| description abstract | Predictions of climate change can significantly affect the optimization of measures reducing the long-term risk for assets exposed to extreme events. Although a single climate model can be represented by a Markov stochastic process and directly integrated into the sequential decision-making procedure, optimization under epistemic uncertainty about the model is computationally more challenging. Decision makers have to define not only a set of models with corresponding probabilities, but also whether and how they will learn more about the likelihood of these models during the asset-management process. Different assumed learning rates about the climate can suggest opposite behaviors. For example, an agent believing, optimistically, that the correct model will soon be identified may prefer to wait for this information before making relevant decisions; on the other hand, an agent predicting, pessimistically, that no further information will ever be available may prefer to immediately take actions with long-term consequences. This paper proposes a set of optimization procedures based on the Markov decision process (MDP) framework to support decision making depending on the assumed learning rate, thus trading off the need for a prompt response with that for reducing uncertainty before deciding. Specifically, it outlines how approaches based on the MDP and hidden-mode MDPs, dynamic programming, and point-based value iteration can be used, depending on the assumptions on future learning. The paper describes the complexity of these procedures, discusses their performance in different settings, and applies them to flood risk mitigation. | |
