Figure 2. Poor fits from models that ignore learning-rate adaptiveness are easily identified in the estimation, but not the four-choice, task.

A & B. Mean log-likelihood associated with a fixed learning-rate model, per simulated trial from the estimation (A) or four-choice (B) task, aligned to change-points in the generative process. Lighter shades of gray represent data from simulated agents with higher levels of learning-rate adaptiveness. C–F. Learning rates (C & D) or inverse temperatures (E & F) inferred from model fits that exclude log-likelihood information from trials occurring 0–10 trials after change-points (abscissa) for estimation (C & E) and four-choice (D & F) tasks. The transient changes in A, C, and E evident for all but the least adaptive simulated agents reflect the fixed learning-rate model's inability to account for behavior just following change-points on the estimation task; no comparable effects are evident for the four-choice task.