University
stringclasses 19
values | Text
stringlengths 458
20.7k
|
|---|---|
Colorado School of Mines | Chapter 4
SOLUTION METHODOLOGIES
Since the by formulation creates a more balanced node tree than the at formu
lation, we use it to explore all potential methods to expedite solution times. All the
methods we examine can also be applied to the at formulation, although additional
work is required to implement some of these methods due to differences in how the
by and at decision variables are defined.
Exact methods are an approach commonly used to solve mixed integer program
ming (MIP) problem instances. The branch-and-bound algorithm is the most common
exact method used to solve MIP problems. Using branch-and-bound to solve MIP
problem formulations guarantees an optimal solution if the algorithm is run to com
pletion. However, this may require a long time, so we often terminate the algorithm
and report the gap between the best integer solution found at the time of termination
and what may be theoretically obtainable. This gap is referred to as a mipgap.
We propose various short-cuts that maintain the exact nature of the branch-
and-bound algorithm, but significantly improve its solution time. These short-cuts
endeavor to limit the search space by eliminating certain variables or a priori setting
the values of other variables. Other exact methods involve the generation of valid
and useful cuts that make the problem formulation more tractable. Lastly, we use
Lagrangian relaxation techniques to solve the problem faster.
Eliminating those variables from the model whose values would necessarily as
sume a value of 0 or 1 in the optimal solution is one exact method. One way to
eliminate variables from the model is to establish an earliest start time for each block
b £ B (we call it ESb). ESb represents the earliest time period that block b can
38 |
Colorado School of Mines | be reached if the mining rate occurs at the upper production or processing bound
(whichever is tighter) without violating sequencing constraints. The earliest start
time allows us to eliminate from the model any variables that would mine block b
before its earliest start time (i.e., the values of these variables are set to 0).
Another way to eliminate variables from the model is to establish a latest start
time for each block beB (we call it LSb). LSb represents the latest time period that
block b must be mined if the mining rate occurs at the lower production or processing
bound (whichever is tighter) without violating sequencing constraints. The latest
start time allows us to pre-determine the values of any variable that would mine
block b after its latest start time (i.e., the values of these variables are set to 1).
We define our variables only in time periods between their earliest and latest start
times. This decreases the number of nodes that the branch-and-bound algorithm must
examine, thus speeding up solution times.
It is important to note that determining earliest and latest start times requires
that the bounds used in the calculations are fixed (i.e., not elasticized). Some prob
lem instances result in infeasibilities which can be resolved by relaxing some of the
constraints. The constraints that are relaxed are elasticized by allowing them to
be violated at a fixed penalty. If a bound (say, the minimum processing bound) is
elasticized, then it cannot be used in the late start calculation.
Exact methods also entail the creation of cuts, or constraints that involve pairs or
groups of variables. Cuts are constraints that are added to the formulation that may
force the linear relaxation of the problem to behave more like an integer program.
The goal is to ensure that none of the cuts eliminates any of the variables in the
optimal solution, but do provide a benefit to the algorithm. As such, we aim to
create cuts that are valid—so they do not remove any optimal integer solutions—and
(theoretically) useful—force decision variables to assume integer values in the linear
relaxation of the problem (i.e., strengthen z£p). Although many cuts may not be
theoretically useful in the strictest sense, they may still provide practically useful
39 |
Colorado School of Mines | information by either limiting the number of blocks that can be accessed by a certain
time period or requiring a certain number of blocks be accessed by a certain time
period. We exploit the structure of our problem to generate these cuts.
Lagrangian relaxation techniques attempt to transform the monolith problem
formulation into a more tractable formulation, solve this new formulation, and then
use that solution’s decision variable values in the monolith. Specifically, Lagrangian
relaxation moves certain complicating constraints to the objective function where
they are weighted with fixed multipliers to discourage violations. We then solve the
Lagrangian relaxation subproblem and use the optimal decision variable values in the
monolith (assuming they are feasible), attempting to bound the objective function
value of the monolith.
4.1 Numerical Example
To help explain the earliest start, latest start, cut generation, and Lagrangian
relaxation methodologies, we use the two-dimensional conceptual model developed in
Section 3.2.1. We present this basic two-dimensional mine again below (see Figure
4.1). With respect to this example, we make the following assumptions:
• The numbers in the blocks are merely used to identify the blocks
• Each block contains 10 tons of material (i.e., n*, = 10 tons)
• Each block contains 10 grams of valuable mineral (i.e., = 10 grams)
• The minimum and maximum processing bounds are 20 tons and 40 tons, re
spectively, and are constant across all time periods (i.e. , C = 20 and C = 40 V t)
• The minimum and maximum production bounds are 20 tons and 40 tons, re
spectively, and are constant across all time periods (i.e., E = 20 and E = 40 Wt)
40 |
Colorado School of Mines | • The fixed cutoff grade is 1 gram of mineral per 1 ton of material
( I
rib if 9b > cutoff grade \
i.e., rb= <
y 0 otherwise J
• The slope requirements are 45°
1 2 3 4 5 6 7
8 9 10 11 12 13 14
15 16 17 18 19 20 21
Figure 4.1. Two-Dimensional Numerical Example. This model provides a numerical
example for use in the various methodologies we propose to expedite solution times.
The first issue to resolve in this model is whether each block is ore or waste.
This is determined by applying the cutoff grade to each block. Since each block in
the model contains 10 grams of mineral, the average grade of each block is 1 gram of
mineral per ton of material. Based on our fixed cutoff grade of 1 gram of mineral per
1 ton of material, each block in this example is considered ore. This means that each
extracted block is sent to a mill.
The 45° slope requirements mean that the sequencing constraints for this example
behave the same way as those described in the two-dimensional conceptual model.
See Section 3.2.1 for more information on the sequencing constraints for this example.
Aside from the geospatial constraints, the mine must also adhere to a set of
operational constraints. These operational constraints set minimum and maximum
bounds on production and processing capacity as well as average grade requirements
during each time period. Because this example assumes a fixed cutoff grade, the
average grade requirements are rendered moot and do not need to be considered. For
simplicity, we assume that both production and processing have the same minimum
41 |
Colorado School of Mines | and maximum per time period bounds (20 tons and 40 tons, respectively); however,
in practice this is rarely the case.
The minimum operational bounds require that a certain amount of work is done
at the mine during each time period. The minimum production bound requires that 20
tons of material are removed from the mine each time period. Likewise, the minimum
processing bound requires that 20 tons of processable material (i.e., blocks that meet
the cutoff grade) are removed from the mine each time period. These minimum
bounds place an upper bound on the life of the mine and allow us to establish a
maximum time horizon for the mine’s operations. Since 20 tons of material and 20
tons of processable material must be removed from the mine each time period, and
due to the fact that the mine contains 210 tons of material (21 blocks each with 10
tons of material) and 210 tons of processable material (all 21 blocks meet the cutoff
grade, therefore each block is an ore block), the life of the mine cannot be longer
than 11 time periods (210 divided by 20 is 10.5, so activity at the mine ends during
the eleventh time period). We use this fact to reduce the number of variables in the
problem via the latest start time idea and to generate cuts.
In the same vein, the maximum operational bounds limit the amount of work
done per time period at the mine. With a maximum production capacity of 40 tons
of material and a maximum processing capacity of 40 tons of processable material,
we can determine a lower bound on the life of the mine (i.e., a minimum time horizon
for operations at the mine). Again, since the mine contains 210 tons of material and
210 tons of processable material, if operations are conducted at maximum production
and processing levels, the earliest time that activity at the mine can finish is by the
sixth time period (210 divided by 40 is 5.25, so activity ceases after the first quarter
of the sixth time period). As was the case with the minimum operational bounds,
we use the maximum operational bounds to reduce the number of variables in the
problem via an earliest start time. We also use these minimum operational bounds
to generate cuts.
42 |
Colorado School of Mines | 4.2 Earliest Starts and Latest Starts
We use the concept of earliest and latest starts to eliminate variable instances
from consideration and fix variable values, respectively. To determine these starts,
we exploit precedence between the blocks along with the minimum and maximum re
source bounds; the former bounds produce late starts while the latter bounds produce
early starts.
4.2.1 Earliest Starts Based on Maximum Production and Processing Bounds
We can determine an earliest start time for each block b € B , i.e., ESb, to reduce
the number of elements in the set TJ, from including the entire time horizon. This
earliest start algorithm eliminates those variables from the model whose values would
necessarily assume a value of 0 in the optimal solution. By using the sequencing con
straints and the upper bounds on production and processing capacity, we determine
the earliest possible time that block b can be reached if we were to mine as quickly
as possible. We can then eliminate any variables that represent mining block b be
fore its earliest possible start time. Assuming that the upper bounds of neither the
production capacity nor the processing capacity are elasticized in our formulation, we
determine an earliest start time based on the production capacity and also an earliest
start time based on the processing capacity for each block in the model. The overall
earliest start time for each block is the later of these two earliest start times. Thus,
the tightest earliest start time is established for each block in the model.
Kuchta, Newman, and Topal (2003) use an early start idea with their work at
LKAB’s Kiruna mine; however, they investigate an underground mine with signifi
cantly different sequencing and operational constraints. Their model does not explic
itly define maximum bounds on processing or production rates, but instead employs
the early start time idea using horizontal and vertical sequencing rules with respect to
operating adjacent machine placements (sites on which load haul dump units operate
43 |
Colorado School of Mines | to remove iron ore from the mine). Boland, Fricke, and Froyland (2006) present a gen
eral format for the earliest start method by combining block precedence constraints
with production constraints and then aggregating them over time for a particular
attribute (such as total ore in each block or amount of usable ore in each block).
Our method is similar to theirs; however, we use all applicable attributes (i.e., pro
duction and processing capacity limits) to independently determine an earliest start
based on each attribute. We then define each block’s overall earliest start as the
most constraining of the block’s independently derived earliest starts based on each
attribute.
The earliest start time reflects how long it takes to reach block b based on its
location in the pit, the maximum production capacity, and the maximum processing
capacity as defined in the problem formulation. Our algorithm, which computes an
earliest start time for every block in the pit, first calculates the support weight of each
block in the pit. This support weight represents the tons of material or the tons of
ore from all the blocks that must be mined (based on the sequencing constraints) in
order to mine the block in question. The support weight also explicitly includes the
tons of material or tons of ore for the given block. We actually calculate two types
of support weights, one with respect to tons of material and another with respect to
tons of processable material. This support weight with respect to tons of material
is then divided by the maximum production capacity and the support weight with
respect to tons of processable material is divided by maximum processing capacity
to arrive two earliest start times for each block in the pit. Since each block has two
earliest start times, the maximum of these two numbers (which represents the later
of the two earliest start times) is the earliest possible time period that the block in
question can be started.
44 |
Colorado School of Mines | Earliest Starts Algorithm
Assumptions We include all the assumptions that we describe with respect to our
model formulation (see Section 3.1).
Definitions
• B = Set of blocks which exists in the data set. Each block b £ B has the
following characteristics:
- An (x, y, z) location in three-space
- A total material content (in tons), %
- A mineral content (in grams),
* If the cutoff grade is met, then the block is considered ore and for that
block the ore weight (r&) is:
n - nb
* If the cutoff grade is not met, then the block is considered waste and
for that block the ore weight (rb) is:
r& = 0
- A precedence set - the set of blocks that must be removed from the pit
due to pit sloping requirements before block b can be accessed
• Sb = Block b and its precedence set (i.e., the set of blocks that block b supports)
• TotalSupportedOreb = Total amount of Ore (in tons) in the set Sb (i.e., block
b and all blocks above block b based on the precedence constraints)
• TotalSupportedMaterialb = Total amount of material (in tons) in the set Sb
(i.e., block b and all blocks above block b based on the precedence constraints)
45 |
Colorado School of Mines | EarlyStartOreb = T otal SupportedOreb + 1
max processing capacity
Early StartMaterialb = T otalSupportedM aterialb + 1
max production capacity
ESb — max (EarlyStartOreb, Early StartMaterialb)
end
output ESb for each block b E B
end
Earliest Starts Numerical Example The idea behind the early starts variable
elimination routine is best explained by examining Figure 4.2 below:
1
1 7
8 9 11 12 13 14
'
15 16 17 19 20 21
Figure 4.2. Earliest Starts Numerical Example. This example depicts the results of
using the earliest starts routine on block 18.
As stated in Section 4.1, the maximum production capacity is 40 tons per period and
each block contains 10 tons of material. This means that in order to reach block 18 in
the figure above, 80 tons of material have to first be removed based on the assumed
45° sloping requirements. Therefore block 18 cannot be removed until period 3 at
the earliest because removing 80 tons of material requires two complete time periods.
Hence, block 18 cannot be reached until time period 3. As such, block 18 has an
earliest start time of 3 based on the maximum production constraint. Because the
maximum processing capacity is also 40 tons per period and each block meets the
cutoff grade (so that each block contains 10 tons of processable material), the earliest
47 |
Colorado School of Mines | start time for block 18 based on maximum processing capacity is also 3. There is no
need to investigate the feasibility of block 18 being part of the optimal solution until
time period 3. Therefore it is not necessary to define decision variables corresponding
to whether or not to mine block 18 during time periods 1 and 2, so, we eliminate
these two variables from the problem formulation.
4.2.2 Latest Starts Based on Minimum Production and Processing Bounds
Similar to the concept of earliest start times, we can compute a latest start time
for each block b E B, i.e., LSb- Generating a latest start time for each block forces
the algorithm to set each block’s value to mined upon reaching its latest start time
by fixing its value to 1 in the optimal solution. Using the sequencing constraints
in conjunction with the lower bounds on production and processing capacity, we
determine the latest possible time that block b can be reached if we were to mine as
slowly as possible. We can set the value of any variable that would indicate mining
block b after its latest start time to 1 (i.e., mined). Assuming that the lower bounds
of neither the production requirement nor the processing requirement are elasticized
in our formulation, we determine a latest start time based on production requirements
and processing requirements for each block in the model. The overall latest start time
for each block is the earlier of these two latest start times. Therefore, the tightest
latest start time is established for each block in the model.
Kuchta, Newman, and Topal (2003) employ the latest start idea in their work
with LKAB’s Kiruna mine. However, their model does not explicitly define minimum
bounds on processing or production rates. Instead, they use horizontal and vertical
sequencing rules regarding adjacent machine placements along with information re
garding which machine placements are active. Boland, Fricke, and Froyland (2006)
do not present any methodology for generating latest starts. Their work only covers
earliest starts.
The latest start time reflects the most time that can pass before block b must
48 |
Colorado School of Mines | be mined based on its location in the pit, the minimum production requirement, and
the minimum processing requirement as defined in the problem formulation. Our
algorithm computes a latest start time for every block in the pit by determining each
block’s holding weight. This holding weight represents the tons of material or the tons
of ore that cannot be mined (based on the sequencing constraints) until the block in
question is mined. The holding weight also explicitly includes the tons of material or
tons of ore for the block in question. Just as with the earliest start times concept, we
actually calculate two types of holding weights, one with respect to tons of material
and another with respect to tons of processable material. The holding weight with
respect to tons of material is then divided by the minimum production requirement
while the holding weight with respect to tons of processable material is divided by
the minimum processing requirement to arrive at two latest start times for each block
in the pit. The overall latest start time for each block is the earlier of the two latest
start times calculated for each block.
Latest Starts Algorithm
Assumptions Just as with the earliest starts algorithm, we include all the assump
tions that we describe with respect to our model formulation (see Section ??).
Definitions
• B represents the set of blocks which exists in the data set. Each block b £ B
has the following characteristics:
— An (x, y, z) location in three-space
— A total material content (in tons),
— A mineral content (in grams), gi>
* If the cutoff grade is met, then the block is considered ore and for that
block the ore weight (r*,) is:
49 |
Colorado School of Mines | Tb — Tib
* If the cutoff grade is not met, then the block is considered waste and
for that block the ore weight (rj>) is:
n = o
— A holding set - the set of blocks that is being held up by block b (i.e., all
the blocks in B that cannot be removed from the pit due to pit sloping
requirements until block b is removed from the pit)
• Hb = Block b and its holding set (i.e., the set of blocks being held up by block
6 from being mined)
• TotalHeldUpOreb = Total amount of ore (in tons) in the set Hb (i.e., block b
and all blocks below block b based on the precedence constraints)
• TotalHeldUpMaterialb = Total amount of material (in tons) in the set Hb (i.e.,
block b and all blocks below block b based on the precedence constraints)
• TotalOrelnPit = Total amount of ore (in tons) in the entire pit (sum of all ore
blocks in B, i.e., sum of for all b € B)
• Total Material InPit = Total amount of material (in tons) in the entire pit (sum
of all material in B, i.e., sum of rib for all b € B)
• LateStartOreb = Latest start time of block b based on the minimum processing
constraint
• LateStartMaterialb = Latest start time of block b based on the minimum
production constraint
• LSb — Latest start time of block 6 based on the most constraining bound
(processing or production)
50 |
Colorado School of Mines | Latest Starts Numerical Example We now present a numerical example of the
latest starts variable elimination routine by examining Figure 4.3 below:
1 3 5 6 7
8 13 14
15 16 17 18 ,» 20 21
Figure 4.3. Latest Starts Numerical Example. This example depicts the results of
using the latest starts routine on block 4.
According to the figure above, block 4 is preventing blocks 10, 11, 12, 16, 17, 18,
19, and 20 from being mined. This only leaves blocks 1, 2, 3, 5, 6, 7, 8, 9, 13, 14,
15, and 21 to be mined before block 4 must be mined. As stated in Section 4.1,
each block contains 10 tons of material. The entire pit contains 210 tons of material
and block 4 is holding up production of 90 tons of material, so the leftover material
that can be mined is 120 tons of material (210 — 90 = 120). Based on the assumed
minimum production capacity of 20 tons per period, the latest that block 4 can be
started is time period 7 (120 divided by 20 is 6, so block 4 must start being mined
at the beginning of time period 7). Since the minimum processing capacity is also 20
tons per period and each block meets the cutoff grade (so that each block contains
10 tons of processable material), the latest start time for block 4 based on minimum
processing capacity is also 7. As such, block 4 has a latest start time of 7. The value
of the decision variables for block 4 during any time period including and after time
period 7 must be 1 (recall, 1 means that the block is considered mined). As a result
of the latest start routine, we can a priori set the values of block 4’s decision variables
after time period 6 in our model formulation to 1.
