Generalized comparison of non-linear expected utility and active inference agency

Determine a generalized, formally precise method for comparing the careful (risk-averse) and explorative (information-seeking) aspects of agency induced by non-linear utility functions U(R(s)) in expected utility theory with the behavior of agents that minimize Expected Free Energy in active inference, within the shared Markov Decision Process and Partially Observable Markov Decision Process frameworks.

Background

The paper evaluates the "subsumption thesis"—the claim that expected utility theory is subsumed by active inference—by comparing agent behaviors in MDPs and POMDPs and presenting examples that challenge a straightforward subsumption. In microeconomics, expected utility commonly employs non-linear utility functions over rewards to capture risk attitudes, which can lead to behavior (e.g., carefulness and exploration) differing from linear expected utility formulations often used in reinforcement learning comparisons.

The authors note that while comparisons with linear utility are tractable, introducing non-linearity leads to an impasse: it is not clear how to compare the resulting agency—particularly the induced careful and explorative aspects—against active inference in a generalizable, formal manner. This motivates the need for a methodological framework to systematically compare such behaviors across shared agent-environment models.