Future Knowledge in Geometallurgical Mining Optimization

Mining and processing involves a lot of decision making, on capacity
building, mine scheduling, blending, process parameters, and
contracted sales. Traditionally stochastic mine planning and
predictive Geometallurgy use stochastic knowledge provided e.g. by
conditional geostatistical simulations of the conditional
distributions of ore properties to infere optimal decisions through
stochastic optimization. Stochastic knowledge is however no fixed
fact, but can rather increase by later aquisition of information,
automatically as a direct consequence of the operation itself, and
optionally through additional exploration.

The contribution shows with simple and easy to comprehend sand box
examples how and why such future knowledge and even the option to
obtain future knowledge already changes, what is an optimal decision
even before this knowledge is obtained. In case of optional knowledge,
the decision to obtain it and when, becomes an integral part of the
decision problem. This radically changes what algorithms can be
feasibly used to compute optimal decissions. Straight forward
stochastic optimization is not yet computationally feasible, for
situations with increasing knowledge. The state of the art for models
using increasing information is to use reinforcement learning based
heuristics.

This contribution explores the idea of making a stochastic
optimizating possible by exploiting certain structures of the mining
related increasing knowledge optimization problem. Possible speedup
are based on 1) inequality relations in stochastic optimization
allowing for advanced branch and bound techniques, 2) exploiting the
fact that certain values are equivalent in different branches which
simplifies comparisions and precomputation, and 3) explicit
computation of conditional expectations in a partial separation of the
processing optimization and the scheduling optimization.