Professional Documents
Culture Documents
Environment
To cite this article: Martha E. Villalba Matamoros & Mustafa Kumral (2018): Underground mine
planning: stope layout optimisation under grade uncertainty using genetic algorithms, International
Journal of Mining, Reclamation and Environment, DOI: 10.1080/17480930.2018.1486692
Article views: 2
1. Introduction
Many underground mining techniques implement production through volumes called stopes. Early
in mine planning, the feasible stope layout is chosen that maximises the profit of the mining venture
under site-specific geotechnical requirements, quality specifications, and accessibility constraints.
This decision-making problem – known as stope layout optimisation – affects the value of the mine
project under the chosen production schedule and can be viewed as equivalent to the ultimate pit
limit problem in open pit mining.
To date, orebody uncertainty is usually ignored in stope layout optimisation. Similar to the petro-
leum industry, the underground mining industry should also consider the impact of uncertainties in
production planning [1]. Some parameters used in stope layout optimisation are estimated or sim-
ulated. Estimation generates the best approximation of an attribute (e.g. grade) at any location, but
does not provide useful information about uncertainty. On the other hand, simulation yields probable
images of the orebody and provides an opportunity to assess uncertainty.
The uncertainty associated with input parameters has not been addressed in underground mine
planning in detail. Furthermore, commercial mining software does not provide tools to deal with data
uncertainty. A plan based on one scenario will be vulnerable to unexpected realisations. In such a case,
the mining company may face financial losses and needs to modify the plans frequently, increasing
costs and mining activities [2]. Therefore, a robust stope layout is required that is insensitive to unex-
pected realisations of the grade.
Although research on stope layout optimisation for underground mining is new relative to open
pit mine planning, significant knowledge has accumulated. For example, Ataee-pour [3] proposed an
approach called the maximum value neighbourhood, where many neighbourhoods by each block are
analysed to select stopes. However, this approach does not consider that different stope layouts could
be generated when the starting point is modified. Topal and Sens [4] developed a heuristic approach
that accounts for several stope sizes to find the optimum stope layout in a three-dimensional block
model. Sandanayake et al. [5] emphasised geotechnical and physical constraints in their analysis of one
predefined stope size, which given the variable shapes of orebodies, provided less flexibility to maximise
profit. Villalba and Kumral [6] addressed minimisation of inherent internal dilution, in addition to
conventional profit maximisation. This approach improved upon previous studies in terms of stope
layout, but the globally optimal solution could not be guaranteed. Manchuk and Deutsch [7] used
simulated annealing to find a solution that maximises profit based on the location of the initial stope
design. Alford and Hall [8] determined the stope layout through a floating stope, where geological
control and mineralisation is tracked, to find the final non-overlapping stope shape through annealing.
See [9] for a comparison of the capabilities and limitations of four of the approaches described above:
maximum value neighbourhood, Topal and Sens, Sandanayake, and floating stope
Starting to account for orebody uncertainty, Grieco and Dimitrakopoulos [10] developed a mixed
integer programming model to solve the stope layout problem under a single input level of risk given
by orebody model fluctuations. The orebody is regularised in minable rings before optimisation,
and the geotechnical and operational constraints are satisfied. The formulation maximises the metal
content times the probability of the rings being greater than the cut-off grade. Verhoeff [11] used
maximum upside potential and minimal downside risk analysis [12] to identify the stope layout that
has the higher upside potential while minimising downside risk. Stope layout is optimised using GAs
for each equal probable orebody realisations at the time. The orebody uncertainty was considered in
the stopes design; however, a single stochastic solution informed about orebody fluctuations is not the
outcome, and probabilistic approaches ignore the joint local uncertainty given by orebody realisations.
This paper proposes a new stope layout approach that is insensitive to unexpected realisations. The
approach integrates orebody uncertainty in stope layout optimisation by employing an adaptation of
genetic algorithms to explore the near-optimal solution. The stope layout evaluation that manages
internal dilution and orebody uncertainty and considers variable stope dimension is a computationally
complex problem. It may require a heuristic or meta-heuristic algorithm to generate near-optimal
solutions within computing times that are practical relative to exact methods.
slices are allowed in the middle of the stope. In other words, slices with internal dilution must be
surrounded by profitable slices. Overall, a stope created by a combination of slices must be profitable.
