European Journal of Operational Research 240 (2015) 825–836

Contents lists available at ScienceDirect

European Journal of Operational Research

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

Innovative Applications of O.R.

Long term production planning of open pit mines by ant colony

optimization
Masoud Soleymani Shishvan a,c,⇑, Javad Sattarvand b,c
a
Resource Engineering, Faculty of Civil Engineering and Geosciences, Delft University of Technology, P.O. Box 5048, 2600 GA, Delft, The Netherlands
b
Mining Engineering Faculty, Sahand University of Technology, Tabriz, Iran
c
Institute of Surface Mining and Drilling (BBKIII), RWTH Aachen University, Germany

a r t i c l e i n f o a b s t r a c t

Article history: The problem of long-term production planning of open pit mines is a large combinatorial problem. Appli-
Received 3 December 2012 cation of mathematical programming approaches suffer from reduced computational efficiency due to
Accepted 25 July 2014 the large amount of decision variables. This paper presents a new metaheuristic approximation approach
Available online 4 August 2014
based on the Ant Colony Optimization (ACO) for the solution of the problem of open-pit mine production
planning. It is a three-dimensional optimization procedure which has the capability of considering any
Keywords: type of objective function, non-linear constraints and real technical restrictions. The proposed process
Metaheuristics
is programmed and tested through its application on a real scale Copper–Gold deposit. The study
Open-pit mine
Combinatorial optimization
revealed that the ACO approach is capable to improve the value of the initial mining schedule regarding
Production planning the current commercial tools considering penalties and without, in a reasonable computational time. Sev-
Ant colony optimization eral variants of ACO were examined to find the most compatible variants and the best parameter ranges.
Results indicated that the Max–Min Ant System (MMAS) and the Ant Colony System (ACS) are the best
possible variants based on the required less amount of memory. It is also proved that the MMAS is the
most explorative variant, while the ACS is the fastest method.
Ó 2014 Elsevier B.V. All rights reserved.

1. Introduction open-pit optimization have been focused on developing new algo-

A global challenge in the years to come is the environmental-
friendly and financially attractive provision of exhaustible
resources (minerals) to meet the ever-increasing demand by the
today’s high-tech society. Currently surface mining accounts for a
significant proportion of the produced minerals. Surface or open
pit mining is a mineral exploitation method by which the deposit
is accessed by digging a large opening in the ground surface, called
a pit, to expose the ore. The mining operation initially starts with a
small pit and develops to a larger pit enclosing the previous one.
The process proceeds until a final shape of the mine called ‘‘the
ultimate pit limit’’ (UPL) is reached. These sequences of pits are
known as mining sequences or pushbacks.
The last 30 years have seen a widely-publicized revolution in
the application of the numerical methods in the mining industry
in order to produce better mine plans on more complicated and
often lower grade deposits. Recent researches in the field of
⇑ Corresponding author at: Resource Engineering, Faculty of Civil Engineering and
Geosciences, Delft University of Technology, P.O. Box 5048, 2600 GA, Delft, The
Netherlands. Tel.: +31 15 2783425.
E-mail addresses: M.SoleymaniShishvan@TUDelft.nl, soleymani@bbk3.rwth-
aachen.de (M.S. Shishvan), sattarvand@sut.ac.ir (J. Sattarvand).
rithms which are (Sattarvand, 2009):
 firstly less complex in terms of comprehensibility and
programming;
 secondly require high computational efficiency in order to be
applicable to the large deposits;
 finally allow the incorporation of real mining complexities such
as variable slopes, working slopes, time value of money, quality
and quantity of planned material and related uncertainties.
A core concept to address this complex and large-scale optimi-
zation problem is the block model, where the ore body is discret-
ized into a three dimensional array of regular size blocks. The
model may have millions of blocks depending on the size of the
deposit and the size of the blocks. A set of attributes such as
tonnage and grade, using different geostatistical techniques and
economic parameters, are assigned to the each block.
The long-term open-pit mine production planning problem can
be defined as specifying the sequence in which the blocks should
be removed from the mine as a certain material type, in order to
maximize the total discounted profit from the mine subject to a
variety of economical and physical constraints.

http://dx.doi.org/10.1016/j.ejor.2014.07.040
0377-2217/Ó 2014 Elsevier B.V. All rights reserved.
826 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836

