JFS E: Food Engineering and Physical Properties

Multiobjective Optimization
Approach: Thermal Food Processing
A. ABAKAROV, Y. SUSHKOV, S. ALMONACID, AND R. SIMPSON

ABSTRACT: The objective of this study was to utilize a multiobjective optimization technique for the thermal steril-
ization of packaged foods. The multiobjective optimization approach used in this study is based on the optimization
of well-known aggregating functions by an adaptive random search algorithm. The applicability of the proposed ap-
proach was illustrated by solving widely used multiobjective test problems taken from the literature. The numerical
results obtained for the multiobjective test problems and for the thermal processing problem show that the pro-

E: Food Engineering &

Physical Properties
posed approach can be effectively used for solving multiobjective optimization problems arising in the food engi-
neering field.
Keywords: aggregating functions, multiobjective optimization, thermal processing of packaged foods

Introduction 1999a, 1999b; Alves and Almeida 2007) and are based on the uti-

I t is well known that the majority of real-life optimization prob-

lems are of a multiobjective nature with conflicting objectives
(particular objective functions), where it is necessary to compute
lization of genetic algorithms. Over the last 15 y, a variety of
evolutionary methods have been proposed for handling multiob-
jective optimization problems. The most important among them
more than 1 optimal solution. These solutions are called non- are:
dominated or Pareto-optimal solutions (Steuer 1985). Each of the r Nondominated sorting genetic algorithms (NSGA-I, NSGA-II)
Pareto-optimal solutions can be considered as a final “compro- (Deb 1999a, 1999b; Deb and others 2004),
mise” solution of a multiobjective optimization (MOO) problem r Niched-Pareto genetic algorithms (NPGA-I, NPGA-II) (Horn and
because it has no a priori advantage over other Pareto-optimal so- Nafpliotis 1993; Erickson and others 2001, 2002; Grandinetti and
lutions. Therefore, the ability to compute the maximum possible others 2007),
Pareto-optimal solutions is very important. r Multiobjective genetic algorithms (MOGA) (Fonseca and Flem-
Various optimization methods to solve MOO problems have ing 1993; Leung and others 2008),
been proposed over the last few decades. These methods can be r Strength Pareto evolutionary algorithms (SPEA, SPEA2) (Zitzler
divided into 2 classes; those belonging to the 1st class are called and Thiele 1999; Zitzler and others 2001; Barán and others 2005),
aggregating functions (Andersson 2000) and consist of transform- r Pareto archived evolution strategy (PAES) (Joshua and Corne
ing the MOO problems into a single global optimization problem 2000),
such that their optimal solutions for several chosen parameters r Pareto envelope-based selection algorithms (PESA) (Corne and
yield 1 Pareto-optimal point. The optimization of the obtained others 2000),
single optimization problem can be done by any of the existing r Micro-genetic algorithms (Coello Coello and Pulido 2001;
optimization methods including deterministic ones, like, for Sardiñas and others 2006).
example, gradient-based or various direct search algorithms The great advantage of the multiobjective evolutionary algo-
(Himmelblau 1972). However, the implementation of such meth- rithm is that its nature allows an entire set of multiobjective so-
ods in practice is usually avoided, in view of the multiextremal na- lutions to be evolved in a single run of the algorithm, instead of
ture of the single optimization problems to be solved. Other global having to perform a series of separate runs, as in the case of the ag-
optimization algorithms such as genetic algorithms (Goldberg gregating function method (Deb 1999a, 1999b; Sarkar and Modak
1989), simulated annealing (Czyzak and Jaszkiewicz 1998; Bandy- 2005). Other disadvantages of the aggregating function method are
opadhyay and others 2008), complex methods (Erdogdu and Bala- the following:
ban 2003), tabu searches (Cavin and others 2004; Jaeggi and others r its sensitivity to the shape of a Pareto-optimal front;
2008), and other types of random search methods are preferable in r the features of the particular multiobjective functions;
this case. r the spread of Pareto-optimal solutions;
The methods of this 2nd class are called multiobjective evo- r the requirements for specific knowledge of the problem to be
lutionary algorithms (Horn and others 1994; Srinivas and Deb solved.
1994; Fonseca and Fleming 1995; Zitzler and Thiele 1998; Deb These could, however, be avoided by choosing the appropriate
optimization algorithm and aggregating functions. It should also
MS 20090344 Submitted 4/17/2009, Accepted 7/4/2009. Authors Abakarov, be noted that aggregating function methods are very easy to im-
Almonacid, and Simpson are with Dept. of Chemical and Environmen- plement and computationally efficient.
tal Engineering, Univ. Técnica Federico Santa Marı́a, P.O. Box 110-V, Val-
paraı́so, Chile, and Centro Regional de Estudios en Alimentos Saludables, A potential disadvantage of the genetic algorithms is the need
Blanco 1623 room 1402, Valparaı́so, Chile. Author Sushkov is with Math- for many more function evaluations in comparison with other
ematics Dept., St. Petersburg Univ., Universitetsky av. 28, Petrodvorets, St. stochastic global optimization algorithms, which is the effect of the
Petersburg 198504, Russia. Direct inquiries to author Abakarov (E-mail:
alik.abakarov@usm.cl). parameterization of critical dependences on genetic algorithms
(Solomatine 1998, 2005). For example, the main criticism of the


C 2009 Institute of Food Technologists
R
Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E471
doi: 10.1111/j.1750-3841.2009.01348.x
Further reproduction without permission is prohibited
 Multiobjective optimization approach . . .

NSGA approach has been the need for specifying the tunable shar- Basic Principles of Multiobjective Optimization
ing parameter σ share . Additionally, this approach has high compu-
tational complexity, including nondominated sorting and a lack of Pareto-optimal solutions
elitism (Sarkar and Modak 2005). A general multiobjective optimization problem can be formu-
Multiobjective optimization has been rarely been implemented lated as follows:
in the food industry. One of the earliest studies concerned with
MOO utilization was published by Nishitani and Kunugita (1979), (x) =  f1 (x) , f2 (x) , . . . , fl (x) → min (1)
x∈X
where the optimal flow pattern in a multi-effect evaporator sys-
tem was determined (Seng and Rangaiah 2008). Over the last few
where:
decades around 40 papers on MOO have been published, where
X ⊂ Rn is a nonempty set of feasible decisions (a proper subset
a variety of applications were proposed for the food industry sec-
of Rn ),
tor (Kopsidas 1995; Kiranoudis and others 1999; Kiranoudis and
x = x1 , x2 , . . . , xn  ∈ X is a real n − vector decision variable, and
Markatos 2000; Krokida and Kiranoudis 2000; Therdthai and oth-
f i :Rn → R are particular multiobjective functions.
ers 2002; Erdogdu and Balaban 2003; Gergely and others 2003;
We assume that all of the constraints are included in the partic-
Hadiyanto and others 2009). The most interesting of such applica-
ular objective functions (1) by utilizing the penalty functions (Him-
E: Food Engineering &

tions was described in Seng and Rangaiah (2008).

