Researchers propose a new method for choosing the best state representation when observations are incomplete or unreliable. The approach balances accuracy with penalties for feature cost and instability. In experiments on three datasets, the method improved performance in some cases but not universally. --- Why it matters: This matters to AI researchers because it addresses a common challenge: how to make models robust when they don't have complete or reliable data. This work provides a new tool for tackling this problem. Source: https://arxiv.org/abs/2608.15095