The optimisation model is formulated as below, this formulation corresponds to an adaptation of
the deterministic heuristic stope layout formulation by Villalba and Kumral [6] to account for orebody
grade uncertainty. This model is run and N layouts are obtained at the end of each realisation. The
notation used in stochastic stope layout optimisation is:
Indexes
Parameters
ς penalty factor associated with recovery loss due to internal dilution ∈ [0, 1]
cutoffdilution
minimum grade for internal slice dilution
R recovery of metal
Rdilution recovery of metal for minimum grade of internal dilution
f (gjk) recovery function of blocks with grade less than the cut-off grade
( )
This function is defined with a linear equation: f gjk = mR × gjk + bR. The slope and y intersect
are calculated as follows:
( )
R−Rdilution
mR = , bR = −cutOff × mR + R
(cutOff−cutOffdilution )
Decision variables
xnijk binary variable, if block k at sector j and realisation n is mined as a part of stope i or not
lnij length of stope i at sector j and realisation n (lnij is between 𝜆min
j , 𝜆j )
max
dnij additional slices to the minimum stope slices i at sector j and realisation n (deif is between 0,
(𝜆max
j j )∕dx)
− 𝜆min
snij number of internal dilution slices for stope i at sector j and realisation n (below the cut-off grade)
bnj maximum allowable amount of internal dilution slices at sector j and realisation n (bnj is between
0 and (𝜆max
j ∕dx) − 2)
The second objective function component – the cost associated with recovery loss due to internal
dilution – facilitates minimising internal dilution due to spatial variability of grades within the orebody
and operational and geotechnical constraints.
J I K(j) ( )
∑ ∑ ∑
xnijk wjk gnjk RP − M − C
j=1 i=1 k=1
Max
J I K(j)
(1)
∑∑∑ ( ) ( ( ))
− xnijk wjk gnjk 1 − Onjk P R − 𝜍(f gnjk , ∀n = 1, ..., N
j=1 i=1 k=1
6 M. E. VILLALBA MATAMOROS AND M. KUMRAL
If the penalty factor is equal to 1, this refers to the lowest penalisation because the second component
uses only the adjusted dilution block recoveries. A factor equal to 0 is the highest penalization since
dilution blocks do not add value. This objective function is subject to some constraints, which ensure
that block k can be mined only once through the sectors and stopes.
I J
∑ ∑
xnijk ≤ 1, ∀k = 1, … , k(j), ∀n = 1, … , N (2)
i=1 j=1
The blocks (ηij) in height and width directions need to be mined as slice predecessors, with block k
being evaluated. The set of slice predecessors for each k is defined as Ωk′:
𝜂
ij
∑
𝜂ij xnijk − 𝜏
xnijk �
� ≤ 0, ∀i = 1, … , I, ∀j = 1, … , J, ∀k = 1, … , K(j), k ∈ Ωk � , ∀n = 1, … , N (3)
𝜏=1
The set of slice predecessors Ωk′ for each block k must have an available number of blocks ηij equal to
the blocks in width and height stope dimensions; otherwise, block k is not considered as part of the
stope being examined.
(𝛿 𝛾j )
j
𝜂ij = × − 1, ∀i = 1, … ,I, ∀j = 1, … ,J (4)
dy dz
The next precedence constraints consider the minimum length each stope may have. The parameter
φij defines the quantity of blocks in the minimum length stope direction. The set of block successors
Ψk′ is given for each block k being evaluated.
𝜑
∑ij 𝜏 �
𝜑ij xnijk − xnijk � ≤ 0, ∀i = 1, … , I, ∀j = 1, … , J, ∀k = 1, … , K(j), k ∈ Ψk , ∀n = 1, … , N
𝜏=1
where (5)
𝜆min
j
𝜑ij = dx
−1
The maximum length of each stope is related to the decision variables (dnij) and a second set of successor
blocks ϒk′. The parameter θij defines the quantity of blocks in the maximum length stope direction
which set of block successors ϒk′ are identified per block k being assessed.