Physical Properties

melblau 1972).
Thermal processing is an important method of food preserva-
If no vector x∗ = x∗1 , x∗2 , . . . , x∗n  ∈ X exists such that x∗ =
tion in the manufacture of shelf-stable packaged foods, and it has
arg minx∈X fi (x), ∀i ∈ 1:l, that is if no vector exists that is optimal
been the cornerstone of the food processing industry for more than
for all objectives concurrently, then there is no unique optimal
a century (Teixeira 1992). The basic function of a thermal process
solution—if it exists we call such a solution a utopian solution—and
is to inactivate bacterial spores of public health significance as well
a concept of acceptable solutions is needed. The subset W P(X) =
as food spoilage microorganisms in sealed containers of food, us-
{xp ∈ X: such that there does not exist an x ∈ X with f i (x) ≤ f i (xp ),
ing heat treatments at temperatures well above the ambient boiling
∀i ∈ 1:l} is called the set of Pareto-optimal solutions of the problem
point of water in pressurized steam retorts (autoclaves). Excessive
(1). Pareto-optimal solutions are the only acceptable solutions of a
heat treatments should be avoided because they are detrimental to
multiobjective optimization problem, since any other solution can
food quality and underutilize plant capacity (Simpson and others
be improved. Pareto-optimal solutions are also known as nondom-
2003).
inated or efficient solutions. The space in E l formed by the points
Thermal process calculations, in which process times at spec-
of the set P(X) = {x | x ∈ W P(X) } is called a Pareto optimal frontier
ified retort temperatures are calculated to achieve safe levels of
or front. Figure 1 provides a visualization of the definitions made
microbial inactivation (lethality), must be carried out carefully to
for the 2-dimensional problem (1) and 2 particular objectives.
assure public health safety and minimum probability of spoilage.
The following definition ensures whether a set of solutions be-
Therefore, the accuracy of the methods used for this purpose is of
long to a local Pareto-optimal set (Deb 1999a, 1999b).
importance to food science and engineering professionals working
Local Pareto-optimal solutions. The set W P ′ (X) is called a set
in this field (Holdsworth and Simpson 2007). Considerable work
of local Pareto-optimal solutions for any x′ ∈ W P ′ (X), when the fol-
has been reported in the literature, showing that variable retort
lowing condition holds:
temperatures (VRT) can be used to marginally improve the quality
of canned food and significantly reduce processing times in com-  
fi x′ ≤ fi (x) , ∀i ∈ 1 : l,
parison to traditional constant retort temperature (CRT) process-     (2)
ing (Teixeira and others 1975; Banga and others 1991, 2003; Chen x ∈ Bε (X) ⊂ X, Bε (X) = x x − x′  ≤ δ
and Ramaswamy 2002).
where δ is a small positive number.
Optimization of thermal sterilization is an optimal control prob-
lem, where it is necessary to search for the best retort temperature
as a function of process time. Banga and others (1991) showed that Methodology
this optimal control problem could be transformed into a nonlin-
ear programming (NLP) problem, and in most cases, the NLP prob- Multiobjective optimization approach
lem became a multimodal optimization problem with several types The multiobjective optimization approach used in this study is
of constraints. This problem has been solved by a series of opti- based on optimizing the following aggregating functions by us-
mization techniques, but only in the context of the single-objective ing the adaptive random search method (Sushkov 1969; Abakarov
optimization problem. These techniques include the gradient- and Sushkov 2002; Simpson and others 2008; Abakarov and others
based methods (Vassiliadis and others 1994), the stochastic method 2009).
known as integrated controlled random search (ICRS) (Banga and Function 1. Linear weighted sum aggregating function
l
Casares 1987), genetic algorithms (Chen and Ramaswamy 2002) 
(x) = λi fi (x) → min
and the adaptive random search method (ARSM) (Simpson and x∈X
i=1
others 2008). l (3)

The purpose of this study was to utilize a multiobjective opti- λi = 1, λi ≥ 0,
mization technique for solving the optimization problem of the i=1

thermal processing of packaged foods. The food quality factors where λi is the weight used for the ith particular objective function
(thiamine and texture retentions of pork puree) (Holdsworth and f i (x).
Simpson 2007) and processing time were considered as particu- As mentioned previously, this and the following aggregating
lar objective functions. The applicability of the proposed approach functions were used to transform the initial MOO problem (1) into
was further illustrated by solving widely used multiobjective test a single global optimization problem such that their optimal solu-
problems taken from the literature (Srinivas and Deb 1994; Deb tions for several chosen parameters give us 1 Pareto-optimal point.
1999a, 1999b). The procedure of computing the Pareto-optimal point, involving

E472 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

aggregating function 1, is simple: choose a value of the random Function 3. The penalty aggregating function,
weight vector λ and optimize the single-objective function (3). If
the solution obtained is optimal, then this solution belongs to the    l
(x) = fk xs + P j xs → min
 
set of Pareto-optimal ones (Deb 1999a). Thus, the aggregating func- (6)
x∈X
j=1
tion (3) allows the computation of any Pareto-optimal solution by
choosing a value of the weight vector λ and optimization of an ob-
tained single objective function, but in the case of convex Pareto- where k, k ∈ 1:l is a randomly chosen number at the 1st step of an
optimal front only (Steuer 1985; Deb 1999a). The approximation of adaptive random search of a particular objective function, f k (xs )
the Pareto-optimal front can be computed by generating the values is the value of the kth particular objective function at step s of the
of the weight vector λ from an appropriate probability distribution; ARSM, and P j (xs ) is the penalty function of the jth particular ob-
for example, from a uniform probability distribution, since a uni- jective function computed at step s of the ARSM. The following for-
form probability distribution does not require a priori knowledge mula is used to compute the penalties P j (xs ), j ∈ 1:l:
of the multiobjective problem to be solved.
P j xs = A  f j xs − f j xs−1  + f j xs − f j xs−1
          
Function 2. Weighted min–max aggregating function (7)

E: Food Engineering &

(x) = min max λi fi (x) where A is a sufficiently large number.

