Askari-Nasab H. & Awuah-Offei K.

101- 1

101

Mixed integer linear programming formulations

for open pit production scheduling

Hooman Askari-Nasab and Kwame Awuah-Offei 1

Mining Optimization Laboratory (MOL)
University of Alberta, Edmonton, Canada

Abstract

We have proposed two mixed integer linear programming (MILP) formulations for large-
scale long-term open pit production scheduling problem. We developed, implemented, and
tested the proposed MILP theoretical frameworks for large-scale open pit production
scheduling.

1. Introduction

Historical assessment of mineral project performances has demonstrated the sensitivity of

projects’ profitability to decisions based upon mine production schedules. The life-of-mine
production schedule defines the complex strategy of displacement of ore, waste,
overburden, and tailings over the mine life. The objectives of long-term production
schedules are to determine the sequence of extraction and displacement of material in order
to maximize the future cash flows of mining operations within the existing economic,
technical, and environmental constraints. Long-term production schedules lead to
definition of reserves and are the backbone of short-term planning and day to day mining
operations. The long-term production schedules resolve mine and processing plant capacity
and their expansion potential; the production schedule, also defines the management
investment strategy. Deviations from optimal plans in mega mining projects will result in
enormous financial losses, delayed reclamation, and resource sterilization.
In this study, we have proposed four mixed integer linear programming (MILP)
formulations for large-scale long-term open pit production scheduling problem. We
developed, implemented, and tested the proposed MILP theoretical frameworks for large-
scale open pit production scheduling.
Current production scheduling methods in the literature are not just limited to, but can be
divided into three main categories: heuristic methods, applications of artificial intelligence
techniques, and operations research methods. Some of these algorithms are embedded into
available commercial software packages.

1
Assistant Professor, Department of Mining & Nuclear Engineering, University of
Missouri-Rolla, USA,
Askari-Nasab H. & Awuah-Offei K. 101- 2

One of the heuristic methods used in mine production scheduling was proposed by
Gershon (1987). XPAC AutoScheduler (Runge Limited, 1996-2009), a commercial mine
scheduling software is developed based on Gershon’s (1987) proposed heuristic. Gershon’s
(1987) algorithm generates cones upward from each block to approximate the shape of a
pit and to determine whether or not the block in question could be part of the schedule. A
list of exposed blocks and a ranking of those blocks based on what makes it more desirable
or less desirable to mine an exposed block at the present time is updated through the
algorithm with an index called the positional weight. This weighted function is used to
determine the removal sequence.
Another popular heuristic used in strategic mine planning software, such as Whittle
(Gemcom Software International, 1998-2008) and NPV Scheduler (Datamine Corporate
Limited, 2008) is based on the concept of parametric analysis introduced by Lerchs and
Grossmann (1965) (LG). The LG algorithm provides an optimal solution to the final pit
outline. There are unlimited numbers of strategies of reaching the final pit, which each has
a different discounted cash-flow. The optimal production schedule is the strategy that
would maximize the discounted cash-flow and meets all the physical and economical
constraints. The parametric analysis generates a series of nested pits based on varying the
price of the product (revenue factor) and finding an optimal pit layout using LG algorithm.
These nested pits then are used as a guideline to identify clusters of high grade ore and to
determine the production schedule. The main disadvantage of heuristic algorithms is that
the solution may be far from optimal and in mega mining projects, this is equal to huge
financial losses.
Various models based on a combination of artificial intelligence techniques have been
developed (Denby and Schofield, 1994; Denby et al., 1996; Tolwinski and Underwood,
1996; Askari-Nasab, 2006; Askari-Nasab et al., 2008; Askari-Nasab and Awuah-Offei,
2009). Tolwinski and Underwood (1996) used a method which combines concepts from
dynamic programming, stochastic optimisation, and artificial intelligence with heuristic
rules to obtain ultimate pit limit and production planning concurrently. The method works
by modelling the development of the mine as a sequence of pits where each pit differs from
the previous pit by the removal of blocks. A probability distribution based on the
frequency with which particular states occur is used to determine the state changes.
Heuristic rules are incorporated to learn these characteristics of the sequence of pits which
produce a good, or poor, result. Denby et al. (1996) employed genetic algorithms and
simulated annealing by generation of random pit population and assessment of a fitness
function to acquire the production schedule and final pit, concurrently. The advantages of
their method are flexibility and solution for the ultimate pit limit and production schedule
at the same time. The major drawback is that the results are not reproducible and there was
no measure of the optimality of the solution.
In a series of publications the authors developed and tested the intelligent agent-based
theoretical framework for open pit mine planning (Askari-Nasab et al., 2005; Askari-
Nasab, 2006; Askari-Nasab et al., 2007; Askari-Nasab and Szymanski, 2007; Askari-
Nasab et al., 2008) comprising algorithms based on reinforcement learning (Sutton and
Barto, 1998) and stochastic simulation. This intelligent open pit simulator (IOPS) (Askari-
Nasab, 2006) has a component that simulates practical mining push-backs over the mine
life. An intelligent agent interacts with the push-back simulator to learn the optimal push-
Askari-Nasab H. & Awuah-Offei K. 101- 3

back schedule using reinforcement learning. The intelligent agent-based mine planning
simulator, IOPS, was successfully used to determine the optimal push-back schedule of an
open pit mine with a geological block model containing 883 200 blocks (Askari-Nasab and
Awuah-Offei, 2009). A number of the artificial intelligence techniques, such as IOPS are
based on frameworks that theoretically will converge to the optimal solution, given
sufficient number of simulation iterations. The main disadvantage however, is that there is
no quality measure to solutions provided comparing against the theoretical optimum.
A variety of operations research approaches including linear programming (LP) and mixed
integer linear programming (MILP) have been applied to the mine production scheduling
problem. The pioneer work of Johnson (1969) used an LP model, which led to the MIP
formulations by Gershon (1983) for the production scheduling problem. Mixed integer
linear programming mathematical optimization models have the capability to consider
multiple ore processors and multiple elements during optimization. This flexibility of
mathematical programming models result in production schedules generating significantly
higher net present value (NPV) than those generated by the other traditional methods.
Every orebody is different, but for a typical open pit long-term scheduling problem, the
number of blocks is in the order of a couple of hundred thousands to millions and the
number of scheduling periods are twenty and more for a life-of-mine yearly schedule.
Evidently, the number of integer and linear decision variables, and the number of
constraints formulating a problem of this size would easily exceed the capacity of current
state of hardware and commercial mathematical optimization solvers.
Various models based on mixed integer linear programming mathematical optimisation
have been used to solve the long-term open-pit scheduling problem (Caccetta and Hill,
2003; Ramazan and Dimitrakopoulos, 2004; Dagdelen and Kawahata, 2007; Boland et al.,
2009). The applications of MILP models result in production schedules generating near
theoretical optimal net present values. In practice, formulating a real size mine production
planning problem by including all the blocks as integer variables will simply exceed the
capacity of the current commercial mathematical optimisation solvers. Various methods of
aggregation have been used to reduce the number of integer variables that are required to
formulate the mine planning problem with MILP techniques. Ramazan and
Dimitrakopoulos (2004) illustrated a method to reduce the number of binary integer
variables by setting waste blocks as continuous variables instead of integer variables.
Ramazan and Dimitrakopoulos (2004) reported a case study on a small single level nickel
laterite block model with 2030 blocks over three periods.
Ramazan et al. (2005; 2007) presented an aggregation method based on fundamental tree
concepts to reduce the number of decision variables in the MILP formulation. The
fundamental tree algorithm has been used in a case study with 38 457 blocks within the
final pit limits, Whittle strategic mine planning software (Gemcom Software International,
1998-2008) has been used to break-down the overall problem into four push-backs.
Subsequently, the blocks within the push-backs were aggregated into 5512 fundamental
trees and scheduled over eight periods using the formulation presented in Ramazan and
Dimitrakopoulos (2004). Information about the run-time of the MILP models are not
presented in Ramazan (2007); also the break-down of the problem into four push-backs
based on the nested pit approach and formulating them as a separate MILP would not
generate a global optimum solution to the overall problem. On the other hand the size of
Askari-Nasab H. & Awuah-Offei K. 101- 4

