Omega 81 (2018) 169–182

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

A new methodology for the open-pit mine production scheduling

problem ✩
Mehran Samavati a,∗, Daryl Essam b, Micah Nehring c, Ruhul Sarker b
a
Australian Centre for Field Robotics, The University of Sydney, Sydney, Australia
b
School of Engineering and Information Technology, The University of New South Wales, Canberra, Australia
c
School of Mechanical and Mining Engineering, The University of Queensland, Brisbane, Australia

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

Article history: The open pit mine production scheduling problem (OPMPSP) consists of scheduling the extraction of a
Received 24 February 2016 mineral deposit that is broken into a number of smaller segments, or blocks, such that the net present
Accepted 30 October 2017
value (NPV) of the operation is maximised. This problem has been formulated as an integer program-
Available online 3 November 2017
ming (IP) model, involving both knapsack and precedence constraints. However, due to the large number
Keywords: of blocks and precedence constraints, this model has remained impractical in real planning applications.
Open pit mining In this paper, we propose a new method to quickly generate near optimum feasible (integer) solutions by
Precedence constrained knapsack problem using the fractional solutions from the linear programming (LP) relaxation of the IP model. To be applica-
Large size scheduling problems ble to real sized problems, a new heuristic that quickly computes a feasible LP solution is also proposed.
Linear programming relaxation Our methodology is tested on a set of both academically designed and real-world mine deposits, and
shows better performance than the heuristic used to tackle the same deposits in the literature. Interest-
ingly, the proposed methodology improves the best known solutions for the majority of the instances.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction using geo-statistic techniques to interpolate between drill-holes.

Blocks may all be uniform in size or may vary in size depending
Open pit mining is a technique for the extraction of mineral de- on the nature of the mineral deposit; however their dimensions
posits close to the earth’s surface. A ‘pit’ or hole is excavated in a are generally compatible with the span and reach of the equipment
series of ‘pushbacks’, or phases, to an appropriate overall angle to that is used in the extraction process.
ensure stability of the pit wall. The ultimate pit limit is reached Blocks containing valuable mineral and waste material are
once the costs associated with removing increasing volumes of scheduled across the life of the mining operation. The objective of
waste with depth exceeds the revenues obtained from processing this strategic mine plan is to maximise Net Present Value (NPV)
and selling the valuable material, thus reducing economic viabil- while complying with all constraints. In the planning of mining
ity. At this point any number of underground mining techniques operations, the major issue is to determine the optimal schedule
may be considered which requires tunnels to be excavated into the for block extractions subject to a variety of operational constraints.
earth in order to facilitate the continued extraction of the deposit This gives rise to the open pit mine production scheduling problem
at depth. (OPMPSP).
For the purpose of planning and scheduling the extraction of a Due to the fact that extraction should be carried out with a spe-
mineral deposit, it is broken into a number of smaller more man- cific angle, known as wall-slope angles, the block extraction pro-
ageable segments or blocks that contain an estimate of several key cess must follow a sequence in which all corresponding blocks lo-
attributes, including the quantity of valuable material within each cated above each block have been removed. This is reflected as a
block. The combination of all blocks that comprise a mineral de- precedence constraint when modelling OPMPSP.
posit is known as a block model. The attributes associated with When scheduling the extraction of valuable mineral and waste
each block are generally estimated from drill-hole data samples material of an open pit mining operation across a predetermined
time horizon it is essential to maintain resource capacities. The
maximum processing capacity limits the amount of valuable min-
✩
This manuscript was processed by Associate Editor Prokopyev. eral that is sent to the processing plant across each time pe-
∗
Corresponding author. riod. The maximum mine production capacity limits the amount of
E-mail addresses: [email protected], [email protected] valuable mineral and waste material that may be extracted in each
(M. Samavati).

https://doi.org/10.1016/j.omega.2017.10.008
0305-0483/© 2017 Elsevier Ltd. All rights reserved.
170 M. Samavati et al. / Omega 81 (2018) 169–182

time period. The physical limitations reflect on site equipment ca- block model were used for processing decisions. The authors ad-
pabilities, haulage distances and the size of the processing plant. dressed a more detailed type of the problem in which block desti-
wSince the 1960s, numerous researchers have been address- nations were considered. The most recent method based on aggre-
ing OPMPSP. Early studies were only focused on the ultimate pit gation and disaggregation has been proposed by Jelvez et al. [10].
limit (UPL) problem, a simplified version of the problem in which Chicoisne et al. [4] developed a methodology consisting of three
only precedence constraints are considered (see Lerchs and Gross- steps. They first solve the linear programming (LP) relaxation of C-
man [1]). OPMPSP was subsequently formulated as an integer lin- PIT with a decomposition method. After solving the LP, they ap-
ear programming (ILP) problem by Johnson [2]. Using binary vari- ply a rounding heuristic (a topological sorting heuristic) to gener-
ables, Johnson considered a number of capacity constraints as well ate feasible solutions. Subsequently, the authors use a local-search
as the precedence constraints. Their binary variable represented a heuristics to further improve the quality of solutions. Moreno et al.
fraction of a particular block sent to a particular destination by a [11] also tackled extremely large instances using the same decom-
particular time period. However, in order to facilitate computation position method and rounding heuristic to quickly generate a very
of the real-world problems, the original model was changed into good feasible solution. Another algorithm based on the topologi-
a straightforward linear programming model, C-PIT, in which the cal sorting heuristic has been recently developed by Liu and Kozan
destination of each block is decided a priori [3,4]. The formulation [12]. In addition to the technique developed by Chicoisne et al.
of C-PIT is presented in Eqs. (1a)–(1f). [4] for solving LPs, Bienstock and Zuckerberg [13] also proposed
a decomposition method for this purpose.
Parameters and sets
Cullenbine et al. [14] developed a combination of Lagrangian re-
T = {1, . . . , T } set of time periods
laxation with a sliding time window heuristic to solve the binary
I = {1, . . . , I } set of blocks
programming model of OPMPSP. Lambert and Newman [15] also
R = {1, . . . , R} set of resources
dri amount of resource r consumed per unit of block i presented an exact approach to the problem with the same con-
straints. Using a sliding time window heuristic, they generate an
Pit profit obtained from block i in time t
initial feasible solution for the use of a tailored Lagrangian re-
crt amount of resource r available in time t
P the set of all precedence relationships in the prob- laxation procedure. Gaupp [16] also took these constraints into
account. The author develops a variable reduction technique to
lem
generate the earliest and latest possible extraction time for every
Decision variables block. Using this information, the author produces special cutting
xit 1 if block i is removed by time period t; 0 otherwise plans for use with the branch-and-cut (B&C) method. Gaupp also
exploits the concept of Lagrangian relaxation in a separate attempt
 to reduce solution times.
Max f (x ) = Pit (xit − xit−1 ) (1a) A topological sort, or a topological ordering of a directed acyclic
i∈I t∈T graph (DAG), is a linear ordering of its vertices such that for ev-
Subject to: ery directed edge jk from vertex j to vertex k, j comes before k in
the ordering [17]. This concept has been used as a heuristic in a
xit ≤ x jt ∀(i, j ) ∈ P, t ∈ T (1b) number of studies dealing with precedence constraints in different
problems, such as the resource constrained project scheduling prob-

