Journal of Natural Gas Science and Engineering 23 (2015) 272e280

Contents lists available at ScienceDirect

Journal of Natural Gas Science and Engineering

journal homepage: www.elsevier.com/locate/jngse

A novel multi-objective estimation of distribution algorithm for

solving gas lift allocation problem
S. Omid H. Miresmaeili a, *, Peyman Pourafshary b, Farhang Jalali Farahani c
a

Institute of Petroleum Engineering, University of Tehran, Tehran, Iran

Department of Petroleum and Chemical Engineering, Sultan Qaboos University, Muscat, Oman
c
School of Chemical Engineering, University of Tehran, Tehran, Iran
b

a r t i c l e i n f o

a b s t r a c t

Article history:
Received 6 November 2014
Received in revised form
2 February 2015
Accepted 5 February 2015
Available online 14 February 2015

Gas lifting is a common practice in the oil industry. Using an appropriate gas lift injection rate can
ensure that the desired oil production rate would be achieved. In the case of an oil eld, the problem
of distributing a certain amount of the available gas among a number of wells is formally known as a
gas lift allocation problem. In this paper, a multi-objective optimization algorithm, based on the
Gaussian Bayesian Networks and the Gaussian kernels, is proposed in order to determine the best
injection points, considering multiple objective functions. Firstly, the problem is solved in a similar
approach to the previous literature with similar gas lift data and similar function approximation
method, to compare the performance of the proposed algorithm with the older ones. Thereafter, an
extended problem is discussed, with minimizing the water production as a new optimization criterion. The developed multi-objective scheme is capable of handling and optimizing a gas-lift problem
with several constraints and conicting objectives such as controlling the gas usage and increasing
the oil production, whereas in the conventional single-objective optimizations, any alteration in
the constraints demands a new optimization process and often there is no place for considering
an additional objective in the gas-lift allocation problem. The results obtained by the proposed
optimization algorithm signicantly overcame those reported in the previous similar literature. For a
single-objective fty-six well problem, the results exhibited 16.24% improvement in the total oil
production.
2015 Elsevier B.V. All rights reserved.

Keywords:
Gas lift allocation
Multi-objective optimization
Gaussian network
Gaussian kernel

1. Introduction
The demands for high oil production rate leads to a gas lift
operation, which consists of injecting a certain amount of highpressure gas through the tubing, and consequently, lowering
the hydrostatic pressure difference along the tube. However
eventually, at some points, the increase in the friction pressure
loss offsets the hydrostatic pressure reduction which, in turn,
limits the increase in the production rate and yields in the convexity of the gas lift performance (GLP) curve. Determining the
optimum injection rates in a multiple well production network is
known as the gas lift allocation problem. Many researchers have
tackled this problem from different points of view by using a

* Corresponding author.
E-mail address: omiresmaeili@alumni.ut.ac.ir (S.O. H. Miresmaeili).
http://dx.doi.org/10.1016/j.jngse.2015.02.003
1875-5100/ 2015 Elsevier B.V. All rights reserved.

wide range of optimization techniques, from the quasi-Newton

(Nishikiori et al., 1989) to the genetic algorithm (Martinez
et al., 1994). Alarcon et al. (2002) improved the quasi-Newton
optimization algorithm and offered a new mathematical model
to t the eld data. Ray and Sarker (2007) challenged the problem as a bi-objective problem through maximizing the oil production level and minimizing the gas usage rate as the
optimization criteria. They suggested a variant of multi-objective
genetic algorithm called, the non-dominated sorting genetic algorithm II (NSGA-II) as the optimization tool. Other approaches,
such as mixed-integer programming (Kosmidis et al., 2005) and
dynamic programming (Camponogara and Nakashima, 2006),
explored the additional aspects of the problem, including the
activation of wells and their interactions. Wang and Litvak (2008)
provided a heuristic method and considered the long-term
reservoir developments by incorporating a reservoir simulator.

