SPE 84191

Well Placement Optimization in Field Development

O. Badru, SPE, Stanford U. and C. S. Kabir, SPE, ChevronTexaco Overseas Petroleum
Copyright 2003, Society of Petroleum Engineers Inc.
This paper was prepared for presentation at the SPE Annual Technical Conference and
Exhibition held in Denver, Colorado, U.S.A., 5 8 October 2003.
This paper was selected for presentation by an SPE Program Committee following review of
information contained in an abstract submitted by the author(s). Contents of the paper, as
presented, have not been reviewed by the Society of Petroleum Engineers and are subject to
correction by the author(s). The material, as presented, does not necessarily reflect any
position of the Society of Petroleum Engineers, its officers, or members. Papers presented at
SPE meetings are subject to publication review by Editorial Committees of the Society of
Petroleum Engineers. Electronic reproduction, distribution, or storage of any part of this paper
for commercial purposes without the written consent of the Society of Petroleum Engineers is
prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300
words; illustrations may not be copied. The abstract must contain conspicuous
acknowledgment of where and by whom the paper was presented. Write Librarian, SPE, P.O.
Box 833836, Richardson, TX 75083-3836, U.S.A., fax 01-972-952-9435.

Abstract
Determination of optimal well locations is a challenging task
because engineering and geologic variables affecting reservoir
performance are uncertain and they are often correlated in a
nonlinear fashion. This study presents an approach where a
hybrid optimization technique based on genetic algorithm,
with polytope algorithm as the helper method, was used in
determining optimal well locations. The Hybrid Genetic
Algorithm (HGA) has been shown to work on synthetic and
field examples alike. The HGA was used to optimize both
horizontal and vertical wells for both a gas injection and water
injection projects with net present value (NPV) maximization
as the objective.
Comparison of results was made between locations
proposed by the HGA and those selected by engineering
judgment. Results showed that horizontal wells performed
better than vertical wells from the recovery standpoint for the
synthetic reservoir. For a real reservoir, however, horizontal
wells performed only marginally better than the vertical wells
owing to low-kv/kh ratio. We also observed that optimal well
locations are a strong function of the anticipated project life.
A method of integrating the HGA with Experimental
Design (ED) was also investigated. For this purpose, a
synthetic reservoir was used and exhaustive runs were made
with increasing well count. In this particular case study, we
observed that the uncertainties in the variables affecting
recovery did not affect the optimal number of wells required
to develop this reservoir. Thus, forehand knowledge of the
well count eliminates the need for the inclusion of this process
variable in the ED matrix.
Introduction
The need for a well placement optimization tool cannot be
overemphasized because reservoir performance is highly
dependent on well locations. Determination of optimal well
locations cannot be based on intuitive judgment because

engineering and geologic variables affecting reservoir

performance are often nonlinearly correlated. Because the
performance of a reservoir is time and process dependent, well
placement decisions cannot be based on static
properties alone.
One way of solving this problem is direct optimization
with the simulator. Numerical models are able to evaluate the
complex interactions of various variables affecting
development decisions, such as reservoir and fluid properties,
well and surface equipment specifications, and economic
factors. Even with these models, the current practice is still the
ad-hoc, single-well-configuration-at-a-time approach when
infill prospects are sought. In each trial, well configuration is
selected based on the intuition of the reservoir engineer. For a
single-well case, this one-well-at-a-time approach may lead to
suboptimal decisions. The problem definitely compounds
when multiple producers and injectors are involved in a field
development scenario. The use of the Hybrid Genetic
Algorithm (HGA) offers a way out.
Gyagler1 and Gyagler et al.2 reported development
and application of HGA in determining optimal well locations
for vertical wells. In this study, we extended Gyaglers
work to include horizontal wells. Thereafter, the tool was
tested with a gas-injection project involving a synthetic
reservoir, and two actual field development scenarios.
Application of infill prospects in a developed field is also
shown. In all cases, comparisons with independent intuitive
solutions are discussed.
Background
Over the years, a lot of research has been done in the field of
optimization. The focus of a significant number of them has
been optimal determination of well locations using numerical
simulation,
coupled
with
an
automated
optimization algorithm.
Detailed descriptions of these methods abound in the
literature. This work uses the Hybrid Genetic Algorithm
(HGA), a member of stochastic optimization algorithms based
on natures way of finding the individual fittest to its
environment. Genetic Algorithm (GA), the core of the HGA
and a member of these evolutionary techniques, was
introduced in 1975 by Holland.3 Genetic Algorithms employ
ideas from the principles of genetics as found in the biological
sciences, as well as in Darwins theory of natural selection.
The use of GAs in optimization is nothing new.
Bittencourt and Horne4 optimized a well-location problem
based on maximization of net present value (NPV). They used
a hybrid technique involving GA, polytope algorithm and a

