A recursive time aggregation-disaggregation heuristic for the

multidimensional and multiperiod precedence-constrained knapsack

problem: an application to the open-pit mine block sequencing problem

Pierre Nancel-Penarda , Nelson Moralesb , Fabien Cornillierc,d,∗

a
Delphos Mine Planning Laboratory, Department of Mining Engineering, Faculty of Physical and Mathematics Sciences,
University of Chile, Beauchef 850, Santiago, Chile; Advanced Mining Technology Center, University of Chile, Avda.
Tupper 2007, Santiago, Chile
b
Dept. of Civil, Geological and Mining Engineering, Polytechnique Montréal, Montréal, Canada
c
Dept. of Industrial Engineering, Universidad de Ingenieria y Tecnologia – UTEC, Lima, Peru
d
CIRRELT, Montreal, Canada

Abstract
A recursive time aggregation-disaggregation (RAD) heuristic is proposed to solve large-scale multi-
dimensional and multiperiod precedence-constrained knapsack problems (MMPKP) in which a profit
is maximized by filling the knapsack in multiple periods while satisfying minimum and maximum
resource consumption constraints per period as well as precedence constraints between items. An
important strategic planning application of the MMPKP in the mining industry is the well-known
open-pit mine block sequencing problem (BSP). In the BSP, a mine is modeled as a three-dimensional
grid of blocks to determine a block extraction sequence that maximizes the net present value while
satisfying constraints on the shape of the mine and resource consumption over time. Large real-life in-
stances of this problem are difficult to solve, particularly with lower bounds on resource consumption.
The advantage of the time aggregation-disaggregation heuristic over a rolling-horizon-based time de-
composition is twofold: first, the entire horizon is considered for the resource consumption from the
first aggregation; and second, only two-period subproblems have to be solved. This method is applied
to a well-known integer programming model and a variant thereof in which blocks can be extracted
in parts over multiple periods. Tests on benchmark instances show that near-optimal solutions for
both of the models can be obtained for extremely large instances with up to 2,340,142 blocks and 10
periods.
Keywords: Heuristics, Integer programming, Time decomposition, Multidimensional and
multiperiod precedence-constrained knapsack, Open pit mine scheduling

1. Introduction

The multiperiod precedence-constrained knapsack problems (MPPCKP), a generalization of the

knapsack problem, has been introduced in Moreno et al. (2010) to solve a well-known scheduling
problem in the mining industry: the open-pit mine production scheduling problem. In a more recent
article by Samavati et al. (2017b), a multidimensional MPPCKP without minimum units of resource

∗
Corresponding author
Email addresses: [email protected] (Pierre Nancel-Penard), [email protected] (Nelson
Morales), [email protected] (Fabien Cornillier )
consumption constraints per period has been studied. In the open-pit mine production scheduling
problem, profit is maximized by selecting blocks for extraction at each period while satisfying re-
source consumption constraints for each period as well as precedence constraints between items.
Multiple resources such as extracting and processing resources are considered and any constraint
corresponds to a knapsack constraint on a resource and a period. In this article, we propose a new
recursive time aggregation-disaggregation (RAD) heuristic to solve the multidimensional and multi-
period precedence-constrained knapsack problem (MMPKP), a multidimensional MPPCKP in which
minimum units of resource consumption constraints per period are considered. In this problem, the
value of a knapsack is maximized by filling it in multiple periods while satisfying minimum and
maximum units of resource consumption per period and precedence constraints between items.
The proposed RAD heuristic is used to solve an important production process scheduling problem
in an open-pit mine to maximize the net profit of the mining operation. In open-pit mines, minerals
are extracted by moving material from the surface and the extracted materials are processed by plants,
stockpiled, or placed in waste dumps depending on the profit potential. To determine the part of
the terrain to be extracted, an open-pit mine is represented as a three-dimensional array of equisized
blocks, known as the block model. Geostatistical methods are applied to the results obtained from
exploratory borehole samples to define the geological attributes and to determine the size of the blocks.
The mineral content (grade) of a block is one of these geological attributes. The value of a block
depends on its geological attributes, its destination, and other parameters and costs related to the
operations. Because the opportunity cost affects the net present value of a block, this value depends on
the moment of its extraction and processing, which can be determined after discretizing the planning
horizon into periods.
The open-pit mine block sequencing problem (BSP) seeks to determine the set of blocks to be
extracted and the extraction time for each block such as to maximize the net present value while satis-
fying operational and geomechanical constraints. The main constraints are the slope angle constraints
that ensure the stability of the open-pit walls by restricting the slope angles to a maximum limit.
These constraints also model the spatial precedence among the blocks by stating that a block cannot
be extracted before extracting its overlying blocks. The operational constraints are related either to the
amount of material to be extracted and processed at each period, or to the time required to transport
and process it. The amount of the material to be extracted and processed, and the time required to
transport and process it, are bounded by lower and upper limits that control the flow of blocks between
the mine and the destinations. When multiple resources are considered (e.g. mining and processing),
this problem corresponds to the MMPKP with time-dependent profits.
Practically, the minimum resource consumption constraints are used to bound the variations of the
quantity of material to extract and process among consecutive periods. By smoothing the intensity
of the operations, these constraints allow controlling the quantity of the mining equipment required
to transport and process the extracted materials. Furthermore, these lower bounds reduce the setup
costs incurred whenever operations are stopped and restarted. As stated by Cullenbine et al. (2011),
“contractual agreements and the chemical and physical properties of the milling process may also
necessitate positive lower bounds on production and processing rates”. They propose an experiment
to analyze the influence of minimum resource consumption constraints on the complexity of the BSP
and show that it takes about twenty times longer to find optimal solutions when considering minimum
extraction capacities.
The origin of the additional complexity is twofold: the larger number of knapsack constraints
for each period and the potential inter-resource conflicts between lower bounds on some resource
consumption (e.g. extraction and processing) and upper bounds on other resources (Samavati et al.,
2017a).

2
 –1 –1 –1 –1 –1

4 2 2

–1 –1 –1 –1 –1 –1 –1 –1 –1 –1

4 2 2 4 2 2

1 1
(a) (b)

–1 –1 –1 –1 –1
Fig. 1. A 2-dimensional instance with a minimum processing resource consumption of 2 blocks per period and a maximum
mining capacity of 8 blocks (numbers represent economic values and grey scale extraction periods) (a) A sliding time
window heuristic that only considers one period at a time will identify blocks in grey to be extracted in period 1; thus, it
will be unable to produce a feasible solution. (b) An optimal solution. 4 2 2

1
Figure 1 presents a 2-dimensional example of conflict between a maximum material-extraction ca-
pacity (hereafter called maximum mining capacity) and a minimum processing resource consumption
constraint. A planning horizon of two periods and a discount rate of 10% per period are considered
in this example. All blocks have weight equal to 1 and values, in monetary units, represented by
the numbers inside the blocks. The minimum processing capacity is 2 per period and the maximum
capacity is 8. A sliding time window heuristic that only considers one period at a time will extract the
8 blocks on the top (light grey in Figure 1a) during period 1, leaving only one block for processing in
period 2 (white in Figure 1a), i.e., will not find a feasible solution. Conversely, the optimal solution
extracts 6 blocks in period 1 (light grey in Figure 1b) and 3 in period 2 (dark grey in Figure 1b) and
has a value of 3.8. This example demonstrates that greedy heuristic approaches that only consider a
few periods at each iteration may unveil such conflicts and fail to identify a feasible solution.
As for this example and Samavati et al. (2017a), in this article we address the problem in which
any extracted block is always processed to maximize its profit or put into waste dumps whenever its
ore concentration is below a fixed cutoff grade, i.e., below the minimum ore concentration required to
process the block.
We aim to develop a recursive time aggregation-disaggregation (RAD) heuristic to solve large
instances of the BSP with minimum extraction and processing constraints. The proposed heuristic is
applied to a well-known integer programming model and a variant proposed by Vossen et al. (2016).
In the variant version, a mixed-integer programming model (MIP) is used to solve problems in which
blocks can be extracted in parts over multiple periods while satisfying the slope angle constraints.
They report improvements of between 1% and 1.9% of the net present value compared with the integer
programming model. As stated by Vossen et al. (2016) “some blocks at a mine’s surface may be
partial at the beginning of period 1 because the undisturbed surface is uneven or because some partial
extraction has already taken place.” Furthermore, extracting entire blocks limits the way in which the
discretized profile of the pit approximates the continuous one. Allowing partially mined blocks may
offer slope angles closer to the maximum values and a better approximation of the varying slope
angles for different depths or sections of the mine. Because the edges of the shapes generated by the
block model are smoothed during the design process with the help of CAD software, the solutions of
either the integer programming model or the variant proposed by Vossen et al. (2016) could be used
as inputs for the design process. Moreover, Morales et al. (2015) compared solutions obtained from
software that consider continuous slopes with integer solutions and show a marked difference in pit
geometries, particularly for the first periods. Parra et al. (2018) also demonstrate that slope angles
may have a considerable impact on the economic value of the first periods. This follows from the fact

3
Fig. 2. Section view of the Newman1 instance of the Minelib library (Espinoza et al., 2013) showing the extraction periods
for the fractional block sequencing problem (FBSP) model in which blocks can be partially extracted at each period.

that steeper angles require less waste to be extracted to comply with the slope.
Figure 2 shows a section view of the extracted periods of a solution for the Newman1 schedul-
ing instance of the Minelib library (Espinoza et al., 2013) resulting from the MIP. In this example,
whenever a block is partially extracted at period t, its remaining fraction is extracted at period t + 1.
In this article, we propose new instances based on the library Minelib (Espinoza et al. 2013)
introducing minimum resource consumption constraints on material extraction and processing.
The remainder of this article is organized as follows. A literature review is provided in Section 2.
Two versions of the problem are modeled in Section 3. The RAD heuristic is described in Section 4,
and the computational results are presented in Section 5. The conclusions are drawn in Section 6.