Physical Properties
x∈X i∈1:l
l
 (4) The optimization process of function (6) could be interpreted as
λi = 1, λi ≥ 0. minimization of the kth particular objective function over x subject
i=1 to the constraints f j (xs ) ≤ ξ j , j ∈ 1:l, where ξ j = f j (xs−1 ). The ap-
The procedure of computing the Pareto-optimal point, involving proximation of a Pareto-optimal front is computed by generating
aggregating function 2 consists of choosing a value of the random the k-values from a uniform probability distribution. The penalty
weight vector λ and optimizing the single-objective function (4). aggregating function is also applicable in the case of a nonconvex
The optimization process for 1 turn consists of minimization over Pareto-optimal front.
x of the particular objective function, which has a maximum value, The aggregating functions (3) and (4) are well known in MOO,
that is, function (4) performs a minimization of the maximum pos- and function (5) is proposed by the authors.
sible weighted loss. Function (4) enables the finding of solutions of
the nonconvex Pareto-optimal fronts, and as before, a priori knowl- Adaptive random search
edge of the system can be implemented by using a probability dis- The adaptive random search method belongs to a specific class
tribution to generate the values of weight vector, λ. The uniform of global stochastic optimization algorithms (Zhigljavsky 1991;
probability distribution might be the most convenient choice. Zhigljavsky and Zilinskas 2008). This class of algorithms is based
The 3rd aggregating function could be considered a modification on generating the decision variables from a given probability dis-
of the following well-known ξ -perturbation (or ξ -constant): tribution, and the term ‘‘adaptive” consists of modifications to the
probability distribution utilized in the searching process, which,
(x) = min fk (x) (5) throughout the whole search process, act as minimum compu-
x∈X
tations of the objective function, locating global solutions. The
subject to:
pedestal probability distribution is utilized in the adaptive random
f j (x) ≤ ξ j , j ∈ 1 : l, k = j search. After every calculation of problem (1), the pedestal distri-
bution of x = x1 , x2 , . . . , xn  is modified so that the probability of
where ξ j are constants. In this case, the single optimization prob- finding the optimal value of the objective function is increased.
lem is constructed by choosing all but 1 particular objective as Let I r ⊂ X be a prospective interval for variable xr , r ∈ 1:n and
constraints. Only one is used as the objective function. The approx- 2q be the width of each prospective interval. Let I be a Cartesian
imation of the Pareto-optimal front can be computed by progres- product of sets I r , r ∈ 1:n. For the random search algorithm I is in-
sively changing the ξ -constant values. An important disadvantage terpreted as a prospective subdomain with center point x 0r , r ∈ 1:n.
in utilizing this function is that the knowledge of an appropriate The procedure of the adaptive random method can be described
range of ξ -constant values for each particular objective function is as follows:
required a priori. This disadvantage was eliminated in the follow- 1. Set the total number of random search iterations Ns .
ing aggregating function by utilization of the penalty functions, as 2. Set the center point x0r , r ∈ 1:n, as the center point of the initial
follows. prospective subdomain I.

Figure 1 --- Visualization of

2-dimensional multiobjective
optimization problem and 2 particular
objectives.

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E473

 Multiobjective optimization approach . . .

3. Generate a new value of vector xs ∈ X from the current

pedestal probability distribution F s (xs−1 ).
4. Compute the value of the objective function s = (xs ) and
calculate a minimal value of the objective function in step s by using
the formula
smin = min s , smin
 
(8)

5. If smin < s−1 0

min set the center point x = x .
s

6. Modify the pedestal probability distribution.

7. s: = s + 1.
8. If s < Ns then go to step 2.
9. End.
To modify the pedestal probability distribution (step 5), the fol-
lowing formulas are used:
r to calculate the probability density hs for step s, s ∈ 1:Ns , the fol-
E: Food Engineering &
Physical Properties

lowing formula is used:

1 − ps
hs = (9)
1 − vs

where vs is the volume of prospective subdomain of random

search which we have in step s (v0 = 1) and ps is the assumed
probability in step s where the optimal solution of problem (1)
belongs to the prospective subdomain I;
r to calculate current volume vs and assumed probability ps , the
basic random search method uses the following expressions:

1 1
qs+1 = qs (2ε) Ns , q0 = (10)
2

 n
vs = 2qs (11)

where ε is the given accuracy;

s s ( pmin − 1)
⎧
⎪
⎪ + 1, if 0 ≤ vs ≤ vmin,
smin
⎨
ps = s (12)
v (1 − pmin ) pmin − vmin
if vmin ≤ vs ≤ 1
⎪
⎪
⎩ + ,
1 − vmin 1 − vmin

Variables pmin and qmin (vmin = (qmin )n ) are 2 heuristic param-

eters in the adaptive random search which allow tuning the algo-
rithm to different types of problems in accordance with the known
physical properties of the system under consideration.
To generate a new value of vector xs ∈ X from the current
pedestal probability distribution (step 2), the following subroutine
is used.
1. ξ : = R AND(), where R AND() is a function which returns a
value uniformly distributed over (0, 1).
2. For r = 1 ton.
3. If ξ < hs then xrs = ar + (br − ar ) × R AND(), where ar and br
are the left and right bounds, respectively, of decision variable xr .
4. Else xrs = x0r − q + 2q × R AND().
5. End If.
6. End For.
Figure 2A, 2B, and 2C shows the example of the 2-dimensional
pedestal frequency distribution transformation during the whole
random search process. Figure 2A shows the pedestal frequency
distribution used for the 1st algorithm iteration. Because there
is no information about the optimization problem to be solved,
any point of the domain X could be chosen with the same
probability.
Figure 2 --- The 2-dimensional pedestal frequency distri-
During the search process, a random search generates ran- bution transformations during the entire random search
0 1 s
dom vector values x , x , . . . , x , calculates problem (1), accu- process.

E474 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

mulates information about the solved problem and transforms ii. C(t f ) ≥ C d , where C d is the desired volume-average final
the pedestal frequency distribution according to the computa- quality retention value and is calculated as shown in Eq. 16:
tions performed. Transformations consist of reducing the devia-
(T − Tre f )
⎡ ⎤
tion of the pedestal distribution (or prospective subdomain I ⊆ L R tf
2 ⎢ ln 10
  
X) around the mean (or the center point x0i , i ∈ 1:n), which is the C(t) = C 0 exp⎣− 10 z ⎦drdz (16)
⎥
L R2 0 0 Dre f 0
current best solution x0 ∈ X, f (x0 ) < f x j , ∀ j ∈ 1 : s. Figure 2B
 

shows that the pedestal frequency distribution can be obtained

in the middle of the search process, and now the points of do- Thermal process optimization problems have been studied by
main X , in terms of probability density, can be divided into 2 several researchers, but not as a multiobjective optimization (Teix-
nonoverlapping subsets, where h and H are the probability den- eira and others 1975; Simpson and others 1993, 2008; Chen and Ra-
sities of the nonprospective and the prospective subdomains, maswamy 2002; Banga and others 2003; Garcı́a and others 2005).
respectively. Figure 2C shows the pedestal frequency distribu- The following types of single thermal process optimization prob-
tion for the final algorithm iteration transformed to the δ − lems were considered by the authors mentioned previously:
function, and it is assumed that for any point x′ of the final 1. Find such a retort function, T rt (t), T low ≤ T rt (t) ≤ T hight , where
prospective domain I, the following condition holds: |x′i − x∗i | ≤ the final quality retention C(t) is maximized, while the final process
lethality, F d0 , is held to a specified minimum.