52 |
Colorado School of Mines | 4.3 Cut Generation Techniques
Generating cuts involves creating valid and useful inequalities that define pairs or
sets of blocks that cannot be mined together. Cuts are constraints that are added to
the formulation that may force the linear relaxation of the problem to behave more
like an integer program. These constructed cuts must not eliminate any optimal
integer solutions (i.e., they must be valid) and should strengthen the formulation by
forcing decision variables to assume integer values in the LP relaxation of the problem
or eliminating the optimal LP relaxation solution (i.e., they should be useful). We
exploit the structure of our problem to create cuts that are valid and useful inequalities
in the form of packing constraints (<) and covering constraints (>).
4.3.1 Cuts in General
All generated cuts must be valid and should be useful. By valid, we mean that
the cut cannot remove any feasible integer solutions. If a cut is not valid, then its
inclusion in the problem formulation may result in a sub-optimal solution. Regarding
usefulness, there is a difference between theoretical usefulness and practical usefulness.
Theoretically, a cut is considered useful if, among other things, it renders infeasible
the optimal solution to the current LP relaxation (Rardin 1998, p. 644). From a
practical standpoint, however, a cut that is not theoretically useful may still make
the model formulation more tractable. For instance, consider a cut that states that
at most one of two binary variables can assume a value of one (i.e., a cut in the
form of a + b < 1, where a and b both represent binary variables). Such a cut may
not be theoretically useful, but from a practical standpoint, it may be very useful
(depending, of course, on the other constraints in the model which may render the
cut redundant). If we know that one variable (say a) has a value of 1, then without
any further computation we also know that the other variable must have a value of 0
(i.e., b = 0). The practical usefulness of these cuts may help the branch-and-bound
53 |
Colorado School of Mines | process to solve our mixed integer programming problem more quickly.
We use a reasonable block selection rule to determine which blocks to investigate
for the generation of cuts. Although any pair or set of blocks can be combined to form
a cut, many cuts created in such an arbitrary manner are not useful (theoretically or
practically). As such, it is important to limit the number of blocks that are used to
create cuts. We wish to pick the best blocks to investigate for cut generation, thus
increasing the likelihood that the blocks create a valid and useful cut.
Our reasonable block selection rule eliminates from contention those blocks which
have little chance of creating a valid and useful cut. The rule examines individual
blocks while the cut generation algorithms involve multiple blocks. As such, our cut
generation procedures ensure that the cuts are valid and are at least practically useful,
while our reasonable block selection rule dictates which blocks the cut generation
procedures investigate. We employ the reasonable block selection rule as a means
of picking the best blocks for inclusion in cuts, realizing that this rule may actually
eliminate some blocks that could form a valid and useful cut. However, the only way
to create every valid and useful cut is to investigate every possible combination of
blocks; a task that is computationally too expensive.
To derive such a reasonable block selection rule, we borrow the support weight
and holding weight ideas explained in the earliest and latest start routines. These
weights allow us to intelligently select blocks for use in creating cuts that are valid
and have a good chance of being useful. We do this by investigating only those blocks
whose supporting weight (or holding weight) is within a certain percent of the block’s
next earliest start weight (or latest start weight). For earliest starts cuts, this rule is:
PercentClosegrade = wtnrp
ES MaxProc
*
'Ujtjnaterial
P ercentClosematerial
—
ES MaxProd
*
where wtare and wtmateriai represent the processable material weight and total material
54 |
Colorado School of Mines | weight of the block and all of its predecessors, respectively; ES is the earliest start for
the block; and MaxProc and MaxProd are the maximum processing and production
capacities per time period, respectively. For latest starts cuts, the rule is slightly
different:
PercentClosegrade = Total0reInPit _ ^ _ 1} * MinProc)
C material TotalMateriallnPit — ((LS — \) * MinProc)
where wt^-e and wtmateriai represent the processable material weight and total material
weight of the block and all of its holders, respectively (i.e., blocks that are being held
up, or prevented from being mined, due to the block in question); LS is the latest start
for the block; TotalOrelnPit and TotalMateriallnPit represent the total amount of
ore and total amount of material in the entire data set; and MinProc and MinProd are
the minimum processing and production requirements per time period, respectively.
We explain this rule with some examples.
By next earliest start weight we mean the amount of production or processing
capacity (whichever is smaller) required to push the block’s earliest start time to the
very next time period. For instance, in the numerical example we have been employ
ing, the maximum production capacity is 40 tons per period. Let us identify a block,
a, with a support weight of 10 tons. This block has an earliest start of 1 ( |_^J +1 = 1)
and is only 25% close to its next earliest start weight x 100% = 25%). As such,
block a would not be a good candidate for use in generating a cut based on our
reasonable block selection rule. Let us identify another block, b, with a support
weight of 35 tons. This block also has an earliest start of 1 (|_|§J + 1 = l), but it is
87.5% close to its next earliest start weight x 100% = 87.5%). Block b would
be a much better candidate for inclusion in a cut than block a. Note, however, that
a cut formed by combining blocks a and b actually would form a valid cut of the
form + w\j,i < 1 because, assuming they share no blocks between their respective
55 |
Colorado School of Mines | predecessor sets, together these two blocks cannot both be mined in time period 1
(because their combined weight is 45 tons and only 40 tons can be mined in time
period 1).
Similarly, by next latest start weight we mean the amount of production or pro
cessing requirement (whichever is larger) left to push the block’s latest start time to
the previous time period. Again, we consider our numerical example where minimum
production capacity is 20 tons per period. We assume our pit contains 210 tons
of material (consistent with the 21 blocks in our example, each weighing 10 tons).
Consider a block, call it a, with a holding weight of 12 tons. This block has a lat
est start of 10 ( [212q-12J +1 = 10) and is 40% close to its next latest start weight
('2io—((f e )*~25) x 100% = 40%^. Block a probably would not be a good candidate for
use in generating a cut. Let us look at another block, b, with a holding weight of 28.
This block also has a latest start of 10 ~] + 1 = 10) but it is 93.3% close to its
next latest start weight ^2ib-((io-i)*20) x 100% = 93.3%^. Picking among these two
blocks, block b would be the better candidate to include in a set of potential blocks
for latest start cut generation.
The percent close numbers here are to illustrate the procedure only. Ultimately,
the user defines the percentage above which a block passes the reasonable block se
lection rule. A higher percentage reduces the number of blocks included in the cut
generation procedure. Our reasonable block selection rule considers individual blocks,
but the generation of cuts involves two or more blocks. Investigating all two-way,
three-way, etc. combinations of blocks would quickly become computationally too
expensive. As such, we use our reasonable block selection rule as a proxy to select
the best individual blocks to include in the generation of valid and useful multi-block
cuts.
Despite limiting the number of blocks we investigate with our cut generation
algorithm, we still examine many combinations of blocks. Examining all of these
block combinations takes a long time, especially as the number of blocks in the cut
56 |
Colorado School of Mines | increases (i.e., there are fewer two-way combinations of a given set of blocks than
there are three-way combinations of the same set of blocks). The time saved by
employing these cuts in our model formulation might be lost due to the time spent
actually creating them. We want to ensure that the reduction in solution time in our
numerical results is not offset by the time required to create the cuts. Determining
the amount of time to spend generating cuts involves a degree of judgment and must
be balanced with the time it takes to solve the monolith. We explore this more in
our numerical results (Section 5.3.2).
Boland, Fricke, and Froyland (2006) present a method of generating cuts by
defining valid knapsack inequalities to serve as cover cuts. However, they only discuss
cuts of the form:
yZ^bt < |B| - 1
beB
where B is the set of blocks involved in the cut (i.e., B = {a, b} for the two-block
examples above using our reasonable block selection rule). We describe cuts of the
form:
Y lWbt ^ i^i -1
beB
also, which Boland, Fricke, and Froyland do not address. Additionally, they do not
employ a reasonable block selection rule to select blocks for cut generation, instead
attempting to generate cuts using all available blocks.
The right-hand-side of cuts involving more than two blocks can have values up
to one fewer than the number of blocks involved in the cut (i.e., |l, 2,..., |B| — 1J).
The cuts generated by Boland, Fricke, and Froyland, however, only permit a right-
hand-side that is exactly equal to the number of blocks involved in the cut minus one
(i.e., they only use a right-hand-side that equals ^|B| — ij). Our cuts are not limited
in this way. These added cuts further speed up solution times.
To generate cuts involving multiple blocks, we investigate super-blocks, which are
formed by combining the blocks in question and their respective precedence sets (or
57 |
Colorado School of Mines | holding sets, as the case may be). When creating these super-blocks, it is important
to use the union operator so that no blocks are double counted in the combined set.
If we consider blocks a and b, then the union of their precedence sets (i.e., 5a,b)
contains the blocks in a s precedence set (5a) and the blocks in b s precedence set
(Sb), without any shared blocks between the two sets counted more than once (i.e.,
Sa,b = Sa U Sb). We then use this super-block in our earliest starts algorithm (or
latest starts algorithm, as the case may be) to determine the earliest possible time
that both blocks a and b can be accessed together as a unit. The earliest starts
algorithm uses the super-block Sa;b instead of the precedence sets Sa and 5b in all
the calculations. The same idea holds for super-blocks formed by combining three or
more blocks and for super-blocks used to determine latest starts.
The cuts generation algorithms allows the user to create cuts based on either
production bounds, processing bounds, or both. As mentioned above, the user con
trols the percent close employed by the reasonable block selection rule. The user also
controls whether the algorithm generates theoretically useful cuts, practically useful
cuts, or both.
4.3.2 Two-Way Earliest Starts Cuts
A potentially valid and useful cut for our model involves allowing at most one of
two blocks to be mined by a particular time period due to the maximum production
and/or maximum processing constraint. Because we are allowing at most one of two
blocks to be mined, this cut takes the following form:
Wa,T-l + 1Vb,T-l 5 1
where a and b are arbitrarily chosen blocks that adhere to the reasonable block selec
tion rule (see Section 4.3.1). Generate the cut by comparing the earliest start time
for each individual block (E5a and A15b, respectively) with the earliest start time of
58 |
Colorado School of Mines | the super-block formed by the union of blocks a and b (ES&^, referred to as r). If r
is greater than both of the single block earliest start times (i.e., r > max(ES8L, ESb)),
then the super-block formed by the union of blocks a and b can only be accessed by
a time period later than the earliest start times for the individual blocks. Because of
this, access is limited to only one of these two blocks by time period t — 1, and an
appropriate cut of the form iCa.r-i + rcb,r-i < 1 can be generated.1
Determining if the Two-Way Earliest Starts Cuts are Valid and Useful
Only cuts that are valid and useful should be included in the model formulation. To
determine if our cuts of the form iya>r-i + ^b,r-i ^ 1 meet these criteria, we must
pay particular attention to the time index r — 1.
Based on the earliest start of the super-block formed by the union of blocks a
and b (ES^b, which we call r), we know that r is the earliest possible time period
that both blocks a and b can be mined together. This means that during any time
period before r, only one of these two blocks can be mined. So, it is valid to limit
access to at most one of the two blocks a and b by time period r — 1.
It may be practically useful to limit at most one of these two blocks a and
b to be accessed by time period r — 1, because if the value of one of the blocks
is known to be mined (i.e., say u^r-i = 1) then the value of the other block is
also known due to the cut (tVb/r-i must equal 0 or the constraint represented by
the cut is violated). To determine the theoretical usefulness of the cut, however,
we must empirically test each cut. We consider a cut theoretically useful if, among
other things, it renders infeasible the optimal solution to the LP relaxation of the
original integer programming formulation. For two-way earliest starts cuts of the
form wa,T_i + Wh,T-i < 1, the cut is useful if the sum of the values of the variables
wa,T-i and mb,T-i in the optimal LP relaxation (we call them û)a,r_i and Wb,T-i) is
1 Recall that our decision variables are defined as Wbt, where the b index identifies the particular
block and the t index identifies a time period by which the block is extracted.
59 |
Colorado School of Mines | greater than 1:
tüa,T—1 H- ^b,T—1 1
Two-Way Earliest Starts Cuts Algorithm
Assumptions We again use all assumptions that we describe with respect to our
model formulation (see Section 3.1).
Definitions
• a = A block (from the set of blocks B) which adheres to the reasonable block
selection rule
• b = Another block (from the set of blocks B and not the same as a) which
adheres to the reasonable block selection rule
• 5a,b — Set of blocks that must be mined (based on the sequencing constraints)
in order to mine blocks a and b (including explicitly mining blocks a and b).
This set contains blocks a and b and the union of all the blocks in each of their
respective precedence sets (i.e., 5a,b = 5a U 5t>, since S& contains block a and
all the blocks in block a s precedence set and 5b contains block b and all the
blocks in block b s precedence set). As a result, no shared blocks between the
precedence sets of blocks a and b are counted more than once in the super-block
represented by 5a,b-
• ES& = Earliest start time for block a (based on either the maximum processing
constraint or the maximum production constraint)
• E5b = Earliest start time for block b (based on either the maximum processing
constraint or the maximum production constraint)
60 |
Colorado School of Mines | • r = ESa,h — Earliest start time for the set of blocks contained in 5a,b (based
on either the maximum processing constraint or the maximum production con
straint)
Inputs
• A set of blocks B
• Maximum processing capacity per time period (in tons of ore) and maximum
production capacity per time period (in tons of material). Note that these
capacity constraints must be hard constraints (i.e., they cannot be elasticized).
Outputs
• Valid cuts of the form:
+ 1
Wqlt—I Wb,T-l —
Algorithm
For each two-way combination of blocks a E B and b E B in which each block
adheres to the reasonable block selection rule with respect to the maximum
production capacity do:
la. Determine the earliest start time for block a (i.e., ESa) based on the
maximum production capacity
2a. Determine the earliest start time for block b (i.e., BBy) based on the
maximum production capacity
3a. Create the set of blocks that represents the union of the precedence sets
for blocks a and b (i.e., 5a,b)
4a. Determine the earliest start time for the set of blocks contained in Sa,b
(i.e., t = B5a>b) based on the maximum production capacity
61 |
Colorado School of Mines | 5a. If t > max(ESa, ESb) then create a cut of the form:
^a,T—1 "t" ^b,r—1 ^ 1
For each two-way combination of blocks a G E and b E E in which each block
adheres to the reasonable block selection rule with respect to the maximum
processing capacity do:
lb. Determine the earliest start time for block a (i.e., ESa) based on the
maximum processing capacity
2b. Determine the earliest start time for block b (i.e., ESb) based on the
maximum processing capacity
3b. Create the set of blocks that represents the union of the precedence sets
for blocks a and b (i.e., Sa,b)
4b. Determine the earliest start time for the set of blocks contained in Sa,b
(i.e.,
t
= ESa,b) based on the maximum processing capacity
5b. If r > max(ESa, ESb) then create a cut of the form:
U?a,T—1 -f" ^b,T—1 — 1
Output all generated cuts
Relative Dominance of Two-Way Earliest Starts Cuts The time index (r) in
the cuts we generate provides information about the relative dominance of different
cuts. Take the following two potential cuts:
• wai,2 + U>b,2 S: 1
• Wa,4 + Wb,4 < 1
62 |
Colorado School of Mines | The first means that at most one of the two blocks a and b can mined by time
period 2, while the second means that at most one of these two blocks can be mined
by time period 4. In this case the latter cut dominates the former. The reason for
this dominance is analogous to the reason that an earliest start of 4 is a stronger
restriction than an earliest start of 2 for any given block. Our two-way earliest starts
cuts algorithm accounts for this dominance and only generates the dominant cut for
any given pair of blocks (assuming such a cut is valid and useful).
Two-Way Earliest Starts Cuts Numerical Example Looking at our two-
dimensional example, we use blocks 10 and 12 to create a two-way earliest starts
cut based on the maximum production capacity (see Figure 4.4).
. I
. , 1
1
m->m .î- ■>4 'ywy "ï*
7
5
A.
8 9 u 13 14
15 16 17 18 19 20 21
Figure 4.4. Two-Way Earliest Starts Cuts Numerical Example. This example depicts
the results of creating a two-way earliest starts cut with blocks 10 and 12.
For this example, we assume that:
• Block a is represented by block 10 in the figure and block b is represented by
block 12 in the figure
• Each block contains 10 tons of material (i.e., rib — 10 for each block a and b)
• The maximum production capacity is 40 tons per time period (for simplicity,
we only use production bounds for this example)
63 |
Colorado School of Mines | Recall that our decision variables are defined as wtt, where the b index identifies the
particular block (blocks 10 and 12 in the figure above for this example) and the t
index identifies the time period by which the block is extracted. We now use the
algorithm to generate a two-way earliest starts cut based on maximum production
capacity:
1. Determine the earliest start time for block a:
ESa = l
2. Determine the earliest start time for block b:
ESb —1
3. Create the set of blocks that represents the union of the precedence sets for
blocks a and b:
&,b = & U Sb = {2,3,4,10} U {4,5,6,12} = {2,3,4,5,6,10,12}
4. Determine the earliest start time for the set of blocks contained in Sa,b-
t = ESa,b = 2
5. Since r > max(ESa, ESb) we can create a cut of the form:
'Wa,T-l + Wb,T-l <1 => Wio.1 + Wi2,i < 1
This means that by the end of time period 1, at most one of the two blocks represented
by block 10 and block 12 in the figure above can be mined.
4.3.3 Two-Way Latest Starts Cuts
Building on the two-way earliest starts cuts idea, we now present the latest starts
version of that cut. This potentially valid and useful cut involves forcing at least one
of two blocks to be mined by a particular time period due to the minimum production
64 |
Colorado School of Mines | and/or minimum processing constraint. Since we are forcing at least one of two blocks
to be mined, this cut takes the following form:
^a,f d" Wb f ^ 1
where a and b are arbitrarily chosen blocks that adhere to the reasonable block se
lection rule (see Section 4.3.1). The cut is generated by comparing the latest start
time for each individual block (LSa and LSb, respectively) with the latest start time
of the super-block formed by the union of blocks a and b (LSa)b, referred to as f). If
f is less than both of the single block earliest start times (i.e., f < min(LSa, LSb)),
then the super-block formed by the union of blocks a and b must be accessed by a
time period earlier than the latest start times for the individual blocks. Because of
this, one of these two blocks must be extracted by time period f , and an appropriate
cut of the form wa,f + Wb,f > 1 can be generated.2
Determining if the Two-Way Latest Starts Cuts are Valid and Useful As
with the earliest starts cuts, only valid and useful cuts should be included in the
model formulation. To determine if our cuts of the form w^r + tUb,f > 1 meet these
criteria, we again pay particular attention to the time index, which is f in this case.