This paper presents a new metaheuristic approximation developed a new algorithm for this problem based on a well-
approach based on the Ant Colony Optimization (ACO) for the solu- known integer programming formulation which called C-PIT
tion of the problem of open-pit mine production planning. The pro- method. The proposed method used a new decomposition process
cess has the proficiency to optimize the UPL and the long-term for solving the linear programming relaxation of the C-PIT when
planning problem simultaneously considering the multi-objective there is a single capacity constraint per time period. They denoted
targets and the complex constraints. It merges the constraints into that this is just a proof of a concept and next step would be extend-
the objective function as a set of the penalties for deviations from ing the critical multiplier method to work explicitly with multiple
the targets. The next section gives a brief review on state of the art, side constraints.
short description of the metaheuristic algorithms and reviews the All these exact approaches are limited by the amount of deci-
former developed metaheuristic approaches in the field of open pit sion variables and can solve relatively small problems, excluding
optimization. Thereafter, fundamental structure of the ACO model- many practical implications.
ing and its applicable variants in the open pit optimization are
described. Finally, the proposed procedure of long-term open pit
2.2. Metaheuristics
planning is explained and the results of its application on a real
scale Copper–Gold mine are presented and discussed.
A metaheuristic is a set of algorithmic concepts that can be used
to improve the heuristic methods to become applicable for difficult
problems. These concepts are usually inspired by biology and
2. Problem statement
nature. The use of metaheuristics has significantly increased the
ability of finding very high quality solutions for large combinatorial
2.1. Mathematical formulation
problems (that are often easy to state but very difficult to solve) in
a rational time. This is particularly true for large problems which
At late 60s, researchers were only focused on the UPL problem.
are difficult to understand. The family of the metaheuristics
Lerchs and Grossmann’s algorithm based on graph theory (Lerchs &
includes, but is not limited to, Genetic Algorithm (GA), Simulated
Grossmann, 1965) and Maxflow algorithm based on network flow
Annealing (SA), Tabu Search (TS), ACO, and Particle Swarm Optimi-
concept (Johnson, 1969) are the first attempts to solve this prob-
zation (PSO).
lem. Subsequent studies motivated to a more general problem
Denby and Schofield (1994) described the process of the appli-
namely the problem of production planning. Gershon (1983) pre-
cation of the GA in optimization of an open-pit mine production
sented a Mixed Integer Linear Programming (MILP) model. The
planning shown in Fig. 1a. The main advantage of their method
model has binary variables and its objective function could be
was in its ability to solve the ultimate pit limit and the long-term
expressed as a maximization of Net Present Value (NPV) of the
planning problem simultaneously. By choosing proper values for
mining operation. The MILP model is subjected to the variety of
the genetic parameters, this method is capable to produce good
technical constraints. For instance, the total tonnage of extracted
results for a small block model in an acceptable time. Later
material, the quantity of each material type and the average grade
Denby and Schofield (1995) continued to consider risk assessment
of each production element should be within predefined ranges of
in their scheduling process. They also extended the algorithm from
limits. Moreover, sequencing constraints are necessary to ensure
2D to 3D (Denby & Schofield, 1996) and used it for a flexible sched-
that a block can be extracted if all immediate successor blocks have
uling operation (Denby, Schofield, & Surme, 1998).
been removed. Finally, reserve constraints are applied to mathe-
Kumral and Dowd (2002, 2005) investigated the solution of the
matically guarantee that a block is mined only once.
open pit mine production scheduling problem by means of the SA
Several approaches have been proposed in literatures to solve
as shown in Fig. 1b. The main advantage of this routine is that it
this MILP model. Dagdelen and Johnson (1986) and Caccetta,
utilizes a multi-objective function comprised of three minimiza-
Kelsey, and Giannini (1998) used Lagrangian parameterization in
tion components. On the other hand, the separate determination
order to relax the mining and milling constraints into the objective
of UPL and production schedule would be counted as a drawback
function. Consequently, the problem could be handled by repeti-
tion of any UPL algorithm such as the (Lerchs & Grossmann,
1965)’s graph theory based algorithm. Later Caccetta and Hill
(2003) proposed a branch and bound technique to solve the formu-
lated scheduling problem. Dowd and Onur (1993) and Onur and
Dowd (1993) formulated the problem as a dynamic programming
model. Ramazan (2007) described the application of fundamental
tree algorithm to reconstruct the mining blocks and decrease the
number of variables in scheduling problems without reducing
the resolution of the model or optimality of the results. They
defined the fundamental tree as any combination of the blocks that
can be profitably mined respecting slope constraints. Boland,
Dumitrescu, Froyland, and Gleixner (2009) proposed an iterative
disaggregation method that refines the aggregates (with respect
to processing) up to the point where the refined aggregates defined
for processing produce the same optimal solution for the Linear
Programming (LP) relaxation of the MILP as the optimal solution
of the LP relaxation with individual block processing. Bley,
Boland, Fricke, and Froyland (2010) presented an integer program-
ming formulation which is strengthened through adding inequali-
ties derived by combining the precedence and production
constraints. The addition of these inequalities decreases the com-
putational requirements to obtain the optimal integer solution. Fig. 1. (a) The process of open pit scheduling by genetic algorithm and (b)
Chicoisne, Espinoza, Goycoolea, Moreno, and Rubio (2012) simulated annealing (Sattarvand, 2009).
 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836
for this method. Godoy and Dimitrakopoulos (2004) used simu-
lated annealing method on the effective management of waste
mining and orebody grade uncertainty. Their objective function
minimizes the chance of deviation from production target for
each period in different scenarios. Recently Lamghari and
Dimitrakopoulos (2012) presented a diversified Tabu search
approach for the open-pit mine production scheduling problem
in the field of metal uncertainty. They used two different diversifi-
cation strategies to generate several initial solutions and then opti-
mized these solutions using the TS method.
2.3. Ant colony optimization
ACO, which is inspired by the foraging behavior of the ant col-
onies, is developed by Dorigo and Stützle (2004). In nature, the
ants walk randomly and upon finding food return to their colony
while laying down chemical trails called pheromone. The phero-
mone trail transmits a message to other members of the colony.
The other ants are likely to follow the trail instead of randomly
traveling. If they eventually find food then reinforce the trail by
depositing more pheromone. Over the time the pheromone trail
starts to evaporate and reduce its attraction. Obviously, magnitude
of the evaporation in longer paths is higher than shorter. Thus, the
intensity of laid pheromone on the shortest path, by comparison,
gradually increases up to the level that balances with the evapora-
tion rate. This makes the shortest path to be marched and followed
by almost all of the ants. The methodology of the ant colony opti-
mization mimics this natural behavior by considering a series of
variables representing and continuously updating the pheromone
values based on the quality of the found solutions, (Dorigo &
Stützle, 2004).
3. ACO methodology of long-term open pit planning
Most of the mathematical programming approaches are limited
by the amount of decision variables, as applications for the long-
term open pit problem become very large and suffer from reduced
computational efficacy. This fact encouraged researchers to use dif-
ferent solvers for this problem. In leading optimization problems,
such as Travelling sales man (TSP), Vehicle routing and Assignment
problem, (Dorigo & Stützle, 2003), ACO approach is applied to find
Fig. 2. The process of long-term open-pit production planning using ACO.
827
the optimal or near optimal solution. The following describes
methodology of application of ACO in the field of open pit mine
planning. Fig. 2 shows the proposed flow chart of the long-term
open-pit production planning using ACO.
3.1. Initial solution
The algorithm consists of saving P variables for each block of the
model, sip, which represent the pheromone value related to the
mining of ith block in pth period. The magnitude of the saved pher-
omones represents the desirability of a block to become the deep-
est point of the mine in that period. The initial values of these
variables are assigned based on a sub-optimal mine schedule gen-
erated by the traditional algorithms. Then the random mining
schedules are constructed according to the initial pheromones.
These schedules deposit an extra pheromone proportional to their
economic quality (fitness value). This action along with consider-
ing the pheromone evaporation in the algorithm, leads to the
optimum boundary of the mining pushbacks.
3.2. Pheromone initialization
The experiments showed that the running time increases dras-
tically using the uniform initial pheromone pattern. Therefore, a
sub-optimal solution for the problem of long-term open-pit sched-
uling is used and initial pheromone trails are assigned to the blocks
according to this sub-optimal solution. Normally, the shape of a
desired pushback does not change significantly from a sub-optimal
solution to the optimal one. Thus, assigning of higher pheromones
to a few numbers of the blocks around the sub-optimal pit depth
could be enough to conduct the algorithm towards the optimal
solution. During the process of pheromone initialization, the pher-
omone values of the ore blocks, close to the pit shape in initial
solution (the highlighted blocks in Fig. 3), are set to relatively
higher values.
3.3. Construction of schedules
In order to construct a mine scheduling solution, a series of fea-
sible pit shapes related to the different mining pushbacks should
be created. Each one of these pits consists of a series of block col-
umns and the shape of each pit could be defined by determination
of the pit depth on these columns. The pheromone value is the
major element in determination of the pit depth on a column.
However, considering the heuristic information such as economic
value of the blocks could sometimes help the efficiency of the
method, either.
3.3.1. Process of depth determination
At the each depth determination step, ant k utilizes a probabi-
listic choice rule, named random proportional rule, to make decision
for the deepness of the pit in that column. Eq. (1) shows the prob-
ability of choosing i as the pit floor by ant k:
Fig. 3. Initialization of pheromone values of the blocks.
828 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836

