Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

International Journal of Surface Mining,

Reclamation and Environment


2004, Vol. 18, No. 2, pp. 99–121

Financially Efficient Dig-Line Delineation


Incorporating Equipment Constraints
and Grade Uncertainty

A.J. RICHMOND1 AND J.E. BEASLEY2

ABSTRACT

Greedy local search heuristics that provide estimates of financially efficient dig-lines are proposed. To
generate alternate dig-lines these heuristics employ a floating circle-based perturbation mechanism that
ensures that mining equipment constraints are satisfied. Distributions of financial payoffs for alternate dig-
lines are constructed from multiple conditional simulations, then evaluated using a mean-downside risk
efficiency model. Computational results show that equipment constraints significantly increase the financial
risk for a given expected payoff and that the heuristic efficient frontier is sensitive to the mining strategy. For
the problem instances that we examined we observed that mining waste material at transitions prior to
mining ore material: minimised the upfront operating costs; maximised the profit for a given level of
financial risk; and freed milling capacity.

Keywords: conditional simulation, portfolio theory, floating cone, combinatorial optimisation,


local search heuristic, open pit mining.

1. INTRODUCTION

In open pit mining, the identification of the optimal (maximum utility) dig-line is
often considerably uncertain. For example, the ore-grade mineralisation may not be
confined to visually observable geological zones; may be highly erratic; the sampling
density may be inadequate; or operational constraints may make the incorporation of
dilution cognitively complex [1]. A risk-averse decision maker may wish to
incorporate the financial implications of this uncertainty into the delineation process.
With perfect knowledge and no mining equipment constraints the dig-line will be a
contour line corresponding to the cut-off grade value. In practice, some material
discarded as waste will, in fact, have a grade that is above the cut-off grade value.

1
Address correspondence to: A.J. Richmond, Earth Science and Engineering Department, Imperial College,
London, SW7 2AZ, UK. E-mail: a.richmond@imperial.ac.uk
2
The Management School, Imperial College, London, SW7 2AZ, UK.

10.1080/13895260412331295376$16.00 # Taylor & Francis Ltd.


100 A.J. RICHMOND AND J.E. BEASLEY

Similarly, some material treated as ore will have a grade value that is below the cut-off
grade value. However, discarding ore-grade material, and treating waste material as
ore does not imply a dig-line that is sub-optimal in terms of profit or utility. The ability
to mine ore-grade material at a profit depends on the size of the equipment being used,
whilst utility is related to the uncertain payoff and the risk-averse characteristics of the
decision-maker.
Local optimisation techniques, such as loss functions [2–4] and utility functions
[5] can take into account uncertainty in the local payoff and specific risk-averse
characteristics. However, these methods fail to consider the potential benefit of
diversifying financial risk that is unique to each location. The diversification of unique
risk in ore delineation can be viewed as a constrained portfolio optimisation problem
that results in a set of local ore selections that are financially efficient [6, 7]. However,
financially efficient solutions contain operationally undesirable transitions between
ore and waste that calls for the free selection of individual blocks.
The automatic determination of dig-lines that satisfy mining equipment constraints
and are optimal in terms of profit and/or financial risk has received little attention in
the literature. Early researchers proposed a profitability-based 2D adaptation of the
floating cone algorithm for generating dig-lines that respected mining equipment
constraints [1]. This approach produces unique, non-optimal (in terms of profit)
solutions that do not account for financial risk. More recently, simulated annealing
heuristics were proposed for the dig-line problem, which was viewed as the
constrained optimisation of a polygon (representing the dig-line) [8, 9]. Practical
implementation required calibration of a non-linear penalty function associated with
the tortuosity of the polygon (which relates to the angles present in the polygon) and a
complex perturbation mechanism. This heuristic was able to generate a smooth
polygon which satisfied equipment constraints in the sense of the working face only.
Other equipment requirements such as the minimum turning radius were not
explicitly considered.
We shall refer to the diversification of financial risk by the suitable choice of local
ore processing options as the ore selection problem (OSP), and to the incorporation of
equipment constraints in the OSP as the dig-line problem (DLP).
Let:
K be the number of mutually exclusive and exhaustive processing options, where
without loss of generality we include treating a block as waste as one of the
processing options and the options are arranged in increasing cost/recovery
order (from the least expensive to most expensive)
B be the total number of blocks
ybk ¼ 1 if we decide to process block b using option k and zero otherwise
S represent the set of block processing options chosen for the entire set of blocks,
with s(b) being the option chosen for block b
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 101

w represent a financial payoff


FS be the distribution of financial payoffs for S
Let us assume that all blocks lie in the same horizontal plane and must be mined. Then
the OSP can be formulated as a mixed-integer non-linear constrained optimisation
problem [7]:
maximise OðSÞ ¼ lEfFS g  ð1  lÞrfFS g ð1Þ
subject to
X
K
ybk ¼ 1 b ¼ 1; . . . ; B ð2Þ
k¼1