2. Related works

In this section, we present relevant articles on strategic open-pit planning problems and the studies
that more specifically consider the difficult minimum resource consumption constraints.
Lerchs & Grossman (1965) were the first to propose an algorithm to solve the final pit problem also
known as the ultimate pit problem. This problem is a simplified version of the BSP in which a value
is allotted to each block without considering the extraction period, and where only slope constraints
are modeled. Picard (1976) demonstrated that the final pit problem is equivalent to the maximum
closure problem, which can be formulated as a minimum cut problem. Hochbaum proposed an effi-
cient polynomial algorithm based on a maximum pseudoflow approach to tackle the final pit problem
(Hochbaum & Chen, 2000; Hochbaum, 2008; Hochbaum & Orlin, 2012). Lerchs & Grossman (1965)
also proposed an approach to schedule block extraction by solving a series of final pit problems while
varying the block revenues to generate a sequence of nested pits. Commercial software programs, in-
cluding Whittle (Gemcom, 2021), still implement this algorithm. A review of open-pit mine designs
with nested pits can be found in Meagher et al. (2014), which focus on the inconsistent sizes that may
be observed between successive nested pits, namely, the gap problem.
Johnson (1968, 1969) presented a linear programming formulation for a multidestination block
scheduling problem under slope, capacity, and blending constraints. Each extracted block containing
ore had to be assigned to a processing plant to maximize its value. Since the real instances of this
problem are difficult to solve, many algorithms were developed such as the Lagrangian relaxation
method proposed by Dagdelen (1986) to solve a problem with fixed thresholds of mineral concentra-
tions (fixed cutoff grades), and where only upper bounds on the capacities are considered.

4
 Caccetta & Hill (2003) proposed a specific branch-and-cut algorithm, which only considered the
upper bounds on resource consumption. The algorithm is based on a “by” formulation, in which the
extraction time of a block is modeled as a binary variable equal to one if the block is extracted by
a given period and equal to zero otherwise. As we can switch from an “at” formulation to a “by”
formulation via variable substitution, both of the models are equivalent. However, despite this equiva-
lence, the numerical experiments performed by Lambert et al. (2014) demonstrate that improved upper
bounds are obtained with the “by” formulation after applying certain variable-reduction operations at
the root node of the branch-and-cut algorithm.
As it is often challenging to solve large real instances, Boland et al. (2009) proposed to aggregate
blocks into bins, where precedence constraints only applied between bins. By reducing the problem
size, this technique helps expedite the resolution time.
Bienstock & Zuckerberg (2009, 2010) achieved breakthrough results using a Lagrangian relaxation-
based method to solve the linear relaxation of the BSP. The method substantially decreases the reso-
lution time compared with the standard linear programming solvers.
Moreno et al. (2010) presented the Critical Multiplier Algorithm to solve the LP-relaxation of the
multiperiod precedence-constrained knapsack problem as a sequence of single-time period problems
applied to the open-pit production scheduling problem.
Bley et al. (2010) used a “by” formulation with upper bounds on resource consumption. Addi-
tional clique and cover cuts based on the capacity constraints strengthen the formulation.
Chicoisne et al. (2012) developed a customized algorithm for the linear programming relaxation
of the problem with only one destination and one capacity constraint per period. They extended
the heuristic method proposed by Gershon (1987) based on expected extraction times to perform
a topological sorting (TopoSort) on the blocks. A feasible integer solution is obtained using this
procedure for a more general problem with multiple capacity constraints per period.
Espinoza et al. (2013) proposed the Minelib library of standardized instances including real and
simulated datasets for three different open-pit mining problems, the final pit problem and the follow-
ing two variants of the open-pit block-scheduling problem: the fixed cutoff grade problem, referred
to as constrained pit limit problem (CPIT), where any extracted block is processed only if its ore con-
centration is equal or greater than the specified minimum grade, and the multidestination optimization
problem with blending constraints, referred to as precedence constrained production scheduling prob-
lem (PCPSP). The PCPSP formulation is a mixed-integer programming model in which different parts
of the same extracted block can be sent to different destinations within the same period. Note that, for
the CPIT and PCPSP scheduling instances of the Minelib library that do not include positive lower
bounds on resource consumption, the null solution is a trivial feasible solution, because it complies
with the upper capacities and does not violate any lower bound.
Jélvez et al. (2016) proposed a spatial aggregation and disaggregation heuristic that improved
some of the published results of the Minelib library for the fixed cutoff grade problem. The best av-
erage relative gaps were obtained with this method for the largest five instances. Others, including
Lamghari et al. (2014), Liu & Kozan (2016), and Samavati et al. (2017b,c), proposed heuristics that
improve the results for some of the fixed cutoff grade Minelib scheduling instances without mini-
mum resource consumption constraints. In particular, Liu & Kozan (2016) obtained good solutions
developing two topological ordering-based algorithms relying on network flow graph and conjunctive
graph theory.
Jélvez et al. (2020) presented a hybrid heuristic algorithm that combined a rolling-horizon de-
composition algorithm with a block-preselection procedure using the concept of expected extraction
times proposed by Gershon (1987). A satisfactory feasible solution was reported for the instance W23
of the Minelib library with blending constraints, an instance for which no feasible solution has been

5
previously reported.
More detailed reviews on general mine planning models and algorithms can be found in Osanloo
et al. (2008), Newman et al. (2010) and Zeng et al. (2021).
As stated in the introduction, lower bounds on resource consumption have practical importance
in the mining operations but make the problem drastically harder to solve with linear programming
models. The articles reviewed in the next paragraphs more specifically address open-pit strategic
planning problems with positive lower bounds on resource consumption.
Ramazan et al. (2005) is one of the first authors to specifically solve open-pit planning problems
with minimum processing consumption constraints and linear programming for a fixed cutoff grade
model. The proposed method, named the fundamental tree algorithm, is based on block aggregation
to lower the number of variables and constraints of the original problem.
Gaupp (2008) presents a branch-and-cut algorithm with reduction techniques to decrease the num-
ber of variables and heuristics based on Lagrangian relaxation to generate feasible solutions for a fixed
cutoff grade model. The earliest and latest possible extraction times available for each block are com-
puted to generate cutting planes. These methods are used to solve instances of up to 10,819 blocks
and 6 periods.
Askari-Nasab et al. (2011) compared several mixed-integer programming formulations of a prob-
lem with a fixed cutoff grade, lower and upper limits on capacity, and blending constraints. They
proposed to aggregate blocks into clusters based on the block attributes to solve instances with up to
2,598 blocks and 12 periods. In this study, lower bounds are only considered on the grade blending
constraints.
To avoid having a single block at the bottom of the pit, Cullenbine et al. (2011) added horizontal
precedence constraints to a “by” formulation of the problem with a fixed cutoff grade. They im-
plemented a rolling time window heuristic that considers the complete horizon while relaxing the
integrality constraints for periods beyond the rolling time window. Using this method, instances with
lower bounds on resource consumption with up to 25,620 blocks and 15 periods could be solved.
Lambert & Newman (2014) solved the problem formulated as in Cullenbine et al. (2011), without
the horizontal precedence constraint. They proposed to identify a sufficient set of blocks to satisfy the
minimum resource consumption constraints. This set of blocks is obtained by solving a series of final
pit problems while varying the ore price. A Lagrangian relaxation method was proposed, which uses
the information obtained while generating the initial solution to select a dualization scheme for some
resource constraints. The performance of this algorithm evaluated for instances with up to 25,000
blocks, 10 periods, minimum and maximum use of the mining and processing resources, shows that
incorporating lower bounds on resource consumption may increase the computational time by over
one order of magnitude.
Vossen et al. (2016) proposed a “hierarchical” Benders decomposition methodology to solve a
BSP variant wherein blocks could be extracted in parts over multiple periods. This new methodology
generalizes a nested benders decomposition using cumulative variables and time-aggregated resource
constraints that consider lower bounds on resource consumption. They solved instances with up to
25,620 blocks and 20 periods.
Samavati et al. (2017a) implemented a local branching procedure that started from a feasible
solution obtained with a greedy method to solve a BSP “at” formulation. The blocks are iteratively
sorted in a random order to feed a topological sort heuristic similar to the Toposort algorithm proposed
by Chicoisne et al. (2012), until obtaining a feasible solution that satisfies the minimum resource
consumption constraints. Instances with up to 50,383 blocks, 10 periods, minimum use of the mining
and processing resource were solved; unfortunately, the data sets have not been publicly released.
A recent study by Letelier et al. (2020) exposes direct block scheduling models for which pre-