𝜃
∑ij 𝜏 � �
𝜃ij xnijk − xnijk � − dnij ≤ 0, ∀i = 1, … , I, ∀j = 1, … , J, ∀k = 1, … , K(j), k ∈ Υ , ∀n = 1, … , N
k
𝜏=1
where (6)
𝜆max
j
𝜃ij = dx
−1
The decision variable dnij allows the stopes to have dimensions between the minimum and maximum
size per sector.
𝜆max
j − 𝜆min
j ≤ dnij × dx, ∀i = 1, … ,I, ∀j = 1, … ,J, ∀n = 1, … , N (7)
The average grade of stope slice predecessors must be greater than the cut-off grade.
𝜂ij +1 (
1 ∑
)
xnijk × wjk × gnjk ≥ cutOff , ∀i = 1, … , I, ∀j = 1, … , J, ∀n = 1, … , N (9)
𝜂ij + 1 k=1
The stope slice predecessors Ωk′ are accumulated until they match the stope size constraints, where
internal slices with average grades between cutOffdilution and cutOff are allowed until they match the
maximum slice number decision variable (bnj) per iteration. Thus, bnj takes values from 0 to maximum
stope slices (two slices per iteration and per sector), and the decision variable snij provides the optimal
amount of internal dilution slices of stope i at sector j and scenario n.
𝜆max
(10)
j
bnj ≤ − 2, ∀j = 1, … ,J, ∀n = 1, … , N
dx
(snij +2∑
)(𝜂ij +1) � �
xnijk × wjk × gnjk
k=1
� � � � ≥ cutOff , ∀i = 1, … , I, ∀j = 1, … , J, ∀n = 1, … , N (13)
snij + 2 × 𝜂ij + 1
Multiplying the quantity of blocks by the maximum number of slices with internal dilution yields the
number of iterations to find the near-optimal solution by each orebody realisation. This first stage
optimised the stope layouts for orebody realisations, then gave a set of solutions that are equally
probable; however, these solutions cannot be assessed independently by GAs because together they
describe the stope layout solution uncertainty.
Figure 4. Variability of N stope scenarios and their average design which two stopes have ≥50% of occurrence of illustrated scenarios.
mining region block when its minimum distance is less than the translation ratio, e.g. block 1 of
stope 1 in dark blue (Figure 4) is clustered with block 1 of scenario 2 and 3 because the Euclidian
distance among them are ≤aj. The maximum stope size could be used as a reference to define this
translation ratio. After all scenarios are assessed, each block has a value associated with its frequency
of occurrence through the regions. To calculate the average design, the blocks that represent more
than 50% of the mining regions are selected (Figure 4).
The clustering process violates geotechnical and operational restrictions. Thus, an additional step
is needed to ensure that stope layout is feasible and can be used in the GA stage. This stage is carried
out with an extra constraint that ensure a minimum percentage r of the average design in each stope.
In this case, each block has a binary variable cjk that identifies if block k is part of the average design,
that is, a feasible stope is made of average design blocks or a least have some average design blocks
which percentage in the stope is ≥r. The percentage r is an input parameter in the formulation.
𝜂ij +1 (
1 ∑
)
xijk × cjk ≥ r , ∀i = 1, … , I, ∀j = 1, … , J (14)
𝜂ij + 1 k=1
The dimension of decision variables xnijk is reduced to xijk because a single average design is used. The
iterations during this optimisation generate robust, feasible solutions that account for the average
design, which contains information about orebody realisations. However, these solutions are consid-
ered suboptimal regarding their profit or fitness value.
search technique and if it is well-designed, it can avoid local optima [17]. Initially proposed by John
Holland in the 1960s, the GA mimics biological evolution and suggests a theoretical framework for
adaptation, where the population of chromosomes moves to a new population by using the concepts
of natural selection and operators, such as crossover, mutation, and inversion [18]. The parents are
randomly chosen to produce children each generation; however, the selection operator selects the
parents or chromosomes in the population that will be able to survive over time. Over multiple gen-
erations, the population trends toward an optimal solution.
The size of this population m is an input parameter that can change through generations. The
selected population size reflects a trade-off between efficiency and effectiveness [19]. Too small a
population may provide insufficient room for exploring the search space effectively, whereas too big
a population may affect the efficiency of the technique, such that no solution is found in a reasonable
computing time.