(dri )(xit − xit−1 ) ≤ crt lem [18] and OPMPSP [4,19]. Gershon [20] used the topological sort
i∈I (1c) concept to develop a heuristic called TopoSort. Chicoisne et al. de-
∀t ∈ T , r ∈ R veloped one of the most impressive methodologies that includes
TopoSort. The first two steps of their methodology generate solu-
xit ≤ xit+1 ∀i ∈ I, t = 1, . . . , T − 1 (1d)
tions within 6% of optimality in seconds even for large instances.
Hence, the purpose of this study is to propose a new methodology
xit ∈ {0, 1} ∀i ∈ I, t ∈ T (1e) that further exploits this feature.
We now describe the contribution of this study. In an attempt
to address C-PIT using the TopSort heuristic, we observed the fol-
xi0 = 0 ∀i ∈ I (1f)
lowing two special properties in the solution of the LP relaxation
In this formulation, the objective function (1a) seeks to max- formulation.
imise the total NPV. Constraint (1b) ensures that no block is ex-
(I) Based on the concept of mathematical expectation (expected
tracted before its predecessors, and constraint (1c) guarantees that
value) in probability theory, an expected value of t, where
the capacity limitations are not exceeded. One of the advantages of
t can be defined as a discrete random variable, and a func-
the C-PIT model is that, by using this model, it is possible to sched-
tion f(t) is the value of its probability distribution, can be de-
ule blocks directly instead of scheduling bench-phases, allowing a
fined for the extraction time of a block. This expected value
more accurate scheduling. However, even in this formulation, the
is obtained by the LP solution. By adding certain cuts to the
huge number of blocks leads to a high number of decision vari-
LP relaxation formulation, the new solution provides a more
ables and constraints, making OPMPSP very difficult to solve.
realistic expected value.
Since the 1980s, a number of methods have been developed to
(II) In the solutions generated by the LP relaxation, the values of
solve the ILP. Dagdelen and Johnson [5] and Caccetta et al. [6] used
the decision variables in the last time periods are generally
Lagrangian parameterization, relaxing the mining and milling con-
identical with those of the optimal solution.
straints. As a result, the problem could be solved by repetition of
any UPL algorithm, such as the Lerchs–Grossman algorithm (LGA) To our best knowledge, these properties have never been dis-
developed by Lerchs and Grossmann [1]. Later Caccetta and Hill covered in the literature. We propose a methodology to employ
[7] proposed a branch and bound technique to solve the formu- these properties in order to directly and efficiently schedule large
lated scheduling problem. Onur and Dowd [8] tried to tackle the size OPMPSP instances.
scheduling problem by formulating it as dynamic programming. In order to test the abovementioned properties, the LP relax-
Boland et al. [9] solved OPMPSP by using aggregated blocks to ation solutions will be evaluated by transforming them to inte-
schedule the extraction process, while individual blocks in the ger schedules. This transformation will be conducted by a heuris-
 M. Samavati et al. / Omega 81 (2018) 169–182 171

Table 1
Characteristics of instances in Samavati et al. [22].

Name Number of blocks Number of periods Number of resources (r)

P75 75 10 2
P196 196 10 2
P8370 8370 10 2
P10240 10,240 10 2
P15059 15,059 10 2
P25392 25,392 10 2 Fig. 1. An example of DAG with six vertices.
P29952 29,952 10 2
P350 0 0 35,0 0 0 10 2
P40 0 0 0 40,0 0 0 10 2 A topological sort, or topological ordering of a directed acyclic
P50562 50,562 10 2
P301088 301,088 10 2
graph (DAG), is a linear ordering of its vertices such that for ev-
ery directed edge jk from vertex j to vertex k, j comes before k
in the ordering. The in- and out-degree of a vertex is specified by
the number of incoming and outgoing edges, respectively. In the
tic that is based on the topological sort concept, i.e., the TopoSort
context of mine planning, blocks can be considered as vertices of
heuristic. In Section 2.1, we will briefly describe this heuristic. In
a DAG, and precedence relationships of the blocks can be regarded
Section 2.2, the first property will be discussed in detail, and in
as the directed edges between vertices. Fig. 1 describes the DAG of
the following section, we will propose a new method, called adding
a block model with 6 blocks. From this figure, block 4, for example,
cut (AC), to exploit this property. This method adds additional con-
cannot be extracted before its predecessors, that is, blocks 1 and 2.
straints to the LP relaxation formulation, and therefore, the existing
In topological sort, an eligible set contains the vertices with
LP techniques developed to solve large size LPs are not applicable.
0 in-degree. The sort process generates a precedence feasible se-
Hence, in Section 2.4, we will develop a new heuristic that quickly
quence of blocks by repeatedly selecting a vertex from the eligi-
computes a feasible LP solution for large size LPs. In Section 3, the
ble set, en-queuing it into a sequence, removing the vertex and its
second property is assessed and the obtained advantages from this
outgoing edges from the DAG, and finally updating the eligible set.
property are merged into the proposed algorithm.
For instance, in Fig. 1, the DAG has three 0 in-degree vertices (1, 2,
The efficiency of the proposed methodology is judged in
and 3). Thus, the eligible set consists of the same vertices. A topo-
Section 4, by testing it on a set of well-known benchmark prob-
logical sort randomly selects one vertex, say vertex number 2, from
lems from an online library known as “MineLib” (see Espinoza
this set and puts it into a sequence. Then, vertex 2 as well as edges
et al. [21]). This library includes both academically designed and
2–4 and 2–5 are removed from the DAG. Finally, the eligible set is
real-world mine deposits. The developed methodology shows bet-
updated and the process is repeated until every vertex is placed
ter performance compared to the heuristic used by the library
into the sequence.
providers, improving the best known objective value for every in-
Because a DAG can have many different topological orderings, a
stance.
weight wi is assigned to every block i, and in the step of selecting
a vertex from the eligible set, the vertex with highest weight is
2. Methodology selected. This means that the ordering is weighted with respect to
w.
In this section, we first describe the well-known TopoSort Given a weighted topological ordering, a feasible schedule can
heuristic. In Section 2.2, we discuss a special property in the LP be simply obtained by the TopoSort heuristic. For the convenience
relaxation solution, and in the following section, propose a new of following the paper, this heuristic is described in Algorithm 1.
method, called adding cut (AC), to exploit this property. Finally, in
Section 2.4, a new heuristic is proposed to quickly generate a fea-
Algorithm 1
sible LP solution for large scale LPs when they include additional
constraints coming from our AC technique. Input: An acyclic graph G = (I, P ) where I indicates the set of blocks
and P denotes the precedence relationship.
Throughout this paper, we refer to two sets of problems on
STEP 1: ti ← 1 ∀i ∈ I, where ti indicates the extraction time of block i.
which our experimental studies have been performed. The first k ← 1.
set contains eleven instances available in an online library known STEP 2: Generate the topological ordering {b1 , b2 , . . . , bI }.
as “MineLib” (http://mansci-web.uai.cl/minelib), compiled by Es- STEP 3: u ← bk and k ← k + 1.
pinoza et al. [21]. These instances will be described in detail in STEP 4: tu ← max{tv : (u, v ) ∈ P }.
tu ← max{tu , min{t : crt ≥ dru ∀r ∈ R}}
Section 4 where we provide computational results (see Table 2).
STEP 5: if tu ≤ T crtu ← crtu − dru ∀r ∈ R (update available resources)
Apart from the instances in Mine Library, we have also used STEP 6: if k < I k ← k + 1 and go to STEP 3, else return ti as the
a set of conceptual instances created by Samavati et al. [22] to extraction time of block i.
perform our experimental studies. The characteristics of these in-
stances are outlined in Table 1. While the data for these instances
Chicoisne et al. [4] proposed a new weight function for TS that
is conceptual in nature, it is however based on real operational
uses LP relaxation solution of a C-PIT formulation. For each i ∈ I
scenarios, making it useful for investigation purposes. For details
and t ∈ T , the authors interpreted Pit = (y∗ it − y∗ it−1 ) where y∗ it is
on how these instances and their corresponding block models are
the LP relaxation solution and Y∗ is the corresponding vector as
constructed, we refer the reader to the original article [22].
the probability that block i will be extracted in period t. Using
this interpretation, they came up with a function that generates
2.1. Topological sort the expected time of extraction for each block. This function is as
follows:
Before explaining our methodology, we describe both the topo-     
logical sort concept and the TopoSort heuristic that have been ap- Ei = t y∗ it − y∗ it−1 + (T + 1 ) 1 − y∗ it−1 (2)
t∈T
plied to several scheduling problems with precedence constraints.
We exploit this heuristic to evaluate the LP relaxation solutions Finally, for each block i, the authors defined wi = −Ei as the
generated in our methodology. weight to the TS heuristic. As stated earlier, in the process of topo-
172 M. Samavati et al. / Omega 81 (2018) 169–182