6
processing, cutting plane techniques, heuristic approaches, and a customized branch-and-bound are
proposed to obtain better bounds and feasible solutions. Instances with minimum use of the pro-
cessing resource of up to 2,340,142 blocks and 20 periods are solved; unfortunately, the minimum
processing constraints have not been publicly released.

3. Mathematical formulations

In this section, the following two problems are formulated: the block sequencing problem (BSP)
in which a block should be either totally extracted or not extracted at all, and the fractional block
sequencing problem (FBSP), based on the model proposed by Vossen et al. (2016), in which a block
can be extracted in parts over multiple periods. The formulations presented here are “by” formulations
where variables represent blocks that are extracted by a specific period. A comparison between the
“at” and “by” formulations of the BSP can be found in Lambert et al. (2014).

3.1. Formulation of the block sequencing problem

Let B denote the set of blocks of the block model, and R denote the set of renewable operational
resources required to extract and process the blocks (i.e., the resources renewed at each period in-
cluding workforce and equipment, as opposed to consumable resources). The profit pbt is obtained
by extracting and processing block b ∈ B at period t ∈ T , where T = {1, . . . , T } denotes the set
of periods. An amount qbr of operational resource r ∈ R is required to extract and process block b,
while a minimum of Lrt or a maximum of Urt of resource r ∈ R must be used at each period t ∈ T .
The open-pit mine block sequencing problem (BSP) is the problem of maximizing the total profit by
determining the subset of blocks to be scheduled and extracted while satisfying the minimum use Lrt
and the maximum availability Urt of resource r at period t. The precedence constraints are defined by
the set P of the slope precedence arcs. This set contains the arcs (b, b′ ) such that block b′ is at a level
above b and must be mined before b to ensure the stability of walls. We set the binary variable xbt
equal to 1 if block b is extracted by period t, and 0 otherwise. Note that xbt − xb,t−1 specifies whether
block b is mined at time t. The BSP can then be formulated as follows:

X
(BSP) max pbt (xbt − xb,t−1 ) (1)
t∈T ,b∈B
xbt ≤ xb′ t ∀ (b, b′ ) ∈ P, t ∈ T (2)
xbt ≥ xb,t−1 ∀b ∈ B, t ∈ T (3)
X
qbr (xbt − xb,t−1 ) ≤ Urt ∀r ∈ R, t ∈ T (4)
b∈B
X
qbr (xbt − xb,t−1 ) ≥ Lrt ∀r ∈ R, t ∈ T (5)
b∈B
xbt ∈ {0, 1} ∀b ∈ B, t ∈ T (6)
xb0 = 0 ∀b ∈ B. (7)

The objective function (1) maximizes the total profit. Constraints (2) correspond to the precedence
constraints given by the slope specifications to ensure the wall’s stability. Under constraints (3) a
block b extracted by period t − 1 remains extracted at period t ≤ T as a block can be extracted only
once. Constraints (4) and (5) state that the maximum and minimum resource consumption constraints
should be satisfied at each period. Constraints (6) and (7) reflect the nature of the variables.

7
 Fig. 3. Example of section view with vertical and oblique precedence arcs

3.2. Formulation of the fractional block sequencing problem

Vossen et al. (2016) proposed a BSP variant in which a block could be extracted in parts over
multiple periods, and proved that any feasible solution of this model satisfies the slope precedence
constraints.
Let P′ denote the set of vertical precedence arcs, and P′′ the set of oblique precedence arcs. In
Figure 3, block b2 is the vertical spatial predecessor of block b0 , i.e., (b0 , b2 ) ∈ P′ , and blocks b1
and b3 are two oblique spatial predecessors of block b0 , i.e., (b0 , b1 ) ∈ P′′ and (b0 , b3 ) ∈ P′′ . Let us
define a continuous decision variable ybt ∈ [0, 1] equal to the fraction of block b extracted by period t,
and a binary variable xbt equal to 1 if block b is entirely extracted by period t, and 0 otherwise. The
FBSP can be formulated as follows:

X
(FBSP) max pbt (ybt − yb,t−1 ) (8)
t∈T ,b∈B
ybt ≤ yb′′ t ∀ (b, b′′ ) ∈ P′′ , t ∈ T (9)
ybt ≤ xb′ t ∀ (b, b ) ∈ P , t ∈ T
′ ′
(10)
ybt ≥ xbt ∀b ∈ B, t ∈ T (11)
ybt ≥ yb,t−1 ∀b ∈ B, t ∈ T (12)
yb0 = 0 ∀b ∈ B (13)
X
qbr (ybt − yb,t−1 ) ≤ Urt ∀r ∈ R, t ∈ T (14)
b∈B
X
qbr (ybt − yb,t−1 ) ≥ Lrt ∀r ∈ R, t ∈ T (15)
b∈B
ybt ∈ [0, 1] ∀b ∈ B, t ∈ T (16)
xbt ∈ {0, 1} ∀b ∈ B, t ∈ T . (17)

Objective function (8) maximizes the total profit of the blocks, whether totally or partially ex-
tracted. Under constraints (9), the fraction of a block b extracted by time t cannot exceed the fraction
of any of its oblique predecessors extracted by time t. Constraints (10) ensure that block b cannot be
extracted before the complete extraction of block b′ located directly above block b, i.e., block b′ is
the vertical predecessor of block b. Constraints (11) reflect the relation between variables ybt and xbt
for any block b at period t. Constraints (12) ensure that if a fraction of a block b is extracted by
period t − 1, at least the same fraction remains extracted by period t ≤ T . Constraints (13) state that no
block can be extracted before the beginning of the first period. Constraints (14) and (15) state that the
maximum and minimum resource consumption should be satisfied at each period. Constraints (16)

8
and (17) reflect the nature of the variables.
As the LP-relaxations of the FBSP and the BSP only differ in terms of the vertical and oblique
predecessors, they are equivalent; therefore, an optimal solution of one of these relaxed models can
be used as an upper bound for the non-relaxed model.

4. Algorithm for the MMPKP

In the above formulations, the number of variables is generally a serious challenge in practice. In
this section, we propose two size-reduction procedures:

1. The first approach to decrease the number of variables is to decompose the problem into a series
of easier-to-solve subproblems. In this section, we propose a recursive time decomposition
algorithm in which the periods are aggregated, then disaggregated. This algorithm considers
the entire horizon from the first subproblem and thereby avoids using more resources than
necessary in the first periods to help to prevent the risk of having an insufficient activity in the
subsequent periods. Another advantage of this aggregation-disaggregation heuristic is that it
generates only two-period subproblems significantly more tractable for the largest instances.

2. The second classical reduction procedure is to only consider the blocks of the final pit, i.e.,
the blocks to extract to maximize the economic value while satisfying only the precedence
constraints. Since strictly positive minimum resource consumption constraints are considered
in this study, the blocks of the final pit may be insufficient to obtain a feasible solution. In such
a case, more blocks have to be added to the final pit. In this section, we describe an integer
programming model to extend the final pit in an attempt to obtain a feasible solution whenever
the overall minimum resource consumption constraints are not satisfied.