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

They also used a multi-objective scheme to reduce the rate oscillations besides maximizing the oil production. However, the
nature of the local search does not even guarantee to nd local
solutions. Furthermore, Sukarno et al. (2009) extended the
operational condition to the case of a dual gas lift system. Zerafat
et al. (2009) presented a comparison of different methods to the
Genetic Algorithm and examined the capabilities of the algorithms in solving the uni-objective gas-lift optimization problem
with the constraints on the lift-gas injection rates. They also
proposed a variant of Ant Colony Optimization to deal with the
problem. It can be witnessed that with the exception of two cases
(Ray and Sarker, 2007; Wang and Litvak, 2008), the literature has
chiey focused on the single objective optimization of the gas lift
problem.
In this study, it is aimed to confront the problem as a multiobjective problem and expand the idea of multiple criteria optimization, since the presence of constraints in the real gas-lift
problems is an inevitable fact. In the conventional single objective approach, for every update in the constraints, the process of gas
lift allocation problem should be re-solved. For example, if the
amount of the available gas or the capacity of compressors changes,
a new solution is needed by the single objective approach. On the
other hand, the multi-objective approach provides a set of solutions
for different criteria, which makes it easier to compare several
production strategies altogether. In a multi-objective approach, it is
possible to simultaneously handle several conicting objectives in
the optimization process.
For instance, by a multi-objective method, it is likely to optimize the total oil production rate and minimize the number of
required wells at the same time, but through a single objective
method, the number of wells should be assumed as a constraint
for the problem.
Hence, since it is necessary to compare different scenarios for
the gas lift design, a multi-criteria optimization approach is an
efcient tool for solving such a multi-dimensional problem with
different constraints. The proposed optimization algorithm is an
estimation of the distribution algorithm (EDA) based on the
~ aga et al., 2000) and the
Gaussian Bayesian networks (P. Larran
Gaussian kernels (Socha and Dorigo, 2008). The general idea of an
EDA is to construct probabilistic models based on a set of promising
solutions, and guide the search process by sampling the probabilistic models. Furthermore, in order to adjust the algorithm for
solving the multi-objective problems, the concept of nondominating sorting (Deb et al., 2002) has been employed. The obtained results from the EDA are compared with those reported by
using the multi-objective genetic algorithm (Ray and Sarker, 2007)
due to the evolutionary nature of both algorithms and lack of any
other similar multi-objective optimization algorithm in the literature on the topic of gas lift optimization. It should be pointed out
that in the eld of petroleum engineering, very few works have
ever employed any variant of the estimation of distribution algorithms. For instance, Abdollahzadeh et al. (2013) utilized an EDABOA (Bayesian Optimization Algorithm) for history matching. To
the best of the authors' knowledge, there exists no study on
applying any multi-objective EDA in the literature of petroleum
engineering.

273

interpolation, should be selected to maintain the convexity of the

objective function.
Another technique is to use a piece-wise linear interpolation,
which is easier to be used, but cannot keep the smoothness of the
convex GLP curve. Generally, in such model selection problems, an
adequate care should be taken in order to avoid an over-tting of
the function approximation. For example, using a high-order
polynomial may seem to describe the given GLP data in a precise
manner, but it amplies the possibility of over-tting. Besides
producing a smooth convex curve, the use of a detailed and
complicated model for approximation of the GLP function can be
computationally expensive. Thus, it should be expected to sacrice
the accuracy of approximation of the shape of the GLP curve at the
expense of computational costs.
Nevertheless, it is worth to indicate that for a fair comparison
between the optimization methods, the objective functions and all
the constraints and conditions should be similar; otherwise,
comparing the results of an algorithm, using a quadratic objective
function with an algorithm by using a piece-wise linear objective
function, would be in vain. This is because of this fact that the
quadratic function may overestimate the function and yield better
solutions, and this does not properly demonstrate the advantage
and efciency of the optimization method.
To this end, the same GLP data (Buitrago et al., 1996) and the
same function approximation method (i.e., piece-wise linear
function) are applied in order to effectively compare the proposed
multi-objective optimization algorithm with the other applications
of the multi-objective optimization (Ray and Sarker, 2007). The
optimization problem in this case is a bi-objective optimization,
which maximizes the oil production rate as one criterion and
minimizes the gas usage as the other one:

max Z1

#ofX
Wells

Qo i

(1)

i1

minZ2

#ofX
Wells

 
Qg i

(2)

i1

Therefore, instead of using a xed constraint on the available

lift-gas, a criterion function for the total gas usage is considered
which, in turn, presents a bi-objective optimization problem.
In this manner, rather than searching for a single optimum solution, a number of solutions will be acquired, which are all
optimal. The resulting set of solutions can be used according to
the availability of the injection gas. Any further criteria can be
introduced in this way to create a vector of objective functions to
be optimized. At the end of this study, in order to extend
the idea of multi-criteria optimization and one-step advance
using its advantages, a case is explored in which another criterion is added to the problem, viz. minimizing the daily water
production. This extra criterion is to serve toward expanding the
case of bi-objective algorithm to the three-objective
optimization.
3. Optimization algorithm