the problem of around thirty thousand blocks over eight periods is more a mid-range
planning problem rather than a long-term life of mine schedule.
Caccetta and Hill (2003) presented a formulation that used binary integer variables, they
developed and implemented a personalized branch-and-cut (Horst and Hoang, 1996)
method in C++ using CPLEX (Bixby, 1987-2009) to solve the relaxed LP sub-problems.
Boland et al. (2009) have demonstrated an iterative disaggregation approach to using a
finer spatial resolution for processing decisions to be made based on the small blocks,
while allowing the order of extraction decisions be made at an aggregate level. Boland et
al. (2009) reported notable improvements on the convergence time of their algorithm for a
model with 96 821 blocks and 125 aggregates over 25 periods. However, combining 96
821 blocks into only 125 aggregates would reduce the freedom of decision variables and
the schedule generated could not be considered as an optimal solution in comparison to the
case that 96 821 blocks had complete freedom. Moreover, in Boland et al. (2009) there is
no representation of the generated schedules in terms of annual ore and waste production,
average grade of ore processed, and cross sections and plan views of the schedules to
assess the practicality of the solutions from mining operational point of view. Boland et al.
(2009) also did not represent enough information on their method of aggregation, they
assumed that an aggregation method similar to Ramazan (2007) would be used.
MineMax (Minemax Pty Ltd, 1998-2009) is a commercially available strategic mine
scheduling software, which uses an MILP formulation solved by ILOG CPLEX (Bixby,
1987-2009) solver. Given that, MineMax is a commercial software we couldn’t find
detailed information about the approach and formulation, but our understanding from the
evaluation of the demo tutorial version of MineMax is that as a general strategy it is
suggested to initially breakdown the final pit into nested pit shells based on parametric
analysis concepts represented by Lerchs and Grossmann (1965). The pit shells define a pit
to pit precedence constrained by the minimum and maximum number of benches by which
the mining of one specified pit shell is to lag behind the previous one. The other option to
define rules for precedence of extraction is either by proportions mined on each bench or
by block precedence based on the overall pit slopes. Next each pit shell is formulated as a
separate MILP model which can contribute to the overall quantity of mining and
processing targets within the grade and precedence constraints; this approach result in
MILP formulations for each pit shell with smaller size which will converge faster, but it
could not be considered a global optimization of the problem since the pit shells are
defined by the parametric analysis first. Another optimization strategy is using sliding
windows which are sub-problems that are tackled on a period by period basis.
Blasor (Stone et al., 2007) and Prober (Whittle, 2007) are other proprietary software which
tackle the strategic mine production scheduling by an MILP. Current MILP formulations
used for open pit production scheduling fail because of: (1) inability to solve large-scale
real-size mining problems as global optimization problem, and (2) inability to quantify the
geological uncertainty inherent within the problem and, as a result, the associated risk with
the mine plans. To overcome these problems the first step is to develop, implement, and
test a theoretical framework that is capable of handling real-size mining problems in a
global optimization framework. The development of an MILP model that can handle a
deterministic large-scale mine production schedule will fulfill this objective.
Askari-Nasab H. & Awuah-Offei K. 101- 5

We have critically reviewed the MILP formulations of the open pit production scheduling
problem. We have implemented, and tested two of these MILP formulations. The
shortcomings and deficiencies of the formulations were documented. We have proposed
two MILP formulations for the long-term open pit production scheduling problem to
overcome the shortcomings of the reviewed methods. We have divided the major decision
variables into two categories, continuous variables representing the portion of a block that
is going to be extracted in each period and binary integer variables controlling the order of
extraction of blocks or the precedence of mining-cuts through a dependency directed graph
using depth-first-search algorithm. The depth-first-search algorithm component, added a
very important practical mining feature to the model. This model allows variable pit slopes
to be integrated in the MILP formulation. As a practical constraint our MILP formulation
also ensures that the fractional extraction of blocks are not going to be split over more than
three periods. We have implemented the optimization formulation in TOMLAB/CPLEX
(Holmström, 1989-2009) environment. An iron ore mine intermediate scheduling case
study over twelve periods was carried out to verify and validate the models and to illustrate
the effectiveness of the generated schedules from mining point of view. The results proved
that the models are accurate and efficient and showed enhanced CPU time performance
comparing to the reviewed models.
The next section of the paper covers the assumptions, problem definition, and the notations
of variables. Section 3 presents four mixed integer linear programming formulations of the
problem, while Section 4 presents the numerical modelling. The next section represents the
numerical experiments used for verification and validation of the models. Finally, Section
6 presents the conclusions and future work followed by the list of references.

2. Assumptions, problem definition, and notation

We assume that the orebody is represented by a geological block model, which is a three-
dimensional array of rectangular or cubical blocks used to model orebodies and other sub-
surface structures. Numerical data are used to represent a single attribute of the orebody
such as: rock types, densities, grades, elevations, or economic data. The size of the block
that is needed for outlining the orebody depends on the shape and size of the ore body,
mining bench height, mining method, and mining equipment. Geostatistics provides
accurate and reliable estimations or simulation of block attributes at locations where no
measurements are available. The most common estimation method used is Kriging (Krige,
1951). However, Kriging results do not capture uncertainty and may be systematically
biased. Uncertainty always exists in presence of sparse geological data. Conditional
simulation algorithms such as Sequential Gaussian Simulation (Isaaks, 1990) are
geostatistical methods used to assess geological uncertainty. The generated realizations are
equally probable and represent plausible geological outcomes. We will tackle the
deterministic problem in this study so Kriging is the estimation method of choice.
We assume that each block is subdivided into smaller regions identified as parcels. A
parcel is part of a block for which the rock-type, tonnage and element content are estimated
or simulated by Geostatistical methods. A block may contain zero or more parcels. The
total tonnage of the parcels may be the same as the tonnage of the block, or it may be less.
If it is less, the difference is called undefined waste, which is waste of unknown rock-type.
If a block has no parcels, the total tonnage of the block is undefined waste. Neither the
Askari-Nasab H. & Awuah-Offei K. 101- 6