½si a ½gi b
Pki ¼ P ; ð1Þ
l2 N ki
½sl a ½gl b

where si is the pheromone value of the block i, gi is the heuristic

information which can be a block value or any information that
causes algorithm obtain better solution, a and b are two parameters
which determine the relative influence of the pheromone trail and
the heuristic information, and N ki is the set of feasible selections
for the ant k. A numerical example of depth determination process
is explained in Table 1.
The set of feasible selections are represented by the upper and
lower boundary of the permitted pit depth. The maximum allowed
depth defines the deepest possible mining depth on that column
whereas the minimum depth is determined according to the shape
of the mine in earlier pushback, Fig. 4.
It should be noted that the process of depth finding is done only
for the columns containing at least one ore block. The depth of the
pit in totally waste columns will be defined based on the neighbor- Fig. 5. Generation of feasible pit according to the selected depths (Normalization).
ing selected depths. Another important concept to consider is that
the initial pheromones are assigned only to the ore blocks. There-
fore, the selected depth will always coincide on an ore block. Normalization process is implemented as the following steps:
Similarly, there will be no pheromone update (evaporation or
deposition) for waste blocks. – Start from the deepest level of the block model and check all the
blocks in this level. If the calculated depth of any column is
equal to this level, flag that block and all the upper blocks as
3.3.2. Normalization an In-Pit blocks.
Normally the consequence of independent depth determination – Move to the upper level and check all blocks. Flag any block as
in each column is not always feasible due to the required slope In-Pit if at least one of the following conditions is satisfied.
angles. Therefore, a normalization stage based on the selected o If the calculated depth in any column containing that block is
depths is necessary in order to generate a feasible pit shape. The equal to or lower than the current level.
normalization step is implemented after determination of o If the minimum depth in any column containing that block is
the depths to ensure that the constructed pit shape covers all the equal to or lower than the current level.
determined depths as well as the outline of the earlier pushbacks. o If at least one of the underlying immediate successor blocks,
The feasible pit shape shown in Fig. 5b is constructed based on the more details in Shishvan and Sattarvand (2012), of that block
set of determined depths and the shape of earlier period displayed is flagged as In-Pit.
in Fig. 5a. – Repeat the previous step up to the uppermost level.

Table 1 3.3.3. Mine schedule construction from generated pits

An example of the depth determination process. Finally constructed individual normalized pits for different min-
ing periods are combined together to produce a mine schedule,
Fig. 6.