2. Problem statement
In order to solve the conventional gas lift allocation problem,
the GLP curve for each well should be determined by using a
proper tting tool, in order to demonstrate the response of wells
to various gas injection rates. As mentioned earlier, the GLP is a
convex function. Hence, a quadratic function or a piece-wise
polynomial function approximation e i.e., cubic spline

An EDA in essence is a stochastic algorithm based on probabilistic models. Contrary to the other evolutionary algorithms
such as Genetic Algorithm, in an EDA, the search operators such
as crossover and mutation are substituted by probabilistic
models. In the conventional genetic algorithm, the solutions are
produced by mutating individual solutions and applying crossover to a pair of solutions (parents). In contrast, an EDA mutates

274

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

the probabilistic distributions and generates new solutions by

sampling the distributions developed during the optimization
process. In this research, an EDA based on the Gaussian kernels
and the Gaussian Bayesian networks has been employed. The
main operator of the algorithm is supposed to t the probabilistic
models to the current promising solutions, and construct new
solutions via sampling the probabilistic models. A Gaussian
network and Gaussian kernel are used as the probabilistic
models. The Gaussian (Bayesian) network can discover the
possible interactions between variables through search by
dening conditional probabilities among decision variables of the
optimization problem, while the Gaussian kernels can assist in
guiding the algorithm in the search space as an extra mechanism
to generate new solutions. This arrangement is an improvement
over the traditional EDAs, which utilizes a single mechanism for
generating solutions in order to guide the search for the optimum, since in this improved version, the relational-based
Gaussian networks with efcient Gaussian kernels have been
combined to produce new solutions for the optimization
problem.
The pseudo-code for the algorithm is as follows:
(a) Set t 1
(b) Initialize the population
(c) Until (the stopping criterion has not been met)
Select a percentage of the current individuals
Construct the probabilistic model (Gaussian network and
Gaussian kernel) based on the selected individuals
Generate new individuals by sampling the constructed
model
tt1
Fig. 1 depicts the owchart of the algorithm. Functions are
evaluated after each step of generating new solutions. Ranking, as a
core concept of the multi-objective scheme, is performed based on
these assessments.
In order to learn the structure of the Gaussian network, a simple
greedy search procedure based on the Bayesian Information Criterion (BIC) metric is employed. In the beginning, the search starts
with an empty graph (adjacent matrix), and then with a single-edge
addition operator, it tries to capture the underlying interactions
between the variables. The search continues until no further edge
addition would improve the metric. The BIC metric is dened according to (Ahn et al., 2006).

BICz

jSj
X

ln PX  0:5lnjSjjqj

(3)

j1

where BIC(z) refers to structure z score, S is the set of samples as

S fx1 ; ; xjSj g, P(X) is the probability distribution function (PDF)
of X, and jqj is the number of the PDF's parameters to be estimated.
The PDF can be encoded using a Bayesian factorization as follows:

PX

n
Y

Pxi jpxi

(4)

i1

where xi is node i in the directed acyclic graph (DAG) of the

Bayesian network and i is its associated set of parent nodes in the
DAG. In order to calculate the conditional probability distributions,
a linear Gaussian model is employed, which denes each node Y as
a function of other nodes with a set of regression coefcients
bo,,bk and the variance s2 (Koller and Friedman, 2009):



PYjx1 ; ; xk N bo b1 x1 bk xk ; s2

(5)

Generating new solutions is accomplished through sampling

the factorizations based on the probabilistic logic sampling (PLS)
(Henrion, 1988), in order to consider the weak ordering implied by
the DAG and guide the generation of the instantiations of the nodes
by their probability.
For an efcient guidance of the optimization algorithm, the idea
of constructing the Gaussian kernel PDFs (a weighted sum of
multiple one-dimensional Gaussian functions) is added to the algorithm. For each kernel, the value of the ith variable of the selected
solutions is used as the mean value for that variable, mi. In addition,
a vector of weights is created according to the following formula:

wl

2
1
pe0:5l1=qN
qN 2p

(6)

In Eq. (6), wl refers to the weight of solution l, q is the selection

pressure, and N is the number of solutions (population size).
Dividing the weights by the sum of weights gives the probability of
choosing each Gaussian function for sampling. Considering the
standard deviations of each variable and each PDF, sli is dened as
the average distance from the chosen solution (Sl) to the other
solutions:

sli





N Si  Si 
X
e
l
e1

(7)

N1

The Gaussian kernel for decision variable i is calculated as:

!2
Gi x

Fig. 1. The owchart of the EDA.

