Inverse Decision Modeling:Learning Interpretable Representations of Behavior
Details
Title : Inverse Decision Modeling:Learning Interpretable Representations of Behavior Author(s): Daniel Jarrett, Alihan Hüyük, Mihaela van der Schaar
Summary
- Provides a unifying perspective on inverse decision modeling: a general framework for learning parameterized representations of sequential decision-making behavior. This is done by first introducing a forward problem \(F\) which is normative and subsumes a lot of control problems, meanwhile the inverse problem \(G\) which is descriptive and subsumes work on imitation and reward learning. The paper provides a formalism on behaviour, the planner and inverse planner both of whom have normative and descriptive parameters, the inverse problem, and other relevant concepts like behaviour project, the inverse decision model.
- Descriptive models are those that capture observable decision-making behavior as-is (e.g. an imitator policy in behavioral cloning), and by “normative” models, are those that specify optimal de- cision-making behavior (e.g. with respect to some utility function).
- Observed behaviour is looked from a bounded rationality perspective.
- Inverse Bounded Rational Control is introduced as a concrete example of the framework.