logical sort, the block (vertex) with the highest weight is selected Subsequently, we randomly selected a non-integer decision
from the eligible set; however, when the LP relaxation solution is variable from the LP relaxation solution of node 2 (decision vari-
considered as the weight of blocks, many of the blocks in the eli- able x96, 1 ). The value of this variable in the optimal integer so-
gible set might have the same weight, and the topological sort se- lution was equal to 1, and thus we added cut x96,1 =1 to the LP
lects randomly as a result. This means that a one-to-one mapping relaxation of node 2. After solving the generated LP relaxation for-
from the set of weights to a topological sort does not exist. For mulation, computing the expected extraction time, and giving it
simplicity of explanation, we refer to this issue as the “TopoSort to the TS heuristic, a schedule with an NPV of $23,778,346 was
random” issue throughout this paper. achieved. The process can be continued in the same way. Every
node can lead to a more improvement in the schedule. In other
2.2. First special property in the LP relaxation words, each node has a better NPV than the father node (i.e., its
previous node). However, it was observed that this did not occur
Based on the concept of mathematical expectation (expected in some nodes, which was due to the following two reasons:
value) in probability theory, it is not hard to see that the above-
mentioned Ei is actually the expected value of t, where t can be (a) It is not guaranteed that the TS heuristic yields a better
defined as a discrete random variable with f (t ) = (y∗ it − y∗ it−1 ) for schedule with a more realistic probability distribution due
each t = 1, . . . , T and f (t ) = (1 − y∗ it−1 ) for t = T + 1, as the value to the aforementioned TopoSort random issue.
of its probability distribution. In fact, if vector Y∗ is replaced with (b) Even though by cutting off the current LP relaxation solu-
the optimal (integer) solution of C-PIT (i.e., Y∗ = X∗ , where X∗ cor- tion (the current non-integer optimal point) the continuous
responds to the optimal integer solution of C-PIT), then Ei is either feasible space of this formulation becomes smaller, and in
equal to the time of extraction of block i, or equal to T + 1 if i is effect the probability of getting the optimal integer solution
not extracted. is increased, there is no guarantee that the new LP solution
Considering this fact, it is clear that if f(t) is a more realis- (i.e., Z∗ ) is more similar to X∗ . In fact, by continuing to cut
tic probability distribution, then Ei will become a more realistic off the non-integer solutions, it is only guaranteed that we
expected value. Therefore, we aim to make f(t) more realistic by will ultimately obtain the optimal integer solution, that is,
adding certain constraints (cuts) to the LP relaxation formulation. Z∗ eventually becomes exactly the same as X∗ .
In the LP relaxation solution of a given problem (Y∗ ), let y∗ bm = For all three instances in Experiment 1, we continued for 20
α for block i = b and period t = m, where 0 < α < 1. It is clear that nodes. As can be seen from Fig. 2, after only five nodes NPV has in-
its corresponding variable in X∗ , i.e., x∗ bm , has a value of either 0 creased from $23,775,561 to $23,898,519. Note that Chicoisne et al.
or 1. We round variable y∗ bm towards the value of x∗ bm , and use [4] gained schedules within 6% of optimality by only giving Ei that
Q, W ⊂ {1, . . . , I} × {1, . . . , T } to represent the index sets for the is obtained by Y∗ , to TS. Our starting schedule is also obtained in
y∗ bm variables that have been rounded to zero or one, respectively. the same process (i.e., the root node). This means that our start-
There is a possibility that the solution of the ES formulation, Z∗ , ing schedule is already a good solution, and a small improvement
can lead to a more realistic f(t). This technique is simply similar to in that can lead to a near-optimum solution. The percentages in
the B&B technique, in which cuts (additional constraints) are added Fig. 3(a) show improvements in the root node after 20 nodes. For
to the LP relaxation formulation to cut off the current non-integer example, a schedule with an NPV of $24,069,812 was obtained at
optimal solution. node 20. This improvement in the root node has been shown by

ES = Max f (x ) = Pit (xit − xit−1 ) 1.24% in Fig. 3(a), which is the gap between the root node’s NPV
i∈I t∈T ($23,775,561) and the NPV of node 20 ($24,069,812). It is also in-
Subject to: (1b)–(1d), (1f), teresting to note that the root node for the same instance was
within 2.9% of optimality, and after only 20 nodes, the solution was
xit = 0, ∀(i, t ) ∈ Q within 1.7% of optimality.
xit = 1, ∀(i, t ) ∈ W In Fig. 2, cuts in all the nodes, except in node 3, have led to
0 ≤ xit ≤ 1, ∀i ∈ I, t ∈ T \(Q ∪ W ) either an identical or a better objective value (NPV) than their fa-
In order to check the assumption whether Z∗ leads to a more ther node. Fig. 3(b) shows how many nodes (out of 20) have led to
realistic f(t), we compute its corresponding Ei and wi , and give wi a better NPV in these three instances. As can be seen, 95% of the
to the TS heuristic to obtain a schedule. If the resulting schedule nodes in both Newman 1 and P196, and 90% in P75 have improved
is better than the previous schedule (i.e., the schedule obtained the schedule. Thus it becomes clear that the issues (a) and (b) have
by giving Y∗ to the TS), then this assumption is true. For this pur- little impact on this process.
pose, we conducted Experiments 1 and 2 on the set of problems In Experimental Study 1, we rounded variables according to
in Table 1. Note that we only considered three instances of this their values in the optimal schedule. In practise where the opti-
table (Newman 1, P75, P196) for Experiment 1, as this experiment mal schedule is unknown, we have to separately add both cuts
required the optimal integer solution of a problem, which is only xit = 0 and xit = 1 to each node, resulting in two separate nodes.
available for these instances. One of the two cuts clearly rounds the variable towards its value
Experimental Study 1. Let Fig. 2 show part of the results ob- in the optimal integer solution, and the schedule obtained by this
tained during performing the experiment on the first instance of cut could be better according to Experimental Study 1.
Table 1, Newman 1. In the first node of this figure, we solved the LP However, it is not possible to recognise this better cut, as the
relaxation of C-PIT, computed the expected extraction time using other cut may lead to an enhancement in the schedule too, which
Eq. (2), and gave it to the TS heuristic. The objective value (NPV) of can be considered as a local optimum. In this case, both of the
the schedule obtained by the TS is $23,775,561. Then, we randomly generated nodes need to add a new cut for further improvement.
selected a non-integer decision variable from the solution of LP re- By continuing to add cuts to both nodes, a node tree will be ob-
laxation (decision variable x42, 1 ). The value of this variable in the tained. In Experimental Study 2, conducted on all the instances in
optimal integer solution was equal to 1 (we obtained the optimal Tables 1 and 2, we aim to analyse how many nodes in a tree have
integer solution using CPLEX), and thus we added cut x42,1 =1 to a better schedule than the root node.
the LP relaxation of C-PIT. Having solved the new formulation, we Experimental Study 2. Having solved the LP relaxation solution
computed the expected extraction time and gave it to the TS. We of C-PIT, we randomly selected a non-integer decision variable xit ,
obtained a schedule with an NPV of $23,778,346 (node 2 in Fig. 2). and separately added both cuts xit = 0 and xit = 1 to the LP re-
 M. Samavati et al. / Omega 81 (2018) 169–182 173

Fig. 2. Results of experiment on testing the special structure in LP relaxation.