N
X

0:5

1
wl pe
i 2p
s
l1
l

xmi
l
si
l

(8)

The sampling is completed by choosing a Gaussian PDF in

accordance with the probabilities, and sampling the chosen
Gaussian function by using a standard normal random number
generator (Socha and Dorigo, 2008).
So far the proposed algorithm can only be used for a singleobjective optimization. In order to enhance the algorithm for a
multi-objective algorithm, some additional congurations must
be taken into account. In other words, in addition to combining
the Gaussian networks and Gaussian kernels for improving the
EDA, the multi-objective concepts are also incorporated in order

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

to enhance the algorithm one more level. It should be mentioned

that even though the traditional EDAs are highly efcient in
solving high-dimensional complex optimization problems, they
cannot be used for tackling their multi-objective problems in their
conventional forms. For this purpose, the idea of non-dominated
sorting is employed, in which, the individuals are ranked based
on non-domination and crowding distance (Deb et al., 2002). The
individuals are compared with each other based on the domination condition. If the individual X1 dominates the individual X2,
f(X1) will not be worse than f(X2) in all objectives and will be better
in at least one objective (denoted X1 _ X2). Assuming that there
are m objectives (fi(x),i2M {1,2,...,m}), X1 _ X2 if and only if
(ci2M:fi(x)  fj(y))(di2M:fi(x) < fj(y)). After the non-dominated
sorting, the parameter (crowding distance) is calculated based on
the position of the corresponding individual in the Pareto front
(distance from the neighboring individuals). In fact, the crowding
distance ensures the diversity in the Pareto front. It should be
noted that these calculations are performed in the objective
function space. Eventually, every individual has a non-dominated
rank and a crowding distance. When comparing two individuals
with different ranks, the lower rank has the priority and if both
individuals belong to the same front, the solutions with the higher
crowding distance (i.e., the non-crowded solutions) are preferred.
In this manner, the EDA would be capable of tackling the uniobjective optimization problems, as well as the multi-criteria
problems.
Some of the advantages of using such an algorithm are as follow:
 It is a derivative-free method, so can be used for any function
whose derivatives are not available or desirable (or affordable)
to be approximated or computed.
 It is very efcient in dealing with the multi-objective problems,
since it can provide a complete Pareto-front in a single run.
 Using probabilistic models, it can extract more information
about the problem being solved, and therefore, can provide a
better solution generating mechanism in comparison to
the older algorithms such as genetic algorithm, which utilizes
xed operators such as mutation and crossover. Using
explicit probability distributions provides a better understanding of the problem domain. To elaborate, the conditional
probability described by the probabilistic model (built
around the current set of solutions) can be used to understand
the relationships and particularly the dependencies
between variables. Using the extra information, the EDA is
able to traverse the search space more intelligently and
efciently.
In general, the main advantages of an EDA over a genetic algorithm (GA) are the absence of multiple parameters to be tuned (e.g.,
crossover and mutation probabilities), and the expressiveness and
transparency of the probabilistic model that guides the search
process (as opposed to the implicit models dened by the other
metaheuristics). Although there are still some parameters, such as
the sample size and the initial number of individuals, present in the
implementation of an EDA, but the core search process of an EDA is
mostly dependent on its probabilistic models and particularly on
the mechanism of generating new solutions (i.e., offspring) from an
estimated distribution. In contrast to the EDAs, the classical genetic
algorithms use simple search operators such as two-point crossover and binary tournament selection, which implicitly dene a
probabilistic distribution model based on the current population
and the generated new population can be seen as samples of that
distribution (Hauschild and Pelikan, 2011). In addition, the EDAs
have been proven to be better suited to some applications than the
GAs, since they are able to achieve competitive and robust results in