To select the initial population, the roulette-wheel method is applied. The objective value is related
to the fitness value and the selection probability of each solution is proportional to its fitness, that is,
the probability distribution of the solutions is used to choose m solutions [17,19]. These solutions
enter the loop of the generation evaluation, as can be seen in Figure 1. Some additional indexes and
variables assist in performing GA:
e a generation, where e = 1, ..., E
h a solution in generation e, where h = 1, ..., H(e)
xijkd binary decision variable, if block k at sector j and solution h is mined as part of stope i or not
m initial population size
t mating pool size
pc crossover probability is a measure of the chance that random blocks at given solution h are
chosen for swapping. pc is also used to select solutions for mating
pm mutation probability is a measure of the likelihood that random blocks at given solution h are
modified
yijd post-processing integer variable, if stope i at sector j and solution d is selected or not as part of
stope layout
Each solution, h, is evaluated as a chromosome or string, blocks are treated as characters in string or
loci in the chromosome, and the alleles of each locus are binary. A locus is related to decision variable
xijkh. Considering that only stope layout blocks have a value of 1 in the string of size Kr(j), the string will
decay in any given generation because the stopes are destroyed by crossover and mutation operators.
To increase the chance of improving fitness through generations, some unfeasible stopes (which were
excluded from the feasible solution) could be added in the string with a character value equal to 1.
These stopes include: (1) profitable stopes that do not match the minimum stope size; (2) stopes that
have less than the minimum percentage (r) of average design; and (3) stopes where internal dilution
made them unprofitable. The blocks of these infeasible stopes have the option to become profitable
using genetic operators because new arrangements of stopes are tested in each generation. This assists
the genetic operators to improve the fitness. A generation is as follows:
(1) Solutions are ranked in descending order of fitness value. Solutions with the highest fitness
value are ranked as first in the population and have a lower probability of selection. The
cumulative distribution function of these solutions is sampled t times to build the mating pool.
(2) The crossover probability selects potential parents from the mating pool. Thus, the mating
pool, with size t, is examined using random numbers between [0, 1] and crossover proba-
bility. If the random number is below pc, the chromosome is recorded as a parent; otherwise
it is discarded. This selection diversifies the population at each generation and reduces the
mating pool size.
10 M. E. VILLALBA MATAMOROS AND M. KUMRAL
(3) The chromosomes of the previously reduced mating pool are used to select couples for repro-
duction. The viability of the couples is tested to diminish the chance of cloning before off-
spring are generated. Two conditions apply: a couple cannot have identical chromosomes
and couples should be unique.
(4) Using the crossover operator, each couple produces two offspring with probability pc. The
crossover block locations at solutions are randomly selected where the block locations of
feasible stopes are locked for crossover, that is, only the block locations of infeasible stopes
and remaining part of solutions are used for crossover.
(5) All offspring generated by crossover are mutated with probability pm. The block locations
to mutate are randomly selected and the block locations of feasible stopes are not used for
mutations.
(6) The fitness of each offspring is assessed by the stope layout algorithm that considers the
average design.
(7) The new set of solutions from offspring fitting evaluations is found and returned to step 1.
It is relevant to choose offspring that have a bias toward the best fitness and maintain the diversity of
solutions in each generation. Thus, crossovers and mutations help to improve fitness each generation
and parent selection at step 2 facilitates preserving diversity.
The sequential Gaussian simulation approach was used to simulate the orebody because it is simple,
flexible, and reasonably efficient [20]. The simulation generates equally probable orebody scenarios
and the average grade of these scenarios is the estimated orebody. The orebody model contains 18,270
blocks with block dimensions of 10 m (EW), 10 m (NS), and 10 m (height) (Table 1). The average gold
grade of the estimated model is 3.15 g/t, and the average grade of the realisations fluctuates between
2.64 and 3.65 g/t.
Each stope must have an average grade ≥3.5 g/t. The stopes can also include internal dilution slices
with average grades <3.5 and ≥0.5 g/t, and barren blocks with grades <0.5 g/t. ( )The blocks with the
grade <3.5 g/t are penalised by adjusting their recovery with a linear function f gjk = 0.12 × gjk + 0.54
where the grade of block k at sector j.