3.4. Pheromone update (evaporation and deposition)

The ACO module in two steps manipulates the initial phero-

mone value of the blacks. The first step, called pheromone evapo-
ration, consists of a uniform reduction in the value of all the
pheromones in order to help the ACO model disregard the bad
solutions. In this stage, the pheromone value of all the blocks cor-
*
Without unit. responding to all the production schedules should be decreased by
a b
**
Based on P ki ¼ P ½si  ½s½gia ½g b formula a = 1, b = 1. a certain percentage. The next step, pheromone deposition, con-
l2N k l l
***
i
Selected depth according to the random number (0.6328). sists of adding additional pheromone to the blocks which have
contributed to the construction of the schedules. It should be noted
that the deposition action is applied only to the ore blocks. In cases
where the pit depth falls on a waste block, the additional phero-
mone is assigned to an imaginary block on the ground surface. This
action creates an imaginary block to compete against the other ore
blocks which have not really contributed in construction of the
schedules.
In general, pheromone update process of ACO could be imple-
mented in a variety of ways. The Ant System (AS) is the first and
simplest method, whereby all of the constructed schedules are
allowed to contribute to the pheromone deposition. In each itera-
Fig. 4. Maximum and minimum depth definition in the process of depth determi- tion of the second method, the Elitist Ant System (EAS), the best
nation according to the slope angles. schedule found up to that iteration (the best-so-far schedule), is
 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836
also allowed to deposit pheromones. Ranked based Ant System
(ASrank) is the third method in which only a few good schedules
are able to add pheromones. The other variants are the MMAS
and the ACS, which allow only the best-so-far schedule to deposit
pheromones and utilize special pheromone limitations in order to
prevent the stagnation in local optimums. Analysis revealed that
MMAS and ACS variants are also counted as the only variants
which could be applied to a large block model in respect to the
amount of memory needed, (Sattarvand & Niemann-Delius, 2013).
3.4.1. Max–Min Ant System (MMAS)
MMAS is one of the most efficient and detailed studied ACO
algorithms, (Stützle & Hoos, 1997, 2000), (Stützle, 1999). The main
characteristics of MMAS are:
– Only the iteration-best ant or the best-so-far ant is allowed to
deposit pheromones.
– Pheromone trail values are limited to the interval [smin, smax].
– Pheromone trails are initialized to the upper pheromone trail
limit.
– Pheromone evaporation rate is very small.
– Pheromone trails are reinitialized each time the system
approaches stagnation or when no improved tour has been
generated for a certain number of consecutive iterations.
Fig. 6. Combination of generated pits to produce a mine schedule.
829
Pheromone evaporation rule of MMAS (Eq. (2)) and deposition
(Eq. (3)) could be mathematically expressed as follow:
si ð1  qÞsi ð2Þ
si si þ DsBest ð3Þ
i
where q is the evaporation rate 0 < q < 1 and DsBest
i is the amount of
pheromone that the best so far ant deposits on the selected depth i.
3.4.2. Ant Colony System (ACS)
The ACS differs from the previous variants from the following
points of view: (Dorigo & Gambardella, 1997b)
– Pseudorandom proportional action choice rule with the
probability of the ants select the depth with the highest
[si]a[gi]b. Whereas it has the (1  q0) probability of using the
same routine as MMAS for the selection.
– Global pheromone update in ACS, after each iteration, only one
ant (the best-so-far ant) is allowed to add pheromones. Addi-
tionally, unlike MMAS, the evaporation process only applies to
the blocks of the best-so-far schedule, not to all the solutions.
– Local pheromone update ants use a local pheromone update
rule and apply it immediately after making the decision on a
block during the schedule construction of ACS.
4. Case study
4.1. Initial solution
Geological model of a Copper–Gold deposit presented by the
Earthworks NPV-Scheduler, (NPV – Scheduler, 2001), is used as
the case study. The model is constructed in Datamine Studio soft-
ware and include 80  120  50 blocks with block dimension of
10  10  10 (m) parallel to the principle geographical directions
and contains just over 350,000 blocks. It also contains the sub-cells
to provide an accurate volumetric model of the topography and ore
boundaries. The Gold (AU), Copper (CU) and density (DENSITY) of
the blocks have been estimated by Kriging based on the borehole
samples and their geological interpretations. Metallurgical tests
advised to distinguish two distinct types of the sulfide ore accord-
ing to their floatability and solubility characteristics. The ore blocks
were classified into SULF1 (mill-smelt-refine) and SULF2 (leach-
SXEW) ore types.
The economic viability of the project and developing a life-
of-mine production schedule based on a gold price of $270 per
ounce and a copper price of 95 cents per pound were assessed in
NPV-Scheduler by Earthworks. Tables 2–4 show the assumed
mining, processing and general and administrative (G&A) costs.
As seen in Table 2, the base mining cost at the reference bench
increases by deepening the mine.
Table 5 illustrates the economic model according to the techni-
cal and the financial condition. Geomechanical studies led to dis-
tinguishing of five separate domains having different stability
behavior. Azimuth and stable slope angles of these regions are
shown in Table 6 (NPV – Scheduler, 2001).
UPL was determined based on a discount rate of 10% and yearly
production of 6 million tonnes of ore. Then, ten pushbacks, each
containing a minimum of 4 million tonnes of mill ore and a mini-
mum access width of 30 meters were constructed Table 7. Any
remnant areas containing less than 200 blocks (approx.
500,000 tonnes) were merged with the adjacent pushback (NPV –
Scheduler, 2001).
As a simple scheduling condition, the following constraints
were considered for each period of this case study:
830 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836

Table 2 o Penalties for lower than base processing rate: for the annual
Mining costs (NPV – Scheduler, 2001). rates less than 4 million tonnes, a 10% reduction in recovery
Rock type Base ($) Reference bench Incremental is considered. This factor diminishes linearly to zero for the
Waste 0.62 2100 0.01 rates from 4,000,000 to 5,110,000.
Ore – SULF1 0.75 2060 0.01 – Average grade of Cu:
Ore – SULF2 0.75 2040 0.01 o Planned grade: 0.2%. For lower grades, recovery changes
according to Eq. (4) whereas for higher grades it is fixed to 88%.
Table 3
Processing costs, (NPV – Scheduler, 2001). Recoveryð%Þ ¼ Cu  440ð%Þ ð4Þ
Ore – SULF1 Ore – SULF2 Consequently, the value of the constructed scheduling solution
Capacity 14,000 tonne per day – receives 18,127,000 dollars of penalty and its economic value drops
Processing costs 3.34 ($) per tonne ore 1.24 ($) per tonne ore from 128,690,000 to 110,563,000 dollars.
Refining and 0.08 ($) per tonne Cu refined –
smelting
4.2. Implementation of the algorithm
SWEX – 0.14 ($) per tonne Cu
cathode
Cu recovery If (Cu < 0.2, R% = 440 * Cu%, 63% In order to analyze the applicability of the ACO in optimizing
R% = 88%) the long-term planning of the open-pit mines and finding the best
Au recovery 91% 0% ACO variants and parameters, a computer program was prepared
in Visual Studio 2010 programming environment and applied on
the case study block model using the only applicable variants of
Table 4 the ACO including the MMAS and the ACS.
Administration and capital costs, (NPV – Scheduler, 2001).