Methodology
To perform well placement optimizations, the HGA is coupled
with an in-house numerical simulator. The HGA comprises
GA, polytope algoritm, and a kriging proxy.
Genetic Algorithm. Genetic Algorithms are search algorithms
based on the mechanics of natural genetics and natural
selection.10 They used survival of the fittest concept
analogous to natural evolutionary mechanisms, combined with
a structured information exchange.10 In many optimization
methods, we move gingerly from a single point in the decision
space to the next using some transition rule to determine the
next point. This point-to-point method is dangerous because it
often locates false peaks in multimodal (many-peaked) search
spaces. By contrast, GAs work from a rich database of points
simultaneously (a population of strings), climbing many peaks
in parallel; thus, the probability of finding a false peak is
reduced over methods that go point to point. In every
generation, a new set of strings is created using bits and pieces
of the old; an occasional new part may be tried for
good measure.
GAs employ natural selection based on fitness, as well as
crossover for information exchange and mutation to introduce
further variety into the search. Thus, a simple genetic
algorithm is composed of three operators: reproduction,
crossover, and mutation.
Usefulness of the Helper Methods. The HGA utilizes
polytope algorithm and a kriging proxy as helper methods.
The use of a proxy requires a significant investment of runs
involving the numerical simulator. The magnitude of this
initial computational investment is unclear because this study
did not show much difference in the efficiency of the HGA

when the kriging option was turned off. Ideally, the proxy
should evolve intelligently as the GA iterates. However
Gyagler1 has shown in Fig. 1 that given sufficient number
of runs, the use of the proxy does reduce the actual number of
simulations and accelerates the optimization process.
300

250

N u m b e r o f S i m u l a ti o n s

Tabu search method. Yeten et al.5 investigated the use of GAs,

a hill climbing search algorithm, and artificial neural networks
in optimizing not only well locations but also well trajectories.
They also based their study on maximizing NPV. Goldberg
and Kuo6 successfully applied GAs in a natural-gas pipeline
optimization problem. Gyagler and Gumrah7 optimized
production rates for a gas storage field using GA, coupled with
a numerical simulator. In another study, on which this work is
based,1,.2 an optimization algorithm was developed combining
GA, polytope algorithm, and a kriging proxy to create the
HGA. In this study, the polytope algorithm serves as a helper
method in improving the efficiency, speed, and exactness of
the GA.
Direct optimization with the numerical simulator is often
infeasible in terms of CPU requirements. As a result,
researchers have also been exploring ways of reducing the
number of simulation runs involved during the optimization
process. The use of proxies in place of the numerical simulator
has been proposed. Gyagler et al.1,2 investigated the use of
both kriging and neural networks as proxies for the simulator.
Pan and Horne8 also used kriging to decrease the number of
simulations required for well-placement optimization. Johnson
and Rogers9 used neural networks in lieu of the numerical
model for a water-injector optimization project. They
preselected 25 candidate locations for injection wells based on
injectivity. However, Gyagler1 showed that a preselection of
locations based on a criterion other than the objective function
could be limiting.