Based on the latest start of the super-block formed by the union of blocks a and
b (LSaib, which we call f), we know that f is the latest possible time period that
both blocks a and b must be mined because as a unit they are holding up access to
the remaining blocks in the pit. By time period f, therefore, at least one of these two
blocks must be removed from the pit. Even if there exists a single block among a
and b that does not need to be extracted (on its own) until a time period later than
f, we still need to remove at least one of the two blocks a or b by time period f to
meet the minimum production and/or minimum processing constraints.
2Recall that our decision variables are defined as Wbt, where the b index identifies the particular
block and the t index identifies a time period by which the block is extracted.
65 |
Colorado School of Mines | It may be practically useful to require at least one of these two blocks a and b
be mined by time period f, because if the value of one of the blocks is determined to
be not mined (i.e., say u/a,f = 0) then the value of the other block is also known due
to the cut (tUb,f must equal 1 or the constraint represented by the cut is violated).
Just as with the earliest starts cuts, to determine the theoretical usefulness of the cut,
we must empirically test each cut. We consider a cut theoretically useful if, among
other things, it renders infeasible the optimal solution to the LP relaxation of the
original integer programming formulation. For two-way latest starts cuts of the form
tna,f + Wb,f > 1, the cut is useful if the sum of the values of the variables w^f and
Wb,f in the optimal LP relaxation (we call them %)a,f and Wb,f) is less than 1:
U^a,f 4" U7b,f ^ 1
Two-Way Latest Starts Cuts Algorithm
Assumptions We again use all assumptions that we describe with respect to our
model formulation (see Section 3.1).
Definitions
• a = A block (from the set of blocks B) which adheres to the reasonable block
selection rule
• b = Another block (from the set of blocks B and not the same as a) which
adheres to the reasonable block selection rule
• Lfa b = Set of blocks that cannot be mined (based on the sequencing constraints)
until blocks a and b are mined (including explicitly mining blocks a and b).
This set contains blocks a and b and the union of all the blocks in each of their
respective holding sets (i.e., = iLa U LTb, since Ha contains block a and all
the blocks in block a’s holding set and H\y contains block b and all the blocks
66 |
Colorado School of Mines | in block b’s holding set). As a result, no shared blocks between the holding sets
of blocks a and b are counted more than once in the super-block represented
by tfa,b.
• LSin = Latest start time for block a (based on either the minimum processing
constraint or the minimum production constraint)
• LSb = Latest start time for block b (based on either the minimum processing
constraint or the minimum production constraint)
• f — LS^b — Latest start time for the set of blocks contained in # a,b (based
on either the minimum processing constraint or the minimum production con
straint)
Inputs
• A set of blocks B
• Minimum processing requirement per time period (in tons of ore) and mini
mum production requirement per time period (in tons of material). Note that
these requirement constraints must be hard constraints (i.e., they cannot be
elasticized).
Outputs
• Valid cuts of the form:
îna,f 'Wb,T ^ 1
Algorithm
For each two-way combination of blocks a £ B and b € B in which each block
adheres to the reasonable block selection rule with respect to the minimum
production requirement do:
67 |
Colorado School of Mines | la. Determine the latest start time for block a (i.e., LSa) based on the mini
mum production requirement
2a. Determine the latest start time for block b (i.e., LSb) based on the mini
mum production requirement
3a. Create the set of blocks that represents the union of the holding sets for
blocks a and b (i.e., Lra,b)
4a. Determine the latest start time for the set of blocks contained in i7a,b (i.e.,
f = LSa,b) based on the minimum production requirement
5a. If f < min(LSa, LSb) then create a cut of the form:
îüa,f VJh,î 2^ 1
For each two-way combination of blocks a € B and b E B in which each block
adheres to the reasonable block selection rule with respect to the minimum
processing requirement do:
lb. Determine the latest start time for block a (i.e., LSa) based on the mini
mum processing requirement
2b. Determine the latest start time for block b (i.e., LSb) based on the mini
mum processing requirement
3b. Create the set of blocks that represents the union of the holding sets for
blocks a and b (i.e., Ba,b)
4b. Determine the latest start time for the set of blocks contained in Ba b (i.e.,
f = LSa,b) based on the minimum processing requirement
5b. If f < min(LSa, LSb) then create a cut of the form:
îi^bjf — 1
68 |
Colorado School of Mines | Output all generated cuts
Relative Dominance of Two-Way Latest Starts Cuts As with the two-way
earliest start cuts, the time index (f) in the cuts we generate provides information
about the relative dominance of different cuts. Take the following two potential cuts:
• Wa,2 + Wb,2 > 1
• Wa,4 + Wb,4 > 1
The first means that at least one of the two blocks a and b must mined by time period
2, while the second means that at least one of these two blocks must be mined by
time period 4. In this case, the former cut dominates the latter. The reason for this
dominance is analogous to the reason that a latest start of 2 is a stronger restriction
than a latest start of 4 for any given block. Our two-way latest starts cuts algorithm
accounts for this dominance and only generates the dominant cut for any given pair
of blocks (assuming such a cut is valid and useful).
Two-Way Latest Starts Cuts Numerical Example Looking at our two-dimensional
example, we use blocks 10 and 12 to create a two-way latest starts cut based on the
minimum production requirement (see Figure 4.5).
1 2 3 4 5 6 7
8 9 11 13 14
15 16 17 21
1
Figure 4.5. Two-Way Latest Starts Cuts Numerical Example. This example depicts
the results of creating a two-way latest start cut with blocks 10 and 12.
69 |
Colorado School of Mines | For this example, we assume that:
• Block a is represented by block 10 in the figure and block b is represented by
block 12 in the figure
• Each block contains 10 tons of material (i.e., = 10 for each block a and b)
• The minimum production requirement is 20 tons per time period (for simplicity,
we only use production bounds for this example)
Recall that our decision variables are defined as where the b index identifies the
particular block (blocks 10 and 12 in the figure above for this example) and the t
index identifies the time period by which the block is extracted. We now use the
algorithm to generate a two-way latest starts cut based on the minimum production
requirement:
1. Determine the latest start time for block a:
LSa = 9
2. Determine the latest start time for block b:
LSb — 9
3. Create the set of blocks that represents the union of the holding sets for blocks
a and b:
#a,b = a u Ffb = {10,16,17,18} U {12,18,19,20} = {10,12,16,17,18,19,20}
4. Determine the latest start time for the set of blocks contained in
f = LSa,b = 8
5. Since f < min(LSa, LSb) we can create a cut of the form:
'Wa.T + Wb,f >1 ^10,8 + ^12,8 1
This means that by the end of time period 8, at least one of the two blocks represented
by block 10 and block 12 in the figure above must be mined.
70 |
Colorado School of Mines | 4.3.4 Three-Way Earliest Starts Cuts
Generating cuts with three blocks is significantly more complicated than gener
ating cuts with just two blocks. With three blocks there are many more combinations
to investigate. Also, the right-hand-side of the constraint can assume two different
values, either 1 or 2. Therefore, the resultant cuts can take either of the following
two forms:
+ tr’b,?'-! + tWc/f'-l or + tUb,f-l + Wc,f-1 5: 2
rya,T-l
where a, b, and c are arbitrarily chosen blocks that adhere to the reasonable block
selection rule (see Section 4.3.1). The first cut allows at most one of the three blocks
to be mined by a particular time period, while the second cut allows at most two of
three blocks to be mined by a particular time period. As with two-way earliest starts
cuts, we employ the maximum production and/or maximum processing constraints
and the support weights of various blocks to construct our cuts.
Let us assume that we can access all three blocks (a, b, and c) by time period
f (i.e., ESa,b,c = t). Additionally, let us assume that the earliest earliest start time
for all the two-way combinations is f' (i.e., f' =min (ES^y,, ES^c, ESb,c))- First, we
need to determine how many blocks are accessible before f. If the earliest earliest
start time for all two-way combinations of the three blocks is less than f (i.e., fz < f),
then by time period (f — 1) at most two of the three blocks are accessible and it is
valid to limit access to at most two of these three blocks. Next, we need to determine
how many blocks are accessible before f'. If the earliest single block earliest start
time for all three blocks is less than f' (i.e., min (ES&, ES\>, ESC) < f'), then by time
period (f7 — 1) at most one of the three blocks is accessible and it is valid to limit
access to at most one of these three blocks.
More specifically, we use the earliest starts algorithm to determine all single
block, pair-wise, and three-way block combination earliest starts (ES) for our three
71 |
Colorado School of Mines | blocks and then define the following:
• f" — min {ES&, ESb, ESC)
• f' = min (E5a,b, ESa^ ESh,c)
• f = -E5a,b,c
We then use these three values (f, f', and f") to generate our cuts.
If f is greater than f', then the blocks that comprise the super-block formed by
the union of blocks a, b, and c can only be accessed by a time period later than the
earliest start times for any of the super-blocks formed by two-way combinations of
blocks a, b, and c. As a result, access is limited to only two of the these three blocks
by time period f — 1 and an appropriate cut of the form ma)f-i + Wb,?-i + wC)f-i < 2
can be generated.
If f' is greater than f", then any two blocks that comprise the super-block formed
by the union of blocks a, b, and c can only be accessed by a time period later than
the earliest start times for any of the blocks a, b, and c individually. As a result,
access is limited to only one of the these three blocks by time period f' — l and an
appropriate cut of the form wa^/_i + Wb,f'-i + wc,f'-i < 1 can be generated.
Determining if the Three-Way Earliest Starts Cuts are Valid and Useful
As with two-way earliest starts cuts, only those three-way cuts that are valid and
useful should be included in the model formulation. We must ensure that both types
of cuts we generate (< 2 and < 1) are valid and useful.
To determine if our cuts of the form u;a)f-i + ^b,f-i + tnc,f-i < 2 meet these
criteria, we must pay particular attention to the time index f — 1. Based on the
earliest start of the super-block formed by the union of blocks a, b, and c (£5a,b,c,
which we call f), we know that f is the earliest possible time period that all three of
these blocks can be mined together. This means that during any time period before
72 |
Colorado School of Mines | f , at most two of these three blocks can be mined. So it is valid to limit access to at
most two of these three blocks by time period f — 1.
It may be practically useful to limit at most two of these three blocks a, b, and
c to be accessed by time period f — 1, because if the values of two of the blocks are
known to be mined (i.e., say u^f-i = iUb,f-i = 1) then the value of the other block
is also known due to the cut (mC)f-i must equal 0 or the constraint represented by
the cut is violated). To determine the theoretical usefulness of the cut, however, we
must empirically test each cut with specific data. We consider a cut theoretically
useful if, among other things, it renders infeasible the optimal solution to the LP
relaxation of the original integer programming formulation. For three-way earliest
starts cuts of the form wa,f-i + ^b,r-i + wc>f-i < 2, the cut is useful if the sum of
the values of the variables Wb,f-i, and wc,f-i in the optimal LP relaxation
(we call them ma,f-i, Wb,f-u and is greater than 2:
^a,f-l + Wb,f-1 + WC)f—1 > 2
To determine if our cuts of the form + tCb,fz-i + wc,t'-i < 1 are valid and
useful, we must pay particular attention to the time index f' — l. Based on the earliest
earliest start of the super-block formed by the union of any two of the blocks a, b,
and c (min (lLFa)b, ÆS'a.c, £Sb,c)5 which we call f'), we know that f' is the earliest
possible time period that any two of the three blocks can be mined together. This
means that during any time period before f', at most one of these three blocks can
be mined. So it is valid to limit access to at most one of these three blocks by time
period f' — l.
It may be practically useful to limit at most one of these three blocks a, b, and
c to be accessed by time period f' — l, because if the value of one of the blocks is
known to be mined (i.e., say — 1) then the values of the other two blocks
are also known due to the cut = Wc,f-i = 0 or the constraint represented
73 |
Colorado School of Mines | by the cut is violated). To determine the theoretical usefulness of the cut, how
ever, we must again resort to empirical tests. We consider a cut theoretically useful
if, among other things, it renders infeasible the optimal solution to the LP relax
ation of the original integer programming formulation. For three-way earliest starts
cuts of the form + Wb.r'-i + Wc,r-i < 1, the cut is useful if the sum of
the values of the variables and wc^ -\ in the optimal LP relaxation
(we call them rDa,f'-i, û)b,f'-i, andû^f'-i) is greater than 1:
tha,fz-l + ?ûb,r'-l + Û)c,t'-1 > 1
It is interesting to note that if f' ^ fz/ (which implies that fz = f” because fz < f"
is impossible), then although the cut is valid (no optimal answers are precluded
from being examined), it is not useful (practically or theoretically). The reason it
is not useful is because fzz tells us the earliest start time that any single block can
be accessed, so if fz = fzz, then by time fz — 1 none of the single blocks will be
accessible due to their single block earliest start times. Essentially, this cut just tells
us something we already know because of each individual block’s earliest start time.
Three-Way Earliest Starts Cuts Algorithm
Assumptions We again use all assumptions that we describe with respect to our
model formulation (see Section 3.1).
Definitions
• a = A block (from the set of blocks B) which adheres to the reasonable block
selection rule
• b = Another block (from the set of blocks B and not the same as a) which
adheres to the reasonable block selection rule
74 |
Colorado School of Mines | • c = Another block (from the set of blocks B and not the same as a or b) which
adheres to the reasonable block selection rule
• Sa,b = Set of blocks that must be mined (based on the sequencing constraints)
in order to mine blocks a and b (including explicitly mining blocks a and b).
This set contains blocks a and b and the union of all the blocks in each of their
respective precedence sets (i.e., S'a.b = Sa U St>, since Sa contains block a and
all the blocks in block a s precedence set and 5b contains block b and all the
blocks in block b’s precedence set). As a result, no shared blocks between the
precedence sets of blocks a and b are counted more than once in the super-block
represented by 5a>b.
• 5a,c = Set of blocks that must be mined (based on the sequencing constraints)
in order to mine blocks a and c (including explicitly mining blocks a and c).
This set contains blocks a and c and the union of all the blocks in each of their
respective precedence sets (i.e., 5a,c = 5a U 5C, since 5a contains block a and
all the blocks in block a’s precedence set and 5C contains block c and all the
blocks in block c’s precedence set). As a result, no shared blocks between the
precedence sets of blocks a and c are counted more than once in the super-block
represented by 5a,c-
• 5b,c = Set of blocks that must be mined (based on the sequencing constraints)
in order to mine blocks b and c (including explicitly mining blocks b and c).
This set contains blocks b and c and the union of all the blocks in each of their
respective precedence sets (i.e., 5b,c = 5b U 5C, since 5b contains block b and
all the blocks in block b’s precedence set and 5C contains block c and all the
blocks in block c’s precedence set). As a result, no shared blocks between the
precedence sets of blocks b and c are counted more than once in the super-block
represented by 5b,c-
75 |
Colorado School of Mines | • S^b.c — Set of blocks that must be mined (based on the sequencing constraints)
in order to mine blocks a, b, and c (including explicitly mining blocks a, b
and c). This set contains blocks a, b, and c and the union of all the blocks
in each of their respective precedence sets (i.e., 5ajb,c = 5a U 5b U 5C, since
5a contains block a and all the blocks in block a s precedence set, 5b contains
block b and all the blocks in block b’s precedence set, and 5C contains block c
and all the blocks in block c’s precedence set). As a result, no shared blocks
between the precedence sets of blocks a, b, and c are counted more than once
in the super-block represented by 5a,b,c-
• ESa — Earliest start time for block a (based on either the maximum processing
constraint or the maximum production constraint)
• ESb — Earliest start time for block b (based on either the maximum processing
constraint or the maximum production constraint)
• ESC — Earliest start time for block c (based on either the maximum processing
constraint or the maximum production constraint)
• f" = min (E5a, ESb, ESC)
• E5a,b = Earliest start time for the set of blocks contained in 5a,b (based on
either the maximum processing constraint or the maximum production con
straint)
• E5a,c = Earliest start time for the set of blocks contained in 5a)C (based on either
the maximum processing constraint or the maximum production constraint)
• E5b,c — Earliest start time for the set of blocks contained in 5b,c (based on either
the maximum processing constraint or the maximum production constraint)
• f' = min (E5a,b, E5a,c, £5b,c)
76 |
Colorado School of Mines | 3a. Determine the earliest start time for block c (i.e., ESC) based on the max
imum production capacity
4a. Create the set of blocks that represents the union of the precedence sets
for blocks a and b (i.e., S^b) and determine the earliest start time for this
set (i.e., ESa,b) based on the maximum production capacity
5a. Create the set of blocks that represents the union of the precedence sets
for blocks a and c (i.e., £a)C) and determine the earliest start time for this
set (i.e., ES^C) based on the maximum production capacity
6a. Create the set of blocks that represents the union of the precedence sets
for blocks b and c (i.e., Sb,c) and determine the earliest start time for this
set (i.e., E'S'b.c) based on the maximum production capacity
7a. Create the set of blocks that represents the union of the precedence sets
for blocks a, b, and c (i.e., £a,b,c) and determine the earliest start time for
this set (i.e., f = ESayb,c) based on the maximum production capacity
8a. Determine the earliest that any two-block set can be accessed:
f' = min (SS'a.b, ES&yC, EShyC)
9a. Determine the earliest that any single block can be accessed:
f" = mm(ESai,EShlESc)
10a. If f > f ' then create a cut of the form:
Wa,f-1 + tUb.f-l + U>c,f-1 2
78 |
Colorado School of Mines | lia. If rz > t" then create a cut of the form:
^a,f'—1 + U7b,f'-1 + ^c.r'-l — 1
For each three-way combination of blocks a E B, b € B, and c E B in which
each block adheres to the reasonable block selection rule with respect to the
maximum processing capacity do:
lb. Determine the earliest start time for block a (i.e., B5a) based on the
maximum processing capacity
2b. Determine the earliest start time for block b (i.e., BSt>) based on the
maximum processing capacity
3b. Determine the earliest start time for block c (i.e., ESC) based on the max
imum processing capacity
4b. Create the set of blocks that represents the union of the precedence sets
for blocks a and b (i.e., Sa,b) and determine the earliest start time for this
set (i.e., BBa,b) based on the maximum processing capacity
5b. Create the set of blocks that represents the union of the precedence sets
for blocks a and c (i.e., 5a)C) and determine the earliest start time for this
set (i.e., ES^c) based on the maximum processing capacity
6b. Create the set of blocks that represents the union of the precedence sets
for blocks b and c (i.e., Bb,c) and determine the earliest start time for this
set (i.e., BSb,c) based on the maximum processing capacity
7b. Create the set of blocks that represents the union of the precedence sets
for blocks a, b, and c (i.e., Ba,b,c) and determine the earliest start time for
this set (i.e., f = BBa,b,c) based on the maximum processing capacity
79 |
Colorado School of Mines | 8b. Determine the earliest that any two-block set can be accessed:
f' = min (ES’a.b, ESa,c: ESh,c)
9b. Determine the earliest that any single block can be accessed:
T" = min (ESa,ESh,ESc)
10b. If f > f ' then create a cut of the form:
wa,f-l + Mb.f-l + WCif-i < 2
11b. If f' > f" then create a cut of the form:
^a,f'—l + lüb.f'-l + WC}f'-l ^ 1
Output all generated cuts
Relative Dominance of Three-Way Earliest Starts Cuts As with two-way
cuts, the time index (f or f') in our generated cuts tells us about the relative domi
nance of different cuts. Again, cuts with a later time index dominate cuts (involving
the same blocks) with an earlier time index.