G&A $38 per oz. Au recovered 4.2.1. Implementation without considering scheduling penalties
Capital – mining $0.08 per tonne mined The final production plan of the case study was imported into
Capital – processing $0.29 per tonne processed
Capital – leach $0.05 per tonne processed
the ACO procedure and the NPV of the resulted solution has been
compared to the initial solution. Despite the fact that the ability
of introducing penalties can be considered as one of the major
advantages of the ACO, this test proves that the ACO is even able
to improve the economic return of a commercially generated plan.
Table 5
Results of building the economic model (NPV – Scheduler, 2001).
4.2.1.1. Max–Min Ant System (MMAS). Fig. 7 shows that the MMAS
Number of ore blocks 52,650 – can improve the value of the mine schedule. After scattering at the
Number of waste blocks 268,428 –
initial iterations, it reaches to its best solution, 134,279,000 ($),
Total ore 131.71 106 * tonnes
Total waste 548.00 106 * tonnes
subsequently at iteration 1430. In comparison to the initial solu-
Total ore value 330.19 106 * $ tion which had a value of 128,690,000 ($), this means about
SULF1 84.11 106 * tonnes 4.35% improvement in the value of the solution.
SULF2 47.60 tonnes
SULF1-AU in situ 56.64 tonnes
SULF2-AU in situ 25.57 tonnes 4.2.1.2. Ant Colony System (ACS). Fig. 8 demonstrates that the ACS
SULF1-CU in situ 206.37 103 * tonnes improves the value of the mine schedule right from the first itera-
SULF2-CU in situ 97.15 103 * tonnes tion, and after 23 iterations it reaches its best solution at
131,995,000 ($). Again, in comparison to the initial solution, this
means about 2.58% improvement in the value of the solution.
Table 6
Slope angles (NPV – Scheduler, 2001). 4.2.2. Implementation with considering scheduling penalties
4.2.2.1. Max–Min Ant System (MMAS). The experiments showed
Azimuth Slope
that there is a high potential of improvement in the solution and
012 41
using slightly higher values of q (0.03 to 0.05) and initial phero-
093 43
128 44
mones as 2/q to 3/q (i.e. 40 to 100) better solutions could be faster
145 41 reached. Although the ratio of smax/smin
pffiffiffiffiffiffiffiffiffiffi is suggested by (Stützle,
pffiffiffiffiffiffiffiffiffiffi
280 40 1999) to be as ð1  n 0:05Þ=ððav g  1Þ: n 0:05Þ for solution of Trav-
eling Salesman Problem (TSP) having n cities and avg choices avail-
able at each step, however, a fixed value of 5 is used for the smin in
– Mining: our case study and the pheromone trails are re-initialized when no
o Base rate: 32,200 tonnes per day. improvement occurs after 10 iterations (Sattarvand, 2009).
o Penalties for higher than base mining rate: 20% increase in In the basic run of the MMAS, the following values were consid-
mining costs for the extra material. ered as the key parameters:
o Penalties for lower than base mining rate: 10% increase in
overhead and mining cost for the rates less than 25,000 ton- – Number of the ants (number of schedules, m) is considered to
nes and linearly increment factor from zero to 10% for the be equal to 50.
values of 25,000–32,200. – Similar to the application of ACO for solution of TSP, the value of the
– Processing: evaporation coefficient, q, had been set to 0.03 in this case, too.
o Base annual rate: 5,110,000 tonnes. – The upper and lower perturbation distance is considered as
o Penalties for higher than base processing rate: extra ore zero. In other words, relatively high pheromone values were
masses have to be considered as waste, because there is no assigned only to the blocks that construct the mining
ore dump or stockpile possibility. pushbacks.
 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836 831

Table 7
Economic statistics (NPV – Scheduler, 2001).

Pushback Profit NPV Total rock Total ore Total waste SULF1-Mill SULF2-Leach AU (gram/ CU Lifetime
(106 * $) (106 * $) (106 * tonnes) (106 * tonnes) (106 * tonnes) (106 * tonnes) (106 * tonnes) tonnes) (%) (year)
1 16.17 15.40 5.62 4.38 1.25 4.19 0.19 0.65 0.27 0.52
2 23.65 20.58 10.25 6.24 4.00 4.73 1.50 0.76 0.31 0.94
3 21.94 17.55 9.63 6.27 3.35 4.21 2.07 0.82 0.27 0.88
4 22.13 16.53 7.81 4.66 3.15 4.45 0.21 0.60 0.37 0.72
5 12.58 8.72 8.51 5.59 2.91 5.06 0.53 0.65 0.20 0.78
6 20.04 12.75 9.80 5.32 4.48 4.97 0.35 0.57 0.32 0.90
7 22.14 12.95 9.63 4.70 4.93 4.26 0.44 0.60 0.38 0.89
8 18.28 9.48 13.71 6.16 7.55 5.68 0.48 0.67 0.27 1.26
9 15.89 7.12 16.73 6.13 10.60 5.75 0.38 0.68 0.26 1.53
10 19.74 7.61 17.12 9.10 8.02 6.70 2.40 0.70 0.23 1.57
Total 192.56 128.70 108.80 58.60 50.24 50.00 8.60 0.68 0.28 10

Fig. 7. Efficiency of the min–max ant system without penalty.

