This work presents a new methodology based on Genetic Algorithms (GA) to solve stochastic optimization problems. The resulting algorithm named GOSP (Genetic Optimizer for Stochastic Problems) implicitly determines at execution time the number of scenarios required to get a reasonable solution provided the uncertainty currently present in the data of the problem to be solved.

One possible approach to solve stochastic problems is to use scenarios that represent realizations of the uncertain values. It has been traditional to determine that number of scenarios before the optimization process is conducted. When the generation of scenarios is cheap (in computational time) then it is not a problem to overestimate this number. Nevertheless, when the computational cost of generating scenarios is prohibitive, the decision about the number of scenarios becomes critical.

The mining industry is characterized by the high levels of uncertainty present in different aspects of its operations, but also by the size of the problems involved. Uncertainty in mining comes from different sources with economic geology being one of the most important. Unfortunately, the generation of scenarios to consider geological uncertainty is computationally expensive and the efforts to incorporate this uncertainty into the decision process suffer from underestimation effects that in turn will affect the quality of the solutions obtained.

The approach considered in this thesis is constructive in the sense that information is generated and considered at execution time rather than before any optimization is conducted. It also uses the learning capabilities of the genetic algorithm to reduce the computational cost of incorporating a new scenario into the process.

Numerical results are used to validate the approach and further experiments performed to characterize the strengths and limitations of the proposed methodology. Some modifications to the basic algorithm are proposed and numerical experiments conducted to check their effectiveness. Finally simplified mining problems are presented that demonstrates the applicability of GOSP and related algorithms for addressing data uncertainty.

The major findings of this thesis are:​

• A novel methodology to incorporate uncertainty within the decision making process in such a way that the number of scenarios to be used for the optimisation algorithm is not predetermined by the decision maker.
• The thesis recommends under which circumstances the GOSP methodology can be applied. GOSP methodology is superior for complex mining problems where the cost of scenario generation is costly. Classical approaches could be preferred in problems with less structural complexity.
• An attempt to generate a faster algorithm, which takes advantage of the learning capabilities of the GA under dynamic environments, produced a faster optimiser with the risk of suboptimality.
• In order to gain insight into the tradeoff between the optimisation and simulation components of GOSP, a new problem, partially based on the well known Traveling Salesman Problem (TSP), has been developed. The new model called City Tour Problem (CTP) was introduced and solved and the numerical results provided quantifiable evidence to establish the proper usage of GOSP.
• The introduction of a learning window concept within GOSP has been demonstrated. Successful implementation has been achieved using several different learning window sizes concluding that calibration is required for each given problem.