position of a parcel within a block, nor its shape, are defined (Gemcom Software
International, 1998-2008). The spatial location of each block is defined by the coordinates
of its center; attributes such as rock-type, quantity of valuable and contaminant elements,
and block tonnage are estimated and grade of each element of interest is calculated based
on the quantity of the attributes and the ore tonnage in that block.
We consider that the geological block model is going to be extracted using open pit mining
techniques. We assume that a classical optimum design of final pit limits is carried out;
either based on graph theory (Lerchs and Grossmann, 1965; Zhao and Kim, 1992) or
network flow algorithm (Johnson and Barnes, 1988; Yegulalp and Arias, 1992). This pit
outline represents reserves that would maximize the profit. We have illustrated in Askari-
Nasab and Awuah-Offei (2009) that a final pit outline obtained directly by using an
optimal long-term scheduling algorithm will result in a pit outline that is a subset of the
conventional final pit outline generated by the Lerchs and Grossmann’s algorithm (1965).
Consequently, we will follow the classical process of open pit long-term scheduling of
first, finding the final pit limits and then generating a production schedule within the final
pit outline.
The basic problem, in its simplest form, is finding a sequence in which ore and waste
blocks should be removed from the predefined open pit outline and their respective
destinations, over the life of mine, so that the net present value of the operation is
maximized. The production schedule is subject to a variety of physical, technical and
economic constraints. The constraints enforce the mining extraction sequence, overall pit
slopes, mining, milling, and refining capacities, blending material to meet head grade
requirements, minimum mining width, and the number of active mining benches in each
production period. The problem presented here involves scheduling of N different ore and
waste blocks within the final pit outline over T different periods of extraction.
Blocks within the same level or mining bench are grouped into clusters based on their
attributes, spatial location, rock type, and grade distribution. We refer to these clusters of
blocks as mining-cuts. Similar to blocks, each mining-cut has coordinates representing the
center of the cut and its spatial location. Fig. 1 illustrates a schematic plan view of a
mining bench. Blocks are aggregated into mining-cuts.

Fig. 1. Schematic plan view of aggregated blocks into mining-cuts on a mining bench.
Askari-Nasab H. & Awuah-Offei K. 101- 7

This grade aggregation methodology will summarize ore data, but will maintain a relevant
separation of lithology. Instead of the common approach of estimating grades for the
element of interest in blocks, the total amount of the element in the block would be
modeled. This tactic will allow to combine the smaller blocks into mining-cuts without
sacrificing the accuracy of the estimated values and to model a more realistic equipment
movement strategy. The mining-cut clustering algorithm developed uses fuzzy logic
clustering (Kaufman and Rousseeuw, 1990) and is not within the scope of this paper and
we will disseminate the clustering approach in another publication.

2.1. Notation

We will present four different MILP formulations for the open pit production scheduling
problem. The notation of decision variables, parameters, sets, and constraints are as
follows:

2.1.1 Sets

N {1,..., N } set of all the blocks in the model.

K {1,..., K } set of all the mining-cuts in the model.
A {1,..., A} set of all directed arcs in the blocks’ precedence directed graph
denoted by Gb (N, A).
B {1,..., B} set of all edges in the mining-cuts precedence directed graph
denoted by Gc (K, B).
D( J ) for each block, n, there is a set D( J ) N which includes all the
blocks that must be extracted prior to mining block n to ensure that
block n is exposed for mining with safe pit slopes, where J is the
total number of blocks in set D ( J ) .
C ( L) for each block, n , there is a set C ( L) D( J ) defining the
immediate predecessor blocks that must be extracted prior to
extraction of block n , where L is the total number of blocks in set
C ( L) .
H (S ) for each mining-cut ck , there is a set H ( S ) K defining the
immediate predecessor cuts that must be extracted prior to extracting
mining-cut k , where S is the total number of cuts in set H ( S ) .

2.1.2 Indices

A general parameter f can take four indices in the format of f ke,,nt . Where:
t {1,...., T } index for scheduling periods.
k {1,..., K } index for mining-cuts.
n {1,..., N } index for blocks.
e {1,..., E} index for elements of interest in each block.
Askari-Nasab H. & Awuah-Offei K. 101- 8

2.1.3 Parameters

d nt d nt is the discounted profit generated by extracting block n in

period t .
vnt the discounted revenue generated by selling the final product within
block n in period t minus the extra discounted cost of mining all
the material in block n as ore and processing it.
qnt the discounted cost of mining all the material in block n as waste.
ck mining-cut k .
e e
g ,g
n k average grade of element e in ore portion of block n and average
grade of element e in ore portion of mining-cut k .
gu e ,t upper bound on acceptable average head grade of element e in period
t.
gl e ,t lower bound on acceptable average head grade of element e in
period t .
on , ok ore tonnage in block n and ore tonnage in mining-cut k .
wn , w k waste tonnage in block n and waste tonnage in mining-cut k .
pu t upper bound on processing capacity of ore in period t (tonnes).
pl t lower bound on processing capacity of ore in period t (tonnes).
mu t upper bound on mining capacity in period t (tonnes).
ml t lower bound on mining capacity in period t (tonnes).
r e ,t processing recovery, is the proportion of element e recovered in
time period t.
p e ,t price in present value terms obtainable per unit of product
(element e ).
cs e,t selling cost in present value terms per unit of product (element e ).
cp e ,t extra cost in present value terms per tonne of ore for mining and
processing.
cmt cost in present value terms of mining a tonne of waste in period t .

2.1.4 Decision Variables

Model 01
znt {0,1} binary integer variable, equal to 1 if block n is to be mined in
period t , otherwise 0.

Model 02
unt [0,1] continuous variable, representing the portion of block n to be
extracted and processed in period t or mined and treated as waste in
period t .
Askari-Nasab H. & Awuah-Offei K. 101- 9

ant {0,1} binary integer variable controlling the precedence of extraction of

blocks. ant is equal to one if extraction of block n has started by or in
period t , otherwise it is zero.

Model 03
xnt [0,1] continuous variable, representing the portion of block n to be
extracted as ore and processed in period t .
ykt [0,1] continuous variable, representing the portion of mining-cut ck to be
mined in period t , fraction of y characterizes both ore and waste
included in the mining-cut.
bkt {0,1} binary integer variable controlling the precedence of extraction of
mining- cuts. bkt is equal to one if extraction of mining-cut ck has
started by or in period t , otherwise it is zero.
Model 04
skt [0,1] continuous variable, representing the portion of mining-cut ck to be
extracted as ore and processed in period t .

Decision variables ykt and bkt with the same definition as in Model
03 are used in Model 04 as well.

2.2. Economic block value modeling

The objective functions of the MILP formulations are to maximize the net present value of
the mining operation. Hence, we need to define a clear concept of economic block value
based on ore parcels which could be mined selectively. The profit from mining a block
depends on the value of the block and the costs incurred in mining and processing. The
cost of mining a block is a function of its spatial location, which characterizes how deep
the block is located relative to the surface and how far it is relative to its final dump. The
spatial factor can be applied as a mining cost adjustment factor for each block according to
its location to the surface. The discounted profit from block n is equal to the discounted
revenue generated by selling the final product contained in block n minus all the
discounted costs involved in extracting block n , this is presented by Eqs. (1) and (2).
discounted profit = discounted revenue - discounted costs (1)

E E
d nt [ on g ne r e ,t ( p e ,t cs e ,t ) on cp e ,t ] [(on wn ) cmt ] (2)
e 1 e 1
discounted revenues discounted cos ts

For simplification purposes we denote:

E E
vnt [ on g ne r e ,t ( p e ,t cs e ,t ) on cp e ,t ] (3)
e 1 e 1
Askari-Nasab H. & Awuah-Offei K. 101- 10

qnt (on wn ) cmt (4)