– Unequal priority was considered for the pheromone trails and
the heuristic information in this case, (a = 1, b = 0.26).
The efficiency of the MMAS is shown in Fig. 9. It proves that the
ACO has strong capability to improve the quality of the initial solution.
The graph also reveals that after the 11th iteration, the ACO algorithm
starts to improve the value of the initial schedule and after 1470 iter-
ations it reaches to its best solution at 122,460,000 ($). In comparison
to the initial solution which had a value of 110,563,000 ($), this means
more than 10.76% improvement in the value of solution.
As described earlier, the value of any schedule has two major
components-profit and penalty. Conventional algorithms of the open
pit planning as that has been utilized for construction of the current
initial solution takes only the first component (profit) into account.
Therefore, their combination could not be optimized. In contrast,
the ACO looks for the solutions having higher total value, even with
lower profit. The variation of the profit and penalty values for the
MMAS is shown in Fig. 10. It shows that the original combination
of profit and penalty values has changed from 193,416,000 ($) and
38,074,000 ($) to 179,146,500 ($) and 8,068,000 ($). In other words,
obtained solution has about 3,827,500 ($) less profit but instead
about 14,353,000 ($) of lower penalties which lead to 10,525,000
($) of improvement in their sum.
4.2.2.2. Ant Colony System (ACS). Here is the most efficient range of
the ACS parameters for the case study:
– Number of the ants: 5.
– The evaporation rate: 0.1.
– The initial pheromone value: 0.01.
– The local pheromone update factor: 0.15.
– The pseudorandom choice probability: 0.7.
The result of using the ACS is shown in Fig. 11. The main notice-
able point in the ACS is the drastic reduction in the number of ants
which has direct effect on the calculation time of each iteration.
Another factor that helps the ACS algorithm to speed up is the fact
that the pheromone evaporation and deposition happen only on
the blocks of the best so far solution. Consequently, comparing to
the other variants of the ACO, the ACS could gain much better
results in a specific time of calculation. This might be very useful
for the large block models.
4.3. Sensitivity analysis of parameters of ACO
The following describes the sensitivity of the MMAS and the
ACS to the variations in its parameters.
4.3.1. Number of ants
The number of the ants is one of the major parameters of the
ACO. The importance of the ant numbers comes from its impact
on the running time of each iteration. In this study, the initial num-
ber of 50 was assigned. The effect of using higher or lower number
of the ants is shown in Table 8.
Although the best number of ants could not be evaluated defi-
nitely according to this table, it is evident that very high numbers
which extremely increase the calculation time cannot result
in considerable solution improvement. The experiments
832 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836

Fig. 8. Efficiency of the ant colony system without penalty.

Fig. 9. Efficiency of the max–min ant system.

Fig. 10. Profit and penalty variations.

 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836 833

Fig. 11. Efficiency of the ant colony system.

demonstrated that the initial number of five ants is relatively Table 9

Effect of the initial pheromone on the solution quality and calculation time.
appropriate and could be altered ±20% without any noticeable dis-
tortion on the algorithm. Initial pher. Value Required time Required
(million $) (minutes) iterations

4.3.2. Initial pheromone values 60 118.7 92 1467

The result of variation in the initial pheromone values is shown 80 119.6 66 1489
100 122.1 61 1464
in Table 9. The lower initial pheromone leads the program to
110 121.3 65 1490
search among the solutions far from the initial schedule. Although 125 121.1 58 1478
it generates slightly better solutions, a growth of 50% in time cal- 150 118.3 51 1375
culation is inevitable. On the other hand, a high initial pheromone
value made the algorithm to stagnate in the solutions around the
initial answer which had the anticipated consequence of the poor evaporation rate and length of perturbation distance on the effi-
results. ciency of the basic the ACS are almost the same as the MMAS.
In the following, the effects of other factors are discussed in
4.3.3. Pheromone evaporation rate details.
The analysis of different values of the evaporation rate, q,
resulted in the rates which are compared in Table 10. It reveals that
the lower values of q, would cause to spend a long time on the poor 4.3.5. Priority factors of pheromone and heuristic information
solutions whereas higher values would prevent the algorithm from Table 12 represents the efficiency of the algorithm with the dif-
adequately discovering far domains from the initial solution. ferent combinations of priority factors for the pheromone values
and the heuristic information. The proposed procedure takes the
4.3.4. Perturbation distance economic values of the blocks as heuristic information. The exper-
The quality of the solutions and the calculation times for the iments showed that (a = 1, b = 0.26) are the best choices for the
different numbers of the perturbation distance are compared in priority factors.
Table 11. The experiments discovered that the main privilege of
the MMAS is due to its explorative nature which lets the algorithm Table 10
to use the higher perturbation distances which may lead to the Effect of the evaporation rate on the solution quality and calculation time.
better solutions. This will take more calculation time and higher Evaporation rate Value Required time Required
scattering iterations prior to the notification of the improvements. (million $) (minutes) iterations
The experiments illustrated that the effects of variations of 0.02 120.1 77 1473
parameters such as number of ants, initial pheromone value, 0.03 122.5 59 1470
0.04 121.5 58 1476
0.05 121.0 57 1268

Table 8
Effect of ant number on the solution quality and calculation time.
Table 11
Ant no. Value Required time Required
Effect of the perturbation distance on the solution quality and calculation time.
(million $) (minutes) iterations
5 122.5 59 1470 Pert. dist. Value Required time Required
10 122.4 152 1454 (million $) (minutes) iterations
25 121.9 290 1441 0 120.2 50 1467
50 123.3 990 1493 1 122 61 1464
1000 112.3 7410 5 2 122.2 88 1479
834 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836

