Figure 4.

(A) The distribution of prediction errors in the semi-Markov TD model (Daw et al., 2006) for three logarithmically spaced cue-reward interval durations (short, medium, long). The timing signal is corrupted by zero-mean Gaussian noise with standard deviation proportional to the interval duration. (B) The average prediction error under each distribution. A symmetric prediction error distribution implies that the average prediction error is invariant to the interval duration. (C,D) Same format as panels A and B, but where the prediction error distribution is rectified slightly below 0. This asymmetric distribution produces average prediction errors that scale approximately linearly with the log interval duration.