SPE 84191

200

150

100

50

0
GA

GA+P

GA+P+NN

GA+P+K

GA+P+K+E

Fig. 1 Comparison of the mean (squares) and the standard

deviation (error bars) for different combinations of Genetic
Algorithm (GA), Polytope method (P), Neural Networks (NN),
1
Kriging proxy (K), and Error analysis (E). (After Gyagler ).

Integration of the Helper Methods. Following the creation

of the initial population of the GA, a polytope is constructed
from the three best individuals of the generation. A potential
new member is then generated by the polytope method. If the
fitness (objective function) of the new member is better than
the worst of the initial population, then the individual with the
lowest fitness is replaced and the algorithm continues. The
polytope method requires at least one function evaluation
per iteration.
To generate the kriging proxy, a database of visited points
is constructed and updated whenever a call is made to the
numerical simulator. The proxy is constructed with this
database. Estimates of unsimulated points obtained after
Genetic Algorithm operators (crossover and mutation) are then
made with the proxy replacing the numerical model. This is
done through a recursive call to the optimization process. If a
maximum point is found through the recursive call, the actual
function evaluation using a simulator is made. If the maximum
is truly better than the best of the generation, the worst
solution is replaced. The proxy method requires one function
evaluation per generation.

SPE 84191

Create initial population of potential well locations

Determine objective functions using the
numerical simulator (CHEARS)
Select individuals based on fitness
Apply GA operators: crossover, mutation

Apply Polytope Method to n+1

points in population
Create kriging proxy from
database

estimated project life being ten years. In this exercise, the

wells were shut-in when the GOR reached 5,000 scf/STB on
a well.
Case 1b. Modified SPE-1. This model was obtained by
doubling the lateral extent of SPE-1, while retaining the
original thickness of 100 ft. A fourth layer was introduced in
this model to allow for detailed sensitivity studies. Thus, the
modified SPE-1 model is a 20,000-ft 20,000-ft reservoir,
with 1,600 grid cells. All other parameters are identical to
those used in the original SPE-1 model. The objective here
was to determine optimal locations for multiple producers
while keeping the injector location fixed at 1,1,1. The progress
of optimization with respect to the number of actual function
evaluations is shown in Fig. 3. Some of the results are
presented in Figs. 4a through 4d. These results show that not
all the HGA-proposed well locations are intuitive. To confirm
correctness of the solution, we moved the wells based on
engineering judgment. Everytime suboptimal recoveries were
realized, thereby confirming the HGA-derived solution.

1400
Optimize with kriging proxy

1300
NPV(MM$)

Verify proxy result with

numerical simulator

No

1200

Yes
Stop?

1100
Return best

10

20

30

40

50

Number of Simulations
Fig. 2 Flowchart of the Hybrid Genetic Algorithm.

Fig. 3 Progress of the optimization for a two-well case.

Example Applications
Synthetic Reservoir
Case 1a. A gas-injection project with SPE-1 model was
optimized. The SPE-1 model is a 3D black-oil simulation
model used for the first SPE comparative solution study11 in
1981. This reservoir is a 10,000-ft 10,000-ft model, with 100
ft pay and is composed of three layers. The layers were
modeled using a 1010 grid making a total of 300 cells. All of
the cells are active, meaning that all of them are suitable
candidates for well completions. This synthetic reservoir is a
layer-cake model and the objective was to find optimal
locations for both an injector and a producer. Because the
answer was known beforehand, this model served as a way of
testing the capabilities of the HGA.
The producer and the injector were originally located at
(10,10,3) and (1,1,1), respectively. These default locations are
the best in terms of recovery and net present value, and was
also proven by the optimization approach. A population size
of 10 was used. The HGA found these locations at the second
iteration; that is, at the 17th call to the simulator, with the

