Fig. 2: Fit of choice behavior of control monkeys using various RL models.
From: Contribution of amygdala to dynamic model arbitration under uncertainty
a Schematic of the RL model with two parallel learning systems, showing an example trial in which stimulus A appeared on the left side. In the static model, a constant ω is assumed to be fixed for each block of trials. In the dynamic models, ω is updated on each trial according to the relative reliability of the two systems. In a more general dynamic model, the fixed parameter ρ (estimated for each subject) adjusts the baseline ratio of two value signals. For ρ = 0.5, the Dynamic ω-ρ model reduces to the Dynamic ω model. The overall value (OV) of a left or right saccade is determined as a weighted combination of action and stimulus values. b Comparison of goodness-of-fit across models. Plotted is the mean negative log-likelihood over all cross-validation instances for each task: What-only (black), What/Where (gray). Numbers in parentheses indicate McFadden R2 (Eq. 20). c An example Where block in the What/Where task and estimated arbitration weight from the Static ω model (dotted line), and arbitration weights (ω, dashed line) and effective arbitration weights (Ω, solid line) from the best model (Dynamic ω-ρ model). In this example, ρ = 0.61, effectively biasing behavior toward a stimulus-based strategy. In this block, rightward action (R) was a better option than leftward action (L) before reversal (rev, horizontal dashed line). d Average trajectory of Ω from the Dynamic ω-ρ model during different tasks and blocks: What-only (solid), What (dashed), and Where (dotted). Different colors correspond to different reward schedules: 80/20 (black), 70/30 (dark gray), 60/40 (light gray). <rev> indicates reversal (horizontal dashed line), with positions normalized across blocks. e Relationship between Ω (block-averaged) and median reaction time for a given block during the What/Where task. Reported are Spearman’s correlation coefficient r and its p-value (two-sided) for all blocks during the What/Where task. Source data are provided as a Source data file.