S ¼ ½ðsðbÞ ¼ kjybk ¼ 1Þ; b ¼ 1; . . . ; B ð3Þ


Z t
rfFS g ¼ ðt  wÞa dFS ð4Þ
1

ybk 2 ð0; 1Þ; k ¼ 1; . . . ; K; b ¼ 1; . . . ; B ð5Þ


where O( ) is the objective function, Ef g is the usual expectation operator,
l(0 l 1) is the weight to be attached to expected payoff, t is the target payoff, and
a (  0) is a measure of the impact of failing to reach the target payoff t. Equation (1)
ensures that the set of processing decisions has the desired tradeoff between expected
payoff and financial risk. Equation (2) guarantees that exactly one processing option
is chosen for each block. Equation (3) identifies the set of block processing options
chosen. Equation (4) assumes that the downside risk function [10] is a suitable
measure of financial risk. The objective function in Equation (1) can then be viewed as
a utility function which reflects risk-averse behaviour for a > 1 [10]. Equation (5)
indicates that the OSP is discrete.
Equation (1) calls for a distribution of financial payoffs associated with S. In
practice, L conditional simulations can provide a multivariate density function
Fð1; . . . ; B; z1 ; . . . ; zL Þ ¼ Prob fzð1Þ z; . . . ; zðBÞ zg that characterises the joint
uncertainty about the B mineral concentration values zð1Þ; . . . ; zðBÞ [11]. Such a
probability distribution can be used to estimate a distribution of financial payoffs
FS ¼ ProbfwðBjS; LÞ wg, derived from the L results:
B X
X K
wðBjS; lÞ ¼ ybk qb ½mpk zl ðbÞ  ck ; l ¼ 1; . . . ; L ð6Þ
b¼1 k¼1

where qb is the quantity (weight) of material contained in block b, m is the mineral


value per concentration unit, pk is the proportion of the mineral recovered using
102 A.J. RICHMOND AND J.E. BEASLEY

processing option k, and ck is the mining and processing cost per unit weight for
processing option k. In practice, all parameters in Equation (6) are uncertain and
should be treated as random variables. For simplicity, in this paper we will consider all
parameters except attribute z to be known and constant. A key point to note here is
that, as the individual block payoffs are conditional on the processing option k, by
considering alternate processing options for block b we change the potential
distribution of financial payoffs for all B blocks FS.
In a modern portfolio theory context, solutions to the OSP for various l identify the
constrained efficient frontier. In other words, a discrete efficient frontier for the OSP
can be constructed by solving the above formulation (Equations (1)–(5)) for varying l
values and eliminating any solutions that are dominated, thereby generating a
discrete curve showing the best possible tradeoff between expected payoff and
financial risk.
Defining a mathematical expression that relates to mining equipment constraints is
not straightforward [8, 9]. Experimental work we carried out on expanding the OSP
objective function (Equation (1)) with an additional component to minimise
transitions in processing decisions between adjacent blocks proved disappointing
(and hence is not reported here). In this paper we propose to consider the DLP as a
special case of the OSP, where, to satisfy mining equipment requirements, the
processing option chosen for selected blocks is dynamically restricted. A two-
dimensional floating object is used to identify those blocks for which the restriction
applies. Thus, the DLP that we will consider in this paper is Equations (1)–(5) with an
additional (dynamic) constraint:

yBk ¼ ybk k ¼ 1; . . . ; K 8 B 2 b ð7Þ

where b defines the successors for block b in a directed graph. When considering the
processing decision for block b, Equation (7) ensures an identical processing decision
is made for any block which is a successor to b in a directed graph.
A simple example of a two-dimensional directed graph is shown in Figure 1, where
the radius of the (floating) circle could represent the minimum turning radius of the
mining equipment. For simplicity, we assume that if the vertex of block B is within
the circle centred at the vertex of block b, then B is a successor to b. For example, in
Figure 1, for b ¼ 13 we have that 13 ¼ {3, 7–9, 11–15, 17–19, 23}. The key idea here
is that whatever processing option is chosen for block b, exactly the same processing
option must be chosen for all the blocks b to which it is connected. In this way we
automatically ensure that the resulting dig-line is actually mineable. Note though that
Equation (7) is not applied to all blocks b ¼ 1, . . . ,B in the entire mining domain.
Were we to do so then (unless the radius of the circle defining b for each block b is
very small) the effect would be that the entire set of blocks would have exactly the
same processing option chosen. Clearly this would be financially unwise and so
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 103

Fig. 1. Graphical representation of floating circle directed graph.

Equation (7) is only selectively applied, as will become clear in our algorithms
presented below.
Section 2 describes the floating circle algorithm for the DLP. Sections 3 and 4
propose two greedy local search heuristics to solve the DLP. The first heuristic
provides rapid estimates of the efficient frontier. The second, a hybrid heuristic,
requires additional computational effort, but improves the quality of the efficient
frontier. In Section 5 we report computational results for various implementations of
the proposed heuristics. We summarise the paper in Section 6.