The deterministic solution based on the estimated model was optimised to be used as the base case
for comparison purposes. A dual-core processor computer (3.70 GHz, 16 GB of RAM) was used. The
programme ran for 33 min to find the deterministic solution, which ignores orebody uncertainty. A
total of 58,810 iterations evaluated all tentative stope locations, the potential number of slices with
internal dilution per stope, and possible stope sizes. The model yielded production of 262,610 oz Au
(Figure 6).
The stope layout uncertainty was quantified as part of the first stage and possible stope locations
were defined (Figure 7). The optimisation of 10 orebody realisations required 8 h and 12 min of
computing time.
In the second stage, the average design of these 10 stope layouts was found. A translation ratio of
40 m was used to find similarities among stope layouts. The gold quantity obtained from this average
design was 265,763 oz (Figure 8), 1.2% higher than the deterministic solution because the stope layout
based on 10 realisations had more flexibility to explore solution space.
The feasibility of the average design was assessed through the procedure where each stope must
have more than 10% of the average design (flagged as 1 in Figure 9, otherwise flagged as 0). A total of
425 feasible solutions from 6426 iterations were found. Twenty solutions were sampled as the initial
population from the cumulative distribution of the 425 solutions. This initial population contains
the heuristic solution, with 294,941 oz Au, as the solution with the highest probability to be selected.
The union of 10 stope layout solutions defined the Kr(j) domain of 2142 blocks, which is 88%
smaller than K(j) domain. This reduction in the domain improves the second and third stages in
terms computing solving time.
The fitness values of the initial population are illustrated in Figure 10 as generation 0, where their
maximum and minimum values are US$124,042,000 and US$88,709,000, respectively. The string size
of each solution was 2142 blocks where stope layout blocks are flagged with 1. The stope layout blocks
ranged between 472 and 592 among the strings. To promote efficient genetic operator performance,
the string switches from 0 to 1 infeasible stope locations, which have the potential to be feasible using
genetic operators. These strings increase their percentage of blocks with value 1 between 30 and 61%
INTERNATIONAL JOURNAL OF MINING, RECLAMATION AND ENVIRONMENT 13
where some infeasible blocks become feasible because the new stope layout is evaluated each gener-
ation. Figure 10 shows that fitness improved through 250 generations.
The input population was sampled 100 times per generation to build the mating pool, which was
examined using random numbers between [0, 1] and crossover rate (0.25) to ensure diversity. Parents
were selected when random numbers were <0.25 and nonrepeating couples were assembled randomly.
The fitness value variability among solutions affects the number of couples to be selected and the next
generation input population size. For instance, the input population fitting values of the 1st generation
had a coefficient of variation of 0.12, whereas this value was 0.08 in the 29th generation.
14 M. E. VILLALBA MATAMOROS AND M. KUMRAL
The roulette-wheel and mating pool building strategy at the 1st generation selected eight couples
given 20 solutions as the input population. The eight couples were solutions 18 and 10, 2 and 11, 4
and 12, 3 and 10, 2 and 10, 11 and 14, 5 and 18, and 4 and 10. The 29th generation had 28 solutions
as the input population and six couples were selected from the mating pool: 14 and 27, 20 and 2, 12
and 15, 27 and 1, 6 and 5, and 2 and 19. The lower input population variability of the 29th generation
led to 33% fewer children, even though the input population size was 31% larger than the 1st genera-
tion. The GAs adaptation was developed in C++ using Microsoft Visual Studio 2010 the same to the
other approaches presented in this paper. This third stage took 6 h and 45 min of computing time to
converge around the 223rd generation and find the best feasible solution, with the maximum profit
and mining recovery and minimum dilution (Figure 11).
Solution of the GA-based algorithm at a gold deposit had 12% more profit, 30% fewer barren blocks,
and 11.6% less dilution than the deterministic solution (Figure 12).
INTERNATIONAL JOURNAL OF MINING, RECLAMATION AND ENVIRONMENT 15
The initial population of the third stage contains the heuristic solution with the highest probability
to be selected. During the heuristic evaluation, the internal dilution percentage is minimised; however,
GA managed better this dilution by reducing it an additional 13.8%. The proposed approach helps
to better explore the solution space by evaluating possible new stope layouts given modifications
produced by crossover and mutations in each child. Figure 13 shows that the GA-based stope layout
solution delivered 13.4% more profit, 36.9% fewer barren blocks, and 15% more average stope grade
than the heuristic solution.