3. Mixed integer linear programming models for open pit production scheduling

We present four different formulations for the open pit production scheduling problem,
with the objective function to maximize the NPV of the mining operation. It is intuitively
apparent that higher NPV’s could be achieved by block models with small block sizes and
high resolution of the orebody model. The block sizes for production scheduling must be
chosen similar to the selective mining size. If the size of the block is not properly defined,
the generated schedule will be simulating the mining operation with a selectivity that could
not be achieved in practice. We have developed, implemented, tested, and compared four
MILP formulations based on different precedence of extraction graphs and extraction and
processing selectivity levels. The four MILP formulations are: (i) Model 01- this
formulation is similar to Ramazan and Dimitrakopoulos (2004), it only consists of binary
integer decision variables, znt , and generates a schedule at block level resolution. We have
extended the model by integrating a variable pit slope component into the implementation;
(ii) Model 02 – we have proposed this model based on the concepts presented in Caccetta
and Hill (2003). The schedule is generated based on a strict temporal sequence of blocks.
Caccetta and Hill (2003) formulation only uses binary integer decision variables, which
makes the size of branch and cut tree intractable for a large scale problem. The proposed
formulation uses continuous variables, unt to model extraction and processing at block
level and binary integer decision variables, ant , are used to control precedence of
extraction; (iii) Model 03 – this formulation is developed based on the concepts of Boland
et al. (2009), processing is controlled at block level with continuous decision variables, xnt ;
where ykt , controls the extraction at mining-cut level. Also, the precedence of extraction of
blocks is controlled at the mining-cut level by means of binary integer variables bkt . The
continuous decision variables ( xnt and ykt ) lead to fractional block extraction, the proposed
model provides control over the maximum number of fractions that each block would
take; (iv) Model 04 – this formulation is based on a combination of concepts presented in
Models 02 and 03. Extraction, processing, and order of block extraction are controlled at
mining-cut level.

3.1. Model 01 – extraction and processing at block level – only binary decision
variables

Objective function:
T N
max d nt znt (5)
t 1 n 1

Subject to:
N N
gl t ,e g ne on znt on znt gu t ,e t 1,..., T , e {1,..., E} (6)
n 1 n 1
Askari-Nasab H. & Awuah-Offei K. 101- 11

N
pl t on znt pu t t 1,..., T (7)
n 1
N
ml t (on wn ) znt mu t t 1,..., T (8)
n 1
J t
J znt z ij 0 n {1,..., N }, t 1,..., T , j D( J ) (9)
j 1 i 1
T
znt 1 n {1,..., N } (10)
t 1

Where Eq. (6) is grade blending constraints; these inequalities ensure that the head grade
of the elements of interest and contaminants are within the desired range in each period.
There are two equations (upper bound and lower bound) per element per scheduling period
in Eq. (6). Eq. (7) is processing capacity constraints; these inequalities ensure that the
total ore processed in each period is within the acceptable range of processing plant
capacity. There are two equations (upper bound and lower) per period per ore type. Eq. (8)
is mining constraints; these inequalities ensure that the total tonnage of material mined
(ore, waste, overburden, and undefined waste) in each period is within the acceptable range
of mining equipment capacity in that period. There are two equations (upper bound and
lower bound) per period. Eq. (9) controls the precedence relationship of block extraction
and pit slopes. For each block there is a set D( J ) n N ,which includes all the blocks that
must be extracted prior to mining block n to ensure that block n is exposed for mining
with safe pit slopes, where J is the total number of blocks in set D( J ) n . The set D( J )1 for
the block labeled as 1 is hatched in Fig. 2. J is the total number of blocks in the set; it is
equal to 28 in this case. Fig. 2 also demonstrates how variable slopes could be modeled by
means of constructing the set D( J ) ; the block labeled as 5 is added to set D( J )1 to
construct a model with flatter slope on the right hand side. There is one equation per block
per period for Eq. (9). Depending on the spatial location of the block in the block model
the set of D( J ) would have different number of blocks. The slope constraints presented by
Eq. (9) is the main reason of increase in the number of constraints and the complexity of
Model 01 formulation. Finally, Eq. (10) defines reserve constraints; we assume that a final
pit limit is superimposed on the block model and we are going to schedule the extraction of
all the blocks within the final pit limit or push-back. Eq. (10) ensures that all the blocks
within the final pit are going to be extracted once.

3.2. Model 02 - extraction and processing at block level – binary and continuous
variables

In this model the mining and processing are at block level resolution; the schedule is
controlled by continuous variables so fractional extraction of blocks may occur. The order
of block extraction is controlled by binary integer variables at block level.
Objective function:
T N
max d nt unt (11)
t 1 n 1
Askari-Nasab H. & Awuah-Offei K. 101- 12

Fig. 2. Precedence of block extraction in the proposed MILP models.

Subject to:
N N
gl t ,e g ne on unt on unt gu t ,e t 1,..., T , e {1,..., E} (12)
n 1 n 1
N
pl t on unt pu t t 1,..., T (13)
n 1
N
ml t (on wn ) unt mu t t 1,..., T (14)
n 1
t
ant uli 0 n {1,..., N }, t 1,..., T , l C ( L) (15)
i 1
t
uni ant 0 n {1,..., N }, t 1,..., T (16)
i 1

ant ant 1 0 n {1,..., N }, t 1,..., T 1 (17)

T
unt 1 n {1,..., N } (18)
t 1

Eq. (12) controls the grade blending constraints. Eqs. (13) and (14) are processing and
mining capacity constraints. Eqs. (15) to (17) control the relationship of block extraction
precedence by binary integer variables at block level. Model 02 only requires the set of
immediate predecessors’ blocks on top of each block to model the order of block extraction
relationship. This is presented by set C ( L) in Eq. (15). Fig. 2 illustrates the set C ( L)1 for
the block labeled as 1, the set includes C ( L)1 {2,3, 4,5} . Where L the number of blocks is
equal to 4; compare this to formulation in Model 01 where J was 28. Eq. (18) ensures that
the fractions of blocks that are extracted over the scheduling periods are going to sum up to
one, which means all the block within the final pit outline are going to be scheduled.
Askari-Nasab H. & Awuah-Offei K. 101- 13

3.3. Model 03 – extraction at mining-cut level and processing at block level – binary
and continuous variables

In this model, processing is at block level and extraction is at mining-cut level. The amount
of ore processed is controlled by the continuous variable xnt , this allows fractional
extraction of blocks in different periods. The order of block extraction which is controlled
by the directed graph Gb (N, A) is transferred into the control of the precedence of mining-
cuts by means of directed graph Gc (K, B). The precedence of mining-cuts relationship is
modeled via the binary integer variable bkt . This is superior to Models 01 and 02, in that,
the amount of ore processed and amount of material mined are controlled by two separate
variables.
Objective function:
T K
max vnt xnt ( qnt ) ykt (19)
t 1 k 1 n ck n ck

Subject to:
N N
gl t ,e g ne on xnt on xnt gu t ,e t 1,..., T , e {1,..., E} (20)
n 1 n 1
N
pl t on xnt pu t t 1,..., T (21)
n 1
K
ml t (on wn ) ykt mu t t 1,..., T (22)
k 1 n ck

xnt ykt ,n n {1,..., N } , n ck , t 1,..., T (23)

t
bkt ysi 0 k {1,..., K }, t 1,..., T , s H ( S ) (24)
i 1
t
yki bkt 0 k {1,..., K }, t 1,..., T (25)
i 1
t
bk bkt 1
0 k {1,..., K }, t 1,..., T 1 (26)

Eqs. (20) to (22) control the grade blending, processing capacity, and mining capacity. In
Model 02 grade blending, processing capacity, and mining capacity are modeled by only
one decision variable, unt . In Model 03 the extraction and processing of ore is controlled by
continuous decision variable xnt at block level, where mining is modeled using a
continuous variable ykt at mining-cut level. This method enables us to have a high
resolution solution for selection of ore and processing. The total amount of material mined
and the order of extraction is then modeled at mining-cut level ykt . This approach reduces
the number of binary integer variables in the model drastically. Fig. 2 illustrates set
H ( S ) {B, C} , the predecessor mining-cuts that must be extracted prior to extraction of
mining-cut A. Eq. (23) represents inequalities that ensure the amount of ore of any block
which is processed in any given period is less than or equal to the amount of rock extracted
Askari-Nasab H. & Awuah-Offei K. 101- 14

from the mining-cut that the block belongs to in the considered time period. For each
mining-cut, k , Eqs. (24) to (26) check the set of immediate predecessor cuts that must be
extracted prior to extracting mining-cut, k .