Table 2 giving the LP relaxation solution to TS, which is now occurring at

Characteristics of instances in MineLib.
our root node. This means that our starting schedule is already a
Name Number of Number of Number of good solution. Therefore, considering the fact that at least 80% of
blocks periods resources (r) nodes result in a better schedule, the starting schedule can be sub-
Newman1 1060 6 2 stantially improved. In the next section, we develop the adding cut
Zuck small 9400 20 2 technique to take advantage of this property.
D 14,153 12 2
Zuck medium 29,277 15 2
P4HD 40,947 10 2 2.3. The adding cut (AC) technique
Marvin 53,271 20 2
W23 74,260 12 3
Zuck large 96,821 30 2 In this section, we propose the adding cut technique to exploit
SM2 99,014 30 2 the first property in the LP relaxation solution. This technique is
McLaughlin limit 112,687 15 1 essentially a branch-and-bound algorithm that applies TS in each
McLaughlin 2140,342 20 1
node of the tree.
Let a root node be the first node in which the LP relaxation so-
lution of C-PIT is solved. Similar to Experiment 2, non-integer deci-
laxation formulation (the root node), resulting in two new nodes. sion variable xit is selected from the solution. Both cuts xit = 0 and
Let us call this step a branch. Both nodes were then solved and xit = 1 are separately added to the current LP relaxation formula-
evaluated by giving the solution to TS. The node with better NPV tion to create two new nodes. The formulations in each node are
was then selected and the process was iterated. In the case where then solved and TS is run on their solutions. The node that pro-
both nodes had a better NPV than the father node (i.e., their pre- vides a better NPV than the father node (i.e., the previous node) is
vious node), the process was iterated in both of them. Both nodes selected and recorded. If both nodes have a higher NPV than the
may also be worse than their father node, due to the issues (a) father node, they are both selected and recorded. If neither of the
and (b) mentioned in Experimental Study 1. In this case, the node nodes provides a better NPV than the father node – this may hap-
with higher NPV was selected. We continued this process for 20 pen due to reasons (a) and (b) mentioned in Experimental Study
branches for each instance. Fig. 4 shows how many of the branches 1 – they are both selected but neither are recorded. The process
led to at least a better schedule than the one in the root node (we is then iterated with selected nodes. Note that one of the stopping
say ‘at least’ because both resulting nodes in a branch can be better criteria can be a predefined gap from optimality, i.e., the gap of the
than the root node). best obtained schedule from a given upper bound. In this study, we
From this figure, at least 80% of the generated nodes have im- stop the algorithm when there is no improvement in the tree for
proved the schedule in the root node. Note that, as stated earlier, two levels in a row (a level of the tree is part of the tree where
Chicoisne et al. [4] gained schedules within 6% of optimality by new nodes are created. For example, the root node is in level 1,
174 M. Samavati et al. / Omega 81 (2018) 169–182

iteratively generate LP solutions and run the TS heuristic on them.

1.4
1.24% For convenience of coding and using this black-box, its pseudocode
1.2 is provided in Appendix A. The black-box starts with several nodes
1 0.92% at the same time and progresses in parallel. As stated in Exper-
0.73% imental Studies 1 and 2, it is better to start from a good solu-
0.8
tion, but the only good solution available at the beginning is the
0.6 first node in which the pure LP relaxation is solved. Therefore, in
0.4 order to have more than one good solution for the black-box to
start with, we continue the node tree for a few levels and use the
0.2 recorded LPs as starting good solutions in the black-box. Note that
0 this is not necessary, but helps obtain a near-optimal solution more
Newman 1 P196 P75 quickly. We will analyse the impact of starting from good solutions
in Section 4.3.
(a)

2.4. A fast heuristic for generating a feasible LP solution

1 95% 95% 90%
0.9 One of the problems in the AC approach is solving the LP relax-
ation formulation of C-PIT. Although MIP solvers are fast and effi-
0.8
cient for solving LP formulations, they may have difficulty solving
0.7
extremely large instances of OPMPSP. Chicoisne et al. [4] proposed
0.6
CMA to solve the LP relaxation of large-sized C-PIT instances. How-
0.5 ever, this technique can only solve the pure LP relaxation of C-PIT,
0.4 and it fails to solve this formulation when we add cuts to it. Now
0.3 let CC-PIT be a formulation of C-PIT that involves cuts based on
0.2 variables in the first period. In Algorithm 2 proposed momentar-
0.1 ily, we develop a heuristic to quickly generate a feasible LP solu-
Newman 1 P196 P75 tion from CC-PIT. It is demonstrated in Section 3 that the larger
the time period, the less the effect that the decision variables have
(b) on improving the quality of solution obtained by AC. Thus we only
consider cuts in the first period.
Fig. 3. (a) improvement in the root node after 20 steps in three instances New-
man1, P196, and P75 in Experimental Study 1; (b) percentage of the nodes that

have led to a better NPV in the same three instances in Experimental Study 1.
CC − PIT = max Pit (xit − xit−1 )
i∈I t∈T

s.t.

the next two nodes are in level 2, the next four nodes after them xit ≤ x jt ∀(i, j ) ∈ P, t ∈ T

are in level 3, and so on). (dri )(xit − xit−1 ) ≤ crt ∀t ∈ T , r ∈ R
i∈I
While this method is fast (simply because LP relaxations can be xit ≤ xit+1 ∀i ∈ I, t = 1, . . . , T − 1
solved quickly using LP solvers for small problems or Critical Mul- xi,t=1 = 0 ∀i ∈ Z
tiplier Algorithm (CMA) by Chicoisne et al. [4] for large ones) the xi,t=1 = 1 ∀i ∈ K
improvement rate (convergence rate) may however become low xit ∈ {0, 1} ∀i ∈ I \(K ∪ Z ), ∀t ∈ T
due to the fact that in each step only one decision variable is con- xi0 = 0 ∀i ∈ I
sidered. Given a set of very good solutions, it is very common in
the context of combinatorial scheduling problems to use Evolution- where Z ⊂ {1, . . . , I} represents the index set for the variables that
ary Algorithms to quickly converge them towards the optimum. have been rounded to zero and K ⊂ {1, . . . , I} dictates the index set
Thus using a black-box that applies an Evolutionary Algorithm, we for the variables that have been rounded to one.

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Newman1

McLaughlin limit
D

P4HD
Zuck small

Zuck medium

P10240
P15059
P25392
P29952
P35000
P40000
P50562
P301088
W23

P196
P8370
SM2

P75
Marvin

McLaughlin
Zuck large

Fig. 4. The result of Experimental Study 2. Percentage of the branches that have led to a better schedule than the root node in all the instances of Tables 1 and 2.
 M. Samavati et al. / Omega 81 (2018) 169–182 175