2. THE FLOATING CIRCLE ALGORITHM

Srivastava, Hartzell, and Davis [1] proposed a profitability-based 2D adaptation of the


floating cone algorithm to generate the maximum payoff dig-line (DLP with l ¼ 1.0),
which we shall refer to as the SHD method. Their idea was to float a circle,
representing the constraint in Equation (7), around the problem area, and if the blocks
that fell within the current circle perimeter were able to be mined as ore profitably,
then the dig-line was extended outwards to the circle perimeter. However, it is well
known that the floating cone algorithm produces sub-optimal results when the cone
evaluations include blocks that already belong to the open pit solution, and when the
test sequence is not re-initiated from the top of the mineral deposit each time a
positively-valued cone is defined [12]. Furthermore, unlike open pit optimisation, in
the DLP there is no option to leave non-profitable material in situ. Thus, we would
expect the SHD method to produce poor solutions to the DLP. Early experimental
work we carried out confirmed this expectation and so we did not pursue the SHD
method any further.
104 A.J. RICHMOND AND J.E. BEASLEY

To account for the deficiencies in the SHD method, we propose several


modifications. We shall refer to the adapted floating circle algorithm for the DLP
(implicitly restricted to l ¼ 1.0) as FCDLP, which proceeds as follows:
1. Select a circle of diameter d that satisfies mining equipment constraints
2. Set k ¼ 1
3. Generate an initial solution S ¼ [s(b) ¼ k, b ¼ 1, . . . , B], i.e. initially choose to
treat all blocks as waste (processing option 1)
4. Define an ordered path for visiting the blocks [bjb ¼ 1, 2, . . . , B] by ordering the
blocks on decreasing average grade of b
5. Set i ¼ 0
6. Set i ¼ i þ 1 and let block b be the i’th block in the ordered path defined at step 4
above
7. Float the circle to b and determine the average grade of b , but excluding from
b any block whose current processing option is k þ 1
8. If the average grade from step 7 is greater than or equal to the optimal cut-off
grade for processing option k þ 1 then:
 update the solution S by changing the ore processing option for all blocks
within the circle b to k þ 1
 go to step 5
9. If i < B go to step 6
10. If k < K  1, then set k ¼ k þ 1 and go to step 5
Using the optimal cut-off grades (rather than profitability) in the floating circle
evaluations at step 8 above accounts for the fact that all blocks must be mined. Blocks
already considered (and accepted) for option k þ 1 are excluded from further
consideration in step 7 above. Note that, the equipment constraint (Equation (7))
applies only to blocks for which s(b)  2 and where the average grade in step 7 above
resulted in a modification to the solution. Furthermore, if K > 2, then there is no
guarantee that this constraint is honoured for blocks for which s(b) ¼ k < K when
considering solution modifications for k þ 1.
By ordering (at step 4 above) the floating circle path from areas of high expected
mineral grades to areas of low expected mineral grades the possibility of a single
(few) high-grade block(s) providing financial support for a large number of non-
profitable blocks is reduced. Computationally, we observed that solutions generated
with such ordered paths tended to have the highest expected payoff values. Unlike the
application of cut-off grades to the ore selection problem, there in no guarantee that
the FCDLP solution is optimal in terms of expected payoff.
The perturbation mechanism ensures that any solution S represents a dig-line(s)
that satisfies mining equipment requirements. Figure 2 plots local ore processing
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 105

Fig. 2. Block processing decisions generated by FCDLP.

decisions generated by FCDLP for a test problem involving just two processing
options that we will consider in more detail in following sections. In this plot, a dig-
line can be considered as the circumference of a contiguous set of blocks for which
the processing decisions are identical. The resolution of such a dig-line varies with the
block dimensions. Note that some waste outlines may be tortuous, implying that ore
material at the transitions must be mined prior to waste material.
The ordering in step 4 above makes the proposed FCDLP algorithm deterministic.
Therefore, for a given set of parameters and model of grade uncertainty the solution is
unique. Increasing d implicitly increases the magnitude of the mining equipment
constraint in the DLP by increasing the number of blocks within the floating circle for
which the processing option must be the same. To generate alternate solutions for a
given d, the floating circle evaluations can be carried out by incrementally varying the
commodity price to create a series of nested dig-lines.
The FCDLP algorithm described above is analogous to the floating cone technique.
Reordering the K processing options in decreasing cost/recovery order (from the most
expensive to least expensive) and altering the criteria in step 8 to accept only average
grades less than the optimal cut-off grade creates an algorithm analogous to the
negative floating cone technique [12].
106 A.J. RICHMOND AND J.E. BEASLEY

3. DESCRIPTION OF THE LOCAL SEARCH HEURISTIC

Both the SHD method and FCDLP cannot account for that part of the DLP where l <
1.0. In this section we propose a local search heuristic (LSDLP) that involves searching
the solution space in the neighbourhood of the current solution for (possibly) improved
solutions by floating the circle around the problem area, and perturbing solution
elements that fall within the circle. Perturbed solutions are accepted only if the objective
function (Equation (1)) is improved, thus, LSLDP is a greedy heuristic. This iterative
process continues until all blocks have been visited, then repeats so long as further
improvement in the objective function can be achieved. LSDLP proceeds as follows:
1. Select the expected payoff weight l and a circle of diameter d that satisfies
mining equipment constraints
2. Set k ¼ 1
3. Generate an initial solution S ¼ [s(b) ¼ k, b ¼ 1, . . . , B], i.e. initially choose to
treat all blocks as waste (processing option 1)
4. Compute the value of the objective function (Equation (1)) corresponding to the
initial solution
5. Define an ordered path for visiting the blocks [bjb ¼ 1, 2, . . . , B] by ordering the
blocks on decreasing average grade of b
6. Set i ¼ 0 and the new solution acceptance counter C ¼ 0
7. Set i ¼ i þ 1 and let block b be the i’th block in the ordered path defined at step 5
above
8. Perturb the current solution S to create a new solution S# in the neighbourhood of
S as follows:
 set S# equal to S
 float the circle to b and determine b
 update S# by setting the ore processing option for any block within the circle
b whose current processing option is k to k þ 1
9. Assess the impact of the perturbations by calculating the objective function value
for this new solution S#
10. Accept the perturbations if O(S#)  O(S) > 0, whereupon,
 the current solution is updated into a new dominant solution, i.e., S ¼ S# and
O(S) ¼ O(S#)
 C¼Cþ1
11. Continue the neighbourhood search as long as new solutions that improve the
objective function may be found, i.e. if i < B then go to step 7, or if i ¼ B and
C > 0 go to step 6
12. If k < K  1, then set k ¼ k þ 1 and go to step 6
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 107