4.1. Recursive time aggregation-disaggregation (RAD) heuristic

Let us refer to a metaperiod as any subset of consecutive periods, and an elementary metaperiod
as any metaperiod of only one period. We propose to solve the BSP and FBSP models by solving
a series of two-metaperiod subproblems. In the RAD heuristic, the initial problem is solved as a
two-metaperiod problem after aggregating all its periods into two metaperiods as well as their corre-
sponding resource constraints. The periods of each metaperiod are in turn aggregated into two other
metaperiods to form a new two-metaperiod subproblem. This procedure is recursively repeated until
solving subproblems with only two elementary metaperiods, i.e., with only two periods. The complete
decomposition results in a binary tree structure in which each node corresponds to a two-metaperiod
subproblem. All two-metaperiod subproblems of the binary tree are solved with a branch-and-cut pro-
cedure to determine the set of blocks (or parts of blocks in the case of the FBSP) to extract within each
of their metaperiods. Note that by solving a subproblem, the sole metaperiod during which a block
should be extracted is known. Its exact extraction period is revealed only after solving its leaf sub-
problems. We also point out that the resource constraints aggregated in each metaperiod are recovered
as we go down the tree. Indeed, the original constraints have to be satisfied whenever a metaperiod
corresponds to a real period.
Figure 4 shows an example of a binary tree generated by the decomposition for an instance of 10
periods. The root node corresponds to the first decomposition of the original problem transformed
into a two-metaperiod problem by aggregating all its periods into two metaperiods {1, 2, . . . , 5} and
{6, 7, . . . , 10}. After solving the two-metaperiod problem P, we obtain two sets of blocks to be ex-
tracted: a set B1 of blocks to be extracted during periods {1, 2, . . . , 5}, and a set B2 of blocks to be

9
extracted during periods {6, 7, . . . , 10}. To determine how the extraction periods of the blocks in B1
are distributed among periods {1, 2, . . . , 5}, a new two-metaperiod subproblem P1 is solved after ag-
gregating the periods of the first metaperiod of the problem P into two metaperiods {1, 2} and {3, 4, 5}.
Consequently, we obtain the sets B1,1 and B1,2 of blocks to be extracted during periods {1, 2} and
{3, 4, 5}, respectively. A new two-metaperiod subproblem P2 is solved in the same way after aggregat-
ing the periods of the second metaperiod of the problem P into two metaperiods {6, 7} and {8, 9, 10} to
obtain the sets B2,1 and B2,2 of blocks to be extracted during periods {6, 7} and {8, 9, 10}, respectively.
When a metaperiod only contains elementary metaperiods, i.e., only periods, it does not require to
be decomposed anymore. Note that to balance the computation time of each subtree a balanced bi-
nary tree can be obtained by setting the number of periods in each metaperiod of a two-metaperiod
subproblem such that it differs by at most one.
For instance, in Figure 4, the root subproblem P is solved to determine the blocks to extract within
each of the two metaperiods {1, 2, . . . , 5} and {6, 7, . . . , 10}. Each of these metaperiods is in turn par-
titioned into two new metaperiods: {1, 2, . . . , 5} into metaperiods {1, 2} and {3, 4, 5}, and {6, 7, . . . , 10}
into metaperiods {6, 7} and {8, 9, 10}. Note that because of the “by” formulation, if a block b is ex-
tracted within the metaperiod {1, 2, . . . , 5} in the solution of the subproblem P, then the same block
should also be extracted by the second metaperiod of the subproblem P1, i.e. the metaperiod {3, 4, 5}.
Formally, xb,1 P = 1 implies xP1 = 1, where xP denotes a binary variable equal to 1 if block b is
b,2 b,m
extracted by metaperiod m ∈ {1, 2} in a subproblem P, and 0 otherwise. Moreover, when the FBSP
formulation is used, if a fraction λ of block b is extracted by the metaperiod {1, 2, . . . , 5} in the so-
lution of the subproblem P, the same fraction λ should be extracted by the second metaperiod of the
subproblem P1, i.e., metaperiod {3, 4, 5}. Formally, if we denote by yb,m P ∈ [0, 1] the fraction of
block b extracted by metaperiod m ∈ {1, 2} of a subproblem P, then yb,1 = λ implies yb,2
P P1 = λ. These

implications are essential to enforce the solution of the subproblems in their children.
The solution of the BSP corresponds to the blocks to be extracted within each of the elementary
metaperiods in the binary tree (in bold in Figure 4).

Illustration of the RAD heuristic. Consider an instance of 4 periods. Figure 5 shows the correspond-
ing binary tree decomposition into subproblems with their respective metaperiods and Figure 6 shows
the iterations of the proposed heuristic. The original problem is decomposed into the two metape-
riods {1, 2} and {3, 4}. The resulting two-period problem P is then solved to obtain two nested pits
that correspond to these two metaperiods, respectively (Figure 6a). A two-metaperiod problem P1 is
solved to obtain the blocks to be extracted at period 1 or 2 (Figure 6b). Subsequently, a two-period
problem P2 is solved to obtain the blocks to be extracted at period 3 or 4 (Figure 6c). In each figure,
the dotted line represents the solution of the solved two-metaperiod problem solved. The solutions
obtained for the subproblems P1 and P2 yield the global solution.

4.2. Consolidating the upper and lower resource consumption constraints

As periods are aggregated into metaperiods, the resource capacities associated with each period
of a metaperiod need to be consolidated. Let Ur,m and Lr,m denote the consolidated upper and lower
capacities of resource r for a metaperiod m. These consolidated capacities are computed as the sum
of the maximum resource r available at metaperiod m and the sum of the minimum use of resource
r at metaperiod m, respectively, over all periods of the metaperiod m, i.e., Ur,m = t∈m Ur,t and
P
Lr,m = t∈m Lr,t .
P
Using the BSP formulation in which all the variables are binary and the blocks cannot be extracted
in parts over multiple periods, more feasibility issues might arise related to the upper and lower re-

10
 {1, 2, …, 5} {6, 7, …, 10}
P

{1, 2} {3, 4, 5} {6, 7} {8, 9, 10}

P1 P2

{1} {2} {3} {4, 5} {6} {7} {8} {9, 10}

P1.1 P1.2 P2.1 P2.2

{4} {5} {9} {10}

P1.2.1 P2.2.1

Fig. 4. Resulting binary tree of subproblems for the same instance of 10 periods. The subproblems are represented with
grey rectangles, and the metaperiods are depicted using round-edged rectangles

{1, 2} {3,4}
P

{1} {2} {3} {4}

P1 P2

Fig. 5. Resulting binary tree of subproblems for an instance of 4 periods.

source consumption constraints compared with the FBSP formulation which offers more flexibility by
allowing blocks to be partially extracted.

Illustration of potential issues related to the upper capacity constraints. Consider an instance of 4
periods with the block model shown in Figure 7. The problem is decomposed as shown in Figure 5.
The final pit (in grey in Figure 7) contains 18 blocks to be extracted over 4 periods, i.e., T = {1, 2, 3, 4}.
For the sake of simplicity, only one resource is considered with constant minimum and maximum
mining consumption of L1,t = 2.0 and U1,t = 4.5 blocks per period t ∈ T , respectively. In Figure 7, the
blocks are represented with their respective undiscounted values, and the present values are computed
with a 10% discount rate. Moreover, any extracted block with a positive value is processed.
The horizon of subproblem P is decomposed into two metaperiods {1, 2} and {3, 4}. For both of
the metaperiods, the consolidated lower and upper mining capacities are 4 and 9 blocks, respectively.
The optimal scheduling of subproblem P is shown in Figure 8 wherein the blocks are represented with
their respective extraction metaperiods: a total of 9 blocks are extracted in the first metaperiod {1, 2}
and 9 blocks in the second metaperiod {3, 4}.
The metaperiods of subproblem P1 are the two elementary metaperiods {1} and {2}. Using the BSP
model, all the 9 blocks extracted within the first metaperiod of subproblem P have to be extracted by
the second metaperiod of P1, and we have therefore xb,2 P1 = 1 for these blocks. However, since no

block can be extracted in part in the BSP model, only 4 blocks can be extracted at each elementary
metaperiod, over the 4.5 blocks allowed by the maximum mining capacity per period (i.e., U1,t ), i.e.,
up to 8 blocks can be extracted in each one of the metaperiods {1, 2} and {3, 4} rather than the 9
blocks per metaperiod scheduled while solving subproblem P. The optimal extraction scheduling for
subproblem P makes the subproblem P1 unfeasible.

11
 {1}
P1
{1, 2}
{2}

P2 P2
{3, 4} {3, 4}

(a) (b)

{1}

{2}
{3}

{4}

(c)

Fig. 6. Illustration of the RAD heuristic for an instance of 4 periods. Figure (a) shows the result of the first subproblem P
solved as a two-metaperiod problem into P1 and P2 after aggregating all the periods into two metaperiods {1, 2} and {3, 4}.
Figures (b) and (c) show the respective results of subproblem P1 and P2 solved as other two-metaperiod problems. As P1
and P2 only contain elementary metaperiods, no more decomposition is needed.