275

~ anzas et al., 2008; Jarboui

the majority of tackled problems (Arman
et al., 2009; Su and Chow, 2012). Nevertheless, it is worth
mentioning that the simpler nature of a conventional GA can be
viewed as an advantage over the more complex structure of an EDA
in solving the small-sized optimization problems (with small
search space), in which applying a more advanced algorithm may
seem as extravagant and unreasonable. It is unquestionable that
most often, a practical gas lift optimization problem involves a large
search space with numerous decision variables. Hence, the simple
GA method is not powerful enough to handle such real large
problems.
4. Results and discussion
Prior to the discussion about the results and inspection of the
performance of the EDA in comparison with the past ndings, it is
important to notice the importance of the Number of Function
Evaluations (NFE) as the main tool for comparing the functioning of
different algorithms. Hence, the results are reported in terms of the
NFE and not the CPU time, since the CPU time is signicantly reliant
on the utilized programming languages, compilers, the programmers' skills, and also the machines used for running the experiments, while in contrast, the NFE allows a fair comparison
between the results obtained by various algorithms (Socha and
Dorigo, 2008).
As discussed earlier, in order to solve the bi-objective optimization problem and demonstrate the efciency of the EDA, the GLP
data for the six wells and fty-six wells from Buitrago et al. (1996) is
used. The piece-wise linear interpolation is applied as the function
approximation tool, owing to the similar approach of Ray and
Sarker (2007). For a three-objective problem, a ten-well model is
constructed and the cubic spline interpolation (piece-wise polynomial interpolation) is employed. In all of the following problems
solved by the EDA, the maximum number of parents for learning
the structure of the Bayesian network DAG is set to 1. The selection
pressure (q) is also adjusted to 0.6.
4.1. The six-well problem
The optimization is performed using the developed EDA for the
uni- and bi-objective problems. The results for the uni-objective
case are presented in Table 1.
It should be pointed out that the attained results are
compared only to Ray and Sarker (2007), because other researchers have used different function approximation methods,
and as a result, the comparison may not provide valid outputs.
The number of function evaluations (NFE) needed for achieving
the EDA results is less than 10,000. The initial population size
(individuals) and the number of samples generated in each
iteration are both 100. The gas injection rate for each well is
reported in Appendix A. It can be deduced from Table 1 that
there is rather no signicant improvement in the results and the
ndings are fairly similar, but as the number of wells increases,
the advantage of the EDA becomes more eminent, as can be
observed in the case for the fty-six-well problem. Additionally,

Table 1
The single objective six-well problem results.
Algorithm

Produced oil (STBD)

Gas usage (Mscf/day)

EDA
GA (Ray and Sarker, 2007)

3665.26
3663.99

4596.9
4597.4

276

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

Fig. 2. Multi-objective solution to six-well problem obtained by the EDA.

the results for the multi-objective function are demonstrated in

Fig. 2. It is worth noting that the principal difference of such a set
of optimal solutions, depicted in Fig. 2 from the conventional
uni-modal single objective gas lift optimization (in which there is
only one single optimum solution), is the existence of a number
of solutions that are all optimal. No solution can be claimed to be
better than the others, however, with a further specication of
the gas-lift problem (e.g., a specied amount of available gas), it
can be decided about the proper allocation of injection gas. In
other words, the main benet of this multi-objective approach is
that the Pareto front (Fig. 2), obtained from the multi-objective
optimization, encompassed the results equal to the case, where
the single-objective algorithm is run for multiple times and with
different constraints. Similar to the uni-objective EDA, for the
multi-objective EDA, the size of population and the sample size
were set to 100. For comparing the quality of multi-objective
solutions, the Pareto points reported by Ray and Sarker (2007)
were applied, as displayed in Table 2. Five Pareto points were
selected from the Pareto front (Fig. 2) to be compared with the
ve Pareto points stated by Ray and Sarker (2007) for the same
problem. The results disclose rather similar results using both
algorithms. Unfortunately, the NFE for these results has not been
reported by the authors; however, the results can be interpreted
in this way that both the EDA and NSGA-II provide acceptable
solutions for the bi-objective problem of the six wells. It will be
shown that for a larger problem (the fty-six-well case), the
advantages of the proposed EDA are more distinguishable.

Fig. 3. EDA performance for single-objective fty-six-well problem.

Table 2
Comparing the Pareto points of the EDA with the NSGA-II (Ray and Sarker, 2007).
Gas usage by
NSGA-II

Gas usage
by EDA

Produced oil
by NSGA-II

Produced
oil by EDA

4550
3139
2613
2351
1893

4446
3120
2628
2326
1900

3657
3445
3294
3214
3026

3646
3436
3324
3227
3048

Table 3
The EDA results for the single-objective fty-six-well problem.
Algorithm

Produced oil (STBD)

Gas usage (Mscf/day)

EDA
GA (Ray and Sarker, 2007)

25612.11
22033.56

22254.86
22376.35

Fig. 4. GA progress for fty-six-well problem (Ray and Sarker, 2007).