The third stage uses the initial population size of 20, a crossover rate of 0.25, and a mutation
rate of 0.05 as main input parameters. Several combinations were tested to choose these parameters
and sensitivity analysis showed that profit fluctuated up to 11%. Variable input population size was
employed through generations, and only the initial population size was an input parameter. The sen-
sitivity analysis of this parameter was performed with 0.25 crossover probability and 0.05 mutation
probability. The outcome of two solutions as initial population size was compared with several initial
population size outcomes (Figure 14).
The sensitivity analysis of the crossover rate considered 20 solutions as the initial population and
a mutation rate of 0.05. The solution with 0 crossover rate was compared with several crossover rate
solutions, and a nonlinear change of profit was observed (Figure 15).
The last input parameter to be examined considered 20 solutions as the initial population and a
crossover rate of 0.25. The mutation rate sensitivity analysis showed that the solution could degenerate
when the mutation rate was >0.05 (Figure 16).
16 M. E. VILLALBA MATAMOROS AND M. KUMRAL
The proposed formulation maximised the profit related to layout design and could be extended
to consider additional geotechnical and mining consideration as locations of pillars if the mining
method required. To have a complete evaluation of the profit, the main access layout development –
which includes level locations, mining directions, and others considerations – and stope sequenc-
ing activities could be incorporated. Simultaneous optimisation of these high correlated consecutive
activities in mining planning will generate solutions that lead to lower mining cost and higher cash
INTERNATIONAL JOURNAL OF MINING, RECLAMATION AND ENVIRONMENT 17
flow of the mining project. The contribution of this paper is related to one activity of underground
mine planning, and this stope layout activity has high impact in the following events due to it is the
first filter to convert resources into reserves.
An important question regarding the use of meta-heuristics is how a solution generates a near-op-
timal answer. There is no universally accepted approach to prove if a solution is near-optimal. Villalba
and Kumral [21] proposed a method to improve quality of a solution produced through GAs based on
experimental design. In this research, population size, mutation and crossover rates were calibrated to
find the best configuration of these parameters – using the full factorial and response surface methodol-
ogy – while maximising the project profit. The other way to improve the quality of a solution is to solve
a small instance with an exact method and the proposed approach, then compare the results of both
methods. If the solutions comply, GAs can be implemented for large instance. However, a successful
comparison cannot guarantee the optimality for every problem because GAs control parameters are
problem specific dependent.
4. Conclusion
Mine plans are generated using estimation or simulation methods based upon sparse field data.
Therefore, the grade at unsampled locations cannot be assessed with certainty. Decisions made under
such circumstances are exposed to significant uncertainty, which adds a risk to a mining enterprise.
Current underground mine planning techniques use a single orebody estimate scenario, which does
not provide knowledge regarding grade uncertainty.
In this research, a new approach is proposed to incorporate orebody uncertainty in the stope layout
optimisation, which permits efficient exploration of the solution space. The presented formulation has
three stages: the first quantifies the stope layout uncertainty given N orebody scenarios, the second
calculates the average design and their feasibility assessment breeds the initial population, and the
third implements GA to improve the initial population through generations.
The proposed GA-based formulation can find near-optimal solutions that converge in a reasonable
computing time. Moreover, their genetic operators assist in searching for a new part of the solution
space where better fitness is obtained. To demonstrate the performance of the proposed approach,
a case study was carried out using data from a gold deposit. The internal dilution was reduced from
36.4 to 32.1% and barren blocks from 15.2 to 10.6%. These were managed by considering variable
stope sizes and internal dilution slices, and penalising the block values with a grade <3.5 g/t (cut-off)
through a lower processing recovery in each fitness evaluation.
The GA solution produced 12% more profit than the deterministic solution and 13.4% more profit
than the heuristic solution. In addition, sensitivity analysis showed that GA input parameters such
as initial population size, crossover rate, and mutation rate affect the profit up to 11%. The solution
satisfies geotechnical and operational constraints and the mining recovery of the economical material
is maximised by: (1) using pieces of previous solutions to build new solutions; (2) selecting a mating
pool strategy; (3) selecting conditions during crossover and mutation operations; and (4) evaluating
the fitness value of each child for each generation. The proposed GA-based approach delivers a robust
stope layout solution; however, the definition of the genetic input parameter will require further
research to simplify the current parameter selection procedure.