–1 –1 –1 –1 –1 –1 –1 –1

–1 –1 –1 –1 –1 –1 –1 –1

–1 –1 9 3 3 3 –1 –1

Fig. 7. A 2-dimensional illustration. Numbers represent the economic value of the blocks. Blocks in grey correspond to the
final pit.

1 1 1 1 1 2 2 2

1 1 1 2 2 2

1 2 2 2

Fig. 8. An optimal solution for the BSP and FBSP models without delta adjustment. Numbers and grey scale represent
extraction metaperiods of blocks. (see Figure 7 for the economic value of the blocks)

12
 1 1 1 1 1 2 2

1 1 1 2 2

2 2 2

Fig. 9. An optimal solution for the BSP model with delta adjustment. Numbers and grey scale represent extraction metape-
riods of blocks. (see Figure 7 for the economic value of the blocks)

1 1 1 1 2 3 4

2 2 2 3 4

3 3 4

Fig. 10. A solution for the BSP model with delta adjustment. Numbers and grey scale represent extraction periods of blocks.
The value of this solution is equal to 1.1. (see Figure 7 for the economic value of the blocks)

The probability to face this issue arising with the BSP model increases with the height of the meta-
period in the binary tree, i.e. the length of the longest downward path from the metaperiod to a leaf,
which also depends on the size of the binary tree. Indeed, if we denote by |m| the number of elemen-
tary periods in a metaperiod m, the height of the metaperiod m is computed as hm = ⌈log2 (|m|)⌉ − 1.
Therefore, to lower the probability of such an issue, we propose to solve the subproblems with more
′ = U
r,m − δr,hm and Lr,m = Lr,m + δr,hm in both of the metaperi-
restrictive consolidated capacities Ur,m U ′ L

ods m = 1 and m = 2, while the adjustment values δU r,hm > 0 and δr,hm > 0 depend on the height hm .
L

Let us solve subproblem P of an example wherein δU 1,hm1 = δ1,hm2 = 1 for the metaperiods m1 =
U

{1, 2} and m2 = {3, 4}. The adjusted maximum capacities are U1,m ′
1
= U1,m
′
2
= 8. In the optimal
scheduling, we observe that only 8 blocks are extracted in each metaperiod (Figure 9), rather than
9 when the maximum capacity is not adjusted. Subproblems P1 and P2 are feasible, and a feasible
global solution is obtained with an objective value of 1.1 (Figure 10).
As a block can be partially extracted with the FBSP model, there is no need for more restrictive
capacity constraints. In the 2-dimensional example, 4.5 blocks are extracted within each period of
subproblems P1 and P2 (see Figure 11), and the optimal value is equal to 2.7.

1 3
1 1 1 1 3 4
2 4
3
2 2 2 3 4
4
3
2 3 4
4

Fig. 11. A solution for the FBSP model. Numbers and grey scale represent extraction periods of blocks or parts of blocks.
The value of this solution is equal to 2.7. (see Figure 7 for the economic value of the blocks)

13
4.3. Final pit extension
The final pit can be extended whenever the sum of resource capacities to extract is below the
sum over the time horizon of the minimum resource consumption for at least one resource. Lam-
bert & Newman (2014) proposed a two-stage extension procedure. In the first stage, the ore price is
increased to generate a larger pit containing a sufficient amount of ore to only satisfy the total pro-
cessing requirements for the time horizon. However, increasing the ore price is not always sufficient
to generate a feasible larger pit. When insufficient extracted blocks are obtained by the first expansion
stage to satisfy the overall minimum resource consumption, they propose a second expansion stage
in which an integer program (PE IP ) maximizes the overall value of new blocks to be added to the
previous selection. In this integer program (PE IP ), the blocks selected in the first stage are forced to
be extracted; their values are not considered anymore, and residual capacity constraints are applied to
the sole new blocks.
Note that the minimum and maximum constraints on extraction consumption are not considered
in the first stage of this algorithm, which may result in unfeasibility in the second stage when too
many blocks are selected. Indeed, in such a case, some overall maximum extraction constraints could
be violated. The inconsistent size of pit sometimes obtained by varying the ore price is referred to as
the gap problem in the literature (Meagher et al., 2014).
We propose a single-stage expansion strategy to prevent such pitfalls. It consists in solving an
integer program (18)-(23) that forces the extraction of all the blocks belonging to the final pit while
satisfying all the resource consumption constraints consolidated over the entire horizon. This strategy
may prioritize the extraction of blocks close to the surface that have the advantage of requiring the
extraction of fewer blocks to satisfy the slope precedence constraints and identify a feasible solution.
Compared with the algorithm proposed by Lambert & Newman (2014), this method generates a more
precisely adjusted set of blocks which helps to prevent the gap problem.
Let FP denote the set of blocks in the final pit, Lr and Ur denote the consolidated minimum
and maximum resource consumption of resource r over the time horizon, that is Lr = t∈T Lrt and
P
Ur = t∈T Urt . Let δrL and δU
P
r denote the capacity constraint adjustment parameters which help to
prevent unfeasibilities arising from the deconsolidated maximum and minimum resource consumption
constraints of the original multiperiod problem, the integer programming model to be solved is as
follows:
X
max pb xb (18)
b∈B
xb ≤ xb′ ∀ (b, b′ ) ∈ P (19)
xb = 1 ∀b ∈ FP (20)
X
qbr xb ≤ Ur − δU
r ∀r ∈ R (21)
b∈B
X
qbr xb ≥ Lr + δrL ∀r ∈ R (22)
b∈B
xb ∈ {0, 1} ∀b ∈ B. (23)

The objective function (18) maximizes the total profit. Constraints (19) correspond to the prece-
dence constraints given by the slope specifications. Constraints (20) state that all blocks of the final
pit FP are extracted. Constraints (21) and (22) state that the adjusted consolidated maximum and min-
imum resource consumption constraints over the time horizon should be satisfied. Constraints (23)
reflect the nature of the variables.

14
 Note that the solution of model (18)-(23) may be a set of blocks for which there is no feasible
solution satisfying the minimum resource consumption constraints for each of the periods. It only
guarantees that the blocks of the extended pit are sufficient to satisfy the consolidated minimum use
of mining and processing capacities without exceeding the consolidated ones over the time horizon.
Indeed, we cannot guarantee that in all cases sufficient ore is available at each period to be extracted
to meet processing constraints. In the worst case, an instance could be infeasible and its lower bounds
would need to be modified to be less restrictive.

5. Computational results

The RAD heuristic was coded in C++ and run on a 64-bit Windows OS workstation with ten 2.6 GHz
Intel Xeon E5 2660v3 processors and 120 Gb RAM. It uses the MineLink library (developed at
Delphos Mine Planning Laboratory at Universidad de Chile), which implements data structures to
store block models and offers an implementation of the pseudoflow algorithm proposed by Hochbaum
& Chen (2000) for the final pit problems. The Gurobi 8.0 library was used to solve the MIP subprob-
lems.
A comparison of the RAD heuristic applied to both BSP and FBSP models with a sliding time
window heuristic (STWH) and the sliding time window heuristic proposed by Lambert & Newman
(2014) which relaxes the integrality constraints on the variables that correspond to the periods beyond
the incumbent time window, thereafter called sliding time window heuristic with relaxation (STWHR),
was performed on 13 modified problems from the 11 instances of the Minelib library (Espinoza et al.,
2013) available at http://mansci-web.uai.cl/minelib. Some of these instances are real (KD, P4HD,
W23, McLaughlin and McLaughlin limit), others are fictitious (Newman, ZuckSmall, ZuckMedium,
ZuckLarge, Marvin, and SM2). Regarding the real instances, KD refers to a copper mine in Arizona,
USA; P4HD to a gold and copper mine in Nevada, USA; W23 to a gold mine in North America; and
McLaughlin to a gold mine in California, USA. McLaughlin Limit refers to the McLaughlin final pit
defined by Somrit (2011).