E: Food Engineering &

ε, ∀i ∈ 1:n.

Physical Properties
2. Find a retort function,T rt (t), T low ≤ T rt (t) ≤ T hight , such that
the final process time tf is minimized subject to the same lethal-
Multiobjective optimization of thermal processing
ity requirement above, while the quality retention must not fall be-
In the particular case of a cylindrical container with radius R and
neath some specified minimum;
height 2L, the mathematical model describing heat transfer by con-
3. Find a retort function,T rt (t), T low ≤ T rt (t) ≤ T hight , such that
duction is a mixed boundary problem, as follows (Teixeira and oth-
the final process time tf is minimized subject to the same lethal-
ers 1969):
ity requirement mentioned above, while the quality retention must
not fall beneath some specified minimum, and energy consump-
∂2T ∂2T

∂T 1 ∂T
=α 2
+ + 2 (13) tion must not exceed a specified maximum; minimum and maxi-
∂t ∂r r ∂r ∂z
mum values are computed at constant retort temperature profiles
(Simpson and others 1993).
where T is temperature, t is time, r and z are radial and vertical lo-
In the general case, the function T rt (t) over t ∈ [0:tf ] can be pa-
cations within the container, and alpha (α) is the thermal diffusivity
rameterized using Np points, and during each time interval tk′ =
of the product.
[tk , tk+1 ), k ∈ 0:(Np − 1), the value of T rt (tk′ ) remains constant at
The model has the following initial and boundary conditions (by
uk (Teixeira and others 1975; Banga and others 1991, 2003). How-
symmetry):
ever, in this case, the use of a cubic spline in approaching global
optimization problems with random search techniques can pro-
T (R, z, t) = Trt (t),
duce superior results over discrete step-wise functions (Simpson
T (r, L, t) = Trt (t), and others 2008), mainly because the cubic spline approximation
∂T allows for significantly reducing the number of decision variables
(0, z, t) = 0, (14) and therefore the necessary number of objective function compu-
∂r
∂T tations to reach the global solution. Therefore, the approximation
(r, 0, t) = 0,
∂z by the cubic spline was utilized in this study to find optimal vari-
T (r, z, 0) = Tin, able retort temperature profiles.
In the general case, the single-thermal-process optimization
where T rt (t), t ∈ [0, tf ] will be the retort temperature as a function problem utilizing the cubic spline approximation can be presented
of time, and T in is the initial temperature at t = 0. as:
The lethality constraint can be specified as follows:
(u1 , u2 , . . . , uNP −1 , t f ) = t f + P1 + P2 + P3 , t f ∈ [tle f t, tright ] (17)
i. F 0 (tf ) ≥ F d0 , where F d0 is the final required lethality and is cal-
culated as a function of time and temperature at the critical
where ui , i ∈ 1:(Np − 1) are the control variables, and tleft and tright
point (cold spot), normally the geometric center of the con-
are obtained from the following expressions:
tainer (in the case of conduction-heated canned foods), ac-
cording to the following equation: (Tlow − Tre f )
 tright
F0d = 10 z dt (18)
 tf (T − Tre f ) 0

F0 (t) = 10 z dt (15)
0 (Thight − Tre f )
 tle f t
F0d = 10 z dt (19)
where T is the temperature at the critical point or cold spot, 0
and T ref is the reference temperature, normally equal to
121.1 ◦ C.
P 1 and P 2 are the penalty functions for lethality and nutrient
Quality retention, on the other hand, is greatly affected by
retention, respectively; the following expressions can be used as
the nonuniform temperature distribution within the package
penalty functions:
from the heated boundary to the cold spot, and it must be in-
tegrated in space over the volume of the container as well as tf

over time. To accomplish this integration over both space and A × F0d − F0 (t) +  F0d − F0 (t)
  
P1 = (20)
time, the following approach was used: t=0

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E475

 Multiobjective optimization approach . . .

tf 2. Convergence to the true Pareto-optimal front can be affected


P2 = A × (C d − C(t) + |C d − C(t)|) (21) by using a difficult g function (multimodal, deceptive, or others);
t=0
thus, this function tests the ability of an algorithm to handle diffi-
culties lateral to the Pareto-optimal front (Deb 1999b).
and P 3 is the penalty function that was used simultaneously with
3. Diversity in the Pareto-optimal front can be affected by choos-
the cubic spline approximation in the search process to hold the
ing an appropriate (nonlinear or multidimensional) f 1 function,
autoclave temperature profile T rt (t) in the given range [T low , T hight ]:
thus this function tests the ability of an algorithm to handle diffi-
tf culties along the Pareto-optimal front (Deb 1999b).
    
P3 = A × |Tlow − Trt (t)| + Thight − Trt (t) − Tlow − Thight (22) All computations were performed on an IBM PC workstation
t=0 with AMD Phenom 8650, Triple-Core CPU, 2.30 GHz, 1.75 GB of
RAM, and Borland C++ Builder 6.0 was utilized to prepare all the
Use of these kinds of penalty functions will lead very quickly to necessaries for the numerical computations program.
finding x ′ ⊆ X such that all the given constraints are satisfied when The number of ARSM runs was set at 1000 for all numerical ex-
the random search is implemented. periments (1000 independent optimizations of each aggregating
In this study, the food quality factors of thiamine content and function were done to compute an approximation of the Pareto-
E: Food Engineering &

texture retention of pork puree were considered as particular ob- optimal fronts).
Physical Properties

jective functions. Each of the quality factors can be computed from Test Problem 1.
expression (16) with its corresponding Dref and z values. The last f1 = 4x1 → min,
x
chosen particular objective is the thermal process time; therefore,
the following multiobjective optimization of the thermal process f2 = g (x) × h ( f1 (x) , g (x)) → min, x1 , x2 ∈ [0, 1] ,
x
optimization problem considered in this study was:

where
1 (u), 2 (u), 3 (u) → min,
u∈U
⎧   
⎪ x2 − 0.2 2
subject to:
⎪
⎪ 4 − 3 exp − , if 0 ≤ x2 ≤ 0.4
0.02
⎪
⎪
⎨
g (x) =
F0 (t f ) ≥ F0d ,
  
x2 − 0.7 2
⎪
⎪
4 − 2 exp − , if 0.4 < x2 ≤ 1.
⎪
⎪
⎪
1 (u) ≥ C 1d
⎩ 0.02
(23)
2 (u) ≥ C 2d ,

T l ≤ 3 (u) ≤ T r , ⎧
f1 (x)
α
⎨1−
⎪ , if f1 (x) ≤ g (x) ,
h ( f (x) , g (x)) = g (x)
where U is the domain of control variables ui , i ∈ 1:(Np − 1), 1 is ⎪
0, otherwise,
⎩
thiamine retention multiplied by −1, 2 is texture retention multi-
plied by −1, 3 is thermal process time, C d1 and C d2 are desired re-
tention values and T l and T r are left and right limits of the process where
time, respectively. ⎧
⎨ 0.25, if x2 < 0.4,
α=
Results and Discussion ⎩ 4, if x2 ≥ 0.4.