A key part of this heuristic is that ore processing decisions for blocks currently within
a dig-line may not be altered during further perturbations until k is increased.
Computationally we found that excluding any blocks whose current processing option
was <k from b in step 8 above resulted in better solutions. This stems from the fact
that, as the processing options are arranged in increasing cost/recovery order, any
block with a current processing option <k has already been considered (and rejected)
for option k and hence is unlikely to improve the solution when processed using
option k þ 1.
For LSDLP the dig-line is built circle by circle, thus, the heuristic can also be
considered as a pseudo-constructive heuristic. It is important to note that, as LSDLP
uses an ordered path, the algorithm is deterministic. Therefore, the solution is unique
for a given expected payoff weight and circle diameter. However, there is no
guarantee that the proposed ordered path will maximise the objective function.
Additional solutions can be obtained by changing the expected payoff weight l,
changing the circle diameter d, re-ordering the K processing options in decreasing
cost/recovery order, or by stochastically varying the floating circle path. Computa-
tionally, we observed that solutions generated with the ordered path as defined in step
5 above tended to have the highest objective function values.

4. DESCRIPTION OF THE HYBRID LOCAL


SEARCH HEURISTIC

A practical analogy for LSDLP is that the blocks are extracted circle by circle. In this
section, we propose a hybrid local search heuristic that includes a ‘‘backfill and
excavate’’ perturbation mechanism. We will refer to this hybrid heuristic as LSHDLP,
which proceeds as follows:
1. Select the expected payoff weight l, a circle of diameter d that satisfies mining
equipment constraints, the backfill parameter values DI, Dmin, and !, and the
stopping parameter value I
2. Define an ordered path for visiting the blocks [bjb ¼ 1, 2, . . . , B] by ordering the
blocks on decreasing average grade of b
3. Set k ¼ 1
4. Generate an initial solution S ¼ [s(b) ¼ k, b ¼ 1, . . . , B], i.e. initially choose to
treat all blocks as waste (processing option 1)
5. Compute the value of the objective function (Equation (1)) corresponding to the
initial solution
6. Set @ ¼  I, the iteration counter C1 ¼ 0, the excavated circle counter C2 ¼ 0, and
the number of circles to backfill j ¼ 0
7. Set C1 ¼ C1 þ 1
108 A.J. RICHMOND AND J.E. BEASLEY

8. If C1 > 1, then determine j using a hybrid function of C1 and C2:


   ! 
C1  1
j ¼ max Dmin ; C2  ðC2  Dmin Þ ð8Þ
DI  1

9. Set C3 ¼ C2  j
10. Set n ¼ 0
11. Perturb the current solution S by backfilling j random floating circles (for which
the current processing option for all blocks within the circle b is k þ 1), i.e. by
setting the ore processing option for all blocks within the circle b to k to create a
backfilled solution Sn
12. Set i ¼ 0 and the excavation acceptance counter C4 ¼ 0
13. Set i ¼ i þ 1 and let block b be the i’th block in the ordered path defined at step 2
above
14. Excavate by perturbing the solution Sn to create a new solution Sn þ 1 in the
neighbourhood of Sn as follows:
 set Sn þ 1 equal to Sn
 float the circle to b and determine b
 update Sn þ 1 by setting the ore processing option for any block within the circle
b whose current processing option is k to k þ 1
15. Assess the impact of this excavation by calculating the objective function value
for this new solution Sn þ 1
16. Accept the perturbations if O(Sn þ 1)  O(Sn) > 0, whereupon, n ¼ n þ 1,
C3 ¼ C3 þ 1, and C4 ¼ C4 þ 1
17. Continue the excavation as long as new solutions that improve the objective
function may be found, i.e. if i < B then go to step 13, or if i ¼ B and C4 > 0 then
go to step 12
18. Assess the impact of this backfill and excavation by calculating the objective
function value for the new solution Sn
19. Accept the perturbations if O(Sn)  O(S) > 0, whereupon,
 the current solution is updated into a new dominant solution, i.e. S ¼ Sn and
O(S) ¼ O(Sn)
 additional iterations are allowed for the search to continue, i.e. @ ¼ @  I
 the excavated circle counter is updated, i.e. C2 ¼ C3
20. If the perturbations are rejected, then reduce additional iterations, i.e.
@ ¼@ þ 1
21. Continue backfilling and excavating until the stopping criterion is met, i.e. if
@ < I then go to step 7
22. If k < K  1, then set k ¼ k þ 1 and go to step 6
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 109