Table 12 addition to the global pheromone trail updating. The local phero-
Effect of priority factors of pheromone and heuristic information on the solution mone update during the schedule construction is applied immedi-
quality and calculation time.
ately after determination of the depth for a column. Table 13
Priority factors Value Required time Required shows the different combinations of these two parameters during
(million $) (minutes) iterations the optimization process.
a = 1, b = 0.2 113.6 62 512 It is evident that the most appropriate values are 0.15 and 0.1
a = 1, b = 0.26 116.5 68 569 for the local and the global pheromone updates, respectively.
a = 1, b = 0.3 114.5 39 370
a = 1, b = 0.4 116.5 59 482
a = 1, b = 0.5 114.3 13 116 4.3.7. Pseudorandom choice probability
a = 2, b = 0.26 116.3 73 787 In fact, by use of the pseudorandom choice probability, the ran-
a = 3, b = 0.26 115.5 32 326 dom behavior of the ACO algorithm is lost. In other respects, with
probability of q0 the ant chooses the block which has the highest
learned knowledge, i.e. [si]a[gi]b, whereas with probability of
(1  q0) it utilizes a probabilistic approach similar to the MMAS.
Table 13
According to Table 14, the most relevant value for the pseudoran-
Effect of local and global pheromone update values.
dom choice probability is 0.7.
Evap. and dep. rate Value Required time Required
(million $) (minutes) iterations
5. Final remarks
Local = 0.1, Global = 0.1 116.2 72 598
Local = 0.1, Global = 0.15 115.1 76 772
Local = 0.1, Global = 0.2 115.9 192 771 The result has been investigated by the means of commercial
Local = 0.15, Global = 0.1 116.5 68 569 software. Any direct comparison with the exact method was not
Local = 0.2, Global = 0.2 115.1 193 790 possible due to the size of the model. The non-availability of the
Local = 0.2, Global = 0.1 116.2 102 726 case studies of other developed metaheuristics was another limita-
tion. Fig. 12 visualizes a section view of the final schedule obtained
by application of the ACO.
The Ore and Waste masses and the corresponding copper and
gold grades of the pushbacks are listed in Table 15. Moreover,
Table 14
the values of the profit, penalty, NPV and lifetime of the phases
Effect of perturbation distance values.
are shown in Table 16. Comparing to Table 7, the improvements
Pseud. prob. Value Required time Required gained by application of the ACO are obvious. It is possible to claim
(million $) (minutes) iterations
that the most of the planning targets have been satisfied by 90%, at
0.6 115.5 45 639 least.
0.7 116.2 72 598
0.8 115.5 113 581
0.9 113.0 182 502 6. Conclusions

Different mathematical techniques such as Integer Program-

ming (IP) and Relaxation methods have been tried to solve the
4.3.6. Local and global pheromone update factor mine production planning problem, but all of them are suffering
Another major difference of the ACS with the former ACO algo- from the fact that they are unable to solve instances of the realistic
rithms is the consideration of a local pheromone update rule in size and the level of detail in a practically acceptable amount of the

Fig. 12. North–South cross section of performed planning.

 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836 835

Table 15
Total ore and waste masses and corresponding copper and gold grades of each pushback.

Pushback Total rock Total waste SULF1-Mill SULF2-Leach AU (gram/ CU AU-R CU-R Lifetime
(106 * tonnes) (106 * tonnes) (106 * tonnes) (106 * tonnes) tonnes) (%) (tonnes) (103 * tonnes) (year)
1 6.71 3.33 2.98 0.41 0.47 0.32 1.28 8.14 0.57
2 8.47 3.76 3.69 1.03 0.66 0.39 2.20 12.17 0.72
3 6.74 2.40 2.93 1.41 0.82 0.28 2.19 6.85 0.57
4 8.76 3.77 3.82 1.19 0.78 0.35 2.71 11.56 0.75
5 8.28 3.62 4.45 0.22 0.56 0.22 2.25 7.97 0.70
6 9.24 4.68 4.03 0.53 0.73 0.38 2.66 13.23 0.79
7 13.73 6.96 5.96 0.81 0.64 0.31 3.47 15.76 1.17
8 12.12 5.51 5.39 1.21 0.67 0.27 3.31 12.18 1.03
9 22.65 11.59 10.23 0.83 0.70 0.21 6.54 17.05 1.93
10 20.81 9.96 8.38 2.47 0.70 0.22 5.31 14.60 1.77
Total 117.54 55.57 5187 10.10 0.67 0.30 31.91 119.51 10