Uncertainty associated with the variables affecting

reservoir performance is well known. This uncertainty, in turn,
affects optimization of well locations. As a result, no single
realization is completely representative of the reservoir.
Experimental Design (ED) is one way of dealing with
uncertainties inherent in various input variables, affecting
reservoir performance. In ED, geologic (e.g., fluid contacts,
net-to-gross pay thickness) and engineering (e.g., well count)
variables are varied simultaneously, unlike the ad-hoc, onevariable-at-a-time approach.12 In quest for a robust method, we
investigated the effect of uncertainties in certain variables
such as kh, kv/kh, perforation interval, and degree of
communication between layers. The underlying thought was
to integrate ED, which accounts for uncertainty, with the
HGA. Table 1 presents the uncertainties associated with the
four variables. ZTRANS represents degree of communication
between layers.

SPE 84191
Table 1 Range of uncertainties for modified SPE-1.

Variable
kh
kv / kh
Producer
perforations
ZTRANS

p-10
0.5
0.1
Layer
3
0.0

p-50
1.0
1.0
Layers
2 and 3
2.0

p-90
2.0
1.1
Layers
2,3, 4
1.0

We had anticipated that with increasing reservoir quality,

the optimum NPV, which was what we wished to maximize,
would skew to the right with p-99 having the highest well
count. Then we would use the optimal well counts for the p-1,
p-50, and p-99 cases in ED, and generate recovery forecasts at
the p-10, p-50, and p-90 levels. The outcome of the study was
quite contrary to what we expected, as shown in Figs. 5 and 6.
Fig. 6 shows that the uncertainty in our input variables does
not change our decision in terms of well count. This
conclusion is, however, case specific.
Nonetheless, the exercise gave a good handle on the
optimal well count for this synthetic reservoir. Forehand
knowledge of this entity will definitely go a long way in
reducing the number of controlled variables in ED. We
explore this point further with the field example discussed in
the next section.
Studies have, however, shown that different geostatistical
realizations do not necessarily give different results when
decisions are made based on recovery. Ref. 12 is a case in
point, wherein Example 2 showed that somewhat different
geologic models and different grid size yielded very similar
recoveries.12 However, presence of major geologic
uncertainties, such as flow barriers or compartmentalization,
will have serious implications on recovery, and business
decisions made without the knowledge of unknown unknowns.
The HGA also showed that determination of optimal well
locations is not only process-dependent, but also timedependent. Fig. 7 shows the optimal well configuration for 5
and 10 years. The reservoir contains a single injector and 5
producers. Our objective was to maximize oil recovery,
keeping the injector location fixed. Results showed that
optimal well configuration at 5 years gave an additional 2.2
MMSTB of oil than the 5-year performance of the optimum
locations at 10 years. The implication of this observation is
that we would be making suboptimal decisions if the timedependency of well locations were not taken
into consideration.

Fig. 4a - Optimum well locations for a single producer.

Fig. 4b - Optimum well locations for two producers.

SPE 84191

Fig. 4c - Optimum well locations for three producers.

Fig. 4d- Optimum well locations for four producers

Producer location

+ Injector location

Fig. 5 Anticipated results.

Fig. 6 Actual results show that the range of uncertainties in

variables does not affect well count decisions for a specific
case study.

SPE 84191

NO_PROD(Q)

XPERM(L)

I_COMP(Q)

W_LENGTH(Q)

I_COMP(L)

KVKH(Q)

KVKH(L)

XPERM(Q)

W_LENGTH(L)

NO_PROD(L)

Fig. 8 Pareto chart showing relative contributions of variables

to recovery.

Fig. 7 Red and black markers represent optimal well locations at

10 and 5 years, respectively.