Multiobjective test problems

Several different types of multiobjective optimization problems This problem has a convex global Pareto-optimal front and 2
constructed from single-optimization problems (Srinivas and Deb local ones, nonconvex and convex. Figure 3 illustrates the nature
1994; Deb 1999a, 1999b) were utilized in this study to illustrate the of Test Problem 1 and was obtained by utilizing the aggregation
workability and practical effectiveness of the proposed approach. function (3); 50 computations of this function were done by ARSM.
The general form of the MOO test problems utilized in this study Figure 4 to 6 show the results computed by utilization of the aggre-
can be written as follows: gating functions (3), (4), and (5), respectively. The number of aggre-
gating function computations for each independent run was equal
f1 = f1 (x) → min to 250 (this number of computations is required for obtaining more
x
(24) than 95% of the Pareto-optimal points by optimization of the ag-
f2 = g(x) × h( f1 (x), g(x)) → min .
x gregating function of best convergence). Figure 4 to 6 depict that
the best convergence was shown by the linear weighted sum ag-
The choosing of appropriate functions for f 1 , g, and h, in MOO gregating function because the single-optimization problem (3) is
problems having specific features can be obtained as follows (Deb easier to solve than the single-optimization problems given by ag-
1999b): gregating functions (4) and (5). Of the Pareto-optimal points, 76.5%
1. Convexity or discontinuity in the Pareto-optimal front can and 81.5% were obtained by using aggregating functions (4) and
be affected by choosing an appropriate h function, thus this (5), respectively (Figure 5 and 6). All aggregating functions used
function tests the ability of the selected algorithm to tackle mul- showed good ability in generating uniformly distributed Pareto-
tiobjective problems having convex, nonconvex or discontinuous optimal points. The maximum computation time spent by aggre-
Pareto-optimal fronts (different shapes of the Pareto-optimal front) gating function (5) for the computation of Pareto-optimal points for
(Deb 1999b). Test Problem 1 did not exceed 6 s.

E476 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

E: Food Engineering &

Physical Properties
Figure 3 --- Graphical illustration of the nature of Test Problem 1.

Figure 4 --- Results obtained from aggregating function (2) for Test Problem 1 after 250 computations.

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E477

E: Food Engineering & Multiobjective optimization approach . . .
Physical Properties

Figure 5 --- Results obtained from aggregating function (3) for Test Problem 1 after 250 computations.

Figure 6 --- Results obtained from aggregating function (4) for Test Problem 1 after 250 computations.

E478 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

Test Problem 2. Test Problem 4.

f1 = x1 → min, f1 = x1 → min,
x
x

f2 = g (x) × h ( f1 (x) , g (x)) → min, x1 , . . . , xd ∈ [0, 1] , f2 = g(x) × h( f1 (x), g(x)) → min,

x x

x1 ∈ [0, 1], x2 , . . . , x10 ∈ [−5.12, 5.12],

where
0.5 where:
⎧
⎨1− f1 (x)
if f1 (x) ≤ g (x) ,
⎪
,
h ( f1 (x) , g (x)) = g (x) 10 10 
⎪ 1  2  xi
⎩
0, otherwise, g (x) = xi − 10 cos √ ,
4000 i=2 i=2
i

⎛ ⎞
d d
  0.5
xil ⎟
⎧
xi − ⎨1− f1 (x)
if f1 (x) ≤ g (x) ,
⎜ ⎪
,

E: Food Engineering &

Physical Properties
⎜ ⎟
⎜ i=2
g (x) = gmin + (gmax − gmin ) ⎜ d
i=2
⎟,
⎟ h ( f1 (x), g(x)) = g(x)
d ⎪
0, otherwise.
⎜  ⎟ ⎩
r l⎠
⎝ x − x i i
i=2 i=2

This test problem also has multimodal Pareto-optimal fronts.

where d = 5, g min and g max are the minimum and maximum values
The results obtained by aggregating functions (3), (4), and (5) with
of the function g(x), respectively, and xil , xir are the minimum and
8000 computations are shown in Figure 13, 14, and 15, respectively.
maximum values of the variable xi .
A lesser ability in generating uniformly distributed Pareto-optimal
The Pareto-optimal front of problem 2 is nonconvex; there-
points was demonstrated by the penalty aggregating function (5),
fore, the weighted sum aggregating function cannot be utilized in
and then 93% of the Pareto-optimal points were obtained; almost
this case. Results obtained by aggregating functions (4) and (5),
the same optimization abilities were shown by aggregating func-
with 2500 computations, are shown in Figure 7 and 8, respectively.
tions (3) and (4). The maximum computation time spent by aggre-
For the chosen number of computations, more than 95% of the
gating function (5) for the computation of Pareto-optimal points for
Pareto-optimal points were obtained by optimization of function
Test Problem 4 did not exceed 40 s.
(3) and 71.5% by using aggregating function (4). Figure 7 and 8
The following conclusions arise from the numerical simulations
show that function (4) has a better ability in generating uniformly
presented previously:
distributed Pareto-optimal points than function (5). The maxi-
1. Utilization of the linear weighted sum function (3) is preferred
mum computation time spent by aggregating function (5) for the
in the case of convex Pareto-optimal fronts because of the ease with
computation of Pareto-optimal points for Test Problem 2 did not
which it is solved in comparison with aggregating functions (4) and
exceed 15 s.
(5).
Test Problem 3.
2. Utilization of the penalty aggregating function (5) can give an
f1 = x1 → min, f2 = g(x) × h( f1 (x), advantage over the weighted min–max function (4); this is impor-
x
tant in the case of a nonconvex Pareto-optimal front, when aggre-
g(x)) → min, x1 ∈ [0, 1], x2 , x3 ∈ [−5.12, 5.12], gating function (3) is completely unusable, and the ξ -perturbation
x
aggregating function, which requires the ξ -constant values a priori
where: for each particular objective function.
3
3. The approach used in this study has shown good ability in
generating uniformly distributed Pareto-optimal points over the

