Missingness-MDPs: A New Framework for Planning Under Missing Data

Researchers have recently unveiled a new kind of partially observable Markov decision process (POMDP), called missingness-MDPs or miss-MDPs, which focuses on missing data theories. In this model, the observation function serves as a missingness function, indicating the probability of certain state features being unseen. They categorize missingness into three types: missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). The main issue is figuring out nearly optimal policies for a miss-MDP without knowing the missingness function, relying on datasets of action-observation trajectories. To achieve guarantees of optimality, understanding the missingness function is crucial, which can be difficult for general POMDPs. The findings are detailed in a paper on arXiv with the ID 2605.12262.