Feasibility of model-based boundary optimization in animals and learning without prior knowledge

Ascertain whether animals possess the detailed knowledge of environmental statistics and computational capacity to perform model-based calculations (e.g., Markov decision processes or dynamic programming) to derive optimal choice thresholds, and determine how a decision boundary can be learned when such knowledge is unavailable.

Background

Model-based solutions for optimal decision boundaries assume that agents know (or can rapidly infer) environmental transition structures and perform complex computations to maximize reward or reward rate. The paper questions whether animals have access to such knowledge or the capacity for these calculations, and emphasizes that, in their absence, the problem of how boundaries are learned remains unsolved—motivating model-free approaches.