Three-way cuts concern themselves with another dominance issue though; the
value to the right of the inequality. Take the following two potential cuts:
• U>a,2 + Mb,2 + Wc,2 < 1
• U>a,2 ^b,2 + WC)2 £ 2
The first means that at most one of the three blocks a, b, and c can be mined by time
period 2, while the second means that at most two of these three blocks can be mined
80 |
Colorado School of Mines | by time period 2. In this case, the former cut dominates the latter. The reason for
this is that the former is more restrictive than the latter. The former restricts access
to only one block, while the latter allows access to any two of the blocks. Our three-
way earliest starts cuts algorithm accounts for this dominance and only generates the
dominant cut for any set of blocks in the same time period (assuming such a cut is
valid and useful).
Three-Way Earliest Starts Cuts Numerical Example Looking at our two-
dimensional example, we use blocks 9, 11, and 13 to create a three-way earliest starts
cut based on the maximum production capacity (see Figure 4.6).
?
14
20 21
Figure 4.6. Three-Way Earliest Starts Cuts Numerical Example. This example de
picts the results of creating a three-way earliest start cut with blocks 9, 11, and
13.
For this example, we assume that:
• Block a is represented by block 9 in the figure, block b is represented by block
11 in the figure, and block c is represented by block 13 in the figure
• Each block contains 10 tons of material (i.e., rib = 10 for each block a, b, and
c)
• The maximum production capacity is 40 tons per time period (for simplicity,
we only use production bounds for this example)
81 |
Colorado School of Mines | Recall that our decision variables are defined as where the b index identifies
the particular block (blocks 9, 11, and 13 in the figure above for this example) and
the t index identifies the time period by which the block is extracted. We now use
the algorithm to generate a three-way earliest starts cut based on the maximum
production capacity:
1. Determine the earliest start time for block a:
ESa = 1
2. Determine the earliest start time for block b:
ESh = 1
3. Determine the earliest start time for block c:
ESC = 1
4. Create the set of blocks that represents the union of the precedence sets for
blocks a and b and determine this set’s earliest start time:
Sa,b = 5a U 5b = {1,2,3,9} U {3,4,5,11} = {1,2,3,4,5,9,11}
ES^b = 2
5. Create the set of blocks that represents the union of the precedence sets for
blocks a and c and determine this set’s earliest start time:
&,C = 5a u 5c = {1,2,3,9} U {5,6,7,13} = {1,2,3,5,6,7,9,13}
ES&jC = 2
6. Create the set of blocks that represents the union of the precedence sets for
blocks b and c and determine this set’s earliest start time:
5b,c = 5b U 5c = {3,4,5,11} U {5,6,7,13} = {3,4,5,6,7,11,13}
#5b,c = 2
82 |
Colorado School of Mines | where a, b, and c are arbitrarily chosen blocks that adhere to the reasonable block
selection rule (see Section 4.3.1). The first cut requires that at least one of three
blocks be mined by a particular time period, while the second cut requires that at
least two of three blocks be mined by a particular time period. As with two-way
latest starts cuts, we employ the minimum production and/or minimum processing
constraints and the holding weights of various blocks to construct our cuts.
Let us assume that we must mine all three blocks (a, b, and c) by time period f
(i.e., LSa,b,c — t). Additionally, let us assume that the latest latest start time for all
the two-way combinations is f' (i.e., f' = max(LSa>b, L5a,c, LSh,c))- First, we need
to determine how many blocks must be mined after f. If the latest latest start time
for all two-way combinations of the three blocks is later than f (i.e., f' > f), then
by time period f at least one of the three blocks must be mined and it is valid to
force at least one of the decision variables representing these three blocks to assume
a value of 1 (i.e., mined). Next, we need to determine how many blocks must start
to be mined after f'. If the latest single block latest start time for all three blocks is
later than f' (i.e., max (L5a, LSb, LSC) > f7), then by time period f7 at least two of
the three blocks must be mined and it is valid to force at least two of the decision
variables representing these three blocks to assume a value of 1 (i.e., mined).
More specifically, we use the latest starts algorithm to determine all single block,
pair-wise, and three-way block combination latest starts (LS) for our three blocks
and then define the following:
• f" = max (LS^ LSb, LSC)
• f7 = max (L5a,b, LSa,c, LSh,c)
• T = LSa,b,c
We then use these three values (f, f7, and f77) to generate our cuts.
If f is less than f7, then the blocks that comprise the super-block formed by the
union of blocks a, b, and c must be accessed by a time period earlier than the latest
84 |
Colorado School of Mines | start times for any of the super-blocks formed by two-way combinations of blocks a,
b, and c. As a result, at least one of these three blocks must be removed by time
period f and an appropriate cut of the form w^f + Wb,f + wc,f ^ 1 can be generated.
If f' is less than f", then any two blocks that comprise the super-block formed
by the union of blocks a, b, and c must be accessed by a time period earlier than the
latest start times for any of the blocks a, b, and c individually. As a result, at least
two of these three blocks must be removed by time period f' and an appropriate cut
of the form w^f' + Wb,f' + wc^ > 2 can be generated.
Determining if the Three-Way Latest Starts Cuts are Valid and Useful As
with two-way latest starts cuts, only those three-way cuts that are valid and useful
should be included in the model formulation. We must ensure that both types of cuts
we generate (> 2 and > 1) are valid and useful.
To determine if our cuts of the form + u/b,f + ttfc.f > 1 meet these criteria,
we must pay particular attention to the time index f. Based on the latest start of the
super-block formed by the union of blocks a, b, and c (L5a,b,c) which we call f), we
know that by time period f all three blocks a, b, and c are holding up access to the
remaining blocks in the pit. Even if there exists a two-way combination of blocks a,
b, and c that does not need to be accessed until a later time period (i.e., its two-way
latest start is later than f), then we still need to remove at least one block during
time period f to meet the minimum production and/or processing requirements. This
means that by time period f, at least one of these three blocks must be mined. So it
is valid to force at least one of these three blocks to be mined by time period f.
It may be practically useful to force at least one of these three blocks a, b, and c
to be accessed by time period f , because if the values of two of the blocks are known
to be not mined (i.e., say rua,f = Wb,f = 0) then the value of the other block is also
known due to the cut (wc,f must equal 1 or the constraint represented by the cut
is violated). To determine the theoretical usefulness of the cut, however, we must
85 |
Colorado School of Mines | empirically test each cut with specific data. We consider a cut theoretically useful if,
among other things, it renders infeasible the optimal solution to the LP relaxation of
the original integer programming formulation. For three-way latest starts cuts of the
form + iUb,f + u>c,f > 1, the cut is useful if the sum of the values of the variables
tUa.f, wb,Ti and Wc,f in the optimal LP relaxation (we call them w^f, û)b,f, and wc,f)
is less than 1:
îÛa,-? d~ vjb,f "t" îÛc,f ^ 1
To determine if our cuts of the form w^f1 + ^b,f' + > 2 are valid and useful,
we must pay particular attention to the time index f'. Based on the latest latest
start of the super-block formed by the union of any two of the blocks a, b, and c
(max (LSa,b, LSa,c, LSb,c), which we call fz), we know that by time period f' at least
two of the three blocks a, b, and c are holding up access to the remaining blocks
in the pit. Even if there exists a block among a, b, and c that does not need to
be accessed until a later time period (i.e., its latest start is later than f'), then we
still need to remove at least two blocks during time period f' to meet the minimum
production and/or processing requirements. This means that by time period f', at
least two of these three blocks a, b, and c must be mined. So it is valid to force at
least two of these three blocks to be mined by time period fz.
It may be practically useful to force at least two of these three blocks a, b,
and c to be accessed by time period fz, because if the value of one of the blocks is
known to be not mined (i.e., say = 0) then the values of the other two blocks
are also known due to the cut (wb,f = wc^> = 1 or the constraint represented by
the cut is violated). To determine the theoretical usefulness of the cut, however, we
must again resort to empirical tests. We consider a cut theoretically useful if, among
other things, it renders infeasible the optimal solution to the LP relaxation of the
original integer programming formulation. For three-way latest starts cuts of the form
wa,f' + Wb,f' + wc,f' > 2, the cut is useful if the sum of the values of the variables
86 |
Colorado School of Mines | Wbtf>, and tüc.f' in the optimal LP relaxation (we call them wa^/, w^f', and wc^) is
less than 2:
za^a,T/ d" U^h,f' d" U^c,f' ^ 2
It is interesting to note that if tz ^ fz/ (which implies that fz = fzz because fz > f/z
is impossible), then although the cut is valid (no optimal answers are precluded from
being examined), it is not useful (practically or theoretically). The reason it is not
useful is because f" tells us the latest start time that any single block must be accessed,
so if fz = tzz, then by time fz all of the single blocks must be mined due to their single
block latest start times. Essentially, this cut just tells us something we already know
because of each individual block’s latest start time.
Three-Way Latest Starts Cuts Algorithm
Assumptions We again use all assumptions that we describe with respect to our
model formulation (see Section 3.1).
Definitions
• a = A block (from the set of blocks B) which adheres to the reasonable block
selection rule
• b = Another block (from the set of blocks B and not the same as a) which
adheres to the reasonable block selection rule
• c = Another block (from the set of blocks B and not the same as a or b) which
adheres to the reasonable block selection rule
• #a,b = Set of blocks that cannot be mined (based on the sequencing constraints)
until blocks a and b are mined (including explicitly mining blocks a and b).
This set contains blocks a and b and the union of all the blocks in each of their
respective holding sets (i.e., i7a,b — Lfb, since Ha contains block a and all
87 |
Colorado School of Mines | the blocks in block a’s holding set and Hb contains block b and all the blocks
in block b’s holding set). As a result, no shared blocks between the holding sets
of blocks a and b are counted more than once in the super-block represented
by #a,b-
H&,c = Set of blocks that cannot be mined (based on the sequencing constraints)
until blocks a and c are mined (including explicitly mining blocks a and c).
This set contains blocks a and c and the union of all the blocks in each of their
respective holding sets (i.e., HatC = HaU Hc, since Ha contains block a and all
the blocks in block a’s holding set and Hc contains block c and all the blocks
in block c’s holding set). As a result, no shared blocks between the holding sets
of blocks a and c are counted more than once in the super-block represented by
#a,c
Hb,c — Set of blocks that cannot be mined (based on the sequencing constraints)
until blocks b and c are mined (including explicitly mining blocks b and c).
This set contains blocks b and c and the union of all the blocks in each of their
respective holding sets (i.e., Hb,c = Hb U Hc, since Hb contains block b and all
the blocks in block b’s holding set and Hc contains block c and all the blocks
in block c’s holding set). As a result, no shared blocks between the holding sets
of blocks b and c are counted more than once in the super-block represented
by Hh,c-
Ha,b,c = Set of blocks that cannot be mined (based on the sequencing con
straints) until blocks a, b, and c are mined (including explicitly mining blocks
a, b, and c). This set contains blocks a, b, and c and the union of all the blocks
in each of their respective holding sets (i.e., /fa,b,c = U # b U # c, since Ha
contains block a and all the blocks in block a’s holding set, Hb contains block
b and all the blocks in block b’s holding set, and Hc contains block c and all
the blocks in block c’s holding set). As a result, no shared blocks between the |
Colorado School of Mines | holding sets of blocks a, b, and c are counted more than once in the super-block
represented by # a,b,c-
• LSa — Latest start time for block a (based on either the minimum processing
constraint or the minimum production constraint)
• LSb — Latest start time for block b (based on either the minimum processing
constraint or the minimum production constraint)
• LSC = Latest start time for block c (based on either the minimum processing
constraint or the minimum production constraint)
• f" = max (LSa, LSb, LSC)
• LSa,b = Latest start time for the set of blocks contained in (based on either
the minimum processing constraint or the minimum production constraint)
• LSa,c = Latest start time for the set of blocks contained in Ha c (based on either
the minimum processing constraint or the minimum production constraint)
• LSb,c = Latest start time for the set of blocks contained in LZb,c (based on either
the minimum processing constraint or the minimum production constraint)
• f' = max (LS'a.b, LSa)C, LSb,c)
• f = L5a,b,c = Latest start time for the set of blocks contained in i/a,b,c (based
on either the minimum processing constraint or the minimum production con
straint)
Inputs
• A set of blocks B
89 |
Colorado School of Mines | • Minimum processing requirement per time period (in tons of ore) and minimum
production requirement per time period (in tons of material). Note that these
capacity constraints must be hard constraints (i.e., they cannot be elasticized).
Outputs
• Valid cuts of the form:
Wa^f -f" '^b,f “t- ^c,f ^ 1
and
m&,f' "h Wb,f' “t- ^c,r' ^ 2
Algorithm
For each three-way combination of blocks a € B, b E B, and c E B in which
each block adheres to the reasonable block selection rule with respect to the
minimum production requirement do:
la. Determine the latest start time for block a (i.e., LSa) based on the mini
mum production capacity
2a. Determine the latest start time for block b (i.e., LSb) based on the mini
mum production capacity
3a. Determine the latest start time for block c (i.e., LSC) based on the mini
mum production capacity
4a. Create the set of blocks that represents the union of the holding sets for
blocks a and b (i.e., Ba,b) and determine the latest start time for this set
(i.e., LSa,b) based on the minimum production capacity
5a. Create the set of blocks that represents the union of the holding sets for
blocks a and c (i.e., B^c) and determine the latest start time for this set
(i.e., LBa,c) based on the minimum production capacity
90 |
Colorado School of Mines | 6a. Create the set of blocks that represents the union of the holding sets for
blocks b and c (i.e., Hb,c) and determine the latest start time for this set
(i.e., LSh,c) based on the minimum production capacity
7a. Create the set of blocks that represents the union of the holding sets for
blocks a, b, and c (i.e., 7fa,b,c) and determine the latest start time for this
set (i.e., f = LS'a.b,c) based on the minimum production capacity
8a. Determine the latest that any two-block set must be accessed:
f' = max (LSa,b, LSa,c, LSb,c)
9a. Determine the latest that any single block must be accessed:
f" = max (LSa, LSb, LSC)
10a. If f < fz then create a cut of the form:
^a,f yJbtf d" 'U/c,t ^ 1
11a. If fz < f" then create a cut of the form:
Wa -p/ ~f" Wb,f' "h ^c,tz ^ 2
For each three-way combination of blocks a € B, b E B, and c E B in which
each block adheres to the reasonable block selection rule with respect to the
minimum processing requirement do:
lb. Determine the latest start time for block a (i.e., LSa) based on the mini
mum processing capacity
2b. Determine the latest start time for block b (i.e., LSb) based on the mini
91 |
Colorado School of Mines | mum processing capacity
3b. Determine the latest start time for block c (i.e., LSC) based on the mini
mum processing capacity
4b. Create the set of blocks that represents the union of the holding sets for
blocks a and b (i.e., i/a,b) and determine the latest start time for this set
(i.e., L5a,b) based on the minimum processing capacity
5b. Create the set of blocks that represents the union of the holding sets for
blocks a and c (i.e., i^a,c) and determine the latest start time for this set
(i.e., LS^c) based on the minimum processing capacity
6b. Create the set of blocks that represents the union of the holding sets for
blocks b and c (i.e., i?b,c) and determine the latest start time for this set
(i.e., LSb,c) based on the minimum processing capacity
7b. Create the set of blocks that represents the union of the holding sets for
blocks a, b, and c (i.e., i7a,b,c) and determine the latest start time for this
set (i.e., f = Z/Sa)b,c) based on the minimum processing capacity
8b. Determine the latest that any two-block set must be accessed:
t — max (Z/iS^bj LS^c,
9b. Determine the latest that any single block must be accessed:
f" = max (L-Sa, LS\>, LSC)
10b. If f < f' then create a cut of the form:
tl?a,r Wh,f d~ yjc,f — 1
92 |
Colorado School of Mines | 11b. If fz < f" then create a cut of the form:
^a,Tz "h H- ^c,fz ^ 2
Output all generated cuts
Relative Dominance of Three-Way Latest Starts Cuts As with two-way cuts,
the time index (f or f') in our generated cuts tells us about the relative dominance
of different cuts. Cuts with an earlier time index dominate cuts (involving the same
blocks) with a later time index.