In Algorithm 2, we develop a fast heuristic, partly based on the same value for this variable. The same is true for
CMA, to produce a feasible LP solution of CC-PIT. Let CP1 (u) be a decision variables in smaller time periods of 19, 18, and
version of CC-PIT for the first time period and with a single re- even 17.
source. (III) As stated earlier, for generating a LP solution by either CMA
 or our developed heuristic, it is necessary that the formu-
C P 1 (u ) = max Pi,t=1 xi,t=1
lations have a single resource constraint per time period,
i∈I
and thus instances with more than one resource must be
s.t.
 divided in separate formulations, each containing one of the
di xi,t=1 ≤ u resource constraints. Our third observation is the fact that
i∈I
in an instance with say two resources, the decision variables
xi,t=1 ≤ x j,t=1 (i, j ) ∈ P
from the two resulting LP relaxation formulations are ex-
xi,t=1 = 0 ∀i ∈ Z
tremely similar in the last time periods.
xi,t=1 = 1 ∀i ∈ K
0 ≤ xi,t=1 ≤ 1 ∀i ∈ I \(K ∪ Z )
Also, suppose CPt (u) is a version of CC-PIT for each time period From these three observations, as well as the fact that the AC’s
t > 1 and with a single resource. LP relaxation solutions are near optimal, there is a possibility that
 the decision variables in the last time periods of these solutions
C Pt (u ) = max Pit xit
are already identical with or similar to optimum. If this is true,
i∈I
then the decision variables of last time periods have little to no ef-
s.t.
 fect on improving the solution quality. In order to test this possibil-
di xit ≤ u ity, an experiment was performed on the instances of Tables 1 and
i∈I
2, as follows:
xit ≤ x jt (i, j ) ∈ P
Experimental Study 3. We conducted this experiment using the
xit ≥ x̄it−1 ∀i ∈ I
black-box that iteratively generates LP solutions and run the TS
0 ≤ xit ≤ 1 ∀i ∈ I
heuristic on them. Let’s first slightly describe the black-box. In this
where x̄it indicates the value of decision variable xit . When t = 1, black-box, the decision variables are divided into different groups
that is in the optimal solution of CP1 (u), and if t > 1 then that and the process of converging decision variables and running TS is
is in the optimal solution of CPt (u). Let xt correspond to xit for separately implemented in each group. This way the groups of de-
∀i ∈ I in period t. Let x̄t represent the optimal solution of CPt (Ut ) cision variables can progress simultaneously and in parallel. Each
for each t > 1, and x̄1 indicate the optimal solution of CP1 (Ut ) group progresses for a predefined number of steps.
for t = 1. Algorithm 2 demonstrates how to generate a feasible We now illustrate our experiment. The decision variables were
CC-PIT’s LP solution by individually solving CPt (Ut ) for each time divided into two categories A and B. Category B contained the deci-
period. sion variables in the last time period, and category A contained the
Note that for generating LP solutions by either the CMA of variables in the remaining columns. The groups of decision vari-
Chicoisne et al. [4] or Algorithm 2, it is necessary that the formu- ables were generated from categories A and B separately. There-
lations have a single resource constraint per time period. There- fore, the groups from the two categories were entirely distinct
fore, in order to address an instance with more than one resource, from each other.
the formulation is solved separately by AC for each resource con- As the black-box progressed, the best NPV found in each group
straint. This does not compromise feasibility as TS generates a fea- at the end of each step was recorded. Fig. 5(a), depicts these NPVs
sible schedule anyway. for instance Zuck Large. In this figure, the X and Y axes show
steps and their NPVs, respectively. The steps that are related to
Algorithm 2 the groups from category B are 469–487 and 954–972. Fig. 5(b)

t and (c) zoom in on these parts of Fig. 5(a). From Fig. 8, although
Input: An instance of CC-PIT with a single resource. Ut = cl for the convergence is quite significant in every single step for groups
l=1
each t ∈ T . from category A, the convergence is almost zero in steps for groups
STEP 1: Solve C P 1 (Ut=1 ) to obtain x̄1 from category B. This demonstrates that the decision variables in
STEP 2: Individually solve CPt (Ut ) for each t > 1 to generate x̄t .
the last time period have little to no effect on the improvement of
STEP 3: Return x̄ = (x̄1 , x̄2 , . . . , x̄T ) as a feasible solution of CC-PIT’s LP
relaxation.
solutions. The experiment was continued by including the last 2,
3, …, 29 time periods into category B. Fig. 6 describes the experi-
ment results for instance Zuck large when the last 2 time periods
were included. From this figure, even in this case, the convergence
3. Second special property in the LP relaxation
is extremely low for the groups coming from category B.
Having analysed the results for other instances, we come to an
It has been stated above that selecting non-integer variables
overall conclusion that the further we get from the first time pe-
from the first period is the best choice in our AC technique due to
riod, the less effective the decision variables are on the conver-
the second abovementioned property. In this section, this property
gence process, and for the last time periods, the convergence is
is assessed. When applying the AC method to different instances,
almost zero. Thus trying to improve the decision variables of the
we observed the following three facts:
last time periods is not computationally reasonable.
(I) A huge proportion of decision variables in a generated LP We believe this conclusion can revolutionise the research area,
relaxation solution has a value of 0 or 1 in the last time pe- as the dimensionality of a problem can be significantly reduced.
riods. The main reason for this observation would be con- Many algorithms can be developed in the future in order to con-
straint (1d) in the C-PIT formulation. verge the AC’s LP relaxation solutions towards the optimal so-
(II) A large number of decision variables in different LP relax- lution, without including some of the last time periods. In the
ation solutions have identical values in the last time peri- next section, we obtain the solutions of the Mine Library in-
ods. For example, if x2,20 = 1 in an instance with a max- stances, by excluding the decision variables of the last 30% of time
imum of 20 time periods, almost every LP solution has periods.
176 M. Samavati et al. / Omega 81 (2018) 169–182

Fig. 5. Behaviour of convergence for instance Zuck large while category B in Experimental Study 3 involves the decision variables of the last time period. (b) and (c) zoom
in on the x axis of Figure (a).
 M. Samavati et al. / Omega 81 (2018) 169–182 177

Fig. 6. Behaviour of convergence for instance Zuck large while category B in Experimental Study 3 involves the decision variables of the last two time periods. (b) and (c)
zoom in on the x axis of (a).
178 M. Samavati et al. / Omega 81 (2018) 169–182

4. Implementation details and results

To evaluate the performance of our AC technique, we test it on

Jelvez et al.
a set of benchmark instances known as “MineLib” compiled by Es-

time (s)

16,304
pinoza et al. [21], and present the computational results.

1229
1825
395

852
771
159
913
45

30
3
4.1. Characteristics of test instances

Jelvez et al. gap

(percentage)
The benchmark instances provided by Espinoza et al. [21] are
available online on http://mansci-web.uai.cl/minelib. Table 2 de-

closed

68.81

60.98
88.81
93.83

88.25
73.95

54.91
56.47

87.30
77.87
82.12
tails the name of each instance alongside its number of blocks and
time periods. Newman1 is an academic data set constructed by the
authors of the library. D is a copper deposit located in North Amer-
ica. P4HD is a gold and copper deposit in North America. Marvin is

Liu and Kozan

Not reported
a well-known conceptual deposit provided with the Whittle min-

time (s)
ing software package. W23 is another gold mine located in North

55,731
10,871
2005
5262
5812

1032
1622
America. SM2 is a conceptual deposit, designed based on a nickel

154
24
1
deposit in Brazil. McLaughlin is a gold mine located in California,
and McLaughlin limit is its final pit computed by Somrit [23]. Zuck

Worse solution

Worse solution
small, Zuck medium, and Zuck large are also conceptually designed

Liu and Kozan

Not reported
(percentage)
mine deposits in this library.

gap closed

66.88

40.94
26.22

83.58
66.21
61.22

57.20

91.18
4.2. Computational results and comparison

The algorithm was coded using MATLAB 7.12.0 (R2011a) and run

AC technique
on personal computers with an Intel Core i7 CPU at 3.40 GHz, 16
Gigabytes of RAM, operating under Microsoft Windows 7. The LP

time (s)
relaxation solutions for all instances were computed in IBM ILOG

4087

6205
3401
2231
1840
1931

3316
943
283
217
36
CPLEX studio 12.6.
The best known NPV of the test instances until 2013 has been
provided by Espinoza et al. [21], and is available in MineLib. These
AC technique

(percentage)
values have been obtained using a modified version of the TS
gap closed

heuristic. Column Best known NPV in MineLib in Table 3 indicates

48.98
95.45
60.92

88.26
88.99
84.31
90.74

87.25
64.10
these values. Column LP upper bound represents the optimal so-

6.94
7.00
lutions of the LP relaxation formulation of C-PIT. This value pro-
vides a good upper bound for the results. The gaps of the so-
lutions obtained in MineLib, shown in the fourth column of the