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

277

Table 4
Three solutions from the Pareto front for the three-objective problem.
Point

Oil production

Gas usage

Water production

1
2
3

9347
6521
10,004

7243
3312
9240

1929
1621
2096

approach. The corresponding Pareto front is shown in Fig. 5. For

this case, the sample size (the number of samples generated in
each generation) was set to 200 and the population size was 100
individuals.

4.3. The three-objective optimization problem

Fig. 5. Pareto front from multi-objective EDA for the fty-six-well problem.

4.2. The fty six-well problem

The results for the case of the uni-objective fty-six-well
problem are expressed in Table 3. A signicant difference (3578
STBD) has been shown between the proposed EDA algorithm and
the Genetic Algorithm (GA), which leads to a 16.24% improvement
in the daily production at a lower gas usage. The detailed injection
plan is presented in Appendix A.
Fig. 3 shows the progress of the proposed EDA in the search
process of the fty-six-well problem. Comparing this gure with
the same gure from Ray and Sarker (2007) portrays a better
understanding about the algorithm's efciency (compare Fig. 3
with Fig. 4). The GA achieved the best objective value of about
22,000 STBD after an NFE of 20,000, while the EDA clearly produces the same solution for less than 1800 function calls. The
conguration of the EDA was adjusted similar to the six-well
problem; i.e., the population size of 100 individuals and the
same size for sampling were used. This superior performance of
the EDA along with the improved nal results (Table 3) gives an
evidence of the exibility of the algorithm in solving the high
dimensional gas-lift optimization problems. Due to the similar
structure of the multi-objective version of the EDA, the algorithm
is subsequently applied to the same problem in a bi-objective

The capability of the multi-objective optimization is next

investigated in a more complex problem, in which minimizing the
water production is viewed as an extra criterion in the gas-lift
optimization problem and also a conicting constraint in the single objective optimization. In this case with a three-objective
problem, a 10-well model is considered, which maximizes the oil
rate as an optimization criterion and minimizes the gas usage and
water rate as the other criteria. Here, the function approximation
method is the piece-wise (cubic) polynomial interpolation
approach (i.e., spline). In addition, the formulation to this problem
is presented as follows:

maximize Z1

#wells
X

Qoil i

(9)

minimize Z2

#wells
X

Qgas


i

(10)

minimize Z3

#wells
X

Qwater i

(11)

Fig. 6 demonstrates the corresponding solutions from the

developed EDA. The three axes of this gure refer to the total oil
production, total gas injection, and total water production.

Fig. 6. Three-objective optimization Pareto from multi-objective EDA.

278

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

Therefore, the multi-objective solutions can be pondered as a set of

single-objective optimization results for two constraints. In order
to check the quality of this problem, three (random) points from the
Pareto front are considered (Table 4).
Upon running the optimization in the single-objective model, as
shown below, similar results are obtained for oil production under
the gas usage and water production constraints:

maximize Z1

#ofX
wells

Qo i

(12)

i1

subject to;

#ofX
wells

Qg


i

AG

(13)

i1
#ofX
wells

Qw i Qw max

(14)

i1

In Eq. (13), AG denotes the available gas for gas lifting (i.e., the
gas usage for the corresponding Pareto point). On the other hand,
(Qw)max in Eq. (14) refers to the amount of water production of the
solution obtained by the multi-objective optimization. The singleobjective optimization, using the EDA for points 1, 2, and 3 in
Table 4, yields an oil production of 8,710, 6,928, and 9,447,
respectively. It should be expressed that the lower oil production
for point 2 of the multi-objective optimization lies in an emphasis
on the diversity of the Pareto front. For a better optimization, a
criteria weighting approach can be employed in order to reinforce
a higher production. In addition, an economic consideration for
these criteria can be recommended. The congurations of the
single- and multi-objective EDA were the same; i.e., 100 individuals for the initial population and 200 samples in each
generation. The maximum gas injection rate for each well was set
to 2000 Mscf/day.