g (x) = 1 + 10 (d − 1) + xi2 − 10 cos (2π xi ),
i=2 global Pareto-optimal frontier.
4. In the case of an a priori unknown Pareto-optimal front, all
aggregating functions presented can be utilized simultaneously
⎧ 0.5
⎨1− f1 (x) to guarantee obtaining uniformly distributed Pareto-optimal solu-
if f1 (x) ≤ g (x) ,
⎪
,
h ( f1 (x) , g (x)) = g (x) tions.
⎪
0, otherwise.
⎩
Multiobjective optimization of thermal processing
Test Problem 3 has a multimodal Pareto-optimal front (Fig- All the parameters utilized in this study are presented in Table 1
ure 9), which is the effect of the multimodal nature of the (Garcı́a and others 2006; Holdsworth and Simpson 2007). As a 1st
function g(x). The results obtained by utilization of aggregating step in the multiobjective optimization problem (23), it was nec-
functions (3), (4), and (5) with 5000 objective function computa- essary to find all those combinations of constant retort tempera-
tions are shown in Figure 10 through 12. Using aggregating func- ture (CRT) and process time, for the conditions listed in Table 1,
tions (4) and (5), 95% and 86.5% of the Pareto-optimal points were which would deliver the same final lethality. In this example, a tar-
obtained, respectively. The aggregating function (5) showed less get lethality of F d0 = 6 min was chosen to produce the iso-lethality
ability in generating uniformly distributed Pareto-optimal points in curve shown in Figure 16. Each point on this curve defines a con-
comparison with functions (3) and (4). The best numerical result stant retort temperature and process time resulting in a final target
was obtained by utilizing the aggregation function (3). The maxi- lethality of F d0 = 6 min.
mum computation time spent by aggregating function (5) for the For each of these equivalent processes, the final level of thiamine
computation of Pareto-optimal points for Test Problem 3 did not and texture retentions was calculated and is presented in Figure 16.
exceed 30 s. Figure 16 shows that the maximum level of thiamine retention with

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E479

E: Food Engineering & Multiobjective optimization approach . . .
Physical Properties

Figure 7 --- Results obtained from aggregating function (3) for Test Problem 2 after 2500 computations.

Figure 8 --- Results obtained from aggregating function (4) for Test Problem 2 after 2500 computations.

E480 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

E: Food Engineering &

Physical Properties
Figure 9 --- Graphical illustration of the nature of Test Problem 3.

Figure 10 --- Results obtained from aggregating function (2) for Test Problem 3 after 5000 computations.

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E481

E: Food Engineering & Multiobjective optimization approach . . .
Physical Properties

Figure 11 --- Results obtained from aggregating function (3) for Test Problem 3 after 5000 computations.

Figure 12 --- Results obtained from aggregating function (4) for Test Problem 3 after 5000 computations.

E482 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

E: Food Engineering &

Physical Properties
Figure 13 --- Results obtained from aggregating function (2) for Test Problem 4 after 8000 computations.

Figure 14 --- Results obtained from aggregating function (3) for Test Problem 4 after 8000 computations.

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E483

 Multiobjective optimization approach . . .

a constant retort temperature (CRT) was approximately 55% over retention were obtained at 128 ◦ C constant temperature and 69 min
the range of possibilities, which could be as low as 36%. In addi- of processing time. Taking into account the value of texture reten-
tion, the maximum and minimum levels of the texture retention tion at 116 ◦ C with a corresponding processing time of 112 min, and
were 46% and 27%, respectively. The optimum CRT process for the the value of thiamine retention at 128 ◦ C and 69 min, we obtained
product chosen in terms of thiamine retention is one in which the 2 solutions for thermal processing (see Table 2).
retort temperature is held constant at 116 ◦ C, with a corresponding As a 2nd step, the Pareto-optimal solutions of the ther-
processing time of 112 min. The optimal conditions for the texture mal processing optimization problem (23) were computed by
E: Food Engineering &
Physical Properties

Figure 15 --- Results obtained from aggregating function (4) for Test Problem 4 after 8000 computations.

Figure 16 --- Iso-

lethality curve for F d0 =
6 min, showing
equivalent process
combinations of retort
temperature and
process time, and
curves showing
thiamine and texture
retentions (quality
attributes).

E484 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

utilizing each of the aggregating functions (3), (4), and (5) The 800 computations of each aggregating function were made
and the VRT profiles. For problem (23) the values 0 and by the ARSM to obtain the nondominated thermal processing so-
115 were taken as left (T l ) and right (T r ) limits of the pro- lutions (profiles). One of the obtained profiles is presented in
cess time, and the values 35% and 40% as the desired re- Figure 17. The values pmin = 0.75 and qmin = 0.35 were selected
tention values C d1 , C d2 , respectively. Table 3 through 5 present for all numerical experiments. The maximum computation time
the Pareto-optimal solutions obtained for thermal processing. spent by the aggregating function (5) for computation of the Pareto-
optimal solutions for thermal processing did not exceed 2000 s.
Table 1 --- Parameters utilized in the thermal processing
simulation study.
Can radius (m) 0.04375 Table 3 --- Summary of results from linear weighted sum
Can height (m) 0.1160 aggregating function.
Thermal diffusivity α(m2 s−1 ) 1 .5443 × 10−7 Linear weighted sum aggregating function
T 0 (◦ C) 45.00
Process Thiamine Texture
Microorganism Bacillus stearothermophilus time (min) retention (%) retention (%)
z M ,ref (◦ C) 10
T M ,ref (◦ C) 121.11 111 58 47
100 57 46

E: Food Engineering &

Physical Properties
Nutrient Thiamine 76 52 47
z N ,ref (◦ C) 25.0 71 50 47
D ref (min) 178.6 60 45 45
T N ,ref (◦ C) 121.11 52 40 44
Quality factor Texture 47 38 42
z N ,ref (◦ C) 45.0
D ref (min) 178.6
T N ,ref (◦ C) 121.11 Table 4 --- Summary of results from min–max aggregating
F d0 (min) 6.0 function.
C d1 (%) 35.0 Weighted min–max aggregating function
C d2 (%) 40.0
T l (min) 45 Process Thiamine Texture
T r (min) 115 time (min) retention (%) retention (%)
112 59 45
Table 2 --- Solutions obtained by CRT utilization. 91 56 47
Process Thiamine Texture 85 55 47
time (min) retention (%) retention (%) 69 49 48
65 47 47
112 55 42 61 45 46
69 48 46 50 39 43

Figure 17 --- Optimum VRT profile for multiobjective thermal processing optimization problem for process time equal
to 100 min.

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E485

 Multiobjective optimization approach . . .