1,646,618,421
1,078,316,053
1,078,419,238
AC technique

408,923,283

858,409,310
246,228,134
673,426,719

table, are computed with respect to this upper bound, where a 399,331,214
848,112,016

56,819,671
24,126,234
value ($)

smaller gap implies a solution closer to optimum. The next col-

umn shows the NPVs obtained by the AC heuristic. From the ta-
ble, it can be seen that AC has substantially enhanced all the NPVs
obtained in 2013. Column labelled AC technique gap closed indi-
Best known gap

cates the percentage of improvement made by AC to the solu-

in MineLib ($)

tions obtained in 2013. For example, consider instance McLaughlin

65,529,796

95,229,995
12,640,362

43,190,083

5,493,989
5,652,304
2,808,309
1,277,034

8,427,136
1002,513

Limit, where the upper bound is 1,078,979,501. Given that the best
611,904

known NPV for this instance in 2013 is 1,073,327,197, the differ-

ence between the upper bound and this NPV is 5,652,304. For the
same instance, the AC heuristic computes a solution with a NPV
NPV in MineLib

of 1,078,316,053. Similarly, the difference between this NPV and

1,073,530,279
1,645,242,774
1,073,327,197
788,652,600

392,226,063
820,726,048
396,858,193

246,138,696
Best known

615,411,415
23,483,671

the upper bound is 663,448. Comparing these two numbers (i.e.,

56,777,190

5,652,304 and 663,448), the gap has been closed by 88.26%. Col-
umn AC technique time represents our technique’s solution times
($)

in seconds.
Computational results and comparison.

Liu and Kozan [12] and Jelvez et al. [10] have also tested their
LP upper bound

algorithms on instances in MineLib. Columns Liu and Kozan gap

1,079,024,268
1,078,979,501
1,648,051,083
409,498,555

400,653,199
854,182,396

863,916,131
247,415,730
710,641,410

57,389,094
24,486,184

closed and Jelvez et al. gap closed show the percentage of improve-
ment made by these studies to the solutions obtained in 2013. In
Column Liu and Kozan gap closed, “worse solution” implies that the
technique obtained a worse solution than MineLib , thus the im-
provement is negative. According to the table, both studies have
Zuck medium

McLaughlin L
McLaughlin

also considerably improved the NPVs reported in 2013. Thus we

Zuck small

Zuck large
Newman1

find it interesting to compare the results of the three algorithms.

Marvin
Name

P4HD
Table 3

W23

SM2

Jelvez et al. [10] obtains a better solution than both Liu and Kozan
D

[12] and our algorithm for instances Newman1, Zuck large, and
SM2. For instances Zuck small, D, Zuck medium, Marvin, McLaughlin
 M. Samavati et al. / Omega 81 (2018) 169–182 179

Table 4
Analysing the effect of starting from LPs on instance Newman 1.

Cycles Run 1 (seven random solutions) Run 2 (seven LP relaxation solutions) Run 3 (six random solutions + one LP relaxation solution)

1 $23,754,732.8253 $23,875,556.6796 $23,775,561.8240

2 $23,799,099.4468 $23,875,556.6796 $23,775,561.8240
3 $23,799,099.4468 $23,799,099.4468 $23,775,561.8240
4 $23,799,099.4468 $23,885,294.9904 $23,880,353.1027
5 $23,826,890.5883 $23,826,890.5883 $23,880,353.1027
6 $23,826,894.4731 $23,826,894.4731 $23,880,353.1027
7 $23,826,894.4731 $23,986,178.5567 $23,918,854.9657
8 $23,826,894.4731 $23,986,178.5567 $23,918,854.9657
9 $23,826,894.4731 $24,024,989.0010 $23,918,854.9657
10 $23,826,894.4731 $24,024,989.0010 $23,918,854.9657

L, and McLaughlin, however, our algorithm achieves a better solu- the LP relaxation formulation of C-PIT. While constraint program-
tion than the other two. Liu and Kozan’s algorithm obtains the best ming solvers are fast and efficient for solving LP formulations, they
solution for P4HD and W23; the authors have not reported any so- may have difficulty solving extremely large instances of OPMPSP.
lution for the last instance. To sum up, Liu and Kozan [12] yield Chicoisne et al. [4] proposed CMA to solve the LP relaxation of C-
the best known solution for two instances of MineLib. Jelvez et al. PIT for large scale instances. However, this technique can only solve
[10] obtain the best known solution for three instances of MineLib. the pure LP relaxation of C-PIT, and it fails to solve an LP involv-
Our algorithm obtains the best known solution for six instances. ing additional cuts. Therefore, we developed a new heuristic that
quickly computes a feasible LP solution for this type of LP formu-
4.3. Analysing the effect of starting with LPs lations.
We tested the performance of our AC algorithm on the re-
As stated in Section 2.3, the AC technique can start with or cent benchmark instances from MineLib [21]. This library provides
without initial ‘good’ solutions. In this section, we describe our both academically designed and real-world instances, thus provid-
conducted experiment to analyse the impact of starting with ini- ing an apt ground to test new algorithms. The computational re-
tial solutions. sults showed that the best known NPVs for every instance in this
We solved instance Newman 1 using AC but with three differ- library have been substantially improved by our algorithm. We also
ent initial solutions: (1) seven random solutions; (2) seven good conducted a comparison of our results against those obtained by
solutions, i.e., LP relaxation solutions recorded in the node tree ex- Liu and Kozan [12] and Jelverz et al. [10] which have addressed
plained in Section 2.3; (3) one recorded LP relaxation solution and the same instances. Both studies managed to considerably enhance
six random solutions. the results in MineLib. Comparing the three studies, it can be seen
Table 4 depicts the NPVs obtained in the first ten cycles of our that Lie and Kozan [12] offers better results than both our algo-
black-box. From this table, although the initial solutions are ran- rithm and Jelvez et al. [10] for only two out of eleven instances. For
dom in the first run, the results are relatively good, which is due the largest example, however, the study reports no result. Jelvez
to the fact that the TS heuristic yields good schedules even with et al. [10] offers the best solution in this comparison for three out
a random input for such a small size problem. This makes New- of eleven instances. Our algorithm, on the other hand, obtains the
man 1 challenging enough for our analysis, and that is actually best solution for six out of eleven instances.
why we selected this instance for this experiment. Nevertheless, Regarding the future research directions, we demonstrated that
the solutions are of low quality compared to Run 2. In Run 1, the a number of decision variables in LP solutions are not fractional
algorithm reaches an NPV of $23,826,894.4731, while an NPV of and can be identical to the optimal solution. We believe this can
$24,024,989.001 is found in Run 2. It is also interesting to note be considerably exploited in the future, as some variables can be
that the convergence is slow in Run 1. By adding one LP relaxation fixed before starting a solution procedure, thereby reducing the
solution in Run 3, the quality of final solution rises, but the con- dimensionality of the problem. However, there is still a question
vergence is still poor. of how many time periods should be fixed in order to reduce di-
mensionality without losing quality of solution. This trade-off be-
5. Conclusions and future research tween quality and computational effort can be considered for fu-
ture work. The proposed AC algorithm showed impressive results;
In this work, we proposed a new algorithm, called AC, in order however, the minimum resource requirements that are of impor-
to address large scale OPMPSPs in a more effective and efficient tance for mining companies should also be taken into consider-
way. We first raised two special properties in the LP relaxation ation in the future. It is believed that the AC technique is more
formulation: (1) by adding some special cuts to the LP relaxation flexible than other successful algorithms in the literature and thus
formulation and solving it, a mathematical expectation (expected can potentially be modified to involve the minimum resource re-
value) for the extraction time of blocks can be obtained. By giving quirements.
this expected value to a topological sort heuristic, a near optimal
schedule is obtained. Our experimental studies showed that con- Appendix A. Mechanism in the black-box
tinuing to add cuts to the LP relaxation can lead to an enhance-
ment in the schedule; (2) a large number of the decision variables As mentioned earlier, we iteratively generate LPs and run TS on
in the generated LP relaxation solutions have identical values to them in a black-box that uses a general Evolutionary Algorithm.
those in the optimal solution. Our experimental studies demon- This can be done by any Evolutionary Algorithm, heuristic or meta-
strated that this generally occurs in the last time periods. heuristic. In this study, we selected the Genetic Algorithm (GA). GA
In order to exploit these properties, we proposed the AC tech- solves optimisation problems based on a natural process mimick-
nique in which we iteratively add cuts to the LP relaxation for- ing biological evolution. This algorithm repeatedly changes a pop-
mulations and run the well-known TopoSort heuristic on them. ulation of individual solutions by modifying the solutions. For de-
One of the concerns in the AC approach is associated with solving tails about GA, readers are referred to Holland [24].
180 M. Samavati et al. / Omega 81 (2018) 169–182