Three-way cuts concern themselves with another dominance issue though; the
value to the right of the inequality. Take the following two potential cuts:
• tUa,2 + Wb,2 + Wc,2 > 1
• ^a,2 + wh,2 + Wc,2 > 2
The first means that at least one of the three blocks a, b, and c must be mined
by time period 2, while the second means that at least two of these three blocks
must be mined by time period 2. In this case, the latter cut dominates the former.
The reason for this is that the latter is more restrictive than the former. The latter
requires that two blocks be mined, while the former requires only one of the blocks
be mined. Our three-way earliest starts cuts algorithm accounts for this dominance
and only generates the dominant cut for any set of blocks in the same time period
(assuming such a cut is valid and useful).
Three-Way Latest Starts Cuts Numerical Example Looking at our two-
dimensional example, we use blocks 9, 11, and 13 to create a three-way latest starts
cut based on the minimum production capacity (see Figure 4.7).
For this example, we assume that:
93 |
Colorado School of Mines | • Block a is represented by block 9 in the figure, block b is represented by block
11 in the figure, and block c is represented by block 13 in the figure
• Each block contains 10 tons of material (i.e., rib = 10 for each block a, b, and
c)
• The minimum production requirement is 20 tons per time period (for simplicity,
we only use production bounds for this example)
Recall that our decision variables are defined as tu#, where the b index identifies the
particular block (blocks 9, 11, and 13 in the figure above for this example) and the
t index identifies the time period by which the block is extracted. We now use the
algorithm to generate a three-way latest starts cut based on the minimum production
requirement:
1. Determine the latest start time for block a:
T& = 9
2. Determine the latest start time for block b:
LSb = 9
3. Determine the latest start time for block c:
1/^ = 9
1 2 6 7
8 14
20
aKSwags
Figure 4.7. Three-Way Latest Starts Cuts Numerical Example. This example depicts
the results of creating a three-way latest starts cut with blocks 9,11, and 13.
94 |
Colorado School of Mines | 11. Since f' < f" we can create a cut of the form:
w&,f' + Wb,f + Wcf' > 2 => + ^11,8 + Wisfi 2 2
This means that by the end of time period 6, at least one of the three blocks repre
sented by blocks 9, 11, and 13 in the figure above must be mined. Also, by the end
of time period 8, at least two of the three blocks represented by blocks 9, 11, and 13
in the figure above must be mined.
4.3.6 Cuts Involving More than Three Blocks
The general ideas we present in our two-block and three-block cut generating
algorithms can be extended to create cuts of more than three blocks. In fact, we can
generate cuts involving as many blocks as can be both produced and processed in any
given time period. However, creating cuts with more than three blocks gets harder
and more time consuming. As the number of blocks involved in the cut increases, the
number of possible block combinations that must be investigated grows exponentially.
This makes it harder to find these cuts yet still requires that they be worth the time
investment with respect to the reduction in overall problem solve time. Our empirical
evidence suggests that cuts become less effective as the number of blocks they contain
increases.
4.3.7 Using the by vs at Formulation in Cut Generation
We employ the by formulation to create our cuts. It is worth mentioning that
if we use the alternative at formulation, then the cuts we describe above are not
sufficient to produce the desired effects. For instance, say a valid and useful cut for
the by formulation is:
^3,3 + ^4,3 < 1 (4.1)
Using the by formulation of the problem, the variable Wbt represents block b being
mined by time period t. Therefore this one constraint implies that by time period
96 |
Colorado School of Mines | 3, at most one of the blocks 3 or 4 can be mined. The by formulation implicitly
accounts for all time periods up to and including the time period represented by the t
subscript, sot he constraint in equation (4.1) above implicitly addresses what occurs
in time periods 1 and 2, along with time period 3 (which is thet subscript). However,
if we use the at formulation, the constraint in equation (4.1) above only requires that
blocks 3 and 4 cannot both be mined at time period 3. For time periods later than
time period 3, this constraint suffices, but for time periods 1 and 2, there is a problem
with the formulation.
In order to get the same result with the at formulation, we need to pursue
one of two approaches. Either we require a set of three constraints or a cut involving
summation. Recall that the decision variable in the at formulation is representing
block b being mined at time period t.
A set of three constraints of the form:
2/3,1 + 2/4,1 < 1
< 1 (4.2)
2/3,2 + 2/4,2
2/3,3 + 2/4,3 < 1
accomplishes the same thing as the constraint in equation (4.1). These constraints
limit the removal of blocks 3 and 4 to at most one at time periods 1,2, and 3. Since
the at formulation includes constraints that permit blocks to be mined no more than
once during the time horizon via constraints (3.1) and (3.2), these three constraints
generate the same cut as equation (4.1). As the t index gets closer to the end of the
time horizon (i.e., approaches T), the number of constraints required to create a cut
using the at formulation increases. With the by formulation, however, we only need
one constraint for each cut, regardless of the t index on the constraint.
97 |
Colorado School of Mines | Another approach involves using summation notation. A constraint of the form:
3 3
suffices to define the constraint represented in equation (4.1). Here again, as the t
index gets close to the end of the time horizon, the number of terms being summed
increases accordingly. In general, using the at formulation adds more complexity
to the generation of valid and useful cuts, thus further justifying our use of the by
formulation.
4.4 Lagrangian Relaxation Methods
Lagrangian relaxation methods attempt to move complicating constraints to the
objective function, thus leaving a set of constraints that are relatively easily adhered
to. The relaxed constraints are dualized and added to the objective function with fixed
penalties (i.e., Lagrange multipliers usually denoted by A’s with various subscripts as
indices).
In the block sequencing problem, the side constraints that enforce minimum
and maximum operational bounds tend to complicate the otherwise simple structure
of the problem and are therefore considered complicating constraints. These side
constraints include:
• Average grade requirements
• Mine production capacity constraints
• Mill processing capacity constraints
With respect to the by formulation presented in Section 3.3.2, these side constraints
correspond to constraints (3.11), (3.12), (3.13), and (3.14), respectively (note that
the average grade constraints are written as two separate constraints for formatting
98 |
Colorado School of Mines | (4.3), each prefixed by its own Lagrange multiplier, A#. This leaves only five sets
of constraints in the problem, thus significantly simplifying the resulting problem’s
structure.
Not all of the side constraints must be moved to the objective function; we can
selectively choose which ones to move. Because there are four side constraints involved
(a minimum and maximum production constraint and a minimum and maximum
processing constraint), there are 16 combinations of scenarios that we investigate
with respect to dualizing these constraints. The simplest scenarios involve moving
only one of these constraints (since there are four side constraints, there are four such
scenarios). There are six scenarios which move two of these constraints, four scenarios
that move three of these constraints, and lastly one scenario that moves all four of
these constraints to the objective function. The scenario that moves none of these
constraints to the objective function is our monolith.
4.4.1 Basic Idea Behind the Lagrangian Relaxation Method
Implementing the Lagrangian relaxation procedure in AMPL and solving it with
CPLEX involves the use of a script to control execution of the program between the
monolith and the Lagrangian relaxation subproblem. Our iterative process attempts
to tighten lower and upper bounds on the optimal objective function value by succes
sively solving the Lagrangian relaxation subproblem of the monolith and using that
solution (if it is feasible) in the monolith to determine an optimal extraction schedule
for that iteration.
Lagrangian relaxation starts by solving the linear programming (LP) relaxation
of the monolith and using that objective function value as an initial upper bound
on the monolith objective function value. Subsequent iterations solve the Lagrangian
relaxation subproblem of the monolith. The optimal decision variable values from the
Lagrangian relaxation subproblem are simply inserted into the monolith (assuming
they are feasible in the monolith) to derive a new monolith objective function value.
100 |
Colorado School of Mines | If the objective function value for the current iteration’s Lagrangian relaxation sub
problem, z*LR, is less than the incumbent upper bound for the monolith, then we
update the upper bound with z*LR. If the current iteration’s monolith objective func
tion value, z^ona is higher than the incumbent lower bound for the monolith, then
we update the lower bound with z^orio. Before the next iteration, the Lagrangian
relaxation procedure updates the Lagrangian multipliers based on the degree of vi
olation incurred by each of the dualized constraints in the monolith. We terminate
the procedure after either reaching an iteration limit or achieving a small enough gap
between the Lagrangian procedure’s lower and upper bounds.
As mentioned above, to obtain the initial upper bound, we solve the LP-relaxation
of the monolith. In our case, this upper bound is actually quite tight, especially as
the data sets get bigger. Since our problem closely resembles a constrained knap
sack, we look at some characteristics of this class of problems to understand this
phenomenon. With bigger data sets, we have the ability to use more heterogeneous
left-hand-side coefficients to fill our knapsack capacity constraints. Two very sim
ple constrained knapsack problems help illustrate this principle. First, examine the
following problem:
max 10xi + 10%2 + lO^s (4.9)
subject to : lOaq + lOzg + 10%3 < 25 (4.10)
Xi E {0,1} (4.11)
The optimal LP-relaxation solution is XiP = 1, X P = 1, x^p = |, resulting in an
2
optimal objective function value of zRP = 25. The optimal integer programming (IP)
solution is x{p = 1 xïf — 1 x^p = 0, resulting in an optimal objective function
, ,
value of ZjP = 20. For this problem, the LP-relaxation does not provide a very strong
101 |
Colorado School of Mines | upper bound. We now present a second problem:
max 3%i + 5x2 + 4x3 + 7x4 + 8x5 (4.12)
subject to : 3xi + 5x2 + 4x3 + 7x4 + 8x5 < 25 (4.13)
Xi E {0,1} (4.14)
The optimal LP-relaxation solution for this problem is xf^ = x ^ = 1, x%p = 1,
x£p = 1, xpp = 1, resulting in an optimal objective function value of z£p = 25.
The optimal IP solution is x{p = 0, XgP = 1, XgP = 1, x4p = 1, XgP = 1, resulting
in an optimal objective function value of z*IP = 24. For this second problem, the
LP-relaxation provides a very strong upper bound. The second problem typifies our
larger data sets; they contain more data that is more heterogeneously distributed
with respect to total material and valuable ore content. As such, for our larger data
sets, the initial LP-relaxation to the monolith provides a very good upper bound.
The key to success when using the Lagrangian relaxation method is selecting the
correct constraints(s) to dualize and then properly setting the values of the multi
pliers for these dualized constraints (the A^’s, in our case). Selecting inappropriate
constraints to dualize may not lead to a simplified Lagrangian relaxation subproblem
or might result in Lagrangian relaxation subproblem solutions that are never feasible
in the monolith. If the dualized multipliers are too high, then the constraints with
which they are associated may have too much slack and the solution may be sub-
optimal. On the other hand, if these multipliers are too low, then their associated
constraints may be violated (since the cost of violation isn’t high enough) and the
solution may be infeasible for the monolith. There are various methods employed
in the literature to update the Lagrangian multipliers between successive iterations,
some of which are discussed in Section 4.4.2.
Among the most troublesome aspects of Lagrangian relaxation is the problem of
infeasibility. Generally speaking, the Lagrangian relaxation subproblem has no diffi
102 |
Colorado School of Mines | culty finding a solution; however, that solution may not be feasible in the monolith.
Because some constraints from the monolith are dualized in the Lagrangian relax
ation subproblem, the solution to the Lagrangian relaxation subproblem may allow
for constraint violations, especially if the cost of doing so (based on the Lagrangian
multiplier values) is low. Despite attempts at modifying the Lagrangian multiplier
values to discourage constraint violations, there may be no way to obtain a feasible
solution to the monolith problem, resulting in what is known as the “condition of
gaps” (Dagdelen 1985, pp. 99-100). The mathematical explanation for the existence
of these gaps has to do with the fact that the mapping of solutions between the
Lagrangian relaxation subproblem and the monolith may not be onto (Everett 1963):
The Lagrange multiplier method therefore generates a mapping of the
space of lambda vectors (components Afc, k = 1, ..., n) into the space of
constraint vectors (components c*, k = 1, ..., n [where ck represent the
constraints in the monolith]). There is no a priori guarantee, however,
that this mapping is onto—for a given problem there may be inaccessible
regions (called gaps) consisting of constraint vectors that are not generated
by any A vectors, (p. 407)
As Fisher points out (1985, p. 18) “In my experience, it is rare in practice that
the Lagrangian solution will be feasible in the original problem. However, it is not
uncommon that the Lagrangian solution will be nearly feasible and can be made
feasible with some minor modifications.” Our experience concurs with this statement
and we find that we rarely obtain a Lagrangian relaxation subproblem solution that is
feasible for the monolith, especially as the number of constraints we dualize increases.
However, following Fisher’s notion that these infeasible solutions can be made feasible,
we create a feasing routine that endeavors to do exactly this.
Dagdelen (1985) solves the block scheduling problem by Lagrangian decomposi
tion techniques. He employs subgradient methods to modify the Lagrangian multipli-
103 |
Colorado School of Mines | ers corresponding to the side constraints consisting of blending and capacity require
ments in his problem formulation. He further reduces the resulting multi-time period
Lagrangian relaxation subproblem into a series of efficiently-solvable single time pe
riod problems. He exploits the network structure of the sequencing constraints in
these single time period problems and uses adjusted block values to solve each of
these problems as an ultimate pit limits problem via the Lerchs-Grossman algorithm,
ultimately creating a block extraction schedule for the entire ore body.. Dagdelen
overcomes the “condition of gaps” by allowing the operational side constraints which
he dualizes in the Lagrangian relaxation subproblem to be violated by “e” in the
primal (i.e., the monolith). Doing this means that the constraints in the formulation
of his monolith are elastic, an idea we do not employ. Without this allowable e error,
his procedure would continue to attempt to create a nonexistent feasible solution and
would result in an endless loop. As a result, the constraint would never be met and
a feasible solution to the monolith would not be found.
Kawahata (2006) expands on the Lagrangian relaxation procedure developed by
Dagdelen (1985). His methodology uses two Lagrangian relaxation subproblems, one
to represent the most aggressive production scheduling case (i.e., working at the maxi
mum production bound) and the other to represent the most conservative production
sequencing case (i.e., working at the minimum production bound), to restrict the
monolith’s optimal solution space. The premise is that the decision variable values
from the solutions to these two Lagrangian relaxation subproblems eliminate vari
ables from the monolith, thus significantly speeding up solve times. However, he
still contends that “a gap problem cannot be avoided as long as the Lagrangian re
laxation method is applied to solve the production scheduling problem.” (2006, p.
64) We show that there are cases in which the optimal decision variable values to the
Lagrangian relaxation subproblem are feasible in the monolith. When the optimal de
cision variable values to the Lagrangian relaxation subproblem are not feasible in the
monolith, we attempt to make them feasible via our feasing routine, thus eliminating
104 |
Colorado School of Mines | the gap problem.
4.4.2 Implementation of the Lagrangian Relaxation Method for our Prob
lem
Unlike Dagdelen (1985) and Kawahata (2006), we treat our constraints as hav
ing rigid right-hand- sides ; thus, we do not allow constraint violations. Although more
realistic, this means we are plagued by an infeasible monolith solution once the La
grangian relaxation subproblem generates a potential solution. As such, we endeavor
not only to intelligently update our multiplier values, but also to employ heuristics
to force our solutions to be feasible. The success of our feasing routine depends on
the characteristics of the data, but ultimately its use aids in resolving the infeasibility
issues with which the Lagrangian relaxation procedure is beset.
Steps in the Lagrangian Relaxation Technique The Lagrangian relaxation
technique we employ consists of an iterative process which attempts to place lower
and upper bounds on our monolith’s optimal objective function value. The steps in
the procedure are:
1. Solve the LP relaxation of the monolith - this serves as the initial upper bound
(UB).
2. Solve the Lagrangian relaxation subproblem (LRSP).
If the LRSP’s optimal objective function value is less than UB, then update
UB.
3. Insert the LRSP’s optimal decision variable values into the monolith (assuming
they are feasible for the monolith).
If the monolith’s optimal objective function value is greater than LB, then
update LB.
105 |
Colorado School of Mines | 4. Update the Lagrangian multipliers (the A^’s).
5. Return to step 2 unless:
• iteration limit is reached
• acceptable LB-UB gap is reached
Successfully implementing the Lagrangian relaxation procedure is dependent on
many issues. First and foremost, we need good initial values for the Lagrangian
multipliers. Then we must have an efficient and effective means to update these
Lagrangian multipliers between iterations. Lastly, we must have a method for gener
ating feasible solutions for our monolith (if the optimal decision variable values from
the Lagrangian relaxation subproblem are not feasible in the monolith). We address
all of these issues and propose ways of resolving them to ensure that the Lagrangian
relaxation procedure converges to an acceptable solution quickly.
Scenarios Our model formulation consists of four sets of side constraints: 1) mini
mum production, 2) minimum processing, 3) maximum production, and 4) maximum
processing. As mentioned in Section 4.4, these four constraints result in 15 scenarios
representing the various one-way, two-way, three-way, and four-way combinations to
dualize them for use in the Lagrangian relaxation procedure. Generally speaking, the
more constraints we dualize, the simpler the structure of the resulting Lagrangian
relaxation subproblem becomes. Dualizing just one constraint simplifies the mono
lith’s constraint set only slightly, because there are still three other complicating side
constraints with which the solver must contend. Dualizing all four constraints means
that the resulting Lagrangian relaxation subproblem has a simplified structure that
solves quickly.
Unfortunately, as more constraints are dualized in the Lagrangian relaxation
subproblem, more constraints are violated in the monolith. If only one constraint
is dualized, then the resulting solution from the Lagrangian relaxation subproblem
106 |
Colorado School of Mines | still has three other constraints that help bound the solution and make it feasible
(or near feasible) for the monolith. When all four constraints are dualized, there
is nothing besides the sequencing constraints (constraints (4.4)) and the four other
auxiliary constraints (constraints (4.5), (4.6), (4.7), and (4.8)) to force the Lagrangian
relaxation subproblem’s solution to create a feasible solution for the monolith. As a
result, the Lagrangian relaxation subproblem’s solution is generally highly infeasible
in the monolith in the sense that many constraints are violated and the extent of
these violations is large.
Using our feasing routine we can force a feasible solution if the infeasibility is
not too great (i.e., the number of infeasible constraints is low and/or the extent to
which they are violated is not too great). Discouragingly, if the infeasibility is great,
even with our feasing routine, we cannot generate a feasible solution for the monolith.
