Figure 6. Walk decisions are stochastic events whose rate accumulates evidence from recent encounters.

(A) Distribution of stop duration following any encounter, or the most recent encounter or clump before a walk. (B) Cumulative encounter counts since walk onset, for various stop durations. (C) Average encounter frequency versus duration of stop bout. (D) In the accumulated evidence model of walk decisions, ${\lambda }_{s\to w}\left(t\right)$ increases at every encounter onset, before decaying to baseline. This is modeled by ${\lambda }_{s\to w}\left(t\right)={\lambda }_0+\mathrm{\Delta }\lambda \int e^{t-t^{\text{'}}/{\tau }_w}w\left(t^{\text{'}}\right)d{t}^{\text{'}}$. Median of estimated parameters are λ₀=0.29s^-1, Δλ=0.41s^-1, τ_w=0.52s (Figure 6—figure supplement 1). (E-G) Analogs of A-C using data generated by the model. (H) Analogs of E-G for the last encounter model, in which the walk rate increases to a fixed value at each encounter before decaying to baseline. (I) Analogs of E-G for the encounter duration model, in which ${\lambda }_{s\to w}\left(t\right)$ switches between a high value during encounters and low value during blanks.

Figure 6—figure supplement 1. Distributions of estimated parameters for stop-to-walk models.

(A) The estimated parameters for the accumulated evidence model of stop-to-walk transitions, using 500 distinct subsets of the data. (B–C) Same for the other two models, which did not explain the data well. In the last encounter model, the $\mathrm{\Delta }\lambda$ parameter was consistently estimated at its bound (effectively zero). In both models, the predictions were poor (Figure 6H-I).