5.1. Test instances with minimum resource consumption constraints

As this study aims to solve difficult instances with positive minimum resource consumption con-
straints, we propose a set of new test instances based on the constrained pit limit problems (CPIT)
instances of the Minelib library. These instances have been modified to include positive lower bounds
on mining and processing tonnage capacities. The instances Newman1, Zuck Medium, P4HD, Zuck
Large, McLaughlin Limit, and McLaughlin have been modified by reducing the horizon (in bold in
Table 1) and by adding or modifying the lower and upper bounds (Table 2).
In the CPIT instances of the Minelib library, each block is associated with two parameters: the
mined and processed tonnages. The mined tonnage of a block corresponds to its actual tonnage, while
the processed tonnage corresponds to the tonnage to be processed if the block is extracted. Since the
decision to process or not process an extracted block is predetermined, the processed tonnage of a
block is equal to the mined tonnage if the block has to be processed once extracted; it is equal to zero
otherwise. In the original Zuck Medium instance, the mined tonnage of several blocks is lower than
the processed tonnage. This issue has been corrected in the modified Zuck Medium instance by setting
the mined tonnage equal to the greater of the mined and processed tonnage of the original instance.
All instances except P4HD and W23 have been modified to generate conflicts between lower
bounds on the processing capacities and upper bounds on the mining capacities.
Table 1 summarizes, for each instance, the size of the block model and the number of variables
and constraints. The names of modified instances are the original Minelib library instance names

15
prefixed with “m” or “m2”. Notice that modified instances of McLaughlin (resp. Marvin) have the
same number of periods, blocks, and variables, but differ in the bounds (see Table 2).
Table 1
Description of the modified instances of the Minelib library (reduced horizons are in bold).

BSP FBSP
Instance Name T Blocks Vars Cons Vars Cons
mNewman1 4 1,060 4,240 19,944 8,480 24,184
mZuckSmall 20 9,400 188,000 3,100,880 376,000 3,288,880
mKD 12 14,153 169,836 2,807,220 339,872 2,977,056
mZuckMedium 14 29,277 409,878 18,206,832 819,756 18,616,710
mP4HD 8 40,947 327,576 6,236,480 655,152 6,564,056
mMarvin 20 53,271 1,065,420 14,078,120 2,130,840 15,143,540
m2Marvin 20 53,271 1,065,420 14,078,120 2,130,840 15,143,540
mW23 12 74,260 891,120 10,068,600 1,782,240 10,959,720
mZuckLarge 25 96,821 2,420,525 28,748,250 4,841,050 31,168,775
mSM2 30 99,014 2,970,420 5,869,800 5,940,840 8,840,220
mMcLaughlinLim 10 112,687 1,126,870 31,481,740 2,253,740 32,608,610
mMcLaughlin 10 2,140,342 21,403,420 742,841,160 42,806,840 774,244,580
m2McLaughlin 10 2,140,342 21,403,420 742,841,160 42,806,840 774,244,580

5.2. Parameter settings

All subproblems of the proposed heuristic were solved in a depth-first search manner. In this sec-
tion, we describe how the capacity constraints were adjusted for the BSP formulation. Moreover, we
explain the preprocessing procedure used before solving the instances, and how the solver parameters
were set.

Capacity constraints adjustments. As reported in Section 4.2, unfeasibilities may occur in the lower
nodes of the binary tree when solving the BSP model. In such cases, we use parameters δUr,hm and δr,hm
L

to set more restrictive capacity constraints for resource r of subproblem m. In the numerical experi-
ments, these parameters were set to:

δU
r,hm = 2
hm −1
maxm qb,r and δr,h
L
= 2hm with hm > 0,

b∈B m

where Bm denotes the block model of the subproblem m, hm is the height of subproblem m in the
graph of subproblems, δU r,0 and δr,0 are set to 0. All parameters δr,hm and δr,hm were set to 0 when
L U L

solving subproblems with the FBSP model.

With respect to the final pit expansion model presented in Section 4.3, the adjustments δrL and
δr were set to the values of the adjustments δU
U
r,hm and δr,hm used in the first subproblem of the RAD
L

heuristic.

Preprocessing procedure. The final pit problem is solved for each instance using a pseudoflow al-
gorithm as proposed by Hochbaum & Chen (2000). The final pit eventually needs to be extended
whenever the sum of resource capacities to extract it is below the sum over the time horizon of the
minimum resource consumption for at least one resource. In such a case, the final pit extension pro-
cedure proposed in Section 4.3 is used.

16
 Table 2
Lower and upper bounds on the mining and processing capacities (in tons) for the modified
instances.

Mining capacity Processing capacity

Instance Periods LB UB LB UB
mNewman1 all 900,000 1,500,000 650,000 900,000
mZuckSmall all 20,000,000 35,000,000 6,300,000 25,000,000
mKD 1 1,800,000 2,100,000 1,000,000 2,000,000
mKD 2-12 1,800,000 2,100,000 1,710,000 2,000,000
mZuckMedium all 7,000,000 18,000,000 5,150,000 8,000,000
mP4HD all 5,000,000 20,000,000 4,000,000 10,000,000
mMarvin all 15,000,000 30,000,000 8,000,000 21,000,000
m2Marvin all 27,000,000 30,000,000 9,000,000 21,000,000
mW23 1 8,000,000 10,000,000 2,000,000 3,610,000
mW23 2-12 5,000,000 15,000,000 2,000,000 3,610,000
mZuckLarge all 1,000,000 2,500,000 700,000 1,200,000
mSM2 1 5,000,000 6,000,000 700,000 800,000
mSM2 2 5,000,000 6,000,000 900,000 1,000,000
mSM2 3 5,000,000 6,000,000 1,250,000 1,300,000
mSM2 4-30 5,000,000 6,000,000 1,100,000 1,300,000
mMcLaughlinLim all 10,000,000 20,000,000 2,500,000 3,300,000
mMcLaughlin all 10,000,000 20,000,000 2,500,000 3,300,000
m2McLaughlin all 11,500,000 20,000,000 2,500,000 3,300,000

Solver parameters. The subproblems of the binary tree were solved with the Gurobi solver by using
a value of 1% for the relative MIP optimality gap parameter. Default values were used for all other
parameters.

5.3. Comparison of the RAD heuristic with the sliding time window heuristics STWH and STWHR
Table 3 shows the size of the subproblems solved for both the BSP and FBSP models. Note that
the proposed methodology successfully decreases the size of the original problems listed in Table 1.
Moreover, we observe that FBSP subproblems are larger than BSP ones. The larger generated sub-
problem corresponds to the root subproblem of the m2McLaughlin instance. For the m2Marvin and
m2McLaughlin instances, the integer programming model proposed in Section 4.3 extends the final
pit to satisfy the overall minimum use of mining resource constraints while respecting the remaining
constraints. For all other cases, only blocks of the final pit are extracted without the need to expand it.

17
 Table 3
Number of variables and constraints in the subproblems solved using the proposed algorithm.

BSP FBSP

Smallest Average Largest Smallest Average Largest

Instance Vars Cons Vars Cons Vars Cons Vars Cons Vars Cons Vars Cons

mNewman1 956 3,698 1,413 7,142 2,120 8,192 1,916 5,142 2,835 7,863 4,240 12,092
mZuckSmall 1,356 4,292 4,317 43,632 18,800 300,688 2,852 7,008 8,953 51,486 37,600 328,888
mKD 1,588 15,926 14,927 234,822 28,306 453,717 1,184 2,626 6,701 51,512 56,612 496,176

18
mZuckMedium 6,702 92,593 15,472 492,121 58,554 2,571,699 10,920 77,884 33,164 527,951 117,108 2,659,530
mP4HD 11,542 89,687 28,181 402,248 81,894 1,518,173 23,056 106,780 54,657 432,275 163,788 1,641,014
mMarvin 1,300 3,704 3,924 27,284 17,032 180,332 3,099 10,655 15,430 65,950 68,120 411,760
m2Marvin 1,423 5,084 4,349 31,785 19,874 188,786 3,364 11,302 18,614 73,617 76,124 423,008
mW23 8,790 39,291 34,735 312,071 148,520 1,603,840 28,144 98,358 81,448 436,699 297,040 1,826,620
mZuckLarge 11,094 65,761 38,448 349,358 193,642 2,203,039 21,252 74,744 76,779 410,402 387,284 2,493,502
mSM2 1,286 1,825 4,868 6,915 36,776 53,500 2,764 4,018 9,780 14,291 73,552 108,664
mMcLaughlinLim 39,866 660,785 84,320 1,911,965 225,374 6,183,661 80,364 721,332 168,909 2,034,149 450,748 6,521,722
mMcLaughlin 39,720 595,014 84,110 1,901,729 221,248 6,077,888 79,765 972,549 172,658 2,158,904 461,832 6,623,507
m2McLaughlin 41,362 621,143 95,746 2,135,640 233,554 6,334,862 99,512 1,190,812 195,148 2,213,745 480,127 6,747,346
 Table 4
Results of the two-period sliding time window heuristic (STWH), the sliding time window heuristic with relaxation (STWHR) for the BSP model,
and the RAD heuristic for the BSP and FBSP models