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
This work was supported by Natural Sciences and Engineering Research Council of Canada (NSERC) and Goldcorp
[Project number: 478371].
18 M. E. VILLALBA MATAMOROS AND M. KUMRAL
References
[1] C. Maschio and D.J. Schiozer, A new optimization framework using genetic algorithm and artificial neural network
to reduce uncertainties in petroleum reservoir models, Eng. Opt. 47 (2015), pp. 72–86.
[2] M. Kumral, Optimal location of a mine facility by genetic algorithms, Min. Technol. 113 (2004), pp. A83–A88.
[3] M. Ataee-Pour, Optimisation of stope limits using a heuristic approach, Min. Technol. 113 (2004), pp. 123–128.
[4] E. Topal and J. Sens, A new algorithm for stope boundary optimization, J. Coal Sci. Eng. 16(2) (2010), pp. 113–119.
[5] D.S. Sandanayake, E. Topal, and M.W.A. Asad, A heuristic approach to optimal design of an underground mine stope
layout, Appl. Soft Comput. 30 (2015), pp. 595–603.
[6] M.E. Villalba and M. Kumral, Heuristic stope layout optimization accounting for variable stope dimensions and
dilution management, Int. J. Min. Miner. Eng. 8 (2017), pp. 1–18.
[7] J. Manchuk and C.V. Deutsch, Optimizing stope designs and sequences in underground mines, SME Transactions
324 (2008), pp. 67–75.
[8] C. Alford and B. Hall, Stope optimisation tools for selection of optimum cut-off grade in underground mine design,
Project Evaluation Conference, Melbourne, 2009.
[9] G. Erdogan, M. Cigla, E. Topal, and M. Yavuz, Implementation and comparison of four stope boundary optimization
algorithms in an existing underground mine, Int. J. Min. Reclam. Environ. (2017), pp. 1–15.
[10] N. Grieco and R. Dimitrakopoulos, Managing grade risk in stope design optimisation: probabilistic mathematical
programming model and application in sublevel stoping, Min. Technol. 116 (2007), pp. 49–57.
[11] R.L.A. Verhoeff, Using genetic algorithms for underground stope design optimization in mining. A stochastic analysis,
Delft University of Technology, 2017.
[12] R. Dimitrakopoulos, L. Martinez, and S. Ramazan, A maximum upside/minimum downside approach to the
traditional optimization of open pit mine design, J. Min. Sci. 43 (2007), pp. 73–82.
[13] M.A. Cuba, J.B. Boisvert, and C.V. Deutsch, Simulated learning model for mineable reserves evaluation in surface
mining projects, SME Trans. 334 (2013), pp. 527–534.
[14] J.R. Birge and F. Louveaux, Introduction to Stochastic Programming, Springer Science & Business Media, New
York, NY, 1997.
[15] R. Dimitrakopoulos and S. Ramazan, Stochastic integer programming for optimising long term production schedules
of open pit mines: Methods, application and value of stochastic solutions, Min. Technol. 117 (2008), pp. 155–160.
[16] M.E. Villalba and R. Dimitrakopoulos, Stochastic short-term mine production schedule accounting for fleet allocation,
operational considerations and blending restrictions, Eur. J. Oper. Res. 255 (2016), pp. 911–921.
[17] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Maching Learning, Addison Wesley Longman,
New York, NY, 1989.
[18] M. Mitchell, An Introduction to Genetic Algorithms, Massachusetts Institute of Technology, Cambridge, MA, 1999.
[19] C. Reeves, Handbook of Metaheuristics, Kluwer Academic, New York, NY, 2003, pp. 55–82.
[20] C.V. Deutsch, Geostatistical Reservoir Modeling, Oxford University Press, New York, NY, 2002.
[21] M.E. Villalba and M. Kumral, Calibrating parameters of generic algorithms used for various mining-related
optimization problems, Nat. Resour. Res. (2018). (under review).