Simulation. Oxford University Press, New
violations. Results indicate that the method without forcing the York.
bound is robust however the CPU-time is more as compared to Babu, B.V., Angira, R., 2002. A differential evolution approach for
the forced bound method. It is important to note that for prob- global optimization of MINLP problems. Proceedings of Fourth Asia
lems where global optimum is located on the bound of deci- Pacific Conference on Simulated Evolution and Learning (SEAL-2002),
Singapore, November 18–22, vol. 2, pp. 880–884. Also available at http://
sion variables forced bound method is found to outperform the discovery.bits-pilani.ac.in/discipline/chemical/bvb/publications.html.
method without forcing the bound (e.g. Problem 7) whereas for Babu, B.V., Angira, R., 2006. Modified differential evolution (MDE) for
problems where local optimum is located on the bound of de- optimization of non-linear chemical processes. Computers & Chemical
cision variables the method without forcing the bound is found Engineering, in press.
to be robust (e.g. Problem 3). Also, the Approach 1 for han- Babu, B.V., Sastry, K.K.N., 1999. Estimation of heat-transfer parameters in a
trickle-bed reactor using differential evolution and orthogonal collocation.
dling binary variables is found to be better than that of Ap- Computers & Chemical Engineering, 23, 327–339. Also available at http://
proach 2. This is because of addition of constraints (as binary discovery.bits-pilani.ac.in/discipline/chemical/bvb/publications.html.
or discrete variable is modeled as continuous variable) and ad- Babu, B.V., Pallavi, G., Chakole, Syed Mubeen, J.H., 2005. Multiobjective
dition of nonconvexities to the problem in case of Approach 2. differential evolution (MODE) for optimization of adiabatic styrene reactor.
It is found that NFE is the least and NRC is highest in MDE as Chemical Engineering Science, 60(17), 4822–4837. Also available at http://
discovery.bits-pilani.ac.in/discipline/chemical/bvb/publications.html.
compared to GA & M-SIMPSA-pen. The NRC in M-SIMPSA Cardoso, M.F., Salcedo, R.L., Feyo de Azevedo, S., Barbosa, D., 1997.
is better than that of GA, M-SIMPSA, and ( + )-ES. The A simulated annealing approach to the solution of MINLP problems.
( + )-ES and MDE are comparable in terms of NFE (rather Computers & Chemical Engineering 21, 1349–1364.
( + )-ES perform slightly better than MDE) but MDE is Chakraborti, N., Kumar, A., 2003. The optimal scheduling of a reversing
found to be robust than ( + )-ES being higher NRC for Prob- strip mill: studies using multipopulation genetic algorithms and differential
evolution. Materials and Manufacturing Processes 18 (3), 433–445.
lems 4∗ and 7 (Table 8). The performance of MDE is found to Chakraborti, N., Mishra, P., Erkoc, S., 2004. A study of the Cu clusters using
be the best among the methods compared for all the problems gray-coded genetic algorithms and differential evolution. Journal of Phase
studied. Equilibria and Diffusion 25, 16.
Ciric, A.R., Gu, D., 1994. Synthesis of nonequilibrium reactive distillation
processes by MINLP optimization. American Institute of Chemical
References Engineers Journal 40, 1479–1487.
Colaco, M.J., Dulikravich, G.S., Martin, T.J., 2004. Optimization of wall
Angira, R., 2005. Evolutionary computation for optimization of selected non- electrodes for electro-hydrodynamic control of natural convection during
linear chemical processes. Ph.D. Thesis, Birla Institute of Technology & solidification. Materials and Manufacturing Processes 19 (4), 719–736.
Science (BITS), Pilani, India. Colaco, M.J., Dulikravich, G.S., Martin, T.J., 2005. Control of unsteady
Angira, R., Babu, B.V., 2003. Evolutionary computation for global solidification via optimized magnetic fields. Materials and Manufacturing
optimization of non-linear chemical engineering processes. Proceedings Processes 20 (3), 435–438.
of International Symposium on Process Systems Engineering and Control Corne, D., Dorigo, M., Glover, F., 1999. New Ideas in Optimization. McGraw-
(ISPSEC’ 03)—For Productivity Enhancement through Design and Hill, Berkshire, England, UK.
Optimization, IIT-Bombay, Mumbai, 3–4 January 2003, paper No. FMA2. Costa, L., Oliviera, P., 2001. Evolutionary algorithms approach to the
Also available at http://discovery.bits-pilani.ac.in/discipline/chemical/ solution of mixed integer non-linear programming problems. Computers
bvb/publications.html. & Chemical Engineering 25, 257–266.
