In signal processing systems, the actions of decision makers or agents often rely on the certainty of state variables. This paper discusses how uncertainty affects decision performance and its trustworthiness. We explore the link between objectives and agents' knowledge, as well as the necessity of uncertainty representation from first principles.
First, under the assumption of a known environment distribution, a risk-neutral agent requires modeling the posterior distribution over the state, while a risk-averse agent can rely on a prediction set and a worst-case decision rule without losing optimality.
Next, in the case of an unknown environment, we identify three complementary approaches to address the resulting epistemic uncertainty: calibration of a fixed predictor, credal (ambiguity) sets with distributionally robust optimization, and Bayesian inference over model parameters. The common thread is that reliable decisions require an uncertainty representation matched to the decision objective and the agent's knowledge profile, along with a guarantee of the utility the agent will actually obtain.
Blogger's Review: This paper delves into the quantification of uncertainty in decision-making, emphasizing the impact of risk preferences on decision strategies. By integrating Bayesian inference with robust optimization, it offers an effective pathway to achieve optimal decisions in complex environments. It serves as a valuable reference for researchers in signal processing and decision theory.