The stopping criterion I has been adopted from the stochastic evolution algorithm [13]
to reward the heuristic with additional iterations whenever improved solutions are
found. The parameter I represents the expected number of iterations the heuristic
needs to find an improvement in the objective function with respect to the current
solution. If I is too small, then the algorithm may not have sufficient time to improve
the solution, and if too large the heuristic may waste too much time in the later stages
of the search [13]. In the computational results reported below the stopping parameter
value I used was 10.
LSHDLP essentially embeds a local search heuristic (LSDLP or the excavation
stage) as part of the perturbation mechanism for a second local search heuristic to
restrict the solution evaluation to local optima of LSDLP. Thus, if I ¼ 0, then
LSHDLP is simply LSDLP. Talbi [14] refers to this type of heuristic as a low-level
relay hybrid. Martin and Otto [15] described a hybrid heuristic of the same class for
the travelling salesman problem in which a large perturbation or ‘‘kick’’ was applied
to the current simulated annealing solution and local search was then employed to find
a new locally optimal solution for objective function evaluation. In LSHDLP, the large
perturbation is supplied by the stochastic backfill stage.
It is important to note that the backfilling strategy in step 11 above may result in
dig-lines that violate mining equipment constraints. For example, consider the dig-
line solution composed of three circles shown in Figure 3(A). If Circle 1 was ran-
domly selected to be backfilled, then the resulting dig-line solution is Circles 2 and 3.
However, simply removing Circle 1 from the dig-line solution, shown in Figure 3(B),
does not produce the correct result. The correct result, shown in Figure 3(C),
requires extending this intermediate solution to the perimeters of Circles 2 and 3. To
overcome this problem in LSHDLP, after backfilling each circle in step 11 above, the
floating circle needs to be re-located to blocks (in close proximity to the backfilled
circle) that previously led to improved solutions and the dig-line extended to the circle
perimeter. Furthermore, backfilling a circle in the middle of a dig-line does not lead to

Fig. 3. Simple example of backfilling a dig-line: (A) initial solution; (B) intermediate solution;
(C) backfilled solution (Note: black represents areas within the dig-line).
110 A.J. RICHMOND AND J.E. BEASLEY

any change in the solution, thus, backfilling occurs pseudo-randomly from the
perimeter of a dig-line inwards.
If the dig-line is completely backfilled at each iteration, then LSHDLP simply
corresponds to a multi-start strategy for LSDLP, with each iteration producing an
identical solution. Furthermore, in later iterations, we observed that backfilling a large
part of the dig-line often involves considerable computational effort with no
improvement in the solution. Consequently, the hybrid function (Equation (8)) in step
8 above was adopted. We found that it considerably enhanced the computational
efficiency of LSHDLP. Equation (8) comprises a power model using a parameter (the
power) ! > 0 to decide the number of circles to be backfilled, the effect of which is:
 for iterations (values of C1) in the interval (2, DI) the number of floating circles
selected for backfilling is iteration dependent
 for iterations (values of C1) > DI a constant number Dmin of floating circles are
selected for backfilling.
Note here, that as shown in step 11 above, when a block b is selected for backfilling
blocks in its floating circle b are also backfilled. Thus, large parts of the dig-line were
backfilled during early iterations (C1 < DI), when significant improvements in the
objective function value were observed. Then, during later iterations (C1 > DI), minor
backfilling (Dmin circles) intermittently permits further small improvements in the
objective function value. In the computational results reported below we used
DI ¼ 100, Dmin ¼ 10, and ! ¼ 0.05.

5. COMPUTATIONAL RESULTS

In this section we illustrate the working of the three heuristics (FCDLP, LSDLP and
LSHDLP) proposed for the dig-line problem on test problems created to represent a
260 m  300 m  5 m bench in a gold mine. The location of blasthole gold assays
(measured in grams per ton) are shown in Figure 4. These values were drawn from the
Walker Lake reference dataset [16] allowing direct quantitative assessment of the
methodologies employed. The mining parameters for this problem are shown in
Table 1. Dump Leach (DL) processing involves placing blasted rock material on a pad
to extract the gold using a cyanide leaching technique. In the Heap Leach (HL)
process the blasted rock material is coarsely crushed prior to being placed on the pad.
Crushing the ore involves an additional cost, but increases the proportion of gold
recovered due to the enhanced surface area exposed to the leaching agent. The vast
majority of gold in finely crushed ore can be extracted using an expensive carbon-in-
pulp (CIP) milling process.
Twenty-five sequential Gaussian simulations of gold grades were generated for
the study area on a 1 m  1 m horizontal support using the blasthole assays as
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 111

Fig. 4. Blasthole gold assays (g/ton).

Table 1. Gold mining parameters (Note: qb ¼ 12.5 tons 8b and m ¼ $9.00/g are constants).

Processing option Waste Dump Leach Heap Leach CIP

Cost ($/ton) 1.50 3.525 5.775 10.275


Recovery (%) 0 45 70 95

conditioning information. Thus, this test problem involves 78,000 blocks, and the
number of feasible solutions (excluding equipment constraints) is K78000. Local
expected gold grades and the probability of exceeding a threshold gold grade (1.20 g/
ton) are plotted in Figures 5 and 6 respectively. The former can be used to visually
assess the local expected payoff, and the latter provides an indication of the local
downside risk. For example, any block with an expected grade greater than or equal to
1.20 g/ton gold can be mined profitably as CIP ore. Given that the expected gold grade
exceeds 1.20 g/ton, Figure 6 plots the probability that the block will be profitable.
Note that there are numerous blocks in the eastern and southern parts of the study area
112 A.J. RICHMOND AND J.E. BEASLEY

Fig. 5. Expected gold grades (g/ton).