R. Angira, B.V. Babu / Chemical Engineering Science 61 (2006) 4707 – 4721 4721
Diwekar, U.M., Rubin, E.S., 1993. Efficient handling of the implicit Nishida, N., Stephanopoulos, G., Westerberg, A.W., 1981. Journal review:
constraints problem for the ASPEN MINLP synthesizer. Industrial & process synthesis. American Institute of Chemical Engineers Journal 27,
Engineering Chemistry Research 32, 2006–2011. 321.
Diwekar, U.M., Grossmann, I.E., Rubin, E.S., 1992. An MINLP process Onwubolu, G.C., Babu, B.V., 2004. New Optimization Techniques in
synthesizer for a sequential modular simulator. Industrial & Engineering Engineering. Springer, Heidelberg, Germany.
Chemistry Research 31, 313–322. Papoulias, S., Grossmann, I.E., 1983. A structural optimization approach
Douglas, J.M., 1985. A hierarchical decision procedure for process synthesis. in process synthesis, Parts I–III. Computers & Chemical Engineering 7,
A.I.Ch.E. Journal 31, 353–362. 695–734.
Fan, H.Y., Lampinen, J., 2003. A trigonometric mutation operation to Rogalsky, T., Derksen, R.W., 2000. Hybridization of differential evolution for
differential evolution. Journal of Global Optimization 27, 105–129. aerodynamic design. Proceedings of the Eighth Annual Conference of the
Floudas, C.A., 1995. Nonlinear and Mixed-integer Optimization. Oxford Computational Fluid Dynamics Society of Canada, pp. 729–736.
University Press, New York. Rudd, D.F., Powers, G.J., Siirola, J.J., 1973. Process Synthesis. Prentice-Hall,
Floudas, C.A., Ciric, A.R., Grossmann, I.E., 1986. Automatic synthesis of Englewood Cliffs, NJ.
optimum heat exchanger network configurations. American Institute of Ryoo, H.S., Sahinidis, N.V., 1995. Global optimization of nonconvex NLPs
Chemical Engineers Journal 32, 276–290. and MINLPs with applications in process design. Computers & Chemical
Floudas, C.A., Aggarwal, A., Ciric, A.R., 1989. Global optimum search Engineering 19 (5), 551–566.
for nonconvex NLP and MINLP problems. Computers & Chemical Salcedo, R.L., 1992. Solving nonconvex nonlinear programming problems
Engineering 13, 1117–1132. with adaptive random Search. Industrial & Engineering Chemistry
Grossmann, I.E., 1985. Mixed-integer programming approach for the Research 31, 262.
synthesis of integrated process flowsheets. Computers & Chemical Stephanopoulos, G., Westerberg, A.W., 1976. Studies in process synthesis:
Engineering 9, 463. part II. Evolutionary synthesis of optimal process flowsheets. Chemical
Grossmann, I.E., Sargent, R.W.H., 1979. Optimum design of multipurpose Engineering Science 31, 195–204.
chemical plants. Industrial & Engineering Chemistry Process Design Storn, R., 1995. Differential evolution design of an IIR-filter with
Development 18, 343. requirements for magnitude and group delay. Technical Report,
Kocis, G.R., Grossmann, I.E., 1987. Relaxation strategy for the structural International Computer Science Institute. Available via Internet:
optimization of process flow sheets. Industrial & Engineering Chemistry ftp.icsi.berkeley.edu/pub/techreports/1995/tr-95-026.pdf.
Research 26, 1869–1880. Storn, R., Price, K., 1997. DE-a simple and efficient heuristic for global
Kocis, G.R., Grossmann, I.E., 1988. Global optimization of nonconvex mixed- optimization over continuous space. Journal of Global Optimization 11
integer nonlinear programming (MINLP) problems in process synthesis. (4), 341–359.
Industrial & Engineering Chemistry Research 27, 1407–1421. Stumberger, G., Dolinar, D., Pahner, U., Hameyer, K., 2000. Optimization of
Kocis, G.R., Grossmann, I.E., 1989. A modeling and decomposition strategy radial active magnetic bearings using the finite element technique and the
for the MINLP optimization of process flowsheets. Computers & Chemical differential evolution algorithm. IEEE Transactions on Magnetics 36 (4),
Engineering 13, 797–819. 1009–1013.
Li, H.L., 1992. An approximate method for local optima for nonlinear mixed Wang, F.S., Chiou, J.P., 1997. Optimal control and optimal time location
integer programming problems. Computers & Operation Research 19 (5), problems of differential-algebraic systems by differential evolution.
435–444. Industrial & Engineering Chemistry Research 36, 5348–5357.
Linnhoff, B., 1981. In: Mah, R.S.H., Seider, W.D. (Eds.), Foundations of
Computer Aided Chemical Process Design: Engineering Foundation, New
York, vol. II, 537–572.