The latter is what frequently happens if we dualize three or four of the constraints.
Additionally, the actual amount of time spent conducting the feasing routine becomes
excessive, thus negating all the time savings achieved in solving the much simplified
Lagrangian relaxation subproblem. Empirically, we see the best results in terms of
quickly converging to an acceptable answer by dualizing only one or two constraints
and then using our feasing routine on the Lagrangian relaxation subproblem’s solution
to create a feasible solution for the monolith.
Multiplier Maximum Values Generally, the only constraints imposed on the
multipliers used in the Lagrangian relaxation method are that they be non-negative.
Essentially, the higher the multipliers’ values, the more penalty there is to the ob
jective function value for violating the multiplier’s associated constraint. However,
there is a high enough multiplier value above which the associated dualized con
straint is not violated because the penalty to the objective function value is greater
than the penalty incurred by violating the associated dualized constraint. Raising the
constraint’s multiplier value beyond this maximum results in over-penalizing the ob-
107 |
Colorado School of Mines | jective function value, leading to a lower objective function and a poorer lower bound.
Excessively punishing the objective function with multiplier values that are too high
leads to wasted iterations, since the multipliers need to be adjusted downward over
the course of subsequent iterations. Therefore, we use information from the problem
formulation to set maximum multiplier values.
When dualizing the maximum production and/or processing constraints, the
multipliers’ upper bounds equal the maximum profit per ton that any accessible
block could possibly achieve. This value is the maximum profit because that is the
most we would be willing to pay to violate the constraint. Any value higher than
this is excessive punishment for violating the constraint. We calculate this value by
determining the maximum ratio between the discounted block value and the total
tonnage of the block ^max > being sure only to include those blocks that are
actually accessible by time period t (based on the block’s earliest start time).
On the other hand, when dualizing the minimum production and/or processing
constraints, the multipliers’ upper bounds equal the most profit per ton foregone for
any accessible block. Again, this value is the most we would be willing to pay to
violate the constraint, but in this case it represents how much we would be willing to
pay to not have to mine undesirable blocks. We calculate this value by determining the
minimum ratio between the discounted block value and the total tonnage of the block
^min î again being sure only to include those blocks that are accessible by time
period t (based on the block’s earliest start time). If all the blocks being investigated
are ore blocks, then this value is some positive number and the minimum processing
and production constraints are never violated. As such, the maximum value for the
multipliers should be 0 since we don’t need to punish the objective for a constraint
that is never violated.
Multiplier Seeding The ultimate goal of the Lagrangian relaxation procedure is
to derive an optimal solution to the Lagrangian relaxation subproblem that is also
108 |
Colorado School of Mines | feasible in the monolith and approaches an optimal solution for the monolith. To
do this, we must discern optimal multiplier values. Through an iterative process,
we adjust the multiplier values based on their effect on their respective dualized
constraints in the monolith (i.e., the amount of slack that the dualized constraints
contain as a result of the solution from the Lagrangian relaxation subproblem). Since
with each iteration we are refining an educated guess, the initial value we use to
seed the multipliers can have a dramatic effect on the number of iterations we must
conduct in order to generate a solution within an acceptable margin or error.
Although zero may be used to seed the multiplier values, we find that this value
is not very helpful. Essentially, seeding the multipliers with zero means that there is
no punishment in the objective function for violating the associated constraints. Since
our objective function is based on a net present value analysis, such a scheme results in
many ore blocks being mined at their earliest possible times, thus severely violating
the maximum production and processing constraints. Because of these constraint
violations, the multiplier values must be increased, and during the early iterations
of the Lagrangian relaxation procedure, the multipliers are constantly alternating to
correct under utilizing and violating the dualized constraint, creating a structure that
is not feasible for the monolith. Our experience shows that an initial value of 1 works
much better for our problem formulation. Seeding the multipliers with a value of 1
means that we incur some degree of punishment for violating a constraint, but that
punishment is not overly severe.
Another scheme is to seed the multipliers with the dual values from their as
sociated constraints in the LP-relaxation. We can generate these dual values when
we solve the LP-relaxation of the monolith to create our initial upper bound for the
problem. However, for large data sets, the time spent finding these dual values is
not trivial. Additionally, if the constraints are not tight in the LP relaxation of the
monolith, the resultant duals are zero, which means that the dualized constraints are
not punished at all (see discussion above). Overall, our experience does not indicate
109 |
Colorado School of Mines | that seeding multipliers with their duals is very useful.
Multiplier Updating Routines As mentioned in Section 4.4.1, there are various
methods available to update the multiplier values. Among the most common is one
called the subgradient method. As Fisher (1981, p. 7) points out, “The subgradient
method is a brazen adaptation of the gradient method in which gradients are replaced
by subgradients.” Given an initial value for the multiplier, Ajt, generate a sequence
of multipliers using the rule:
= 4 + - 6) V U ,k
where tk is a positive scalar representing the step size at iteration k and the term
(Axk — b) represents the slack in the dualized constraint at iteration k as a result
of finding the optimal solution to the Lagrangian relaxation subproblem at iteration
k. In the same paper, Fisher points out two other popular approaches for updating
Lagrangian multipliers: 1) those employing the simplex method via column generation
techniques and 2) multiplier adjustment methods. The former do not see much use
because they tend to converge slowly and are rather difficult to program. The latter,
which are problem-specific, may afford great benefits if used properly.
To solve our problem via the Lagrangian relaxation method, we use the subgra
dient method (which we call the traditional approach to multiplier updating) and also
attempt some multiplier adjustment methods employing either a percentage change
(i.e., we increase or decrease the multiplier by a fixed percentage at each iteration) or
the traditional approach with a historical look-back (i.e., we set the new multiplier
value equal to the weighted average of its value in the previous iteration as well as
what it traditionally would be in the current iteration).
The percentage change multiplier adjustment method works as follows: If an
inequality constraint is satisfied, then the multiplier value encourages the objective
110 |
Colorado School of Mines | function to utilize slack in this constraint. On the other hand, if the constraint is not
met, then the value of the multiplier for that constraint is raised by a fixed percentage
as a way of discouraging the Lagrangian relaxation subproblem’s objective function
from violating that constraint.
The traditional approach with a historical look-back initially uses the subgradi
ent method described above to create a new multiplier value. However, the resultant
multiplier value is not used in its entirety. A fixed percentage of the previous it
eration’s multiplier value is included along with a fixed percentage of the current
iteration’s multiplier value (calculated via the subgradient method) to create the up
dated multiplier value for the current iteration:
4= (curr%) + (hist%) A^”1 V i,t,k
For example, 75% of the current iteration’s multiplier and 25% of the previous iter
ation’s multiplier may be used to create the current iteration’s multiplier value. As
a result, this method employs information from the previous iteration to help pre
vent large changes in multiplier values, especially in the first few iterations of the
Lagrangian relaxation method.
Held, Wolfe, and Crowder (1974) present another multiplier updating scheme
based on four different scaling parameters; however, we do not find their methods
very promising.
4.4.3 Feasing Routines
Fisher’s idea of modifying an infeasible solution so that it becomes feasible is
what leads us to propose a feasing routine for the open pit scheduling problem. Given
enough spare blocks not in the current Lagrangian relaxation subproblem’s optimal
solution (i.e., blocks that are not mined in the current optimal solution), we can
selectively add or remove blocks from the optimal Lagrangian relaxation subproblem’s
111 |
Colorado School of Mines | solution to create a solution that is feasible for the monolith. When infeasibilities
occur because of not meeting minimum production and/or processing bounds, we
simply add the best blocks to the solution to ensure that these lower bounds are met.
On the other hand, when infeasibilities occur because of exceeding the maximum
production and/or processing bounds, we remove the best blocks from the solution
to ensure that the upper bounds are met. In both cases, by best we mean that
we pick blocks that help us meet the various constraint bounds using as few blocks
as possible and avoid violating other constraint bounds in the process. We start
the feasing process in the first time period with a constraint violation and then
check all subsequent time periods to ensure that our feasing actions in previous time
periods do not have adverse affects. If our feasing actions from previous time periods
cause subsequent time period’s constraints to become infeasible, we use our feasing
routine on these later time periods also. Once our feasing routine is complete, we
ensure that the solution to the Lagrangian relaxation subproblem is feasible for all
time periods before passing the decision variable values back to the monolith. Our
empirical experience shows that employing this feasing routine significantly increases
our ability to use the Lagrangian relaxation method and determine feasible solutions
for the monolith (see Section 5.3.2 for results).
When using the feasing routine to eliminate infeasibilities due to violating the
maximum processing or maximum production, by best, we mean removing those
blocks that best help meet the maximum constraints while not violating the mini
mum constraints or the sequencing constraints in the process. Our goal is to generate
a feasible solution for the monolith. For example, if our current Lagrangian relax
ation subproblem solution violates maximum production constraints, then the feasing
routine finds the heaviest waste block(s) to remove. Removing waste blocks ensures
that we do not violate the maximum processing constraints (because waste blocks
have no processable material in them). Picking the heaviest waste blocks means that
we do not spend extra time searching for more blocks than necessary in order to
112 |
Colorado School of Mines | meet the maximum production constraint. To ensure that we do not violate any of
the sequencing constraints, we remove blocks from among those that represent the
bottom-most mined blocks in the current time period’s optimal solution to the La
grangian relaxation subproblem. By remove we mean that we mine the block one
time period later, unless we are at the end of the time horizon, in which case the
block is not mined at all. Selecting a correct block in the current time period whose
extraction we shift to one time period later ensures that we do not adversely affect
the sequencing constraints for subsequent time periods. Also, we are careful not to
isolate a block on any level, thus violating the sixth sequencing constraint. If our
actions result in leaving a block completely alone on a given level, we also move that
block to the next time period so that the we do not violate the sixth sequencing
constraint. The example in Section 4.4.3 clarifies this concept.
When we conduct the feasing routine to remove infeasibilities as a result of
minimum processing or minimum production constraint violations, by best we mean
adding blocks that best help meet the minimum constraints without violating the
maximum constraints in the process. For example, if the current Lagrangian re
laxation subproblem solution violates the minimum processing constraints, then the
feasing routine finds the heaviest ore block(s) to add. Adding ore blocks ensures
that we make the violated minimum processing constraint feasible while not adding
useless waste blocks that may potentially create a violation of the maximum produc
tion constraint. Again, picking the heaviest ore blocks ensures that we do not search
for more blocks than necessary to satisfy the minimum processing constraint. The
blocks are added at the top of the pit, picking among those blocks that are not in the
optimal solution by the current time period (i.e., we mine unmined blocks that are at
the highest level in the pit). Adding blocks at the top of the pit guarantees that we
do not violate the sequencing constraints and preserves the feasibility of the solution
with respect to the sequencing constraints when used in the monolith. Again, we
ensure that no blocks are isolated on any level so that we do not violate the sixth
113 |
Colorado School of Mines | sequencing constraint. If our added block is isolated on a given level, then we also add
a neighbor block to preclude violating the sixth sequencing constraint. We terminate
the feasing routine when we obtain a feasible solution, or when there are no more
blocks to shift. An example in Section 4.4.3 clarifies this concept.
Although our feasing routine helps produce feasible solutions, there are some
caveats. First and foremost, there are some Lagrangian relaxation subproblem so
lutions that contain constraint violations to such a degree that our feasing routine
cannot correct them. This is especially true as the number of time periods increases
and/or the number of dualized constraints increases. The feasing routine may also
take a long time to execute, especially for large data sets.
Feasing Routine for Maximum Constraints Algorithm
Assumptions We include all the assumptions that we describe with respect to our
model formulation (see Section 3.1).
Definitions
• T = number of time periods in the horizon
• t = time period in which a maximum constraint is violated
• B*ll9%ble = the set of all blocks that are on the lowest level (with respect to
the z-axis) of the Lagrangian relaxation subproblem’s optimal solution in time
period t
• kl™ — the 2-coordinate of the set B flgible in time period t
• wlïst = the variable representing the best block b (i.e., bbest) to remove from the
Lagrangian relaxation subproblem’s optimal solution in time period t for the
current iteration
114 |
Colorado School of Mines | Inputs
• Maximum processing capacity per time period (in tons of ore) and maximum
production capacity per time period (in tons of material) - note that these
capacity constraints must be hard constraints (i.e., they cannot be elasticized)
• An optimal solution for the Lagrangian relaxation subproblem, {w^tR}
• A set of blocks not in the optimal solution for the Lagrangian relaxation sub
problem, {Wft1}
Outputs
• A feasible solution for the monolith based on the current optimal solution from
the Lagrangian relaxation subproblem
Algorithm
For all time periods t = 1...T repeat while the solution is infeasible in the
monolith or until further feasing routine actions cannot be taken, i.e., until
is empty:
1. Determine B flglble and k\ow for time period t.
2. Determine the best block to remove, whst\
If any of the following three scenarios occurs:
• the maximum production constraint is violated
• the maximum production constraint is violated by a greater per
centage than the maximum processing constraint
• the maximum production constraint and the minimum processing
constraint are both violated
Then:
115 |
Colorado School of Mines | let wltst = the heaviest block (with respect to weight) of all the
blocks in B f%gible that contain the least amount of usable material
in them
If any of the following three scenarios occurs:
• the maximum processing constraint is violated
• the maximum processing constraint is violated by a greater per
centage than the maximum production constraint
• the maximum processing constraint and the minimum production
constraint are both violated
Then:
let whst = the heaviest ore block (with respect to usable material)
of all the blocks in B figtble that contain the least amount of total
material in them
3. Set wblst = 0 (i.e., not mined based on the definition of our variables)
4. lit <T then mine block bbest in the next time period (i.e., set wjj+i = 1)
5. Ensure that the sixth sequencing constraint is not violated by moving the
block whst to the next time period.
If the block wb^st is alone on level k\ow in time period t (i.e., it has no
neighbors) then it can be moved without violating the sixth sequencing
constraint. Otherwise, check if each one of its plus sign neighbors has
a neighbor. If moving w^st to the next time period isolates any one
of its neighbors, then the isolated neighbor block must also be moved
to the next time period.
6. Check the current solution to ensure it is feasible for the monolith in time
period t, i.e., it satisfies the operational (side) constraints in time period t.
If the solution is feasible, increment t by 1 and return to step 1. Otherwise,
go to step 7.
116 |
Colorado School of Mines | 7. Update the set B figible by removing bhest from it, and correspondingly
update k[ow, if applicable. Return to step 2.
If the algorithm produces a feasible solution for the monolith, use this solution
in the monolith to attempt to update its lower bound.
Feasing Routine for Maximum Constraints Numerical Example Using our
two-dimensional example, we employ our feasing routine for maximum constraints to
create a feasible solution for the monolith from an infeasible solution generated by
the Lagrangian relaxation subproblem (see Figure 4.8 below).
*
1
- .•v
8 9 12 14
1 'f
15 16 17 20 21
Figure 4.8. Feasing Routine for Maximum Constraints Numerical Example. This ex
ample depicts the idea of using the feasing routine to render an infeasible Lagrangian
relaxation subproblem solution feasible for the monolith by removing the best block
among blocks 18 and 19.
For this example, we assume that:
• Each block contains 10 tons of material (i.e., rib = 10)
• The maximum production constraint is 40 tons per time period
• t = 3 and t < T
• The z-coordinate index runs from 1 at the bottom to 3 at the top of the pit
Recall that our decision variables are defined as Wbt, where the b index identifies
the particular block and the t index identifies a time period by which the block is
117 |
Colorado School of Mines | extracted. With a maximum production capacity of 40 tons per time period, by
the end of time period 3 at most 120 tons of material can be mined. However, the
diagram above depicts 130 tons of material being mined by the end of time period 3
(all the light grey and dark grey blocks), so the solution to the Lagrangian relaxation
subproblem violates the maximum production constraint in the monolith. We use
the feasing routine to find the best block that is part of the optimal solution to the
Lagrangian relaxation subproblem in time period 3 and remove it from the solution:
1. Determine B f%gible and kltow for time period 3:
^eligible _ 19} and therefore k1™ = 1.
2. Determine w^st:
Since the maximum production constraint is the only violated constraint,
represents the heaviest block (with respect to weight) of the the blocks in
^eligible por this example, however, each block weighs the same (10 tons),
so we arbitrarily choose block 18 as the best block: wffi =>
3. Set = 0:
«% = 0
4. If t < T then mine block bbest in the next time period:
Since we assume t <T, then wblsl = 1.
5. Ensure that the sixth sequencing constraint is not violated by moving wlfst to
the next time period:
Moving block 18 to be mined in time period 4 now isolates block 19, so in order
to obey the sixth sequencing constraint, we must move block 19 to be mined in
time period 4 also.
6. Now we have a feasible solution in time period 3. Let t — 4 and return to step
1 (in the case that the solution is infeasible in time period 4 or later).
118 |
Colorado School of Mines | After eliminating infeasibilities for time periods 4 through T, we can use the
modified optimal solution to the Lagrangian relaxation subproblem in the monolith
to generate an objective function value (an NPV) and update the lower bound if the
resultant value is greater than the incumbent lower bound.
Feasing Routine for Minimum Constraints Algorithm
Assumptions We again include all the assumptions that we describe with respect
to our model formulation (see Section 3.1).
