Overview
Active feature acquisition asks a deliberately practical question: when every feature has a cost, which information should a model buy before it predicts? AFABench provides a common experimental framework for evaluating that question across static, myopic, and reinforcement-learning-based policies.
The benchmark treats feature acquisition as a sequential decision problem. At each step, a policy observes the features already acquired, chooses whether to purchase more information, and is evaluated by the resulting trade-off between predictive performance and acquisition cost. This framing makes it possible to study non-myopic behavior, where a feature can be valuable because it changes what should be acquired next rather than because it is immediately predictive.
My Role
I worked on the benchmark and the surrounding research questions as part of my completed MSc thesis in the Machine Learning and Decision Making Lab at Chalmers. The project shaped my current interest in sequential decision-making under partial observability, especially in settings where information itself is an action with a cost.