gorithm was further applied to a multi-objective problem. It was

observed that due to its enhanced searching mechanism, the EDA
can handle the high dimensionality of the gas-lift allocation
problem with rather fewer objective function calls. Moreover,
it is worth mentioning that the same idea can be applied to
the other similar evolutionary algorithms such as ant colony
optimization. In addition, it should be stated that the main
disadvantage of an EDA in comparison with the classical GA is
the complexity of the explicit denition of the probabilistic
models.
The similar results of the EDA and the GA in smaller problems
reveal the sufciency of the GA (and NSGA-II) for these problems.
As the dimensions of the problem increase, the capabilities of
the highly advanced EDA lead to better results for the optimization process. Real gas lift problems often deal with several
dimensions and conicting criteria. Hence, the developed
method can be an appropriate tool to handle these complicated
problems.
The framework for the multi-objective method enables introducing new criteria in a straight-forward manner, such as minimizing the water production. A further extension of the multiobjective gas lift optimization problem is recommended since
many other criteria can be investigated. For instance, minimizing
the produced gas can be another aspect of this problem, although
it should be considered that this produced gas can be re-used for
gas lifting and therefore, a gas recycling factor might be considered. The idea of multi-criteria optimization can be employed to
transform the gas lift constraints into extra criteria in the optimization. As mentioned earlier, one run of a multi-objective algorithm gives information equal to several runs of a singleobjective optimization algorithm for different constraints.
Furthermore, investigating the economic considerations of the
multi-objective gas lift optimization can be the subject of future
studies, because of this fact that the importance of higher oil
production is not often the same as the importance of lower gas
usage, and so on.

5. Conclusion
Appendix A
In this paper, a novel multi-objective optimization algorithm
based on the Gaussian (Bayesian) networks and the Gaussian kernels has been developed in order to solve a gas lift allocation
problem. The proposed algorithm can deal with the uni-as well as
the multi-objective gas-lift optimization problems. The problem of
optimizing two conicting criteria, minimizing the total injected
gas, and maximizing the total produced oil, was addressed by using
the algorithm. On the other hand, minimizing water production
was also considered as the third criterion in order to extend the biobjective optimization.
This method and its efciency were discussed by comparing
the results of the developed algorithm in the single- and multiobjective models with the other similar researches, in terms
of the number of tness function calls and the corresponding
outputs. The attained ndings exposed a signicant improvement over the NSGA-II (non-dominated sorting genetic algorithm
II) employed by Ray and Sarker (2007). Following the
promising results obtained in the single-objective runs (16.24%
increase in the oil production of the fty-six well case), the al-

The lift gas injection rates for the six-well problem in Mscf/day:
490.4942, 875.1588, 1249.3807, 836.3851, 1145.4807, 0.0113.
The lift gas injection rates for the fty-six well problem in Mscf/
day:
350.4196, 246.0045, 433.1768, 30.3901, 17.3274, 202.6782,
2.6593, 221.3926, 1291.2546, 17.0620, 635.5761, 858.0295, 17.7965,
230.7689, 451.2857, 349.5096, 18.9914, 0, 0, 601.9636, 748.0263,
327.1792, 0, 1022.7790, 112.5235, 4.5944, 0, 0, 0, 0, 4.0442, 0,
274.8627, 0, 228.3654, 76.0302, 14.1148, 145.1779, 293.8401,
96.6771, 11.5083, 0, 760.7527, 73.5790, 7.9738, 12.7657, 5997.8564,
9.0431, 5.0588, 9.3884, 0.4073, 5993.5012, 24.6471, 16.0582, 0,
7.8220.

Appendix B
Ten well production data:

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

279

280

S.O. H. Miresmaeili et al. / Journal of Natural Gas Science and Engineering 23 (2015) 272e280

