Beruflich Dokumente
Kultur Dokumente
solve LP-based models find a mathematically provable optimum solution. Few mine
optimization applications can make the same claim; they may produce a superior
solution to one produced by hand and they may refer to optimisation, but they
cannot demonstrate they have found the true optima.
When comparing options or conducting sensitivity analysis, the users of LOBOS can
be assured that the differences in value between options or the impact on value or
productivity associated with a change in a parameter are due to the option or
parameter change themselves. For this reason, LOBOS is ideal for strategic
optimisation and project sensitivity analysis, providing a powerful tool for evaluating
project risk.
Solutions based on non-exact optimisers must be used with caution when applied to
sensitivity and option analysis. Options or capacities can be compared from many
repeated runs, but there is no guarantee that the comparison is being made with a
true global optima. Instead, the comparison may be between alternative local optima
and not reflect the change in option value.
LOBOS is not application specific. Models are formulated based on your input and
then solved to a verifiable level of optimality. The technologies imbedded in other
mine optimization packages must be written for very specific applications. For
example, consider the common case of finding a production schedule that
maximizes NPV. The user of a scheduler not based on LP/MIP who wishes to apply
a new type of constraint cannot do so, not until changes in the underlying algorithm
have been made by the developer, most likely at the users cost. In contrast, since
LOBOS is a platform for LP/MIP formulation and solution, the user simply adds the
new constraint, adding new variables and relationships between variables. If you
require new modeling elements, MineSmith can rapidly provide the required
modeling interface as this involves no changes to the underlying technology.
the value of all tonnes mined). Once formulated, the emphasis in optimization is to
find the most suitable solution algorithm, yielding the best result in the least time. In
most LP/MIP Solvers, this process is automated, allowing the user to focus on the
model.
Some problems do not lend themselves to solution as a mathematical program. In
these cases specific algorithms have to be developed. An example is the Steiner
Tree problem in which there is a network in which only the position of some nodes is
known and the objective of the optimisation is to find the location of the Steiner
Points such that some aspect of the resulting network is optimal. In the case of
decline optimization, only the access points are known and the problem is to solve
for intermediate points in the declines path such that the total length is minimized.
This problem cannot be directly solved using a Network LP. Instead, application
specific algorithms have been developed using both Simulated Annealing (SA) and
Dynamic Programming (DP) techniques.
There is a fundamental difference in optimisation applications using LP/MIP and
other technologies:
For LP/MIP applications a library of standard solution algorithms are available
that can be applied to all LP/MIP problems. Some libraries are open source.
Others are commercial and proprietary, but in all cases the underlying LP/MIP
technology is well understood and published.
For specific applications that do not lend themselves to formulation as an
LP/MIP, specific algorithms must be designed using technologies such as
Genetic Algorithms (GA), Dynamic Programming and Simulated Annealing.
The resulting engine is application specific, highly proprietary and delivered as
a black box.
Therefore, the focus of LP/MIP based optimisation for industrial applications is on
model formulation as the underlying algorithms to optimise the model are not
dependent on the formulation. This is not to say that the formulation is not important
in terms of algorithmic efficiency, but that while the formulation might be improved,
the underlying Solver does not need to be altered, regardless of the variables,
constraints and objective included in the model.
In contrast, for applications not solvable by mathematical programming (or not
efficiently solvable), the emphasis is on development of the Algorithm itself using a
toolbox of optimisation techniques including GA, DP and SA. The emphasis in these
applications is not on the formulation. In fact, the formulation will not be evident to
users who will only see data entry and solution results. There is no addition or
modification of variables and constraints as these are fixed in the application. If new
forms of variables, constraints or objective are required, then the underlying
algorithm must be modified before they can be accounted for.
than a true optimum. A solver, such as Gurobi, that can tell you how close a
solution is to an upper limit on the true optimal solution is an Exact optimiser. This is
a key concept; most optimisation applications in the industry cannot actually
provide any estimate of how close they are to true optimality.
But isnt this just academic hair splitting? If an application provides a significantly
improved solution over one done by hand, isnt that good enough?
The answer depends on the problem being solved. Specifically, it depends on the
size and complexity of the solution space. A purpose built non-exact algorithm
designed for a problem that is of limited size and highly constrained might
consistently manage to find a solution that is close to optimal for that specific
problem, but when applied to a much larger and more loosely constrained problem, it
may fail miserably. A good example of this is the application of Genetic Algorithms to
activity scheduling to support detailed mine project management.
In a sequencing and scheduling problem a GA seeks a local maximum (remember,
we never know about optimality with a GA, which will simply continue to solve until it
can get no better solution in the allotted time). To be effective, the GA must include
another heuristic search algorithm that will move into other areas of the solution
space to see if a better solution can be found in that locality. For the final solution to
be optimal, all localities would have to be visited by the GA. In a relatively small
problem that is tightly constrained, the probability of visiting all important localities
might be high and the resulting solution may be globally optimal.
Unfortunately, with a GA there is no way of knowing if the local solution is optimal
and certainly no way to judge the adequacy of the search heuristic in hitting all the
local optima. If the GA is being applied to improving an activity schedule, this
limitation is not a real problem; as long as it is producing an improved solution it is
adding project value. But, serious problems could arise if this same GA was used for
a different application, especially one with a much larger and less tightly constrained
solution space.
A prime example of misuse of a GA-based activity scheduler is to apply it to option
analysis. GAs should never be used for option analysis as you have no idea what
you are actually comparing: the value of the solution at different local optima or the
actual value difference of, for example, two cutoffs at the global optimum. In fact, the
two top populations of solutions might be coming from the two best localities
examined, and these maybe significantly suboptimal in comparison to the true
optimal solution. A GA has no way of evaluating how far any solution is from the true
optimum. Therefore, while is perfectly reasonable to use the GA to improve an
activity schedule, there is significant risk associated with using it for strategic
analysis, especially when you have an alternative in LOBOS which can evaluate the
distance from the true optima. For further discussion of this issue we refer you to:
http://en.wikipedia.org/wiki/Genetic_algorithm#Criticisms
fixed and the resulting MIP can be applied to the optimization of activity scheduling
with all the power and flexibility offered by LP technology.
A feature of LP/MIP optimisation that differentiates LOBOS from purpose built
activity schedulers is the ability to optimise rates. This enables LOBOS to assign
high rates to objects whose rapid completion contributes to increased project value.
Likewise, LOBOS can slow or even halt an activity if this results in improved blending
or when production resources are needed elsewhere.
The value of varying rates is more apparent when LOBOS is applied to aggregations
of production objects such as a stoping block. Since a stoping block may include
many stopes and numerous production faces, rates will vary depending on how
many faces can be brought into production in parallel. LOBOS resolves this difficulty
by determining the rate of production and development of a stoping block based on
both the blocks capacity (if assigned) and the optimal allocation of production
resources throughout the operation.
Purchase a LOBOS license and take your strategic and tactical planning to
the next level
We provide consulting services where we apply all our professional expertise
and experience to develop a strategic or tactical plan for your business
A combination of the above, where we come in as consultants and
additionally provide a LOBOS installation and scenario tailored to your
business, and then step back and let you take it from there.