Fig. 6. Probability of block gold grade exceeding 1.20 g/ton.


FINANCIALLY EFFICIENT DIG-LINE DELINEATION 113

Table 2. Mining scenarios.

K Ordering from k ¼ 1 to k ¼ K

Scenario 1A 2 Waste, CIP


Scenario 1B 2 CIP, Waste
Scenario 1C 2 Mixeda
Scenario 2A 4 Waste, DL, HL, CIP
Scenario 2B 4 CIP, HL, DL, Waste

Note. aWaste, CIP in eastern half and CIP, Waste in western half.

that are expected (but not guaranteed) to be profitable. The simulated gold grades
together with the information shown in Table 1 were used to calculate FS.
In this paper we consider two problem instances: DLP1 in which CIP is the only
ore processing option available; and DLP2 in which all three ore processing options
(DL, HL and CIP) are available. Unless specifically stated otherwise, a circle of
diameter 10.4 m was used to satisfy mining equipment requirements. Furthermore, as
the DLP solutions are sensitive to the ordering of the K processing options, we will
consider the specific orderings listed in Table 2. It is important to note here that our
computational experience has been that inferior results are obtained when the ore
processing options were not ordered on increasing or decreasing cost/recovery.
In the first problem instance DLP1, for a gold price of $9.00/g, the marginal cut-off
gold grade for CIP is 1.20 g/ton. For the second problem instance DLP2, the optimal
cut-off gold grades are 0.5 g/ton, 1.0 g/ton, and 2.0 g/ton for DL, HL, and CIP
respectively. Using these cut-off gold grades and assuming free selection the expected
payoffs for DLP1 and DLP2 are $4,121,927 and $4,559,189 respectively. These
expected payoffs are optimal for a risk-neutral decision-maker when no mining
equipment constraints exist. The corresponding block processing decisions are plotted
in Figure 7. Note that operationally undesirable transitions between block processing
decisions are common in these plots and justifies the use of the proposed mining
equipment constraint (Equation (7)).

5.1. Efficient Frontiers


To generate the efficient frontier with the heuristics, fifty different values of l were
considered, varying from 0.99 to 0.99999998. The risk function parameters were fixed
at a ¼ 2 and t ¼ $4,000,000 for DLP1, and a ¼ 2 and t ¼ $4,400,000 for DLP2. The
computation times shown in Table 3 are for our heuristics as coded in FORTRAN and
run on a PC with a 1.7 GHz processor. Note that LSHDLP is significantly slower
computationally than LSDLP.
The minimum risk efficient solutions generated by LSDLP and LSHDLP are
plotted in Figures 8 and 9 respectively. Note that, for Scenarios 1A and 1B, the
114 A.J. RICHMOND AND J.E. BEASLEY

Fig. 7. Cut-off grade solutions: (A) DLP1; (B) DLP2.

Table 3. Computational times (seconds) for various LSDLP and LSHDLP implementations.

Scenario LSDLP LSHDLP

1A 5,192 99,151
1B 3,505 87,148
1C 4,434 –
2A 11,931 –
2B 9,437 –

solutions generated by the two heuristics appear visually similar. However, LSHDLP
solutions tend to contain fewer locations selected as CIP. Figure 8(C) shows
the solution for Scenario 2A, in which the CIP dig-line is nested within the HL dig-
line, which in turn is nested within the DL dig-line. Consequently, to honour mining
equipment constraints, this solution implies that CIP material at the transitions must
be mined prior to HL material, which in turn must be mined prior to DL material, and
so on. For the solution shown in Figure 8(D) the implicit mining order is reversed.
Figure 8(E) demonstrates the flexibility of perturbation mechanism in the heuristics,
in that the ordering of the K processing options can vary in different regions of the
problem area.
Figure 10 presents the heuristic efficient frontiers for LSDLP. Note that the
heuristic efficient frontiers vary with the mining scenario. In Figure 10(A), the
heuristic efficient frontier associated with Scenario 1B dominates all Scenario 1A and
1C solutions. Similarly, in Figure 10(C), Scenario 2B solutions dominate Scenario 2A
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 115

Fig. 8. LSDLP minimum risk efficient solutions: (A) Scenario 1A; (B) Scenario 1B; (C) Scenario 2A;
(D) Scenario 2B; (E) Scenario 1C.
116 A.J. RICHMOND AND J.E. BEASLEY

Fig. 9. LSHDLP minimum risk efficient solutions: (A) Scenario 1A; (B) Scenario 1B.