Field Examples
Case 2. Reservoir-1. The second case is a water injection
project. The reservoir is unsaturated with an established oilwater contact. The reservoir was modeled as a 2305218
grid, with areal cell sizes of 50 m 50 m and an average
thickness of 2 ft. Of the 215,280 cells, only 172,203 cells were
active. Previous work12 on this reservoir using ED showed that
well count was the most important factor affecting oil
recovery as shown on the Pareto chart in
Fig. 8. This result
has since been verified by bringing well count outside of the
ED matrix and performing a three-level, partial-factorial
design with the other variables. Fig. 9 shows that the range of
uncertainty in the other variables does not significantly affect
reservoir performance. Having established this fact, we wished
to optimize well locations for this reservoir with all other input
variables at their p-50 level.
Previous work showed that 4 producers and 1 injector
would suffice for optimum development of this reservoir. All
the wells preselected during the initial ED study were
horizontal. Our objective was to find the optimum well
configuration for this reservoir based on NPV
maximization.Table 2 presents the parameters used for NPV
calculations. We started with vertical wells to see the impact
of well orientation on NPV and recovery. Thereafter, we
optimized well location configuration with all the wells being
horizontal, and compared our results to what we had before
the implementation of the HGA.
An independent study was done on the same reservoir
with a different optimization algorithm (Algorithm-2).5 The
objective was the same; that is, maximization of NPV. The
results for all the cases considered are discussed below.

Fig. 9 Contributions of variables to recovery factor (RF)well

count excluded.

We also optimized horizontal well locations for the same

reservoir. The well locations proposed by the HGA performed
better in terms of recovery, as well as NPV. These results are
presented in Table 3. The large number of simulation runs
seen in the implementation of Algorithm-2 was possible
because the model was scaled up to a 992212 grid, and the
runs were made using parallel processing. From the results
obtained, we may conclude that well locations based on
reservoir-engineering judgment compare quite favorably with
those using the GA-based algorithms. Guaranteeing an
absolute optimal solution is probably infeasible, time-wise, but
a slightly better solution often leads to a higher profit margin.
The various well locations used for comparison are shown
in Fig. 10.

SPE 84191

Table 2Parameters used for NPV calculations.

Discount rate (%)

Oil price ($/bbl)
Gas price ($/Mscf)
Water handling cost ($/bbl)
Operating cost ($/bbl)
Capital expenditures ($)
Vertical wells
Injector cost (MM$/injector)
Producer cost (MM$/producer)
Horizontal wells
Injector cost (MM$/injector)
Producer cost (MM$/producer)

10
22
0
0.5
3
3
4
3.9
5.2

Table 3- Comparison of solutions for Reservoir-1.

Well
Proposed
No.
of Cum. Oil
Type
Develop.
Simulations (MMSTB)
Initial
Study
HGA
HGA
HGA
Algorithm
-2

NPV
(MM
$)

Horiz

4 prod, 1inj

< 10

38.54

418

Vertical
Horiz
Horiz
Horiz

4 prod, 1inj
4 prod, 1inj
3 prod, 1inj
4 prod, 1inj

30
30
15
200

38.85
40.62
35.28
37.94

396
448
414
421

of 21 MMSTB. A single well optimization was done because

of the relatively small size of the oil-bearing zone.
Selection of optimal well location for a producer based on
reservoir-engineering judgment gave a cumulative oil
recovery of 5.03 MMSTB at the end of 23 years. Using the
HGA, we had 5.16 MMSTB of oil. Both approaches
completed the well in the seventh layer because that was the
most favorable layer in the reservoir. Fig. 11 shows the well
locations proposed by the two methods. One should bear in
mind that finding optimal well locations with the HGA is a
stochastic process. In this particular case study, the run was
terminated at the 20th simulation run. The higher the number
of runs, the greater is the quality of the solution as already
shown in Fig. 3. Table 4 shows that both approaches compare
well in terms of NPV and recovery, although locations are
quite different.

Eng.
Judg.
HGA

Table 4- Comparison of solutions for Reservoir-2.