Sliding time window heuristics

for the BSP model RAD heuristic

STWH STWHR BSP FBSP

LP Gap Time Gap Time Gap Time Gap Time Partial blocks SubPb
Instance Upper bound % [s] % [s] % [s] % [s] % #

mNewman1 23,740,544 1.79 11 2.72 6 2.14 2 2.01 4 14.64 3

mZuckSmall 929,861,634 Not feasible > 1 day 1.70 250 3.46 242 27.72 19

19
mKD 81,125,108 > 1 day > 1 day - - 8.25 5,031 29.75 11
mZuckMedium 710,409,606 Not feasible > 1 day 5.00 6,703 6.88 39,522 11.30 13
mP4HD 235,019,284 0.03 1,287 > 1 day 0.17 4,220 0.23 6,688 0.25 7
mMarvin 875,461,209 Not feasible > 1 day 2.31 292 4.81 151 17.50 19
m2Marvin 868,404,340 Not feasible > 1 day - - 3.96 92 23.21 19
mW23 399,975,146 Not feasible > 1 day 0.60 7,327 0.53 18,794 9.14 10
mZuckLarge 57,388,739 > 1 day > 1 day 1.59 14,135 0.90 23,486 1.18 24
mSM2 1,247,672,349 Not feasible 0.78 44,480 1.46 20 0.34 32 0.41 29
mMcLaughlinLim 1,078,802,179 > 1 day > 1 day 0.25 38,533 0.37 21,871 0.38 9
mMcLaughlin 1,078,847,800 > 1 day > 1 day 0.56 56,613 0.28 70,166 0.57 9
m2McLaughlin 1,076,305,701 > 1 day > 1 day - - 0.39 22,566 0.26 9
 Table 4 shows a comparison of the two-period sliding time window heuristic (STWH), the sliding
time window heuristic with relaxation (STWHR) for the BSP model, and the proposed RAD heuristic
for the BSP and FBSP models. The sliding time window heuristics were applied for the BSP model
with a two-period sliding time window for the STWH, and with a one-period sliding time window for
the STWHR.
For each instance, we report:

• The upper bound corresponding to the LP-relaxation (LP upper bound). The Bienstock-Zuckerberg
algorithm (Bienstock & Zuckerberg, 2010) was used to compute the LP upper bound for the
mNewman1, mZuckMedium, mP4HD, mW23, mZuckLarge, mMcLaughlinLimit, mMcLaugh-
lin, and m2McLaughlin instances. Since the LP-relaxations of both BSP and FBSP formulations
are equivalent, the same bound is used to compute the relative integrality gaps for both models.

• The relative integrality gap (Gap), between the solution obtained by each tested heuristic and
the LP upper bound, computed as follows:
LP upper bound − objective value
Relative integrality gap =
LP upper bound

• The wall clock computing time (Time) for each tested heuristic.

• The percentage of partially mined blocks (Partial blocks) corresponding to the number of par-
tially mined blocks divided by the number of extracted blocks in the solution obtained with the
proposed heuristic on the FBSP model.

• The number of subproblems (SubPb) for the BSP and FBSP models, each of which corresponds
to a node of the binary tree. This number is the same for both models.

RAD heuristic. When a feasible solution is obtained, the RAD heuristic is generally more efficient
in terms of solution quality and computing time with the BSP model than with the FBSP model.
The computing time efficiency of the BSP formulation may be explained by the fact that FBSP mod-
els are larger than BSP ones, as shown in Table 3, and the tighter upper bounds on the mining and
processing capacities (Section 4.2) that may decrease the number of conflicts between knapsack con-
straints. However, no feasible solutions were obtained using the BSP model for mKD, m2Marvin,
and m2McLaughlin instances; an unfeasibility was reported for one of their subproblems. No unfea-
sibility was reported with the FBSP model, the proposed heuristic always obtained feasible solutions
with a relative integrality gap of 2.49% on average over the 13 instances. The percentage of partially
mined blocks varies from 0.25% for the mP4HD instance to 29.79% for the mKD instance. For a flat
block model as the mSM2, only 0.41% of partially mined blocks are generated for 29 periods. These
percentages seem to depend more on the 3-dimensional shape of the block model than on the number
of periods of the instance.

Sliding time window heuristics. Only three instances could be solved with the sliding time window
heuristics STWH and STWHR within one day of computational time. The mNewman1 instance has
been solved with both methods, while the mP4HD and mSM2 instances could only be solved with
one method, STWH and STWHR, respectively. We believe that the sliding time window heuristic
(STWH) fails to generate feasible solutions because of the lack of information on the later periods
in the initial iterations. In the sliding time window heuristic with relaxation (STWHR) the relaxed
variables carry information about the future, but we observe that this method has only been able to

20
solve the smallest instances mNewman1 and mSM2 after applying the preprocessing procedure. For
all the others instances, it could not find any solution within one day of computational time. Indeed,
we observe that the size of the subproblems prevents the mP4HD instance from being solved in a
reasonable time with the STWHR, while the lack of information on the latter periods prevents the
mSM2 from being solved with the STWH.

Comparing the RAD heuristic with the sliding time window heuristics STWH and STWHR. As the
results show, the RAD heuristic addresses both issues: the lack of information on the later periods in
the case of the STWH and the size of the subproblems to be solved in the case of the STWHR. Indeed,
it reported feasible solutions with low relative integrality gaps even for the largest and most difficult
modified instances of the Minelib. Therefore, the RAD heuristic is an interesting alternative to the
classical sliding time window heuristics when lower bounds on resource consumption constraints are
considered.

6. Conclusions

A recursive time aggregation-disaggregation (RAD) heuristic was introduced to solve the fol-
lowing two versions of the open-pit mine scheduling problem with minimum resource consumption
constraints: the block sequencing problem (BSP) and the fractional block sequencing problem (FBSP)
proposed by Vossen et al. (2016). Using a binary tree structure, it recursively aggregates and disaggre-
gates the scheduling periods to sequentially solve two-period subproblems. Since the application of
the proposed heuristic to the BSP model may generate unfeasible solutions, we proposed tightening
the father subproblem’s constraint capacities while considering the height of the binary tree. However,
if the proposed heuristic is applied to the FBSP model, wherein blocks can be extracted in parts, the
lower and upper bounds of the resource consumption constraints need not be adjusted.
For the mMcLaughlin instance, the RAD heuristic with both of the models provided feasible
solutions. With 2,340,142 blocks, 10 periods, and 2 minimum resource consumption constraints for
each period, mMcLaughlin is, to our knowledge, the largest public instance with lower bounds on
resource consumption for which a feasible solution is reported.
While comparing with the more classical sliding time window heuristics STWH and STWHR, the
RAD heuristic addresses both pitfalls of the lack of information on the later periods encountered with
the STWH, and of the large size of the subproblems encountered with the STWHR.
A single-stage procedure is also introduced to extend the final pit whenever the total use of the
mining and processing resources is below the respective required minimums over the time horizon.
This procedure prevents the well-documented gap problem pitfall.
Moreover, we introduced 13 new instances based on the constrained pit limit problem (CPIT) in-
stances of the Minelib to include lower and upper bounds on resource consumption. Sliding time win-
dow heuristics have been implemented, showing the difficulty of solving the scheduling instances with
conflicting knapsack constraints. Feasible solutions were obtained by using the proposed heuristic for
all 13 instances with the FBSP model, and 10 instances with the BSP model. However, compared
with the FBSP model, better average relative integrality gaps were obtained with the BSP model, i.e.,
1.58% for the BSP model and 1.98% for the FBSP model over the 10 instances for which feasible
solutions were obtained with both of the models.
In the future, the computational times may be reduced by strengthening the formulation of the
subproblems by possibly adding clique and cover cuts, as suggested by Bley et al. (2010). To im-
plement those constraints, the algorithm proposed by Bienstock & Zuckerberg (2009) may be used
to solve the LP-relaxations of scheduling instances with minimum resource consumption constraints.

21
The consideration of stochastic versions of these problems is also a relevant research direction for
further study.

Acknowledgements

This work was partially funded by the ANID / PIA Project AFB180004. We thank José Fernando
da Costa Oliveira and the three anonymous reviewers for their valuable comments and suggestions on
earlier versions of this paper.