Fig. 10. LSDLP efficiency plots: (A) expected payoffs for DLP1; (B) actual payoffs for DLP1; (C) expected
payoffs for DLP1; (D) actual payoffs for DLP1.
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 117

solutions. Experimental observations indicated that varying the ordering of the K


processing options in different regions of the problem area (Scenario 1C) produced
heuristic efficient frontiers that are bounded by the two extreme cases, Scenarios 1A
and 1B. We would comment here that, as far as we are aware, this paper is the first in
the literature to present algorithms for dig-line optimisation that allow the information
in Figure 10 to be produced. If the bench were to be mined according to the solutions
corresponding to points on the efficient frontiers, then the actual payoffs can be
calculated from the reference dataset and are shown here in Figures 10(B) and 10(D).
In both of these diagrams, note that the maximum actual payoff for a given scenario is
not associated with the corresponding highest risk (expected payoff) solution.
Note here that in Figure 10 if we examine the relative magnitudes of the values
given for downside risk (of the order of thousands of millions) and expected payoff (of
the order of millions) we might at first sight wonder why anyone would ever bother to
do any mining at all in this gold mine, since risk seems to dramatically outweigh
payoff. The reason for the disparity in values here lies in the definition of downside
risk (Equation (4)) which, since we use a ¼ 2, is actually expressed in squared dollars,
whereas expected payoff is expressed in dollars.
Figure 11 presents the LSDLP and LSHDLP efficient frontiers and the FCDLP
solutions for Scenarios 1A and 1B. Alternate FCDLP solutions were generated by
varying the gold price from $6.50/gram to $11.00/gram in $0.25 increments. Note that
some inferior FCDLP solutions are not shown in Figure 11. In Figure 11(A), the
LSHDLP heuristic efficient frontier for Scenario 1A dominates all LSDLP solutions,
which in turn dominate most FCDLP solutions. Similar relationships can be seen in
Figure 11(C) for Scenario 1B. The actual payoffs associated with points on the efficient
frontiers are plotted in Figures 11(B) and 11(D). In both of these diagrams, note that,
except for FCDLP in Scenario 1A, the maximum actual payoff for a given scenario is
not associated with the corresponding highest risk (expected payoff) solution.

5.2. Financial and Production Implications


Solutions associated with the various mining scenarios discussed above have further
significant financial and production implications. Detailed production and cost
outcomes for the lowest and highest risk LSDLP efficient solutions for each mining
scenario are shown in Table 4. With regard to Table 4 and Figure 10 we observed that,
for the two dig-line test problems examined, mining waste material at transitions prior
to mining ore material:
 minimises ore dilution,
 maximises profit for a given level of financial risk,
 frees milling capacity, and
 minimises upfront operating costs as a direct result of lower CIP tonnages.
118 A.J. RICHMOND AND J.E. BEASLEY

Fig. 11. Efficiency plots: (A) expected payoffs for Scenario 1A; (B) actual payoffs for Scenario 1A;
(C) expected payoffs for Scenario 1B; (D) actual payoffs for Scenario 1B.

Table 4. Production details for various efficient solutions and mining strategies – circle diameter ¼ 10.4 m.

Mining Expected Total operating Expected cost Expected ore Expected ore CIP
strategy production (oz) costs ($) per ounce ($/oz) dilutiona (%) lossb (%) tonnage

Maximum profit
1 32,497 5,074,071 156.1 45.33 9.74 411,575
2 31,687 4,835,061 152.6 43.36 10.53 384,338
3 32,081 4,948,259 154.2 44.31 10.15 397,238
4 32,469 4,633,560 142.7 29.20 5.76 207,750
5 31,611 4,380,761 138.6 26.39 6.71 183,238
Minimum risk
1 32,402 5,075,058 156.6 45.54 9.88 411,688
2 31,302 4,759,377 152.0 42.97 10.95 375,713
3 31,753 4,890,125 154.0 44.11 10.54 390,613
4 32,373 4,626,818 142.9 29.21 5.93 205,388
5 31,130 4,276,915 137.4 25.68 7.28 172,825

Note. aPercentage of locations classified as CIP that are not profitable based on the CIP marginal cut-off
grade 1.20 g/ton.
b
Percentage of locations classified as waste that are profitable based on marginal cut-off grades.
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 119

For the same mining scenario, low-risk solutions tend to have lower gold production
and operating costs when compared to high-risk solutions.

5.3. Circle Diameter


To assess the impact of utilising mining equipment of varying dimension, heuristic
efficient frontiers for Scenario 1 were also generated with LSDLP using circle
diameters of 5.0 m and 14.5 m. Figure 12 plots the minimum risk efficient solutions
for Strategy 1A. Figure 13 summarises the results in an efficiency plot. Due to
increased dilution, the larger the circle diameter d employed (the greater the mining
equipment constraint) the lower the expected payoff and the higher the financial risk
of the corresponding solutions. This is demonstrated in Figure 13(A) by the heuristic
efficient frontiers associated with a circle of diameter 5.0 m dominating solutions
associated with a circle of diameter 10.4 m, which in turn dominate solutions
associated with a circle of diameter 14.5 m. In practice, if mining equipment of
different capacity is available, these types of plots could be used to determine the
equipment size which is financially efficient for a given DLP. However, in this paper,
the mining and processing costs for the K options were kept constant for all LSDLP
implementations. In reality, due to economies of scale, there is commonly a negative
correlation between the size of the mining equipment employed (circle diameter d)
and the mining cost per unit weight for processing option k (ck). These variable
operating costs would therefore need to be incorporated into the optimisation and

Fig. 12. Minimum risk efficient solution for Strategy 1A: (A) circle diameter ¼ 5.0 m; (B) circle
diameter ¼ 14.5 m.
120 A.J. RICHMOND AND J.E. BEASLEY

Fig. 13. LSDLP efficiency plots for DLP1 with various circle diameters: (A) expected payoffs; (B) actual
payoffs.

decision process. The actual payoffs associated with points on the efficient frontiers
are plotted in Figure 13(B). Note that, the maximum actual payoff for a given scenario
is not associated with the corresponding highest risk (expected payoff) solution.