Well Type
Completion
No. of
Cum. Oil
NPV
Length (ft)
Sim.
(MMSTB)
(MM$)
Runs
Horizontal
1,500
< 10
5.03
29.94
Horizontal

1,500

20

5.16

30. 79

Case 3. Reservoir-2. This reservoir is saturated with wellestablished fluid contacts, with an OOIP (original-oil-in-place)

Fig. 10 Comparison of the different approaches: HGA, Algorithm-2, and reservoir- engineering judgment. The blue region represents the
aquifer and the crosses are injector locations.

SPE 84191

Engineering
Judgment

HGA

Fig. 11 Schematic of Reservoir-2 showing fluid distributions in

th
the 7 layer. Green represents the gas cap, red represents the
oil bearing zone, and blue stands for the aquifer.

Case 4. Reservoir-3. This example was considered earlier in

Ref. 13, while discussing automated-history matching. In
this history-matched model, we sought infill-drilling
propects with horizontal wells. Here, a thin-oil column
(< 40 ft) is overlain by a sizable gas cap and underlain by a
fair size aquifer.
Initial investigation with engineering judgment
identified only two potential sweet spots, designated as H1
and H2 in Fig. 12. As Table 5 shows, the H2 location
appeared promising, but the H1 location yielded slightly
under the economic threshold of 2 MMSTB. We invoked the
HGA seeking a more attractive prospect for the H1 location,
while keeping the H2 location unchanged. Counter to our
intuition, the HGA found a new prospect (H1-HGA) toward
the gas-cap. The location H represents the prolific horizontal
well currently on production.
Table 5 shows that the HGA clearly outperformed our
intitial guess. Overall, the field recovery improved by 2.7
MMSTB when wells H1-HGA and H2 were put on stream,
compared to those selected by intuition, H1 and H2. Note
that production redistribution occurs when well locations are
changed. Therefore, one cannot simply add individual well
contributions, as presented in Table 5, to compare the cases.
These recoveries were realized with gravity-stable
flooding, by setting the producing rates at 1,000 STB/D.
Earlier, we found that drilling injectors has an adverse
impact owing to preferential fluid movement toward a few
take points as depletion matures in this 35-year-old
producing field. This finding was subsequantly confirmed by
the HGA.
Discussion
Well placement by reservoir-engineering judgment is often a
time-consuming, trial-and-error approach, in which the
enginner has to consider numerous variables affecting
reservoir performance. The use of the HGA eliminates this
need, once the process variables are established. This tool
has been shown to reduce the number of man-hours needed

to determine the best locations for wells. The entire process is

stochastic and the quality of solution improves with the
number of runs. At the very least, the HGA can serve as a
guide to the reservoir engineer in determining the sweet spots
in a reservoir.
Use of this algorithm will definitely go a long way in
expediting any reservoir development plan. Even though this
study used a single geologic realization for the field examples
considered, various researchers have shown that a single
realization of a reservoir suffices for most business decisions,
in the absence of significant flow barriers.
Let us point out one reality. Process variables, such as
completion intervals, production and injection rates, and
project life for any well orientation have to be determined by
the reservoir engineer beforehand, prior to invoking the
HGA. That is because HGA is merely a tool that provides an
objective solution regardless of the experience of the
engineer. However, the depletion process and its associated
variables must be established independently.
Table 5- Comparison of solutions for Reservoir-3.
Well Type
Cum Oil, MMSTB
Well H1
Eng. Judg.
HGA

Horizontal
Horizontal

Well H2

1.86
6.34

4.08
3.92

H1

H1-HGA

H2

Fig. 12 Well locations in layer 17 showing oil saturation at the

start of project life.

