Counterfactual equivalence for POMDPs, and underlying deterministic environments

Stuart Armstrong

Partially Observable Markov Decision Processes (POMDPs) are rich environments often used in machine learning. But the issue of information and causal structures in POMDPs has been…