References

Askari-Nasab, H., Pourrahimian, Y., Ben-Awuah, E., & Kalantari, S. (2011). Mixed integer linear
programming formulations for open pit production scheduling. Journal of Mining Science, 47(3),
338–359. https://doi.org/10.1134/S1062739147030117

Bienstock, D. & Zuckerberg, M. (2009). A new LP algorithm for precedence constrained production
scheduling. Optimization Online. http://www.optimization-online.org/DB_FILE/2009/
08/2380.pdf. Accessed September 25, 2021

Bienstock, D. & Zuckerberg, M. (2010). Solving LP relaxations of large-scale precedence constrained

problems. Integer Programming and Combinatorial Optimization, 1–14. https://doi.org/10.
1007/978-3-642-13036-6_1

Bley, A., Boland, N., Fricke, C., & Froyland, G. (2010). A strengthened formulation and cutting
planes for the open pit mine production scheduling problem. Computers & Operations Research,
37(9), 1641–1647. https://doi.org/10.1016/j.cor.2009.12.008

Boland, N., Dumitrescu, I., Froyland, G., & Gleixner, A. (2009). LP-based disaggregation approaches
to solving the open pit mining production scheduling problem with block processing selectivity.
Computers & Operations Research, 36(4), 1064–1089. https://doi.org/10.1016/j.cor.
2007.12.006

Caccetta, L. & Hill, S. (2003). An application of branch and cut to open-pit mine scheduling. Journal
of Global Optimization, 27, 349–365. https://doi.org/10.1023/A:1024835022186

Chicoisne, R., Espinoza, D., Goycoolea, M., Moreno, E., & Rubio, E. (2012). A new algorithm for
the open-pit mine production scheduling problem. Operations Research, 60(3), 517–528. https:
//doi.org/10.1287/opre.1120.1050

Cullenbine, C., Wood, R., & Newman, A. (2011). A sliding time window heuristic for open pit
mine block sequencing. Optimization Letters, 5(3), 365–377. https://doi.org/10.1007/
s11590-011-0306-2

Dagdelen, K. (1986). Optimum open pit mine production scheduling by lagrangian parameteriza-
tion. Proc. 19th Applications for Computers and Operations Research in the Minerals Industries
(APCOM) Symposium, 127–142.

Espinoza, D., Goycoolea, M., Moreno, E., & Newman, A. (2013). Minelib: a library of open pit
mining problems. Annals of Operations Research, 206(1), 93–114. https://doi.org/10.1007/
s10479-012-1258-3

22
Gaupp, M. (2008). Methods for improving the tractability of the block sequencing problem for open
pit mining. Colorado School of Mines. Ph.D. thesis.

Gemcom (2021). Gemcom Whittle™ Strategic Mine Planning software. https://www.3ds.com/

products-services/geovia/products/whittle/. Accessed September 25, 2021

Gershon (1987). Heuristic approaches for mine planning and production scheduling. Int. Journal of
Mining and Geological Engineering, 5(1), 1–13. https://doi.org/10.1007/BF01553529

Hochbaum, D. S. (2008). The pseudoflow algorithm: A new algorithm for the maximum-flow prob-
lem. Operations Research, 56(4), 992–1009. https://doi.org/10.1287/opre.1080.0524

Hochbaum, D. S. & Chen, A. (2000). Performance analysis and best implementations of old and
new algorithms for the open-pit mining problem. Operations Research, 48(6), 894–914. https:
//doi.org/10.1287/opre.48.6.894.12392

Hochbaum, D. S. & Orlin, J. B. (2012). Simplifications and speedups of the pseudoflow algorithm.
Networks, 61, 40–57. https://doi.org/10.1002/net.21467

Johnson, T. (1968). Optimum open-pit mine production scheduling. Operations Research Department,
University of California, Berkeley. Ph.D. thesis.

Johnson, T. (1969). Optimum open-pit mine production scheduling. A Decade Of Digital Computing
In The Mineral Industry, 539–562.

Jélvez, E., Morales, N., Nancel-Penard, P., & Cornillier, F. (2020). A new hybrid heuristic algorithm
for the precedence constrained production scheduling problem: A mining application. Omega, 94,
102046. https://doi.org/10.1016/j.omega.2019.03.004

Jélvez, E., Morales, N., Nancel-Penard, P., Peypouquet, J., & Reyes, P. (2016). Aggregation heuristic
for the open-pit block scheduling problem. European Journal Of Operational Research, 249(3),
1169–1177. https://doi.org/10.1016/j.ejor.2015.10.044

Lambert, W., Brickey, A., Newman, A., & Eurek, K. (2014). Open-pit block-sequencing formulations:
A tutorial. INFORMS Journal on Applied Analytics, 44(2), 127–142. https://doi.org/10.
1287/inte.2013.0731

Lambert, W. & Newman, A. (2014). Tailored lagrangian relaxation for the open pit block sequenc-
ing problem. Annals of Operations Research, 222(1), 419–438. https://doi.org/10.1007/
s10479-012-1287-y

Lamghari, A., Dimitrakopoulos, R., & Ferland, J. A. (2014). A hybrid method based on linear pro-
gramming and variable neighborhood descent for scheduling production in open-pit mines. Journal
of Global Optimization, 63(3), 555–582. https://doi.org/10.1007/s10898-014-0185-z

Lerchs, H. & Grossman, H. (1965). Optimal design of open-pit mines. Transactions C.I.M., 58,
47–54.

Letelier, O., Espinoza, D., Goycoolea, M., Moreno, E., & Muñoz, G. (2020). Production schedul-
ing for strategic open pit mine planning: A mixed-integer programming approach. Operations
Research, 68(5), 1425–1444. https://doi.org/10.1287/opre.2019.1965

23
Liu, S. & Kozan, E. (2016). New graph-based algorithm to efficiently solve large scale open pit
mining optimization problems. Expert Systems With Applications, 43, 59–65. https://doi.org/
10.1016/j.eswa.2015.08.044

Meagher, C., Dimitrakopoulos, R., & Avis, D. (2014). Optimized open pit mine design, pushbacks
and the gap problem—a review. Journal of Mining Science, 50(3), 508–526. https://doi.org/
10.1134/S1062739114030132

Morales, N.and Jélvez, E., Nancel-Penard, P., Marinho, A., & Guimaraes, O. (2015). A comparison
of conventional and direct block scheduling methods for open-pit mine production scheduling.
Proceedings de APCOM 2015, Fairbanks, Alaska, USA, 1040–1051.

Moreno, E., Espinoza, D., & Goycoolea, M. (2010). Large-scale multi-period precedence constrained
knapsack problem: A mining application. Electronic Notes in Discrete Mathematics, 36, 407–414.
https://doi.org/10.1016/j.endm.2010.05.052

Newman, A., Rubio, E., Caro, R., Weintraub, A., & Eurek, K. (2010). A review of operations research
in mine planning. INFORMS Journal on Applied Analytics, 40(3), 222–245. https://doi.org/
10.1287/inte.1090.0492

Osanloo, M., Gholamnejad, J., & Karimi, B. (2008). Long-term open pit mine production planning: a
review of models and algorithms. International Journal of Mining, Reclamation and Environment,
22(1), 3–35. https://doi.org/10.1080/17480930601118947

Parra, A., Morales, N., Vallejos, J., & Nguyen, P. M. V. (2018). Open pit mine planning considering
geomechanical fundamentals. International Journal of Mining, Reclamation and Environment,
32(4), 221–238. https://doi.org/10.1080/17480930.2017.1278579

Picard, J. (1976). Maximal closure of a graph and applications to combinatorial problems. Manage-
ment Science, 22(11), 1268–1272. https://doi.org/10.2307/2630227

Ramazan, S., Dagdelen, K., & Johnson, T. (2005). Fundamental tree algorithm in optimising
production scheduling for open pit mine design. Mining Technology, 114(1), 45–54. https:
//doi.org/10.1179/037178405X44511

Samavati, M., Essam, D., Nehring, M., & Sarker, R. (2017a). A local branching heuristic for the
open pit mine production scheduling problem. European Journal of Operational Research, 257(1),
261–271. https://doi.org/10.1016/J.EJOR.2016.07.004

Samavati, M., Essam, D., Nehring, M., & Sarker, R. (2017b). A methodology for the large-scale
multi-period precedence-constrained knapsack problem: an application in the mining industry.
International Journal of Production Economics, 193, 12–20. https://doi.org/10.1016/j.
ijpe.2017.06.025

Samavati, M., Essam, D., Nehring, M., & Sarker, R. (2017c). A new methodology for the open-pit
mine production scheduling problem. Omega, 81, 169–182. https://doi.org/https://doi.
org/10.1016/j.omega.2017.10.008

Somrit, C. (2011). Development of a new open pit mine phase design and production scheduling al-
gorithm using mixed integer linear programming. Dissertation, Colorado School of Mines, Golden,
CO.

24
Vossen, T. W. M., Wood, R. K., & Newman, A. M. (2016). Hierarchical Benders Decomposition for
Open-Pit Mine Block Sequencing. Operations Research, 64(4), 771–793. https://doi.org/10.
1287/opre.2016.1516

Zeng, L., Liu, S. Q., Kozan, E., Corry, P. & Masoud, M. (2021). A comprehensive interdisciplinary
review of mine supply chain management. Resources Policy, 74, 102274. https://doi.org/10.
1016/j.resourpol.2021.102274

25