Table 5 --- Summary of results from penalty aggregating Chen C, Ramaswamy H. 2002. Modeling and optimization of variable retort temper-
function. ature (VRT) thermal processing using coupled neural networks and genetic algo-
rithms. J Food Engr 53(3):209–20.
Penalty aggregating function Coello Coello C, Pulido G. 2001. A micro-genetic algorithm for multiobjective opti-
mization. In: Zitzler E, Deb K, Thiele L, Coello Coello CA, Corne D, editors. First
Process Thiamine Texture international conference on evolutionary multi-criterion optimization. Zurich,
time (min) retention (%) retention (%) Switzerland: Springer-Verlag. Lecture Notes in Computer Science Nr 1993. p
126–40.
106 58 45 Corne D, Knowles J, Oates M. 2000. The Pareto envelope-based selection algorithm for
90 55 48 multiobjective optimization. In: Schoenauer M, Deb K, Rudolph G, Yao X, Lutton E,
80 53 47 Merelo JJ, Schwefel HP, editors. Proceedings of the Parallel Problem Solving from
76 51 48 Nature VI Conference. Paris, France: Springer. Lecture Notes in Computer Science
nr 1917. p 839–48.
56 45 43
Czyzak P, Jaszkiewicz A. 1998. Pareto simulated annealing—a metaheuristic tech-
49 38 43 nique for multiple-objective combinatorial optimization. J Multi-Criteria Decision
Anal 7:34–47.
Deb K. 1999a. Evolutionary algorithms for multi-criterion optimization in engi-
neering design. In: Miettinen K, Mäkelä M, Neittaanmäki P, Périaux J, editors.
Proceedings of Evolutionary Algorithms in Engineering and Computer Science
From Table 2 to 5, we can see that the VRT profiles are capable (EUROGEN-99). p 135–61.
of achieving higher levels of quality retention with lower process Deb K. 1999b. Multiobjective genetic algorithms: problem difficulties and construc-
tion of test problems. Evol Comput 7(3):205–30.
time, and thereby the VRT optimization involved in the multiobjec-
E: Food Engineering &

Deb K, Mitra K, Dewri R, Majumdar R. 2004. Towards a better understanding of

Physical Properties

tive optimization of thermal processing of food allows for obtaining the epoxy-polymerization process using multi-objective evolutionary computa-
tion. Chem Engr Sci 59:4261–77.
global Pareto-optimal solutions. Erdogdu F, Balaban M. 2003. Complex method for nonlinear constrained optimiza-
Depending on the real-world application, expert food engineers tion of thermal processing multi-criteria (muliti-objective function). Food Sci Hum
Nutr 26(3):303–14.
can choose 1 suitable “best” solution from among those presented Erickson M, Mayer A, Horn J. 2001. The niched Pareto genetic algorithm 2 applied
in Table 3 to 5. to the design of groundwater remediation systems. In: Zitzler E, Deb K, Thiele L,
Coello Coello CA, Covne D, editors. First international conference on evolutionary
multi-criterion optimization, Zurich, Switzerland: Springer-Verlag. Lecture Notes
Conclusions in Computer Science nr 1993.

T he multiobjective optimization approach was utilized to

demonstrate a technique for solving the problem of optimiza-
tion of the thermal sterilization process for packaged foods. A set of
Erickson M, Mayer A, Horn J. 2002. Multi-objective optimal design of groundwater
remediation systems: application of the niched Pareto genetic algorithm (NPGA).
Adv Water Res 25:51–65.
Fonseca C, Fleming J. 1993. Genetic algorithms for multiobjective optimization: for-
mulation, discussion and generalization. In: Stephanie F, editor. Proceedings of the
particular criteria that could be most important was chosen, and Fifth International Conference on Genetic Algorithms. San Mateo, Calif.: Morgan
the nondominated thermal processing profiles were computed. Kaufmann Publishers. p 416–23.
Fonseca CM, Fleming P. 1995. An overview of evolutionary algorithms in multiobjec-
With the expertise of food engineers, any of them can be chosen as tive optimization. Evol Comput 3:1–16.
a suitable “best” solution. It should be noted that the set of partic- Garcı́a M, Balsa-Canto E, Alonso A, Banga J. 2006. Computing optimal operating poli-
cies for the food industry. J Food Engr 74(1): 13–23.
ular criteria used in this study cannot be considered to be unique, Gergely S, Bekassy-Molnar E, Vatai G. 2003. The use of multiobjective optimization to
and, depending on a practical situation, this set can be changed, improve wine filtration. J Food Engr 58:311–6.
and the processing profiles can be recomputed. Goldberg D. 1989. Genetic algorithms for search, optimization, and machine learning.
Reading, Mass.: Addison-Wesley.
Applicability of the aggregating functions (3), (4), and (5) for dif- Grandinetti L, Guerriero F, Lepera G, Mancini M. 2007. A niched genetic algorithm
ferent types of Pareto-optimal fronts and the ability of the adaptive to solve a pollutant emission reduction problem in the manufacturing industry: a
case study. Comput Oper Res 34:2191–214.
random search method to solve the obtained single-optimization Horn J, Nafpliotis N. 1993. Multiobjective optimization using the niched Pareto ge-
problems in a reasonable computation time make the proposed netic algorithm. Technical report nr 93005, Univ. of Illinois at Urbana-Champaign,
U.S.A.
approach usable and effective for solving multiobjective food en- Horn J, Nafploitis N, Goldberg D. 1994. A niched Pareto genetic algorithm for multi-
gineering problems. objective optimization. Proceedings of the first IEEE Conference on Evolutionary
Computation. p 82–7.
Hadiyanto M, Boom R, Straten G, Boxtel A, Esveld D. 2009. Multi-objective optimiza-
Acknowledgments tion to improve the product range of baking systems. J Food Process Engr. DOI:
10.1111/j.1745-4530.2008.00240.x.
Authors Simpson and Abakarov are grateful for the financial sup- Himmelblau D. 1972. Applied nonlinear programming. New York: McGraw-Hill Book
port provided by CONICYT through the FONDECYT project nr Co., 1972.
Holdsworth D, Simpson R. 2007. Thermal processing of packaged foods. 2nd ed. New
1090628. York: Springer.
Jaeggi D, Parks G, Kipouros T, Clarkson P. 2008. The development of a multi-objective
Tabu Search algorithm for continuous optimisation problems. Eur J Oper Res
References 185:1192–212.
Abakarov A, Sushkov YU, 2002. The statistical research of random search. Mathemati- Joshua D, Corne D. 2000. Approximating the nondominated front using the Pareto
cal models. Theory and application. Saint-Petersburg: Saint-Petersburg State Univ. archived evolution strategy. Evolutionary Computation 8(2):149–72.
p 70–101. Kiranoudis C, Markatos N. 2000. Pareto design of conveyor-belt dryers. J Food Engr
Abakarov A, Sushkov YU, Almonacid S, Simpson R. 2009. Thermal processing opti- 46:145–55.
mization through a modified adaptive random search. J Food Engr 93:200–9. Kiranoudis C, Maroulis Z, Marinos-Kouris D. 1999. Product quality multi-objective
Alves M, Almeida M. 2007. MOTGA: a multiobjective Tchebycheff based genetic algo- dryer design. Drying Tech 17:2251–70.
rithm for the multidimensional knapsack problem. Comput Oper Res 34:3458–70. Knowles J, Corne D. 2000. Approximating the nondominated front using the Pareto
Andersson J. 2000. A survey of multiobjective optimization in engineering design. Re- archived evolution strategy. Evol Comput 8(2):149–72.
ports of the Dept. of Mechanical Engineering, LiTH-IKP-R-1097, Linkoping Univ., Kopsidas G. 1995. Multiobjective optimization of table olive preparation systems. Eur
Linkoping. J Oper Res 85:383.
Bandyopadhyay S, Saha S, Maulik U, Deb KA. 2008. Simulated annealing-based multi- Krokida M, Kiranoudis C. 2000. Pareto design of fluidized bed dryers. Chem Engr J
objective optimization algorithm: AMOSA. Evol Comput, IEEE Trans 12(3):269–83. 79:1.
Banga J, Casares J. 1987. ICRS: application to a wastewater treatment plant model. Leung S, Wong W, Mok P. 2008. Multiple-objective genetic optimization of the spatial
In: “Process Optimisation” (IChemE Symposium Series Nr 100), The Institution of design for packing and distribution carton boxes. Comput Ind Engr 54:889–902.
Chemical Engineers. Ed. Oxford, U.K.: Pergamon Press. p 183–92. Nishitani H, Kunugita E. 1979. Optimal flow-pattern of multiple effect evaporator sys-
Banga J, Perez-Martin R, Gallardo J, Casares J. 1991. Optimization of thermal process- tems. Comp Chem Engr 3:261.
ing of conduction-heated canned foods: study of several objective functions. J Food Sardiñas R, Reis P, Davim J. 2006. Multi-objective optimization of cutting parameters
Engr 14:25–51. for drilling laminate composite materials by using genetic algorithms. Compos Sci
Banga J, Balsa-Canto E, Moles C, Alonso A. 2003. Improving food processing using Technol 66:3083–8.
modern optimization methods. Trends Food Sci Technol 14(4):131–44. Sarkar D, Modak J. 2005. Pareto-optimal solutions for multi-objective optimization of
Barán B, Christian von Lücken, Sotelo A. 2005. Multi-objective pump scheduling op- fed-batch bioreactors using nondominated sorting genetic algorithm. Chem Engr
timisation using evolutionary strategies. Adv Engr Software 36(1, January):39–47. Sci 60(2, January):481–92.
Cavin L, Fischer U, Glover F, Hungerbuehler K. 2004. Multiobjective process design in Seng C, Rangaiah S. 2008. Multi-objective optimization in food engineering. Chapter 4
multi-purpose batch plants using a Tabu Search optimization algorithm. Comput in Taylor & Francis Book “Optimization in Food Engineering,” Ed. Ferruh Erdogdu.
Chem Engr 28(4):393–430. 800 p.