Algorithm 3

Input: A set of initial LP relaxation solutions to start with. Each LP solution is an individual in GA. Let nPop indicate the number of solutions and Pop
= {Pop (1 ), Pop(2 ), . . . , Pop(nPop)} be the set of these solutions. Each solution Pop (i) contains the following three levels: (1) Pop.LP (2) Pop.schedule (3)
Pop.NPV.
STEP 1: Find the solution with the highest NPV, e.g., Pop (i), and select it as the best solution, i.e., Best ← Pop (i).
I×N
STEP 2: Cycle ← 1, and Ns ← SubSize , where SubSize is the number of variables in groups, and I and N are the number of blocks and time periods, respectively.
STEP 3: Randomly divide the decision variables into Ns groups of size SubSize.
s ← 1, n ← 1.
STEP 4: Construct Ns subproblems each including nPop different solutions as follows:
4.1 Assign the values of decision variables in solution Pop (n) to the decision variables in group s in the nth solution of subproblem s, i.e.,
Sub(s).Pop(n).LP(s) ← Pop(n).LP(s),
4.2 Assign the values of the decision variables of the best solution Best to the remaining decision variables in the nth solution of subproblem s, i.e.,
Sub(s).Pop(n).LP(remaining varialbes) ← Best.LP(remaining varialbes),
4.3 Run the TS heuristic to obtain the solution’s schedule and NPV,
4.4 if n < nPop
n ← n + 1, and go to STEP 4.1.
STEP 5: if S < Ns
s ← s + 1, n ← 1, and go to STEP 4.1.
STEP 6: Evolve each subproblem s separately using GA operators as follows:
6.1 Number _o f _generations ← 1;
6.2 Apply crossover and mutation operators only to decision variables of group s of the solutions,
6.3 Run the TS heuristic to obtain the schedule and NPV of the solutions in the new generation,
6.4 if Number _o f _generations < the predefined number of generations
Number _o f _generations ← Number _o f _generations +1, and go to STEP 6.2.
else Best (s) ← the best solution in subproblem s.
STEP 7: Let k = {Best (1 ), Best (2 ), . . . , Best (Ns ) } indicate the set of best solutions obtained in Step 6.4 for every subproblem. Let Best (i)be the solution that
offers the highest NPV in set k.
7.1 update the best global solution, Best, as follows:
if Best.NPV < Best (i).NPV
Best ← Best (i);
STEP 8: if Cycle < predefined number of cycles
Cycle ← Cycle + 1, and go to STEP 9;
else Return Best as the best solution.
STEP 9: Update the initial LP relaxation solutions to start a new cycle:
9.1 s ← 1, n ← 1;
9.2 Assign the values of the decision variables in group s of the nth solution of subproblem s to the decision variables in solution Pop (n), i.e.,
Pop(n).LP (s) ← Sub(s).Pop(n).LP (s).
9.3 if n < nPop
n ← n + 1, and go to STEP 9.2;
if s < Ns
s ← s + 1, n ← 1, and go to STEP 9.2;
9.4 go to STEP 3.

Instead of starting from one node (LP solution), our black-box able for a continuous domain. The authors analysed these variants
starts with several nodes at the same time and implements the of operators and compared their performances by solving a variety
process of generating LPs and running TS in parallel. In each node, of benchmark problems. The authors stated that the best crossover
the decision variables are divided into different groups. The groups operator for the continuous domain is SBX with index parameter
progress simultaneously and in parallel too. Each group progresses n = 3, and the best mutation operator is the non-uniform one with
for a certain number of steps, and afterwards, the black-box takes coefficient b = 5. According to them, SBX is best applied with a 0.9
the LP with the best obtained NPV and starts the process all over probability to individuals selected by a 3-element tournament, and
again from that LP. We call this a ‘cycle’. Thus, if cycle is equal to the best performance for the non-uniform mutation is with a 0.1
5, it means that the black-box does the process of taking the best mutation rate. Hence, in this study, we picked SBX and the non-
LP and starting from it, for 5 times. We do this procedure by GA. uniform mutation for the GA.
For convenience of coding and using our black-box, we provide the A generational scheme is the second aspect that needs to be
pseudocode in Algorithm 3. In this Algorithm, each solution con- considered when using GA. It is related to the way the generated
tains the three levels Pop.LP, Pop.schedule, and Pop.NPV. Level Pop.LP offspring are inserted in the population after applying crossover
holds the value of decision variables. Pop.schedule and Pop.NPV hold and mutation. The most basic way is that offspring directly replace
the schedule and NPV of the solution, respectively. the parents. In this paper, the elitism rate is 3, implying that each
In Step 4, several different subproblems are generated. Each population holds the best three individuals and replace the oth-
subproblem includes a few LP solutions. The aim of each subprob- ers with generated offsprings. This means populations in each sub-
lem is to only evolve a specific group of decision variables. The problem in Algorithm III contain different solutions, whilst concur-
subproblems progresses separately and in parallel. rently evolving steadily to better NPVs (fitness function values).
Step 6 describes the process of evolving each subproblem us-
ing GA operators. Over the last few decades, several crossover and
Appendix B. Calibration of the GA
mutation operators have been introduced for optimising problems
with a continuous search space, such as blend crossover (BLX-α ),
Similar to any other optimisation algorithms, our algorithm in-
simulated binary crossover (SBX), simplex crossover (SPX) parent
volves control parameters that need to be calibrated. In the pro-
centric crossover (PCX), polynomial mutation, non-uniform muta-
posed algorithm, two parameters require such calibration: (1) sub-
tion etc. However, finding the best operator is a tedious task. To
problem size (SubSize) (2) number of generations (Ng).
help researchers and practitioners, Elsayed et al. [25] have con-
For the purpose of calibration, the Design of Experiments ap-
ducted an extensive review of the variants of GA operators suit-
proach has been applied in this study (see Mongomery [26]). More
 M. Samavati et al. / Omega 81 (2018) 169–182 181

Table 5 Table 6
Three categories for set of instances. Parameter values used for the algorithm after calibration.

Category Name Small size problems Medium size problems Large size problems

Small size instances Zuck small SubSize: 50 0 0 SubSize: 60 0 0 SubSize: 90 0 0

D Ng: 5 Ng: 5 Ng: 10
Zuck medium
P4HD
Medium size instances Marvin
W23 Having conducted the ANOVA test on each category separately,
Zuck large it was noticed that the highest and lowest F-ratio corresponded to
SM2 factors SubSize and Ng, respectively, in all three sizes of problems,
McLaughlin limit
and thus they were fixed in the same order using LSD intervals.
Large size instances McLaughlin
Table 6 shows the values fixed for the studied parameters.