Definitions
• T = number of time periods in the horizon
• t = time period in which a minimum constraint is violated
• B figlble = the set of all blocks that are on the highest level (with respect to
the z-axis) of the Lagrangian relaxation subproblem’s optimal solution in time
period t
• k^igh — the ^-coordinate of the set B ftgible in time period t
• w^st = the variable representing the best block b (i.e., bbest) to add to the
Lagrangian relaxation subproblem’s optimal solution in time period t for the
current iteration
Inputs
• Minimum processing capacity per time period (in tons of ore) and minimum
production capacity per time period (in tons of material) - note that these
capacity constraints must be hard constraints (i.e., they cannot be elasticized)
• An optimal solution for the Lagrangian relaxation subproblem, {wfctR}
119 |
Colorado School of Mines | • A set of blocks not in the optimal solution for the Lagrangian relaxation sub
problem, {w^tR}
Outputs
• A feasible solution for the monolith based on the current optimal solution from
the Lagrangian relaxation subproblem
Algorithm
For all time periods t = 1...T repeat while the solution is infeasible in the
monolith and there are blocks that can be added to the solution in time period
t (i.e., there are blocks in the data set not mined that can be mined in time
period t based on their earliest start times):
1. Determine B fl9%ble and k^igh for time period t.
2. Determine the best block to add, w^st:
If any of the following three scenarios occurs:
• the minimum production constraint is violated
• the minimum production constraint is violated by a greater per
centage than the minimum processing constraint
• the minimum production constraint and the maximum processing
constraint are both violated
Then:
let WfoSt = the heaviest block (with respect to weight) of all the
blocks in B ftgMe that contain the least amount of usable material
in them
If any of the following three scenarios occurs:
• the minimum processing constraint is violated
120 |
Colorado School of Mines | • the minimum processing constraint is violated by a greater per
centage than the minimum production constraint
• the minimum processing constraint and the maximum production
constraint are both violated
Then:
let WfoSt = the heaviest ore block (with respect to usable material)
of all the blocks in B figMe that contain the least amount of total
material in them
3. Set whst = 1 = £ + 1,..., T (i.e., mined based on the definition of our
variables)
4. Ensure that the sixth sequencing constraint is not violated by adding the
block w\lst to the optimal solution for time period t.
If adding the block wblst to the optimal solution for time period t
means that the block is alone (i.e., has no plus-sign neighbors)
on level k^9*1, then its addition to the optimal solution violates the
sixth sequencing constraint. To avoid violating the sixth sequencing
constraint when adding w\lst to the optimal solution, also add the best
block from among wblst,s plus-sign neighbors to the optimal solution.
5. Check the current solution to ensure it is feasible for the monolith in time
period t, i.e., it satisfies the operational (side) constraints in time period t.
If the solution is feasible, increment t by 1 and return to step 1. Otherwise,
go to step 6.
6. Update the set B ftgible by removing bbest from it, and correspondingly
update h1™, if applicable. Return to step 2.
If the algorithm produces a feasible solution for the monolith, use this solution
in the monolith to attempt to update its lower bound.
121 |
Colorado School of Mines | Feasing Routine for Minimum Constraints Numerical Example Using our
two-dimensional example, we employ our feasing routine for minimum constraints to
create a feasible solution for the monolith from an infeasible solution generated by
the Lagrangian relaxation subproblem (see Figure 4.9 below).
Figure 4.9. Feasing Routine for Minimum Constraints Numerical Example. This ex
ample depicts the idea of using the feasing routine to render an infeasible Lagrangian
relaxation subproblem solution feasible for the monolith by adding the best block
among blocks 8,9, and 14.
For this example, we assume that:
• Each block contains 10 tons of material (i.e., rib — 10)
• All blocks have an earliest start time of t = 1
• The minimum production constraint is 20 tons per time period
• t = 7
• The ^-coordinate index runs from 1 at the bottom to 3 at the top of the pit
Recall that our decision variables are defined as w^, where the b index identifies
the particular block and the t index identifies a time period by which the block is
extracted. With a minimum production capacity of 20 tons per time period, by the
end of time period 7 at least 140 tons of material must be mined. However, the
diagram above depicts 130 tons of material being mined by the end of time period
122 |
Colorado School of Mines | 7 (the light grey blocks), so the solution to the Lagrangian relaxation subproblem
violates the minimum production constraint in the monolith. We use the feasing
routine to find the best block that is not part of the optimal solution to the Lagrangian
relaxation subproblem in time period 7 and add it to the solution:
1. Determine BfwMe and k^tgh for time period 7:
^eligible
__ g 14} and therefore h1?9*1 = 2.
2. Determine
Since the minimum production constraint is the only violated constraint,
represents the heaviest block (with respect to weight) of the the blocks in
^eligible ' por this example, however, each block weighs the same (10 tons),
so we arbitrarily chose block 8 as the best block: =>
3. Set wbhtst = 1:
1
=
4. Ensure that the sixth sequencing constraint is not violated by adding w\lst to
the optimal solution for time period t:
Adding block 8 to the optimal solution in time period 7 means adding an isolated
block, thus violating the sixth sequencing constraint. As a result, we must also
add block 9 to the optimal solution for time period 7 in order to obey the sixth
sequencing constraint.
5. Now we have a feasible solution in time period 7. Let t = 8 and return to step
1 (in the case that the solution is infeasible in time period 8 or later).
After eliminating infeasibilities for time periods 8 through T, we can use the
modified optimal solution to the Lagrangian relaxation subproblem in the monolith
to generate an objective function value (an NPV) and update the lower bound if the
resultant value is greater than the incumbent lower bound.
123 |
Colorado School of Mines | Chapter 5
NUMERICAL RESULTS
5.1 Data
To examine the methods and procedures described, we use a master data set
that represents an open pit mine consisting of 19,320 blocks to empirically test our
methodologies. Associated with this data set are minimum and maximum bounds
on the per time period production and processing constraints at the mine. The mine
follows 45° sloping rules.
We reduce this 19,320 block data set by an order of magnitude into an envelope
of blocks that contains 1,060 blocks. We create yet another data set two orders of
magnitude smaller called a micro-pit. This micro-pit includes 196 blocks. With such
a small data set, we are able to obtain an optimal solution and graphically investigate
the results.
To further investigate our methodologies, we create two additional data sets from
the master data set. The first of these is a data set containing the 1,980 blocks found
in a 13 by 13 by 12 block subset of the original 19,320 blocks. The second of these
data sets is one containing the 2,880 blocks in an 18 by 17 by 12 block subset of the
original 19,320 blocks. Note that the original 19,320 block data set is not a uniform
cube of blocks, so the subsets we create are also not uniform cubes (this is why
13 x 13 x 12 7^ 1,980 and 18 x 17 x 12 ^ 2,880). When describing our computational
results, we refer to these various data sets by the number of blocks they contain.
To obtain additional large data sets, we perturb the mineral content of each of the
blocks in the 19,320 block data set by ±5% to create seven additional instances of this
large data set. We refer to these data sets as A, £, C, D, E, F, and G perturbations.
124 |
Colorado School of Mines | Lastly, we examine a much more complicated open pit mining model with variable
cut-off grades, stockpiles, and blocks that can be partially mined. We refer to this
data set as Newmont and use it to show how our methodologies work in a more
general setting.
Below is a table of all the various data sets we use and their pertinent charac
teristics, including the number of time periods in the horizon:
name # blocks # binary variables # constraints # time periods
1,060 1,060 6,360 32,748 6
1,980 1,980 11,880 71,562 6
2,880 2,880 17,280 105,504 6
10,819 10,819 64,914 395,885 6
10,819A 10,819 64,914 395,885 6
10,819B 10,819 64,914 395,885 6
10,8190 10,819 64,914 395,885 6
10,819D 10,819 64,914 395,885 6
10,819E 10,819 64,914 395,885 6
10,819F 10,819 64,914 395,885 6
10,819G 10,819 64,914 " 395,885 6
Newmont 61 1,391 55,022 25
Table 5.1. Data Sets Used to Empirically Test our Methodologies. This table sum
marizes the pertinent characteristics of the various data sets we employ to test our
solution methodologies
It is important to note that the Newmont formulation also includes 162,934 continuous
variables.
5.1.1 Data Pre-processing
We find that examining the data we use in our model formulations before actually
running the optimization routines provides some very enlightening insights. Some of
the data we use has many individual datum that we can effectively remove from the
data set without sacrificing optimality in any manner. In practice, using data blindly
without investigating its characteristics can either lead to erroneous results or extra
computation.
125 |
Colorado School of Mines | Our largest data set containing 19,320 blocks is 25 blocks wide in the ^-coordinate
direction by 26 blocks long in the y-coordinate direction by 60 blocks deep in the z-
coordinate direction (note that the z-coordinate runs from the bottom up, i.e., a lower
number represents a deeper level in the pit). Although a complete block structure
with these dimensions would have 39,000 blocks in it (25 * 26 * 60 = 39,000), this
data set has only half that many blocks indicating that some pre-processing has been
done to eliminate blocks that are not part of the orebody.
Examining the resulting 19,320 blocks even further yields the observation that
there are absolutely no ore blocks on any of the bottom 17 levels of the pit and
therefore no reason to include any of these 5,474 blocks in our data set since they
will never be mined. Because of this, we reduce our 19,320 block data set to a 13,846
block data set. Next, we discover that dispersed throughout the rest of the ore body
are 3,027 phantom blocks which are completely empty (they contain no material
and no mineral content). Removing these blocks leads us to a data set containing
10,819 blocks, thus nearly halving the size of our original 19,320 block data set.
This reduction in the data set pays huge dividends for all our solution methodologies,
especially since integer programming is notoriously plagued by exponential solve times
with respect to the size of problem instance.
5.2 CPLEX Parameter Settings
We use the AMPL programming language, version 2006.06.26 (2006) to formulate
our model. We then enter this formulation into the CPLEX solver, version 10.1
(2006). CPLEX offers many parameter settings that can be altered by the user when
solving mixed integer programming problems. Varying these parameter settings can
dramatically change the problem’s solution time. Unfortunately, there are many
parameters to explore and no one combination of settings works for all problems.
The model’s performance depends on the combination of parameter settings used in
the CPLEX solver. As such, we must either determine the best parameter settings
126 |
Colorado School of Mines | for each problem instance or find a set of parameters that works well for different
instances of the same class of problems. The most straightforward way to determine
which parameter settings to use is to try each setting available. However, this quickly
becomes an enormous task. To efficiently discern which parameter settings work
best (including how the various settings interact with each other), we rely on past
programming experience and knowledge of what seems to work well with similarly
formulated problems to investigate those parameters that have generated the most
promising results. Based on the problem being investigated here, the most promising
parameters (and their associated definitions according to the AMPL CPLEX 10.0
User’s Guide by ILOG 2006) are:
• baropt - used to specify the barrier (i.e., interior point) method to solve linear
programming problems
• branch - used to specify a branching direction on the fractional decision variable
value (i.e., strong or weak branching)
• heurfreq - frequency with which CPLEX applies a rounding heuristic at the
nodes
• mipcuts - used to specify the level of aggression CPLEX uses to generate cuts
based on different combinatorial constructs
• mipemphasis - used to guide CPLEX’s branch and cut strategy
• probe - used to determine the amount of solution probing CPLEX conducts
• rinsheur - used to determine how often to apply the relaxation induced neigh
borhood search heuristic (RINS heuristic)
These seven parameters and their various settings result in 216 different combinations
we explore to discern the best parameter settings. The results are different for each
êôBKjœ&s'w»
127
GOLDEN, CO 80401 |
Colorado School of Mines | data set used. However some general trends do emerge. The following parameter
settings produce the fastest solution times for the class or problems we investigate:
• monolith without earliest starts, latest starts, or cuts — branch -1 mipcuts -1
mipemphasis 1 probe 1 rinsheur 40
• monolith with earliest starts and latest starts — branch -1 mipcuts -1 mipem
phasis 1 probe 1 rinsheur 40
• monolith with earliest starts, latest starts, and cuts — mipcuts -1 mipemphasis
1 rinsheur 40
• Lagrangian relaxation procedure — baropt1 branch -1 heurfreq 20 mipemphasis
1 rinsheur 40
The newest version of CPLEX, version 11, has a parameter tuning feature which
intelligently selects the best parameter settings to use for each problem instance.
Future work on this problem would benefit from its use.
5.3 Computational Results
We use a Sun Fire V240 with 2GB of RAM to conduct all computations in
CPLEX. We use an IBM Thinkpad with an Intel 2.13 GHz processor and 2.0 GB of
RAM and a LENOVO desktop with dual core AMD ATHLON64 5000+ processors
and 2.0 GB of RAM to run all earliest starts, latest starts, and cuts algorithms.
5.3.1 Visual Depiction of an Extraction Sequence
As discussed in Section 5.1, we create a micro version of the data to investigate
different aspects of the model. Since the micro-pit is so small, graphing the results
of a two time period problem instance is relatively easy and representing the actual
1Note that the baropt parameter only pertains to the initial LP relaxation. We do not use it to
solve any of the Lagrangian relaxation subproblems.
128 |
Colorado School of Mines | three-dimensional pit outlines during each time period is possible (see Figure 5.1
below). This dynamic depiction of how the actual extraction operations occur at the
mine helps mine engineers communicate the schedule to their employees.
at beamnina of operations after 1 time period after 2 time periods
Figure 5.1. Visual Depiction of Micro Pit Results. This figure depicts the optimal
solution of a two-period extraction sequence using the micro-pit data.
5.3.2 Computational Results for Earliest Starts, Latest Starts, Cuts, and
the Lagrangian Relaxation Procedure
We use our earliest starts procedure to calculate a complete predecessor list for
each block in the data set and then determine each block’s earliest start time period.
Next, we use our latest starts procedure to calculate a complete holder list for each
block in the data set and then determine each block’s latest start time period. Lastly,
we use these predecessor and holder lists, along with their associated earliest and
latest starts to generate cuts. When generating cuts, we use our reasonable block
selection rule to empirically determine how many blocks to include for the various
types of cuts we generate. Generating cuts involves a degree of judgment and must be
balanced with the amount of time it takes to solve the monolith. Generally speaking,
we aim to generate cuts to such a degree that their generation time is no more than
about 20% of the time it takes to solve the monolith without any earliest or latest
129 |
Colorado School of Mines | starts or cuts. We summarize these generation times in Table 5.2 below:
problem instance preds time ES time holders time LS time cuts time
(sec.) (sec.) (sec.) (sec.) (sec.)
1,060 10 3 11 3 62
1,980 14 4 15 4 46
2,880 22 8 24 8 284
10,819 1,059 61 N/A N/A 1,869
10,819A 1,130 65 N/A N/A 1,816
10,819B 1,090 63 N/A N/A 1,834
10,819C 1,128 65 N/A N/A 1,803
10,819D 1,120 64 N/A N/A 1,802
10,819E 1,125 65 N/A N/A 1,798
10,819F 1,305 75 N/A N/A 2,471
10,819G 1,285 74 N/A N/A 2,404
10,819 AVG 1,155 67 N/A N/A 1,975
Newmont ~ 0 ~ 0 N/A N/A 4
Table 5.2. Summary of Generation Times for Predecessor Lists, Earliest Starts,
Holder Lists, Latest Starts, and Cuts. This table summarizes the time spent to
generate predecessor lists (preds) and the associated earliest starts (ES), holder lists
(holders) and the associated latest starts (LS), and cuts (cuts). All times are in
seconds. The penultimate row in the table (10,819 AVG) presents the average results
for all the 10,819 block data set instances. We do not create holder lists nor latest
starts for the 10,819 data set instances and the Newmont data set.
Looking at Table 5.2, we notice that for our largest data set, on average we
spend one minute to calculate the earliest starts for all 10,819 blocks. We do not
create holder lists nor latest starts for any of the 10,819 data set instances because
empirical evidence shows us that these values are not useful due to the characteristics
of these data sets.
When applying the Lagrangian relaxation procedure, it is important to dualize
the correct constraint (s). Our empirical evidence indicates that dualizing more than
one constraint results in a Lagrangian relaxation subproblem whose optimal decision
variable values are not feasible in the monolith. Additionally, we are not able to
make the optimal solution to the Lagrangian relaxation subproblem feasible with our
feasing routine, rendering the entire Lagrangian relaxation procedure in its current
130 |
Colorado School of Mines | implementation ineffective. As such, we conclude that dualizing only one constraint
works best. The actual constraint to dualize is dependent on the nature of the data.
Our results indicate that dualizing the constraints with the most slack in them works
best. However, it is easy to dualize each constraint and then run the Lagrangian
procedure in parallel on four separate machines. The first instance to converge to an
acceptable solution gap indicates which single constraint to dualize.
We use our feasing routine in any iteration that involves a Lagrangian relaxation
subproblem solution that is infeasible in the monolith. Additionally, we conduct our
feasing routine until we either find a feasible solution to the monolith or run out of
blocks with which we can conduct the feasing routine (i.e., if there are no more blocks
to add to or remove from the model, the feasing routine terminates).
Using the by formulation we describe in Section 3.3.2 above, we calculate a
solution within 2% of optimality to determine the extraction sequence for the data
sets presented in Table 5.1 above. We present detailed results in the appendix and a
summary of the execution times in Table 5.3 below.
Examining Table 5.3, it is apparent that our methodologies drastically improve
solution times. We compare the monolith’s solution time (column monolith in Table
5.3) with the solution times of using just earliest and latest starts (column ES & LS
in Table 5.3), earliest and latest starts with cuts (column ES & LS & cuts in Table
5.3), and the Lagrangian relaxation procedure with earliest and latest starts (column
Lagrangian Relaxation with ES & LS in Table 5.3). Overall, using earliest and latest
starts reduces computer solve times by 80.2%. Including cuts with earliest and latest
starts also results in an average solution time reduction of 80.2%. Implementing the
Lagrangian relaxation procedure provides an average reduction of 84.0%.
Taking the size of the data sets into account, we see that the Lagrangian re
laxation procedure significantly improves solution times for bigger data sets, while
earliest and latest starts with cuts seem to work best with smaller data sets. First
and foremost, none of the eight instances of the 10819 data set solve to 2% mipgap
131 |
Colorado School of Mines | problem instance monolith ES & LS ES & LS & cuts Lagrangian
solution solution solution Relaxation
time (sec.) time (sec.) time (sec.) with ES & LS
solution time (sec.)