H1

SPE 84191

Conclusions
A method for optimizing the NPV of a reservoir
development plan using the hybrid genetic algorithm is
presented. Within the limits of this investigation, the
following conclusions are made:
1. The hybrid genetic algorithm in combination with a
reservoir simulator is a powerful and useful tool for
determining optimal well locations.
2. Optimized number of wells and their locations are
not absolute. That is because optimization is
dependent on various process variables, such as the
selected completion interval, recovery process,
production and injection rates, and the project life.
3. The HGA is suitable for both new and existing
fields. For new field developments, the optimal
number of wells (producers and injectors) can be
established a priori, before invoking ED. That is
because the number of wells appears insensitive to
the range of input variables; that is, p-1, p-50, and
p-99 cases, for a given geologic model. On the
other hand, for a history-matched model, search for
additional infill opportunities may be made over a
selected area.
4. When compared to the conventional, ad-hoc
method of selecting well locations, the HGA turns
out to be a laborsaving tool for reservoir engineers.
However, experiences with the field examples
showed that the locations selected with engineering
judgment are quite comparable to those established
by the HGA, in terms of recovery or NPV.
Acknowledgments
We thank colleague Baris Gyagler for his various
assistances with his code. We express our gratitude to Burak
Yeten of Stanford U. for his contributions to the Reservoir-1
case study.
Many other colleagues provided logistical and
computational support, most notably Jason Lewis, Beverly
Pope, and Alan Bernath. We are grateful to the Nigeria/MidAfrica Business Unit of ChevronTexaco for encouragement
and support.

References
1.
2.

3.
4.

5.

6.
7.

8.

9.

10.
11.
12.

13.

Gyagler, B.: Optimization of Well Placement and

Assessment of Uncertainty,
Ph.D. Thesis, Stanford
University, Stanford, CA, 2002.
Gyagler, B., Horne, R. N., Rogers, L., and Rosenzweig,
J. J.: Optimization of Well Placement in a Gulf of
Mexico Waterflooding Project, SPEREE (June 2002)
229.
Holland, J.: Adaptation in Natural and Artificial Systems,
MIT Press, Cambridge, 1975.
Bittencourt, A.C., and Horne, R. N.: Reservoir
Development and Design Optimization, paper SPE
38895 presented at the 1997 SPE Annual Technical
Conference and Exhibition, San Antonio, Texas, 5-8
October.
Yeten, B., Durlofsky, J., and Aziz, K.: Optimization of
Nonconventional Well Type, Location and Trajectory,
paper SPE 77565 presented at the 2002 SPE Annual
Technical Conference and Exhibition, San Antonio, TX,
Sept. 29-Oct. 2, 2002.
Goldberg, D. E., and Kuo, C. H.: Genetic Algorithms in
Pipeline Optimization, Journal of Computers in Civil
Engineering (1987) 1(2), 128-141.
Gyagler, B. and Gumrah, F.: Comparison of Genetic
Algorithm with Linear Programming for the Optimization
of an Underground Gas-Storage Field, In situ (1999) 23,
No. 2, 131.
Pan, Y., and Horne, R.N.: Improved Methods for
Multivariate Optimization of Field Development
Scheduling and Well Placement Design, paper SPE
49055 presented at the 1998 SPE Annual Technical
Conference and Exhibition, New Orleans, Louisiana, 2730 September.
Johnson, V.M. and Rogers, L.L.: Applying Soft
Computing Methods to Improve the Computational
Tractability of a Subsurface Simulation-Optimization
Problem, J. Pet. Sci. and Eng., special issue on Soft
Computing (2001) 153.
Goldberg, D.: Genetic Algorithms in Search, Optimization
and Machine Learning, Addison-Wesley, (1989) 1.
Odeh, A.S.: Comparisons of Solutions to a ThreeDimensional Black-Oil Reservoir Simulation Problem,
JPT (January 1981) 13.
Kabir, C.S., et al.: Developing New Fields Using
Probabilistic Reservoir Forecasting, paper SPE 77564
presented at the 2002 SPE Annual Technical Conference
and Exhibition, San Antonio, TX, Sept. 29-Oct. 2, 2002.
Kabir, C.S., Chien, M.C.H., and Landa, J.L.: Experiences
With Automated History Matching, paper SPE 79670
presented at the 2003 SPE Reservoir Simulation
Symposium, Houston, TX, 3-5 February.