Abstract
We propose an effective algorithm for black-box optimization problems without derivatives in the presence of output constraints. The proposed algorithm is illustrated using a realistic short-term oil production case with complex functions describing system dynamics and output constraints. The results show that our algorithm provides feasible and locally near-optimal solutions for a complex decision-making problem under uncertainty. The proposed algorithm relies on building approximation models using a reduced number of function evaluations, resulting from (i) an efficient model improvement algorithm, (ii) a decomposition of the network of wells, and (iii) using a spectral method for handling uncertainty. We show, in our case study, that the use of the approximation models introduced in this paper can reduce the required number of simulation runs by a factor of 40 and the computation time by a factor of 2600 compared to the Monte Carlo simulation with similarly satisfactory results.