Table 16 References
Economic statistics of each pushback.
Bley, A., Boland, N., Fricke, C., & Froyland, G. (2010). A strengthened formulation and
Pushback Profit Penalties NPV Cumulative cutting planes for the open pit mine production scheduling problem. Computers
(106 * $) (106 * $) (106 * $) lifetime & Operations Research, 37(9), 1641–1647.
(year) Boland, N., Dumitrescu, I., Froyland, G., & Gleixner, A. M. (2009). LP-based
disaggregation approaches to solving the open pit mining production
1 9.35 0.24 8.63 0.57
scheduling problem with block processing selectivity. Computers & Operations
2 21.21 0.01 18.75 1.29
Research, 36(4), 1064–1089.
3 15.60 0.01 13.04 1.87 Caccetta, L., Kelsey, P., & Giannini, L. (1998). Open pit mine production scheduling.
4 22.78 0.02 17.74 2.61 Paper presented at the APCOM 98: Computer applications in the mineral
5 9.48 2.15 5.34 3.32 industries international symposium (pp. 65–72).
6 23.26 0.11 15.66 4.10 Caccetta, L., & Hill, S. P. (2003). An application of branch and cut to open pit mine
7 23.55 0.02 14.24 5.27 scheduling. Journal of Global Optimization, 27(2–3), 349–365.
8 18.90 0.52 10.08 6.30 Chicoisne, R., Espinoza, D., Goycoolea, M., Moreno, E., & Rubio, E. (2012). A new
9 25.57 1.26 11.10 8.23 algorithm for the open-pit mine production scheduling problem. Operations
10 21.58 1.12 7.89 10.00 Research, 60(3), 517–528.
Dagdelen, K., & Johnson, T. B. (1986). Optimum open pit mine production
Total 191.26 5.45 122.46 10 scheduling by Lagrangian parameterization. Paper presented at the
Proceedings of 19th APCOM symp., Littleton, Colorado (pp. 127–142).
Denby, B., & Schofield, D. (1996). Genetic algorithms for open pit scheduling-
extension into 3-dimensions. Paper presented at the MPES conference, Sao
time. On the other hand, the metaheuristics have the capability to Paulo, Brazil (pp. 177–185).
Denby, B., Schofield, D., & Surme, T. (1998). Genetic algorithms for flexible
get near optimal solutions for large cases but in small cases they
scheduling of open pit operations. Paper presented at the Application of
have not too much robustness. This study has presented a new computer in mineral industry. International symposium, London (pp. 605–616).
long-term open pit production planning technique based on the Denby, B., & Schofield, D. (1994). Open pit design and scheduling by use of genetic
concept of the Ant Colony Optimization (ACO). Its application has algorithms. Transactions of the Institution of Mining and Metallurgy (Section A:
Mining Industry), 103, A21–A26.
been tested on a real size Copper–Gold deposit using two variants Denby, B., & Schofield, D. (1995). Inclusion of risk assessment in open pit design and
of the ACO, including the MMAS and the ACS. The analysis revealed scheduling. Transactions of the Institution of Mining and Metallurgy (Section A:
that the ACO can successfully improve the value of the initial min- Mining Industry), 104, A67–A71.
Dorigo, M., & Gambardella, L. M. (1997b). Ant colony system: A cooperative learning
ing schedule generated by the traditional algorithms by up to 12% approach to the traveling salesman problem. IEEE Transactions on Evolutionary
in a reasonable computational time. Although, this improvement Computation, 1(1), 53–66.
was mainly due to the consideration of the penalties in the devia- Dorigo, M., & Stützle, T. (2003). The ant colony optimization metaheuristic:
Algorithms, applications, and advances. In Handbook of metaheuristics
tions from planning targets, however, the magnitude of the pure (pp. 250–285). Springer.
enhancements was also considerable. Dorigo, M., & Stützle, T. (2004). Ant colony optimization. New Dehli: Prentice-Hall of
The main advantages of the proposed the ACO algorithm are: India.
Dowd, P. A., & Onur, A. H. (1993). Optimization of open-pit mine design – Part 1:
Optimal open-pit design. Transactions of the Institution of Mining and Metallurgy
– The ability to consider any kind of objective functions. (Section A: Mining Industry), 102, A95–A104.
– The capability of considering variable slope angles. Gershon, M. E. (1983). Optimal mine production scheduling: Evaluation of large
scale mathematical programming approaches. International Journal of Mining
– The computation time is less sensitive to the size of the model.
Engineering, 1(4), 315–329.
– The amount of required memory is acceptable. Godoy, M., & Dimitrakopoulos, R. (2004). Managing risk and waste mining in long-
term production scheduling of open-pit mines. SME Transactions, 316, 43–50.
Moreover, its drawbacks could be considered as having no guar- Johnson, T. B. (1969). Optimum production scheduling. Paper presented at the
Proceedings of 8th APCOM symp., Salt Lake City (pp. 539–562).
antee to reach the global optimum schedule and being sensitive to Kumral, M., & Dowd, P. A. (2002). Short-term mine production scheduling for
the settings of the ACO parameters. In moving forward, a high industrial minerals using multi-objective simulated annealing. Paper presented
potential of incorporation of uncertainties related to the metal at the APCOM 2002-International symposium on the application of computers
and operations research in the minerals industry (30th: 2002: Littleton,
grades and commodity price fluctuations are seen. Colorado).
Kumral, M., & Dowd, P. A. (2005). Simulated annealing approach to mine production
Acknowledgment scheduling. Journal of the Operational Research Society, 56, 922–930.
Lamghari, A., & Dimitrakopoulos, R. (2012). A diversified Tabu search approach for
the open-pit mine production scheduling problem with metal uncertainty.
Authors appreciate the Sungun Copper Mine for providing the European Journal of Operational Research, 222(3), 642–652.
required software packages. The manuscript was significantly Lerchs, H., & Grossmann, I. F. (1965). Optimum design of open pit mines. Canadian
Institute of Mining Transactions, 68, 17–24.
improved by the careful reviews of two anonymous reviewers.
NPV – Scheduler (2001). Tutorial of Earthworks NPV-Scheduler software version
We thank them for their comments and suggestions. 3.4. Copyright Ó 2001 Mineral Industries Computing Ltd.
836 M.S. Shishvan, J. Sattarvand / European Journal of Operational Research 240 (2015) 825–836
Onur, A. H., & Dowd, P. A. (1993). Open pit optimization – Part 2: Production
scheduling and inclusion of roadways. Transactions of the Institution of Mining
and Metallurgy (Section A: Mining Industry), 102, A105–A113.
Ramazan, S. (2007). The new fundamental tree algorithm for production
scheduling of open pit mines. European Journal of Operational Research, 177,
1153–1166.
Sattarvand, J. (2009). Long-term open-pit planning by ant colony optimization
(Ph.D.), Ph.D. Thesis, RWTH Aachen University, Aachen, Germany.
Sattarvand, J., & Niemann-Delius, C. (2013). A new metaheuristic algorithm for
long-term open-pit production planning. Archives of Mining Sciences, 58(1),
107–118.
Shishvan, M. S., & Sattarvand, J. (2012). Modeling of accurate variable slope angles
in open-pit mine design using spline interpolation. Archives of Mining Sciences,
57(4), 921–932.
Stützle, T. (1999). Local search algorithms for combinatorial problems: analysis,
improvements, and new applications (Vol. 220): Infix Sankt Augustin, Germany.
Stützle, T., & Hoos, H. (1997). The MAX–MIN ant system and local search for the
travelling salesman problem. Paper presented at the Proceedings of the IEEE
international conference on evolutionary computation, Indianapolis, USA.
Stützle, T., & Hoos, H. (2000). MAX–MIN ant system. Future Generation Computer
Systems, 16(8), 889–914.