6. CONCLUSIONS

The computational results presented in the previous section demonstrate a number of


methods proposed for the DLP. A 2D adaptation of the floating cone algorithm (FCDLP)
produced a series of nested dig-line solutions by varying the commodity prices, a
common practice in open pit optimisation. These dig-line solutions were subsequently
evaluated for the DLP, identifying most solutions as being financially inefficient.
Downside risk was explicitly incorporated into the DLP optimisation process in a local
search algorithm (LSDLP) to define a heuristic efficient frontier. However, the ordered
floating circle path and perturbation mechanism employed in LSDLP significantly
restricts the solution space explored. A hybrid local search algorithm (LSHDLP) was
proposed to evaluate alternate LSDLP solutions generated from partial (backfilled) dig-
line solutions. Computationally, it was observed that stochastically backfilling the dig-
line and floating the circle on a path ordered on decreasing/increasing metal content
resulted in the highest objective function values. For the dig-line test problems
examined, the LSHDLP heuristic efficient frontiers dominated the corresponding
LSDLP heuristic efficient frontiers. However, under practical mining conditions, the
computational effort (24 h and greater of CPU time) for LSHDLP in the DLP case study
may be considered excessive. In such circumstances, the number of expected payoff
weights l used to define the heuristic efficient frontier could be reduced. If the risk
averse characteristics of the decision maker are known, then only a single solution needs
to be generated [6]. Alternatively, the stopping parameter I could be reduced to a level
which provides solutions in an acceptable timeframe.
FINANCIALLY EFFICIENT DIG-LINE DELINEATION 121

The attraction of the proposed heuristics is that they are effectively independent of
the adopted objective function and floating object. Alternate efficiency models can be
considered by substituting the corresponding risk measure in Equation (4). If circle
geometry is not suitable for the ore delineation problem in question, for example in
stope delineation, then other three-dimensional floating objects may be appropriate.
Finally, we would comment here that, as far as we are aware, the LSDLP and
LSHDLP algorithms presented in this paper are the first in the literature that explicitly
consider financial risk in the optimisation of dig-lines. Furthermore, these algorithms
identify a heuristic efficient frontier for the dig-line problem in a computationally
effective manner.

REFERENCES

1. Srivastava, R.M., Hartzell, D.R. and Davis, B.M.: Enhanced Metal Recovery Through Improved Grade
Control. In: Proceedings of 23rd APCOM, Littleton, USA, 1992, pp. 243–249.
2. Srivastava, R.M.: Minimum Variance or Maximum Profitability? CIM Bull. 80(901) (1987), pp. 63–68.
3. Glacken, I.C.: Change of Support and Use of Economic Parameters for Block Selection. In: E.Y. Baafi,
N.A. Schofield (eds.): Geostatistics Wollongong 1996. Kluwer Academic Publishers, London, 1997,
pp. 811–821.
4. Richmond, A.J.: Maximum Profitability With Minimum Risk and Effort. In: Proceedings of 29th
APCOM, Beijing, China, 2001, pp. 45–50.
5. Richmond, A.J.: Applying Four Different Risk Models in Local Ore Selection. Nat. Resourc. Res. 11(4)
(2002), pp. 299–314.
6. Richmond, A.J.: Financially Efficient Ore Selections Incorporating Grade Uncertainty. Math. Geol.
35(2) (2003), pp. 195–215.
7. Richmond, A.J. and Beasley, J.E.: An Iterative Constructive Heuristic for the Ore Selection Problem.
J. Heur. (submitted for publication).
8. Norrena, K. and Deutsch, C.V.: Determination of Dig Limits Subject to Geostatistical, Economic, and
Equipment Constraints. In: Proceedings of SME Meeting, Denver, CO, February 27–28, 2001.
9. Norrena, K.P. and Deutsch, C.V.: Optimal Determination of Dig Limits for Improved Grade Control.
In: Proceedings of 30th APCOM, Colorado, USA, 2002, pp. 329–339.
10. Fishburn, P.C.: Mean-Risk Analysis With Risk Associated With Below-Target Returns. Am. Econ. Rev.
67(2) (1977), pp. 116–126.
11. Goovaerts, P.: Geostatistics for Natural Resources Evaluation. Oxford University Press, New York,
1997.
12. Lemieux, M.: Moving Cone Optimizing Algorithm. In: A. Weiss (ed.): Computer Methods for the 80’s
in the Mineral Industry. SME of AIMMPE, New York, 1979, pp. 329–345.
13. Saab, Y.G. and Rao, V.B.: Combinatorial Optimization by Stochastic-Evolution. IEEE Trans. Comp.
Des. Integr. Circ. Sys. 10(4) (1991), pp. 525–535.
14. Talbi, E.G.: A Taxonomy of Hybrid Metaheuristics. J. Heur. 8 (2002), pp. 541–564.
15. Martin, O.C. and Otto, S.W.: Combining Simulated Annealing with Local Search Heuristics. Ann.
Oper. Res. 63 (1996), pp. 57–75.
16. Isaaks, E.H. and Srivastava, R.M.: An Introduction to Applied Geostatistics. Oxford University Press,
New York, 1989.

You might also like