Novos algoritmos de programação dinâmica aplicados à cadeia de suprimento de mineração

Abstract

This work presents a stochastic optimization model for the mining supply chain, involving all its links, from mines to final clients. Due to the sequential nature of the decisions process and the long terms involved, Markov Decision Processes (MDP) modeling was used. Comparing with the existing works, it was found that many new stochastic optimization in mining papers were recently published, but none encompassed all supply chain links simultaneously, mainly given its complexity that makes it difficult to obtain the solution in reasonable computational times (curse of the dimensionality). To circumvent this problem, it was necessary to develop two new dynamic programming algorithms to reduce the state space and action space. TABA is based on temporal aggregation and expands the application of this type of algorithm to a new class of MDP. LSPSI works to reduce the action space, improving the search process introduced by in the policy evaluation step by PSI algorithm. These algorithms, when combined, brought great performance gains and overlapped the efficiencies of both more traditional, as value iteration and policy iteration, and more modern algorithms, as SDDP. In the end, the model and the algorithms are tested in a numerical problem of the mining chain and the results obtained prove the importance of stochastic optimization considering supply chain complete view and the efficiency of the proposed algorithms