3.4. Model 04 – extraction and processing at mining-cut level – binary and continuous
variables

In this model, mining and processing are both at mining-cut level. The blocks are
aggregated prior to schedule optimization and the ore processing and mining are controlled
by two continuous variables.
Objective function:
T K
max (vkt skt qkt ykt ) (27)
t 1 k 1
K K
gl t ,e g ke ok skt ok skt gu t ,e t 1,..., T , e {1,..., E} (28)
k 1 k 1
K
pl t ok skt pu t t 1,..., T , e {1,..., E} (29)
k 1
K
ml t (ok wk ) ykt mu t t 1,..., T (30)
k 1

skt y t
k k {1,..., K } , t 1,..., T (31)
Equations (24) to (26)
Eqs. (28) to (30) control the grade blending, processing capacity, and mining capacity
constraints at mining-cut level with fractional extraction from mining-cuts. The extraction
from mining-cuts is assumed to be uniform among all the blocks which belong to that
mining-cut. Eq. (31) ensures that the amount of ore extracted and processed from any
mining-cut in any given period is less than or equal to the amount of rock extracted from
that mining-cut. Eqs. (24) to (26) are similar to those demonstrated in Model 03 for
precedence of block extraction.

4. Numerical modeling

In most linear optimization problems, the variables of the objective function are continuous
in the mathematical sense, with no gaps between real values. To solve such linear
programming problems, ILOG CPLEX implements optimizers based on the simplex
algorithms (Winston, 1995) (both primal and dual simplex) as well as primal-dual
logarithmic barrier algorithms.
Branch and cut is a method of combinatorial optimization for solving integer linear
programs. The method is a hybrid of branch and bound and cutting plane methods (Horst
and Hoang, 1996). Refer to Wolsey (1998) for a detailed explanation of the branch and cut
algorithm, including cutting planes. In recent years there has been significant
improvements in mathematical programming optimizers such as ILOG CPLEX (Bixby,
1987-2009). This optimizer uses branch and cut techniques to solve MILP models and it
makes the latest theory in optimization of large-scale industrial problems available
Askari-Nasab H. & Awuah-Offei K. 101- 15

commercially. In this study we used TOMLAB/CPLEX version 11.2 (Holmström, 1989-

2009) as the MILP solver. TOMLAB/CPLEX efficiently integrates the solver package
CPLEX (ILOG Inc, 2007) with MATLAB environment (MathWorks Inc., 2007). An
important termination criterion that the user can set explicitly in CPLEX is the MILP gap
tolerance. We have used the relative MILP gap tolerance, which indicates to CPLEX to
stop when an integer feasible solution has been proved to be within the gap tolerance of
optimality.

4.1. Size and complexity

One of the major obstacles in using the MILP formulations for mine production scheduling
is the sheer size of the problem. The number of blocks, N, in the model is usually between
tens of thousands to millions which will lead to a formulation with an objective function
with many variables. Moreover, the main physical constraint in open pit mining is the
block extraction precedence modeled by binary integer variables. This set of constraints
also controls the overall pit slope in different regions. The numbers of blocks that must be
extracted prior to mining each block are numerous and will result in formulations with
many constraints. Therefore, we are dealing with an MILP formulation with many
variables and many constraints.
The most common difficulty with MILPs is the size of the branch and cut tree. The tree
becomes so large that insufficient memory remains to solve an LP sub-problem. The
number of binary integer variables in the formulations determines the size of the branch
and cut tree. As a general strategy in our formulations we aimed at reducing the number of
binary integer variables, we also focused on developing formulations that will mainly use
continuous optimization techniques rather than discrete optimization. Table 1 shows the
number of decision variables and the number of binary integer variables required for the
proposed MILP formulations as a function of number of blocks, N, number of mining-
cuts, K, and number of scheduling periods, T. The goal has been to reduce the number of
binary integer variables in the models by introducing mining-cuts as the means of
controlling the precedence of extraction of blocks rather than having one binary integer
variable per block.
Table 1 – number of decision variables in the MILP formulations

MODEL Number of decision variables Number of integer variables

Model 01 N T N T
Model 02 2 N T N T
Model 03 (2 K N) T K T
Model 04 3 K T K T

5. Results and discussions

We have developed, implemented, and tested the proposed MILP models presented in
section 3 in TOMLAB/CPLEX environment (Holmström, 1989-2009). We compare the
performances of the proposed models based on net present value generated, practical
Askari-Nasab H. & Awuah-Offei K. 101- 16

mining production constraints, smoothness of the generated schedules, size of the

mathematical formulations, the number of integer variables required in formulation, and
computational time required for convergence.
All the developed formulations are verified by numerical experiments on a synthetic data
set containing 120 blocks and a real mining push-back including 2 598 blocks with seven
mining benches. We have also validated Model 03 and Model 04 with an iron ore life-of-
mine schedule with 26 000 blocks over a 20 year scheduling horizon. We tested our
models on a Quad Core Dell Precision T7400 computer at 3.00 GHz, with 3.25 GB of
RAM. Since we aimed at a comparative analysis of the proposed models a relative
tolerance of 2% on the gap between the best integer objective and the feasible integer
solution was chosen.
Table 2a and b show the numerical results of the tests of MILP models with the data set
containing 120 blocks over four periods of extraction. To reach a feasible solution in
different models we have used a mining capacity upper bound of 64 to 66 million tonnes
per period, whereas the processing capacity varies between 4.9 to 5.3 million tonnes per
period. We were forced to set different upper and lower bounds for different models,
because the tighter boundaries on some models resulted in infeasible solutions. As
expected Model 02 generated the highest NPV; in this formulation processing and mining
are both at block level with continuous variables so fractional extraction of blocks are
allowed. The variables in Model 02 have the highest resolution among all the other
models. Model 01 had the longest runtime with 42.53 seconds as expected; this is due to
the numerous numbers of constraints that are generated by Eq. (9). Also, the model is a
pure MIP with no continuous variables, the formulation searched 4 432 branch and bound
nodes to reach an optimized solution within a 2.25% gap. Ramazan et al. (2005) proposed
to reduce the number of binary integer variables required by Model 01 by defining only the
ore blocks as binary variables; this approach will reduce the number of binary integer
variables but since the formulation assigns blocks to periods of extraction, rather than
determining a strict temporal sequence of blocks, still the size of the problem and the
runtime even for a small number of blocks is not within a reasonable timeframe and simply
this formulation is not a practical tool. In Models 03 and 04 the concept of mining-cuts are
introduced and the number of binary integer variables and the computational time required
is reduced drastically compared to Model 01 and 02 (>72% and >99.6%, respectively). On
the other hand, the NPV of Model 03 and 04 are reduced by 0.72% and 1.11% when
compared to Model 02, as in Model 03 and Model 04 the variables are aggregated and
have less freedom in the MILP formulations.
Table 2a – Inputs and numerical results for the synthetic data set containing 120 blocks
Blocks – Cuts Periods mu t / ml t pu t / pl t NPV
MODEL
(N K ) (T ) (MT) (MT) ($M)
01 120 - 0 4 64 / 0 5.3 / 0 387.88
02 120 - 0 4 64 / 0 4.9 / 0 391.34
03 120 - 21 4 66 / 0 5.0 / 0 388.53
04 120 - 21 4 66 / 0 5.0 / 0 387.00
Askari-Nasab H. & Awuah-Offei K. 101- 17