E486 JOURNAL OF FOOD SCIENCE—Vol. 74, Nr. 9, 2009

Multiobjective optimization approach . . .

Simpson R, Almonacid-Merino S, Torres J. 1993. Mathematical models and logic Teixeira A, Zinsmeister G, Zahradnik J. 1975. Computer simulation of variable retort
for the computer control of batch retorts: conduction-heated foods. J Food Engr control and container geometry as a possible means of improving thiamine reten-
20(3):283–95. tion in thermally-processed foods. J Food Sci 40(3):656–9.
Simpson R, Almonacid S, Teixeira A. 2003. Optimization criteria for batch retort bat- Therdthai N, Zhou W, Adamczak T. 2002. Optimization of the temperature profile in
tery design and operation in food canning-plants. J Food Proc Engr 25(6):515–38. bread baking. J Food Engr 55:41.
Simpson R, Abakarov A, Teixeira A. 2008. Variable retort temperature optimization Vassiliadis V, Pantelides C, Sargent R. 1994. Solution of a class of multistage dynamic
using adaptive random search techniques. J Food Control 19(11):1023–32. optimization problems. Problems without path constraints. Ind Engr Chem Res
Solomatine D. 1998. Genetic and other global optimization algorithms—comparison 33(9):2111–22.
and use in model calibration. Proc. 3rd Intl. Conf. Hydroinformatics, Copenhagen, Zhigljavsky AA. 1991. Theory of global random search. Dordrecht: Kluwer Academic
1998. p 1021–8. Publishers.
Solomatine D. 2005. Adaptive cluster covering and evolutionary approach: compar- Zhigljavsky A, Zilinskas AG. 2008. Stochastic global optimization. Ser. Springer op-
ison, differences and similarities. Proc. IEEE Congress on Evolutionary Computa- timization and its applications, Vol. 9. Berlin: Springer-Verlag. xvi+362 p. ISBN:0-
tion, Edinburgh, U.K., 1959–1966. 387-36720-0.
Srinivas N, Deb K. 1994. Multi-objective function optimization using non-dominated Zitzler E, Thiele L. 1998. Multiobjective optimization using evolutionary
sorting genetic algorithms. Evol Comput 2(3):221–48. algorithms—a comparative case study. In: Eiben AE, Back T, Schoenauer M,
Steuer R. 1985. Multiple criteria optimization: theory, computation and application. Schwefel HP, editors. Parallel problem solving from nature, V. Berlin, Germany:
New York: John Wiley & Sons. Springer. p 292–301.
Sushkov YU. 1969. Method, algorithm and program of random search. Leningrad: Zitzler E, Thiele L. 1999. Multiobjective evolutionary algorithms: a comparative
VNII Transmash. 43 p. case study and the strength Pareto approach. IEEE Trans Evol Comput 3(4):257–
Teixeira A. 1992. Thermal process calculations. Chapter 11 in Heldman DR, Lund 71.
DB, editors. Handbook of food engineering. New York: Marcel Dekker, Inc. p 563– Zitzler E, Laumanns M, Thiele L. 2001. SPEA2: improving the strength Pareto evo-
619. lutionary algorithm. Technical report 103, Computer Engineering and Networks

E: Food Engineering &

Teixeira A, Dixon J, Zahradnik J, Zinsmeister G. 1969. Computer optimization of nu- Laboratory (TIK), Swiss Federal Inst. of Technology (ETH) Zurich, Gloriastrasse 35,

Physical Properties
trient retention in thermal processing of conduction-heated foods. Food Technol CH-8092 Zurich, Switzerland.
23(6):137–42.

Vol. 74, Nr. 9, 2009—JOURNAL OF FOOD SCIENCE E487