Extended Data Fig. 3: Testing alternative foraging strategies.

(a) Illustration of the logistic regression model for predicting the switching decision of mice using a combination of the two main DVs, ‘Consecutive failures’ and ‘Negative value’, as well as additional DVs. Specifically, we tested 3 classes of additional DVs: 1) those relying on absolute time, 2) those relying on average reward rates, and 3) those that weigh recent evidence more strongly. The design matrix of the model thus consisted of the two main DVs, the time of each lick relative to the first lick of each bout (class 1), the average reward rate over 1, 3 and 10 previous bouts (class 2) and a version of the negative value DV that weighs recent evidence more heavily than the past ones (for class 3), such as: x_t+1 = (1 − α)·g(o_t+1)·x_t + α·c(o_t+1), with α = 0.8. (b) Deviance explained from a logistic regression model that predicts choice behavior based only on the 2 main DVs (left) and from the full model that also includes the additional DVs in (a). The central mark indicates the median across behavioral sessions (n = 42 sessions), and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. The whiskers extend to the most extreme data points. There was no significant difference between the deviance explained of the two models (two-sided Wilcoxon signed rank test: p = 0.22), indicating that the additional DVs do not improve the performance of the model. (c) Relative variance explained by each predictor of the full model for each behavioral session (n = 42 sessions across 21 mice, 2 sessions per mice). The dominant DV (the one with the maximum relative variance explained) was most often the ‘Consecutive failures’ (18 sessions), followed by the ‘Negative value’ (17 sessions), and finally the additional DVs (2 session for the absolute time, 2 sessions for average reward rate, 3 sessions for the weighted negative value).