Table 2b – Inputs and numerical results for the synthetic data set containing 120 blocks
Coefficient Matrix No.
Root Node CPU TIEM No. Integer No. of B & B
MODEL A(rows col) decision
Gap % (S) variables nodes visited
variables
01 2.25 42.53 372 × 480 480 480 4432

02 2.37 0.61 3004 × 960 960 480

38
03 1.29 0.17 920 × 648 648 84 0

04 1.40 0.15 530 × 252 252 84 0

Table 3a and Table 3b show the numerical results of the MILP models for an iron ore
push-back data set containing 2 598 blocks over twelve scheduling periods. Fig. 3 and 4
illustrate cross sections of the final pit limits including the orebody and rock type model.
The push-back studied includes seven benches of the final pit demonstrated in Fig. 3 and 4
from elevation 1500m to 1590m. The blocks represent a volume of rock equal to
50m 25m 15m. The model contains 155 million tonnes of material with 84 million
tonnes of iron ore with an average grade of 73% magnetic weight recovery (MWT%).
Sulfur and phosphor are present as deleterious elements and their grades need to be
controlled within an acceptable range in the processing plant feed.
Table 3a - Inputs and numerical results for the data set containing 2598 blocks

MODEL
Blocks -Cuts Periods mu t / ml t pu t / pl t gu e ,t / gl e ,t gu e ,t / gl e ,t NPV
(N K ) (T ) (MT) (MT) (MWT%) (S% & P%) ($M)
01 2598 - 0 12 13 / 0 7.15 / 0 No bounds 0 / 1.8 & 0.14 -
02 2598 - 0 12 13 / 0 7.15 / 0 No bounds 0 / 1.8 & 0.14 3011.65
03 2598 - 148 12 13.2 / 0 7.15 / 0 65 / 80 0 / 1.8 & 0.14 2947.68
04a 2598 - 148 12 13.5 / 0 7.15 / 7.0 65 / 80 0 / 1.8 & 0.14 2947.09
04b 2598 - 239 12 13.5 / 0 7.15 / 7.0 65 / 80 0 / 1.8 & 0.14 2985.02
04c 2598 - 436 12 13.5 / 0 7.15 / 7.0 65 / 80 0 / 1.8 & 0.14 2991.55

Table 3b – Inputs and numerical results for the data set containing 2598 blocks
Root Node CPU TIME Coefficient Matrix No. Of nonzero No. Integer No. of B & B
MODEL
GAP % (S) A(rows cols) elements in A variables nodes visited

01 No Integer Solution - 5 668 788 31 176 -

02 0.05 14 748.32 280 944 × 62 352 2 091 882 31 176 340

03 1.08 1 168.60 43 202 × 34 728 226 774 1 776 70

04a 0.52 26.73 11 232 × 5 325 75 640 1 776 50

04b 0.35 34.37 17 580 × 8 604 115 898 2 868 40

04c 0.58 34.94 30 264 × 15 696 198 082 5 232 30

Askari-Nasab H. & Awuah-Offei K. 101- 18

Fig. 3. Cross section 98400 N of the final pit including the orebody and rock type model looking
west (meters).

Fig. 4. Cross section 600640 E of the final pit including the orebody and the rock type model
looking north (meters).

The maximum allowable average grade for sulfur is 1.8% and for phosphor is 0.14% per
period. It is also desirable to keep an average head grade between 65% and 80% of
magnetic weight recovery (MWT). Our goal was to generate a schedule with a uniform
feed within a range of 7 to 7.15 million tonnes of ore per period. Furthermore, we intended
to keep a steady (1.89 to 1.93) stripping ratio over the scheduling horizon, so we chose a
maximum mining equipment capacity of 13.5 million tonnes per period. This would ensure
that the mining equipment capacity required is not going to fluctuate over time.
To schedule 2598 blocks over 12 scheduling periods, Model 01 generated a huge
coefficient matrix with more than 5.6 million nonzero elements and an integer solution for
the formulation did not exist (Table 3b). Model 02 generated a coefficient matrix with
Askari-Nasab H. & Awuah-Offei K. 101- 19

more than 2 million nonzero elements and ran for over four hours. As expected, Model 02
generated the highest NPV of $3011.65 million of all the models, with a 0.05% gap
tolerance. Fig. 5a to 5c show the plan view and cross sections of the generated schedule by
Model 02. A closer examination of Fig. 5a to 5c reveals a tight schedule from a practical
mining point of view. Fig. 5b shows that in the second period mining occurs on six active
benches. This requires considerable equipment movement. Implementation of such a
schedule is a challenge in the field, if limited numbers of mining shovels are available.
In Model 03 we have used clustering techniques to aggregate the blocks into 148 mining-
cuts, clustering has reduced the number of integer variables to 1 776 from 31 176 in Model
02. Although in Model 03 the processing is at block level, the NPV has dropped to
$2947.68 million, a 2.1% reduction (Table 3b). This is because mining occurs at the
mining-cut level and has reduced flexibility when compared to Model 02. The run time has
been reduced to half an hour from four hours.
Figs. 6a to 6c illustrate the plan view and cross sections of the schedule generated by
Model 03. This schedule is more practical from a mining point of view since there are only
three active benches in the second period compared to the six active benches of Model 02.
We have investigated the effect of number of mining-cuts on the quality of solutions in
terms of NPV and the run-time on three different cases with Model 04. We clustered the
2598 blocks into 148, 239, and 436 mining-cuts using the clustering algorithm. We refer to
these as Model 04a, 04b, and 04c, respectively. A model with fewer mining-cuts will have
fewer number of integer variables in the MILP formulation and result in reduced run-time,
as shown in Table 3b.
Alternatively, models with more mining-cuts imply more freedom for the decision
variables and higher NPV would be expected. The NPV for Model 04a with 148 mining-
cuts is $2947.09 million, while the NPV for the Model 04c with 436 cuts has increased to
$2991.55 million. One of the very important improvements with the Model 04
formulations is the drastic drop in runtime. Compare the four hour runtime for Model 02
to the less than 35 seconds for Model 04. This is a very significant improvement; one must
take into account that we are examining a very small model with only 2598 blocks in this
study. The increase in the number of blocks would very quickly make Model 02 intractable
and the model is not going to converge for larger real size models. Scrutinizing Fig. 7a to
7c shows that the number of active benches for different periods of extraction has
decreased to two or three benches. The reduction of number of active benches is because of
clustering blocks into mining-cuts. A smaller number of mining-cuts would generate
clusters with more blocks. On the other hand, we are constructing the mining-cuts on a
bench by bench basis; which tends to generate schedules that expand the mine outlines
more horizontally rather than vertically. More research is required to develop a framework
that will optimize the number of blocks included in each mining-cut in terms of making a
balance in aiming for the highest NPV possible, while the maximum number of active
benches included in the schedule would be practical from a mining point of view.
One of the features that make MILP formulations a robust platform for mine planning is
not only the NPV maximization but also the control that the mine planner would have on
upper and lower bounds of ore and waste production targets.
Askari-Nasab H. & Awuah-Offei K. 101- 20

Fig. 5a. Model 02, plan view of bench 1575m.