References
precisely, a factorial design was conducted, in which all possible
[1] Lerchs H, Grossman FI. Optimum design of open-pit mines. Operations re-
combinations of the following factors were tested. search. Institute of Operations Research Management Sciences; 1964. 901
SubSize: 7 levels (40 0 0, 50 0 0, 60 0 0, 70 0 0, 80 0 0, 90 0 0, 10,0 0 0); Elkridge Landing Rd, Ste 400, Linthicum Hts, Md 21090-2909.
Ng: 4 levels (5, 10, 20, 30). [2] Johnson T.B. Optimum open pit mine production scheduling 1968; DTIC Docu-
ment.
Parameter tuning using factorial design has been taken into ac- [3] Albach H. Long range planning in open-pit mining. Manag Sci 1967;13(10)
count in several studies [27–30]. In this technique, all the alter- B-549-B-568.
native algorithms yielded by different combinations of the stud- [4] Chicoisne R, Espinoza D, Goycoolea M, Moreno E, Rubio E. A new al-
gorithm for the open-pit mine production scheduling problem. Oper Res
ied factors are run on one or a series of examples. The results of 2012;60(3):517–28.
these runs are considered as the response variable of the exper- [5] Dagdelen K, Johnson TB. Optimum open pit mine production scheduling by
iment. The resulting experiment is then analysed by means of a Lagrangian parameterization. In: Proceedings of the 19th APCOM symposium
of the society of mining engineers, AIME; 1986.
multifactor analysis of variance (ANOVA) method. The F-ratio is a
[6] Caccetta L, Kelsey P, Giannini L. Open pit mine production scheduling. In: Pro-
component of the ANOVA that is computed as the ratio between ceedings of the computer applications in the mineral industries international
the unexplained variance and the variance explained by a factor symposium, APCOM 98; 1998.
[7] Caccetta L, Hill SP. An application of branch and cut to open pit mine schedul-
(see Montgomery and Runger [31]). The magnitude of the F-ratio
ing. J Global Optim 2003;27(2–3):349–65.
is an indicator of significance. This means a greater F-ratio implies [8] Onur AH, Dowd P. Open-pit optimization – part 2: production scheduling and
that the factor has more effect on the response variable. inclusion of roadways. Trans Inst Min Metall Sect A Min Ind 1993;102:105–13.
Thus, the procedure of calibration by means of factorial design [9] Boland N, Dumitrescu I, Froyland G, Gleixner AM. LP-based disaggregation ap-
proaches to solving the open pit mining production scheduling problem with
is as follows: it begins by running all the alternative algorithms, block processing selectivity. Comput Oper Res 2009;36(4):1064–89.
yielded by different combinations of studying factors, on one or a [10] Jélvez E, Morales N, Nancel-Penard P, Peypouquet J, Reyes P. Aggrega-
series of instances. The results (the response variable of the exper- tion heuristic for the open-pit block scheduling problem. Eur J Oper Res
2016;249(3):1169–77.
iments) are then analysed by means of the ANOVA method. This [11] Moreno E, Espinoza D, Goycoolea M. Large-scale multi-period precedence con-
analysis yields an F-ratio for each factor. Subsequently, the factor strained knapsack problem: a mining application. Electron. Notes Discret Math
with the greatest F-ratio is chosen, and a multiple comparison test 2010;36:407–14.
[12] Liu SQ, Kozan E. New graph-based algorithms to efficiently solve large scale
is carried out between the studied levels of that specific factor (for open pit mining optimisation problems. Expert Syst Appl 2016;43:59–65.
example the 4 levels (5, 10, 20, and 30) for factor Ng). This mul- [13] Bienstock D, Zuckerberg M. Solving LP relaxations of large-scale precedence
tiple comparison test is normally done with the least significative constrained problems. In: Proceedings of international conference on integer
programming and combinatorial optimization. Springer; 2010.
difference (LSD) intervals (see Ruiz et al. [28]). Using this compari-
[14] Cullenbine C, Wood RK, Newman A. A sliding time window heuristic for open
son, the best level is found for the factor, and the factor is fixed at pit mine block sequencing. Optim Lett 2011;5(3):365–77.
this desired level. Then, the second greatest F-ratio is selected and [15] Lambert W, Newman A. Tailored Lagrangian relaxation for the open pit block
sequencing problem. Ann Oper Res 2014;222(1):419–38.
the same procedure is applied.
[16] Gaupp MP. Methods for improving the tractability of the block sequencing
The procedure continues this way until all the factors are fixed, problem for open pit mining 2008; DTIC Document.
that is, all the parameters of the algorithm are tuned. [17] Pang C, Wang J, Cheng Y, Zhang H, Li T. Topological sorts on DAGs. Inf Process
We applied the same procedure to the Mine Library problems Lett 2015;115(2):298–301.
[18] Cheng R, Gen M. An evolution programme for the resource-constrained project
to calibrate our algorithm’s parameters. In our case, we had a total scheduling problem. Int J Computer Integr Manuf 1998;11(3):274–87.
of 4 × 7 = 28 different combinations to test. Note that we divide [19] Zhang M. Combining genetic algorithms and topological sort to optimize
the Mine Library set of problems into three different categories, open-pit mine plans. In: Proceedings of the 15th mine planning and equip-
ment selection; 2006. p. 1234–9.
based on the number of decision variables, as illustrated in Table 5. [20] Gershon M. Heuristic approaches for mine planning and production schedul-
This is because the sizes of OPMPSPs are substantially different and ing. Int J Mining Geol Eng 1987;5(1):1–13.
calibration should be performed separately. [21] Espinoza D, Goycoolea M, Moreno E, Newman A. MineLib: a library of open pit
mining problems. Ann Oper Res 2013;206(1):93–114.
The first instance of the Mine Library set of problems was not [22] Samavati M, Essam D, Nehring M, Sarker R. A local branching heuris-
considered in any of the categories for calibration, as the size of tic for the open pit mine production scheduling problem. Eur J Oper Res
this academically designed instance is extremely small and would 2017;257(1):261–71.
[23] Somrit C. Development of a new open pit mine phase design and produc-
cause great influence to the calibrated precision of parameters.
tion scheduling algorithm using mixed integer linear programming. Golden,
Regarding the response variable for the experiments, the fol- CO: Colorado School of Mines; 2011.
lowing performance measure is considered for every instance: [24] Holland J. Adaptation in natural and artificial systems: an introductory analysis
with applications to biology, control, and artificial intelligence. University of
Michigan Press; 1975.
Al gsol − Bestsol
Relative percentage deviation (RPD ) = × 100 [25] Elsayed SM, Sarker R, Essam D. A comparative study of different variants of
Bestsol genetic algorithms for constrained optimization. In: Simulated evolution and
learning. Springer; 2010. p. 177–86.
where Algsol is the solution found by a given algorithm alternative [26] Montgomery DC. Design and analysis of experiments. John Wiley & Sons;
2008.
on a given problem and Bestsol is the best solution obtained by the [27] Ruiz R, Stützle T. A simple and effective iterated greedy algorithm for the per-
algorithm. We ran five duplicates of each treatment. mutation flowshop scheduling problem. Eur J Oper Res 2007;177(3):2033–49.
182 M. Samavati et al. / Omega 81 (2018) 169–182

[28] Ruiz R, Maroto CA, lcaraz J. Two new robust genetic algorithms for the flow- [30] Vallada E, Ruiz R. A genetic algorithm for the unrelated parallel machine
shop scheduling problem. Omega 2006;34(5):461–76. scheduling problem with sequence dependent setup times. Eur J Oper Res
[29] Vallada E, Ruiz R. Genetic algorithms with path relinking for the minimum 2011;211(3):612–22.
tardiness permutation flowshop problem. Omega 2010;38(1):57–67. [31] Montgomery DC, Runger GC. Applied statistics and probability for engineers.
John Wiley & Sons; 2009.