References
Abdollahzadeh, A., Reynolds, A., Christie, M., Corne, D.W., Williams, G.J.J.,
Davies, B.J., 2013. Estimation of distribution algorithms applied to history
matching. SPE J. 18, 508e517. http://dx.doi.org/10.2118/141161-PA.
Ahn, C.W., Ramakrishna, R.S., Goldberg, D.E., 2006. Real-coded Bayesian optimiza~ aga, P., Inza, I., Bengoetxea, E. (Eds.), Totion algorithm. In: Lozano, J.A., Larran
wards a New Evolutionary Computation Advances in the Estimation of
Distribution Algorithms. Springer, Berlin Heidelberg, pp. 85e124. http://
dx.doi.org/10.1007/3-540-32494-1_3.
Alarcon, G.A., Torres, C.F., Gomez, L.E., 2002. Global optimization of gas allocation to
a group of wells in articial lift using nonlinear constrained programming.
J. Energy Resour. Technol. 124, 262. http://dx.doi.org/10.1115/1.1488172.
~ anzas, R., Inza, I., Santana, R., Saeys, Y., Flores, J.L., Lozano, J.A., Van de Peer, Y.,
Arman
~ aga, P., 2008. A review of estimation of
Blanco, R., Robles, V., Bielza, C., Larran
distribution algorithms in bioinformatics. BioData Min. 1, 6. http://dx.doi.org/
10.1186/1756-0381-1-6.
Buitrago, S., Rodriguez, E., Espin, D., 1996. Global optimization techniques in gas
allocation for continuous ow gas lift systems. In: Proceedings of SPE Gas
Technology Symposium. Society of Petroleum Engineers. http://dx.doi.org/
10.2118/35616-MS.
Camponogara, E., Nakashima, P.H.R., 2006. Solving a gas-lift optimization problem
by dynamic programming. Eur. J. Oper. Res. 174, 1220e1246. http://dx.doi.org/
10.1016/j.ejor.2005.03.004.
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective
genetic algorithm: NSGA-II. IEEE Trans. Evol. Comput. 6, 182e197. http://
dx.doi.org/10.1109/4235.996017.
Hauschild, M., Pelikan, M., 2011. An introduction and survey of estimation of distribution algorithms. Swarm Evol. Comput. 1, 111e128. http://dx.doi.org/
10.1016/j.swevo.2011.08.003.
Henrion, M., 1988. Propagating uncertainty in Bayesian networks by probabilistic
logic sampling. In: Uncertainty in Articial Intelligence, vol. 2, pp. 149e163.
Jarboui, B., Eddaly, M., Siarry, P., 2009. An estimation of distribution algorithm for
minimizing the total owtime in permutation owshop scheduling problems.

Comput. Oper. Res. 36, 2638e2646. http://dx.doi.org/10.1016/j.cor.2008.11.004.

Koller, D., Friedman, N., 2009. Probabilistic Graphical Models: Principles and
Techniques. MIT Press.
Kosmidis, V.D., Perkins, J.D., Pistikopoulos, E.N., 2005. A mixed integer optimization
formulation for the well scheduling problem on petroleum elds. Comput.
Chem.
Eng
29,
1523e1541.
http://dx.doi.org/10.1016/
j.compchemeng.2004.12.003.
~ aga, P., Etxeberria, R., Lozano, J.A., Pen
~ a, J.M., 2000. Optimization in continLarran
uous domains by learning and simulation of Gaussian networks. In: Proc. 2000
Genet. Evol. Comput. Conf. Work. Progr, pp. 201e204.
Martinez, E.R., Moreno, W.J., Moreno, J.A., Maggiolo, R., 1994. Application of genetic
algorithm on the distribution of gas-lift injection. In: Proceedings of SPE Latin
America/Caribbean Petroleum Engineering Conference. Society of Petroleum
Engineers. http://dx.doi.org/10.2118/26993-MS.
Nishikiori, N., Redner, R.A., Doty, D.R., Schmidt, Z., 1989. An improved method for
gas lift allocation optimization. In: Proceedings of SPE Annual Technical Conference and Exhibition. Society of Petroleum Engineers. http://dx.doi.org/
10.2118/19711-MS.
Ray, T., Sarker, R., 2007. Genetic algorithm for solving a gas lift optimization problem. J. Pet. Sci. Eng. 59, 84e96. http://dx.doi.org/10.1016/j.petrol.2007.03.004.
Socha, K., Dorigo, M., 2008. Ant colony optimization for continuous domains. Eur. J.
Oper. Res. 185, 1155e1173. http://dx.doi.org/10.1016/j.ejor.2006.06.046.
Su, W., Chow, M.-Y., 2012. Performance evaluation of an EDA-based large-scale
plug-in hybrid electric vehicle charging algorithm. IEEE Trans. Smart Grid 3,
308e315. http://dx.doi.org/10.1109/TSG.2011.2151888.
Sukarno, P., Saepudin, D., Dewi, S., Soewono, E., Sidarto, K.A., Gunawan, A.Y., 2009.
Optimization of Gas injection allocation in a dual Gas lift well system. J. Energy
Resour. Technol. 131, 033101. http://dx.doi.org/10.1115/1.3185345.
Wang, P., Litvak, M., 2008. Gas lift optimization for long-term reservoir simulations.
SPE Reserv. Eval. Eng 11, 147e153. http://dx.doi.org/10.2118/90506-PA.
Zerafat, M.M., Ayatollahi, S., Roosta, A.A., 2009. Genetic algorithms and ant Colony
approach for Gas-lift allocation optimization. J. Jpn. Pet. Inst. 52, 102e107.
http://dx.doi.org/10.1627/jpi.52.102.