Fig. 5b. Model 02, cross section 98300m looking east.

Fig. 5c. Model 02, cross section 600675m looking north.

Askari-Nasab H. & Awuah-Offei K. 101- 21

Fig. 6a. Model 03, plan view of bench 1575m.

Fig. 6b. Model 03, cross section 98300m looking east.

Fig. 6c. Model 03, cross section 600675 looking north.

Askari-Nasab H. & Awuah-Offei K. 101- 22

Fig. 7a. Model 04a, plan view of bench 1575m.

Fig. 7b. Model 04a, cross section 98300m looking east.

Fig. 7c. Model 04a, cross section 600675m looking north.

Askari-Nasab H. & Awuah-Offei K. 101- 23

Fig. 8a. Model 02 tonnage of ore and waste per period.

Fig. 8b. Model 03 tonnage of ore and waste per period.

Askari-Nasab H. & Awuah-Offei K. 101- 24

Fig. 8c. Model 04a tonnage of ore and waste per period.

Also, there is an inherent task of blending the run-of- mine materials before concentration.
The objective is to mine in such a way that the resulting mix meets the quality and quantity
specifications of the processing plant. The blending problem becomes more important as
we get more into detailed planning in short/medium range plans.
Fig. 8a to 8c illustrate the yearly tonnage of ore processed, waste mined, and the total
tonnage of material mined in each period of production. In Model 02 we set a maximum
mining capacity of 13 million tonnes per period and a processing capacity of 7.15 million
tonnes per period (Table 3a) with no lower bound on mining and processing capacities.
Fig. 8a illustrates the results of Model 02, the generated schedule is smooth with very little
fluctuations in the tonnage of feed and stripping ratio. Meanwhile, referring Fig. 9 to 11 for
the blending results show that the average grade of MWT, sulphur, and phosphor are
within the acceptable range defined in Table 3a (plots are for Models 02, 03, and 04a). A
very interesting phenomenon that should be noticed is how in Model 02 (Fig. 9), the MILP
high grades for iron ore in the early periods and then the MWT average grade starts to get
lower every year. The high grading phenomenon in early periods is completely in
accordance with the highest NPV generated by Model 02.
In Model 03 we set a maximum mining capacity of 13.2 million tonnes per period and a
processing capacity of 7.15 million tonnes per period (Table 3a) with no lower bound on
mining and processing capacities. We were forced to increase the mining upper bounds
since the 13 million tonnes upper bound did not generate a feasible integer solution; this is
because of introduction of mining-cuts into the model, which reduces the freedom of the
variables. Assessment of Fig. 8b shows that the ore feed is not as smooth as Model 02
Askari-Nasab H. & Awuah-Offei K. 101- 25

schedule. Grade blending constraints are all honored except for phosphor in the first period
that exceeded to 0.145% rather than 0.14%.
In Model 04a we set a maximum mining capacity of 13.5 million tonnes per period and a
processing capacity of 7.15 million tonnes per period (Table 3a) with a lower bound of 7
million tonnes on the processing capacity. All the constraints presented in Table 3a are
honored and the ore production schedule is the smoothest among all with an iron ore grade
of 65% to 80% average for the MWT.

Fig. 9. Average iron ore (MWT%) grade per period.

Fig. 10. Average sulphur grade per period.

Askari-Nasab H. & Awuah-Offei K. 101- 26

Fig. 11. Average phosphor grade per period.

6. Conclusions and future work

The paper investigated the shortcomings of the current mixed integer linear programming
(MILP) models used for open pit production scheduling, particularly the inability to solve
large-scale real-size mining problems. In this study, we have developed, implemented, and
tested MILP theoretical frameworks for large-scale open pit production scheduling. The
developed models proved to be able to handle deterministic large-scale mine production
problems.
To reduce the size of the open pit production scheduling problem we introduced the
concept of mining-cuts into the MILP formulations. Blocks within the same level or
mining bench are grouped into clusters based on their attributes, spatial location, rock type,
and grade distribution. Four MILP formulations are presented: Model 01- only consists of
binary integer decision variables and generates a schedule at block level resolution; Model
02 – the schedule is generated based on a strict temporal sequence of blocks. The
formulation uses continuous variables to model extraction and processing at block level.
Binary integer decision variables, are used to control precedence of extraction; Model 03 –
processing is controlled at block level with continuous decision variables and the
precedence of extraction of blocks is controlled at the mining-cut level by means of binary
integer variables; Model 04 –Extraction, processing, and order of block extraction are
controlled at mining-cut level. We have implemented the optimization formulations in
TOMLAB/CPLEX (Holmström, 1989-2009) environment.
An iron ore mine intermediate scheduling case study over twelve periods was carried out to
compare, verify and validate the models. The summary of the comparative analyses
revealed that: (i) an MILP formulation at block level resolution (Model 02) is not suitable
for long-term scheduling. Although, the block level formulation generates a higher NPV
Askari-Nasab H. & Awuah-Offei K. 101- 27

compared to the rest of the models, they would quickly go out of memory and it is almost
impossible to generate a life-of-mine schedule for a real-size mine; (ii) block level
resolution models (Model 02 and Model 03) are more appropriate for short-range
scheduling where the number of blocks are in the order of thousands to ten thousand and
the scheduling periods are in the order of ten to twelve periods. These formulations could
be used to break-down the long-term yearly schedule into a monthly schedule; (iii)
introduction of mining-cuts as the mining units (Models 03 and 04) and as a means to
control the precedence of block extraction would drastically reduce the size and runtime of
the MILP formulations and the number of binary integer variables required; (iv) MILP
formulations based on processing and extraction at mining-cut level (Model 04) leads to
practical mathematical tools addressing life-of-mine schedules of large-scale open pit
operations which was impossible to solve with the previous formulations. These models
provides control over all the mining, processing, and blending constraints, to the mine
planner while maximizing the NPV; (vi) clustering algorithms and the number of mining-
cuts affect the computational efficiency of developed MILP formulations and the generated
optimal NPV.
Further focused research is underway to develop and test different clustering techniques
that would generate an optimized clustering approach for mining-cuts. Also the next step is
to extend the mixed integer linear programming frame work into stochastic mathematical
programming domain to address the geological uncertainty issue.

7. References

[1] Askari-Nasab, H., (2006), "Intelligent 3D interactive open pit mine planning and
optimization", PhD Thesis Thesis, © University of Alberta, Edmonton, Canada,
Pages 167.

[2] Askari-Nasab, H. and Awuah-Offei, K., (2009), "Open pit optimization using
discounted economic block value", Transactions of the Institution of Mining and
Metallurgy. Section A, Mining industry, Vol. 118, 1, pp. 1-12.

[3] Askari-Nasab, H., Frimpong, S., and Awuah-Offei, K., (2005), "Intelligent optimal
production scheduling estimator", in Proceedings of 32nd Application of
Computers and Operation Research in the Mineral Industry, © Taylor & Francis
Group, London, Tucson, Arizona, USA, pp. 279-285.

[4] Askari-Nasab, H., Frimpong, S., and Szymanksi, J., (2007), "Modeling open pit
dynamics using discrete simulation", International Journal of Mining, Reclamation
and Environment, Vol. 21, 1, pp. 35- 49.

[5] Askari-Nasab, H., Frimpong, S., and Szymanksi, J., (2008), "Investigating the
continuous time open pit dynamics", The Journal of the South African Institute of
Mining and Metallurgy, Vol. 108, 2, pp. 61-73.