1,060 1,082 147 82 105
1,980 202 148 95 193
2,880 1,481 366 419 660
10,819 > 24 hr s. 16,326 3,801 708
10,819A > 24 hr s. 1,570 7,803 769
10,819B > 24 hr s. 4,225 29,545 1,261
10,819C > 24 hr s. 3,054 1,570 698
10,819D > 24 hr s. 4,355 1,633 1,075
10,819E > 24 hrs. 1,426 2,794 680
10,819F > 24 hrs. 3,680 7,273 14,159
10,819G > 24 hrs. 838 14,121 3,687
10,819 AVG > 24 hrs. 4,434 8,567 2,880
Newmont 11,476 9,719 8,696 N/A
Table 5.3. Summary of Results from Implementing Earliest Starts, Latest Starts,
Cuts, and the Lagrangian Relaxation Procedure. This table compares the results
(in seconds of CPLEX solve time) of using our algorithms on the various data set
instances. The column labeled monolith represents the raw data. The column labeled
ES & LS is the raw data with earliest and latest starts implemented. The column
labeled ES & LS & cuts depicts the raw data with earliest and latest starts and an
appropriate level of cuts included. Lastly, the column labeled Lagrangian Relaxation
with ES & LS presents the results from implementing the Lagrangian relaxation
procedure on the data set with earliest and latest starts. The penultimate row in
the table (10,819 AVG) presents the average results for all the 10,819 block data set
instances. The structure of the Newmont formulation is different enough from our
model formulation to preclude us from using the Lagrangian relaxation procedure on
it.
132 |
Colorado School of Mines | within 24 hours (i.e., 86400 seconds). For these large data sets, we observe that
earliest and latest starts reduce solve times by 94.9%, while adding cuts actually
decreases this time savings to about 90.1%. However, the Lagrangian relaxation pro
cedure works extremely well, reducing solve times by an average of 96.7%. For six
of the eight 10819 data instances, the Lagrangian relaxation procedure is the fastest
method, reducing solve times by an astonishing 99%.
5.3.3 Comparison of Results with Commercial Software
To gain an appreciation for how well our methodologies work, we compare our
results with those of a commercially available mine scheduling software package, Mi-
neSight Economic Planner - MSEP (2006). MSEP uses the traditional approach
to solve the block sequencing problem, so it first determines the ultimate pit limits
and then generates nested pits and pushbacks which it uses to schedule the block
extraction of the ore body. The software suffers from many limitations, including
1) an inability to specify a time horizon, 2) not being able to include lower bounds
on the operational constraints, and 3) a failure to adhere to upper bounds on the
operational constraints. The software claims to use a dynamic cutoff grade approach,
which theoretically provides a better schedule with respect to maximizing NPV.
We use the 1060, 1980, 2880, and 10819 (original case only) to test MSEP’s
performance. We present the results from using MSEP for these four data instances
in Table 5.4 below: In each instance, the software violates the upper bound on the
production constraint in time period 1. Looking at Table 5.4, these violations are
not trivial, amounting to over a ten-fold increase in required production capacity for
the 10,819 data set instance. Additionally, depending on the data set, the software
arbitrarily chooses a time horizon: 4 time periods for the 1060 data set, 7 time
periods for the 1980 data set, 7 time periods for the 2880 data set, and 10 time
periods for the 10819 data set. Despite using a dynamic cutoff grade, disregarding
maximum production constraints, not adhering to minimum operational constraints,
133 |
Colorado School of Mines | and arbitrarily setting a longer time horizon in all but the 1060 block data set, MSEP’s
optimal NPVs are lower for the 1060, 1980, and 2880 block data sets, and only
marginally higher for the 10819 data set. Recall, though, that the 10819 block data
set is run for 10 time periods by MSEP, while ours is only 6 time periods.
The actual algorithm that MSEP uses is a complete mystery. Granted, the soft
ware runs remarkably quickly (solution times are on the order of 20 seconds or fewer),
but we have absolutely no confidence in the quality of these solutions. The heuris
tic MSEP employs has serious drawbacks that result in unimplementable extraction
schedules. Specifically, the ten-fold violation of the maximum production constraint
in time period 1 assumes that a certain amount of waste can be removed, i.e., pre
stripped, during a “pre-production” year. This requires that resources are available
for such an activity. Additionally, there is no indication that the software includes
the cost of conducting this initial work in its NPV calculation.
We are confident in the quality of our solutions. We use deterministic operations
research methods that have withstood the test of time. Our solution times may be
longer, but the resultant block extraction schedule adheres to all operational and
problem our NPV MSEP NPV # time time period 1
instance ($106) ($106) periods production constraint
violation (tons)
1,060 19.0 13.5 4 773,000
1,980 17.5 17.3 7 600,000
2,880 15.5 14.2 7 3,000,000
10,819 9.1 9.8 10 10,500,000
Table 5.4. MSEP’s Results for the 1,060, 1,980, 2,880, and 10,819 Data Set Instances.
This table summarizes the results of implementing the 1,060, 1,980, 2,880, and 10,819
data set instances in MSEP. The column labeled Our NPV shows the NPV we achieve
using the same data set in our monolith MIP formulation. The column labeled MSEP
NPV presents MSEP’s final NPV for the extracted blocks from the pit. The column
labeled # time periods shows how many time periods MSEP uses to calculate its
NPV. The last column shows how many excess tons of production capacity (over the
1,000,000 tons stipulated by the maximum production capacity constraint) MSEP
requires to achieve the NPV it reports.
134 |
Colorado School of Mines | Chapter 6
LIMITATIONS, EXTENSIONS, AND CONCLUSION
6.1 Limitations and Extensions
Our model is first and foremost a deterministic model. All inputs are known
with absolute certainty. In real-world mines, this assumption is often not valid, espe
cially with regard to block characteristics deep underground. To effectively address
the random nature of block content or the value of the ore being removed from the
mine, stochastic programming is a better approach for solving realistic open pit min
ing problems. As Dimitrakopoulos (1998) shows, the optimal solution to the block
sequencing problem is affected by uncertainties in many of the input parameters such
as: 1) in-situ grade uncertainty, 2) uncertainty in the operational mining specifica
tions such as production and processing capacities or sloping rules, and 3) economic
uncertainties with respect to operating costs or the value of the mineral being ex
tracted. In order to address these potential uncertainties, Achireko and Frimpong
(1996) use neural networks to resolve the randomness in block characteristics while
Ramazan and Dimitrakopoulos (2004b) directly address in-situ grade variability in
their model formulation.
Variable cutoff grade models should also be investigated. Such models add an
other dimension to the decision variables (in the form of a location index (I) indicating
where each block is sent in the optimal solution), but more accurately reflect reality
and handle in-situ ore variability better. However, adding another index significantly
increases the number of decision variables that must be investigated by the model.
The extra problem detail this affords comes at a cost of larger problem size.
There may be ways to generate earliest and latest start times based on the
136 |
Colorado School of Mines | minimum or maximum number of blocks that can be removed in a certain time
period. Tighter problem formulations may be achieved using more aggressive cut
generation schemes that investigate larger sets of blocks. A stronger reasonable block
selection rule or quicker cut generation methods would allow us to include more cuts
in the formulation without significantly increasing overall solution time (i.e., the time
spent generating the cuts would not be absorbed by the time saved in using them).
With respect to the Lagrangian relaxation procedure, there are many additional
tactics that may be employed. Nemhauser and Wolsey (1988) provide some alter
natives to the subgradient method for multiplier updating. Among these alternative
methods is one that uses a constraint generation idea which could also lead to the
introduction of more cuts via cutting planes, thus further tightening the formulation.
Additionally, the use of the interior point method for the solution of the LP relax
ation of the monolith or for any other linear programs may provide solutions with a
different and promising algebraic structure.
The feasing routine we create may also be improved. Determining how long to
conduct the feasing routine and how often (with respect to iteration count) it should
be used are items warranting further investigation. The feasing routine might only
be applied every n iterations, or only if the Lagrangian relaxation solution is not “too
infeasible” based on both the number of constraints violated in the monolith and the
extent to which these violations occur. Lastly, placing a limit on the amount of time
spent conducting the feasing routine may preclude using it for excessive amounts of
time with little or no success.
A completely different feasing routine based on something other than manually
adding or removing blocks from the Lagrangian relaxation subproblem solution may
also be worth examining. One such feasing routine may be to find the first time
period with a constraint violation and impose the violated constraint as a hard con
straint, then resolve the Lagrangian relaxation subproblem and check if these new
decision variable values are now feasible in the monolith, imposing the violated con
137 |
Colorado School of Mines | straint as another hard constraint if the solution is not feasible in the monolith. Such
an incremental approach, however, may take a long time to implement because we
must resolve the Lagrangian relaxation subproblem after each constraint is added.
Ultimately, any routine that endeavors to render feasible those Lagrangian relaxation
subproblems that are infeasible in the monolith must allow the Lagrangian relaxation
procedure to converge to within an acceptable margin of error faster than solving the
monolith outright.
Aggregating time periods or blocks to determine strategic mine schedules may
also be useful. Other relaxation and decomposition techniques, such Dantzig-Wolfe
decomposition and column generation methods, may provide fruitful results.
Additional research to reduce solution times should focus on methods to either
limit the number of variables in the problem or methods that do not necessitate the
use of branch-and-bound algorithms to solve MIP problem instances. Heuristics based
on genetic algorithms or artificial neural networks may provide better solution times.
Regardless of the methodology used, any procedure that can reduce the solu
tion time required to determine the efficient block extraction schedule is of benefit to
mine engineers in their quest to efficiently sequence the extraction of profitable ma
terial from their mines. The alternative is to either suffer with slow algorithms and
long solution times, or use intuition to guess the best extraction schedule. Neither
alternative is attractive for the complex mines that we see in the world today.
6.2 Conclusion
Efficiently scheduling the extraction* of ore from an open pit mine helps ensure
that the mine maximizes the net present value of the minerals in the orebody. Solving
the block sequencing problem results in a time-indexed schedule of when any given
block in the orebody should be removed (if it is removed at all) that maximizes the
NPV of the ore in the pit subject to all sequencing and operational constraints.
Mine planners use two approaches to solving the block sequencing problem: one
138 |
Colorado School of Mines | based on the ultimate pit limits and another based on a comprehensive approach.
The former divides the process into three separate stages that are solved sequentially,
while the later takes a global view of the problem. Although more difficult to solve,
the comprehensive approach provides more flexibility and ultimately creates a better
schedule. In our research, we pursue solving the block sequencing problem using this
approach. We propose various methodologies that make the problem more tractable.
We limit the solution space by defining decision variables only between their earliest
and latest possible start times. We present a series of cut generation algorithms that
produce valid and useful cuts to tighten the problem formulation. Lastly, we employ a
Lagrangian relaxation technique with a feasing routine to make infeasible Lagrangian
relaxation subproblem solutions feasible for the monolith.
Employing our methodologies significantly reduces solve times while not com
promising the optimal solution. Although our earliest starts idea appears in the
literature, no one employs a latest starts idea in open pit mining. Our cuts are
much more aggressive. The Lagrangian relaxation procedure we use does not require
soft constraints, but instead uses our feasing routine to ensure feasible Lagrangian
relaxation subproblem solutions for the monolith.
The techniques we present: 1) earliest and latest starts, 2) cuts, and 3) La
grangian relaxation, serve as tools to expedite solution times for the block sequencing
problem. Just like any handyman knows, one tool is never sufficient for all jobs. In
the same vein, our three tools complement each other and serve as different techniques
to aid in arriving at solutions to the block sequencing problem. Our empirical results
show that using our tools reduces solve times by well over 95% without sacrificing any
confidence in the answers achieved. In today’s finicky commodities markets, being
able to adapt to changing market conditions and incorporate the latest mine-specific
data to update operating schedules is paramount to ensure a profitable mining ven
ture.
139 |
Colorado School of Mines | APPENDIX A
The following tables present detailed results of employing our solution method
ologies. Each table compares our techniques with the solution time if none of our
techniques is used. The monolith column presents the results using the raw data
with no modifications. The ES & LS column shows the results using the raw data
with earliest and latest starts implemented. The ES & LS & cuts column depicts
the raw data with earliest and latest starts and an appropriate level of cuts included.
Lastly, the column labeled Lagrangian Relaxation with ES & LS presents the results
from implementing the Lagrangian relaxation procedure on the data set with earliest
and latest starts. Each table represents a separate data set, so we show each of the
twelve data sets depicted in Table 5.1.
monolith ES & LS ES & LS Lagrangian Relaxation
& cuts with ES & LS
# cuts 0 0 7,039 0
cut generation time (sec.) 0 0 62 0
# binary variables 6,360 5,025 5,025 5,025
# constraints 31,688 23,801 30,840 23,795
MIP simplex iterations 81,449 27,691 18,796 13,417
branch- and-bound nodes 360 120 60 0
computer time (sec.) 1,082 147 82 105
NPV ($10*) 19.0 19.0 19.1 18.8
Table A.I. Detailed Results for the 1,060 Data Set. This table shows the detailed
results for the data set called 1,060.
146 |
Colorado School of Mines | ABSTRACT
In a mining project, the orebody is located and outlined using the exploration
data. After determining an outline of a massive deposit, a block model is developed to
represent the deposit and an ultimate pit limit containing the set of blocks which
maximum total dollar value is found. In long term open pit mine planning, the major
problem after the ultimate pit limit determination is to define the periods and destinations
for the blocks to be mined in order to maximize the total discounted profit from the mine
subject to a series of operational constraints. Mixed Integer Linear Program (MILP)
formulations for optimum mine planning have been proposed for many years but the
computational time in solving the formulations is a major problem. We investigate the
optimization techniques that are being used in the mining industry for solving the
production scheduling problem for open pit mines. In this thesis, new strategies and a
related algorithm will be developed to solve the large MILP open pit mine phase design
model using the economic parameters and geologic block model information as the basic
input for determining the designed phases that lead to the optimum long term open pit mine
schedules. The computational techniques will be programmed and applied to realistic
mining projects coming from actual operations. The results coming from the new phase
design algorithm will be discussed and compared to the results generated by using available
commercial software used by the industry. |
Colorado School of Mines | ACKNOWLEDGMENTS
I wish to express sincere appreciation to my advisor. Dr. Kadri Dagdelen for his
guidance and continuous support. Special thanks to all committee members: Dr. Alexandra
Newman of the Division of Economics and Business, Dr. Vaughan Griffiths of the
Engineering Department, Dr. John Grubb of the Mining Engineering Department, and Dr.
Hugh Miller of the Mining Engineering Department for their valuable suggestions and
assistance.
Special thanks to Dr. Thys Johnson, my advisor’s advisor for his interest and time
in discussing my research. He also helped proof read this thesis.
Thanks also to my fellow graduate students, especially Larry Clark, Kazuhiro
Kawahata, Songwut Artittong, Hendro Fujiono, Yasser Akbarzadeh, Andrea Brickey,
Mehmet Cigla, and Benito Perez for our friendship and the good times we had together.
Furthermore, I acknowledge that my research was partially supported by Newmont
Corp. USA. I sincerely appreciate the reliable support from being a teaching assistant since
2005 for the following professors: Dr. Masami Nakagawa (Mine plant design), Dr.
Manohar Arora (Statics), Tracy Barnes (Geostatistics), and again Dr. Kadri Dagdelen
(Surface mine design. Geostatistics, and Mine systems analysis). I appreciate Dr. Ramona
Graves of the Petroleum Engineering Department for an opportunity to help teach
Geostatistics in her department.
Lastly, I am thankful to my wife, my daughter, and my parents for providing moral
support and love. |
Colorado School of Mines | CHAPTER 1
INTRODUCTION
In a mining project, the first step is exploration, and the deposit’s tonnage and
grade is determined by a sampling process (such as drilling). The mineral deposit is
located and outlined using the exploration data. After determining an outline of a
massive deposit, a block model is developed to represent the deposit. The size of a block
for a typical open pit mine varies depending on geology and mining method. The size of
the blocks is generally considerably smaller than the drill hole spacing. The grade of
each block in the model is estimated using a technique such as distance weighting or
Kriging (Dagdelen, 1985). The value of each block is calculated based on the grade in
that block using economic parameters such as commodity price, mining cost, processing
cost, and recovery.
An ultimate pit limit containing the set of blocks which has the maximum total
dollar value for a given block model can be found using current computer optimization
techniques. These techniques are based on the 3D Lerchs-Grossmann (Lerchs &
Grossman, 1965) and the Johnson’s network flow (Johnson, 1968) methods. Both
methods guarantee to find the optimum pit in three dimensions (3D) regardless of block
height, width, and length proportions and give the ultimate pit limit which maximizes the
value of the identified set of ore and waste blocks that can be mined at proper slope
angles (Barnes R., 1980).
After the ultimate pit limits are defined, the development and design of phases or
pushbacks that will be mined during the progression of the pit is the next crucial step in
long range open pit mine planning. The phase designs serve as the basis to obtain Life-
of-Mine (LOM) plans and schedules that define the future cash flows of a given project.
The phases within the ultimate pit are traditionally obtained by finding successively
larger ultimate pit limits with respect to economic block models that are generated by
using successively increasing prices. By changing the commodity prices, the pit size can
1 |
Colorado School of Mines | be decreased or enlarged. This traditional price parameterization approach is based on
the ultimate pit limit method, and has been used as a phase design technique for decades.
From phase designs, the next step is to find the production schedule that optimizes
the cutoff grade and maximizes the net present value (NPV) of a given project. The mine
production scheduling problem can be formulated as a mixed integer linear programming
(MILP) problem with constraints related to the mining extraction sequence, mining and
milling capacities, grades of mill throughput, and various operational requirements such
as minimum pit bottom width. During the last 20 years, the Mining Engineering
Department at the Colorado School of Mines (CSM) has put considerable effort into
developing and implementing the MILP models that optimize production schedules and
cutoff grades to maximize the NPV of complex mining operations. The outcome of the
development, a program called OptiMinetm, was developed as a production scheduling
and cutoff-grade optimization tool to universally handle complex mining operations
(Dagdelen & Kawahata, 2007). However, the program uses bench-by-bench block
aggregation techniques which combine blocks on the same bench and same phase into a
single object in order to reduce the solution time. As such, the production schedule
obtained using this approach depends on the phase design. Mine engineers currently
employ the traditional phase design method; we demonstrate in this thesis that this
method does not guarantee the phase design that maximizes the NPV. Therefore, we
develop a new phase design algorithm to be used as a mine plan optimization tool that
improves the NPV of an open pit mine.
1.1. Thesis Objectives
The objective of this thesis is to determine a phase design method that will generate
phases to be used as the basis for obtaining production schedules that maximize the NPV of
a given open pit mining project. The thesis will focus on developing a new phase design
method for optimization of open pit mine production schedules by formulating and solving
the problem as a mixed integer linear program (MILP) in a reasonable amount of time.
2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.