2024

Moss J. Robert; Kozlova Mariia; Corso Anthony; Caers Jef

To solve sequential decision-making problems in practice, modelling assump-tions are made to make the problems tractable. A partially observable Markov decision process (POMDP) is a mathematical framework for sequen-tial decision-making and is solved using planning or reinforcement learning algorithms. In this work, we develop a framework to analyze the sensitiv-ity of POMDPs across various model fidelities. We introduce the POMDP model-fidelity framework (PMFF), where fidelities of the state uncertainty (i.e. belief), environment, planner, and inference algorithm are analyzed. We use Simulation Decomposition (SimDec) to quantitatively study the sensitiv-ity in the variance of the output planning performance metrics (e.g. returns and accuracy) given different input model fidelities. We apply PMFF and Sim-Dec to a real-world case study of critical mineral exploration, and results for this example suggest that planning with lower-fidelity models may be sufficient for decision-making.