[6] Askari-Nasab, H. and Szymanski, J., (2007), "Open pit production scheduling using
reinforcement learning", in Proceedings of 33rd International Symposium on
Askari-Nasab H. & Awuah-Offei K. 101- 28

Computer Application in the Minerals Industry (APCOM), © GECAMIN LTDA,

Santiago, Chile, pp. 321-326.

[7] Bixby, R. E., (1987-2009), "ILOG CPLEX", ver. 11.0, Sunnyvale, CA , USA:
ILOG, Inc.

[8] Boland, N., Dumitrescu, I., Froyland, G., and Gleixner, A. M., (2009), "LP-based
disaggregation approaches to solving the open pit mining production scheduling
problem with block processing selectivity", Computers and Operations Research,
Vol. 36, 4, pp. 1064-89.

[9] Caccetta, L. and Hill, S. P., (2003), "An application of branch and cut to open pit
mine scheduling", Journal of Global Optimization, Vol. 27, November, pp. 349 -
365.

[10] Dagdelen, K. and Kawahata, K., (2007), "Oppurtunities in Multi-Mine Planning

through Large Scale Mixed Integer Linear Programming Optimization", in
Proceedings of 33rd International Symposium on Computer Application in the
Minerals Industry (APCOM), © GECAMIN LTDA, Santiago, Chile, pp. 337-342.

[11] Datamine Corporate Limited, (2008), "NPV Scheduler", ver. 4, Beckenham, United
Kingdom: Datamine Corporate Limited,.

[12] Denby, B. and Schofield, D., (1994), "Open-pit design and scheduling by use of
genetic algorithms", Transactions of the IMM Section A, Vol. 103, January - April
1994,, pp. A21-A26.

[13] Denby, B., Schofield, D., and Hunter, G., (1996), "Genetic algorithms for open pit
scheduling - extension into 3-dimensions", in Proceedings of 5th International
Symposium on Mine Planning and Equipment Selection, ©
A.A.Balkema/Rotterdam/Brookfield, Sao Paulo, Brazil, pp. 177-186.

[14] Gemcom Software International, I., (1998-2008), "Whittle strategic mine planning
software", ver. 4.2, Vancouver, B.C.: Gemcom Software International.

[15] Gershon, M., (1987), "Heuristic approaches for mine planning and production
scheduling", Geotechnical and Geological Engineering, Vol. 5, 1, pp. 1-13.

[16] Gershon, M. E., (1983), "Mine scheduling optimization with mixed integer
programming", Mining Engineering, Vol. 35, 4, pp. 351-354.

[17] Holmström, K., (1989-2009), "TOMLAB /CPLEX", ver. 11.2, Pullman, WA, USA:
Tomlab Optimization.

[18] Horst, R. and Hoang, T., (1996), "Global optimization : deterministic approaches",
© Springer, Berlin ; New York, 3rd ed, Pages xviii, 727 p.
Askari-Nasab H. & Awuah-Offei K. 101- 29

[19] ILOG Inc, (2007), "ILOG CPLEX 11.0 User’s Manual September", ver. 11.0:
ILOG S.A. and ILOG, Inc.

[20] Isaaks, E. H., (1990), "The application of Monte Carlo methods to the analysis of
spatially correlated data, Ph.D. thesis," PhD Thesis, © Stanford University,
Stanford, CA, USA, Pages 226.

[21] Johnson, T. B., (1969), "Optimum open-pit mine production scheduling", in

Proceedings of Proceedings, 8th International Symposium on Computers and
Operations Research, © Salt Lake City, Utah, USA, pp. 539-562.

[22] Johnson, T. B. and Barnes, R. J., (1988), "Application of the maximal flow
algorithm to ultimate pit design", in Engineering design : better results through
operations research methods, Vol. 8, Publications in operations research series, R.
R. Levary, Ed. New York, © North-Holland, pp. xv, 713.

[23] Kaufman, L. and Rousseeuw, P. J., (1990), "Finding groups in data : an

introduction to cluster analysis", © Wiley, New York, Pages xiv, 342 p.

[24] Krige, D. G., (1951), "A statistical approach to some basic mine valuation and
allied problems at the Witwatersrand", MSc Thesis, © University of
Witwatersrand, South Africa,

[25] Lerchs, H. and Grossmann, I. F., (1965), "Optimum design of open-pit mines", The
Canadian Mining and Metallurgical Bulletin, Transactions, Vol. LXVIII, pp. 17-
24.

[26] MathWorks Inc., (2007), "MATLAB Software", ver. 7.4 (R2007a): MathWorks,
Inc.

[27] Minemax Pty Ltd, (1998-2009), "MineMax Scheduler", ver. 4.2, West Perth,
Western Australia: Minemax Pty Ltd.

[28] Ramazan, S., (2007), "Large-scale production scheduling with the fundamental tree
algorithm - model, case study and comparisons", in Proceedings of Orebody
Modelling and Strategic Mine Planning, © The Australian Institute of Mining and
Metallurgy, Perth, Western Australia, pp. 121-127.

[29] Ramazan, S., Dagdelen, K., and Johnson, T. B., (2005), "Fundamental tree
algorithm in optimising production scheduling for open pit mine design", Mining
Technology : IMM Transactions section A, Vol. 114, 1, pp. 45-54.

[30] Ramazan, S. and Dimitrakopoulos, R., (2004), "Traditional and new MIP models
for production scheduling with in-situ grade variability", International Journal of
Surface Mining, Reclamation & Environment, Vol. 18, 2, pp. 85-98.

[31] Runge Limited, (1996-2009), "XPAC Autoscheduler", ver. 7.8: Runge Limited.
Askari-Nasab H. & Awuah-Offei K. 101- 30

[32] Stone, P., Froyland, G., Menabde, M., Law, B., Pasyar, R., and Monkhouse, P. H.
L., (2007), "Blasor - Blended Iron ore mine planning optimization at Yandi,
Western Australia", in Proceedings of Orebody Modelling and Strategic Mine
Planning, © The Australian Institute of Mining and Metallurgy, Perth, Western
Australia,

[33] Sutton, R. S. and Barto, A. G., (1998), "Reinforcement learning, an Introduction",

© The MIT Press, Cambridge, Massachusetts, Pages 432.

[34] Tolwinski, B. and Underwood, R., (1996), "A scheduling algorithm for open pit
mines", IMA Journal of Mathematics Applied in Business & Industry, 7, pp. 247-
270.

[35] Whittle, G., (2007), "Global asset optimization", in Proceedings of Orebody

Modelling and Strategic Mine Planning, © The Australian Institute of Mining and
Metallurgy, Perth, Western Australia, pp. 331-336.

[36] Winston, W. L., (1995), "Introduction to mathematical programming : applications

and algorithms", © Duxbury Press, Belmont, Ca., 2nd ed, Pages xv, 818, 39 p.

[37] Wolsey, L. A., (1998), "Integer programming", © J. Wiley, New York, Pages xviii,
264 p.

[38] Yegulalp, T. M. and Arias, J. A., (1992), "A fast algorithm to solve the ultimate pit
limit problem", in Proceedings of 23rd APCOM Symposium, © AIME, Littleton,
Colorado, pp. 391-397.

[39] Zhao, Y. and Kim, Y. C., (1992), "A new optimum pit limit design algorithm", in
Proceedings of 23rd APCOM Symposium, © SME, Littleton, Colorado, University
of Arizona, pp. 423-434.
Askari-Nasab H. & Awuah-Offei K. 101- 31

8. Appendix

MATLAB and TOMLAB/CPLEX code and documentation for Model 03.

The code is for processing at block level and mining at mining-cut level.