Fig. 5
Fitting the multi-timescale model of Fig. 2 to the data. a A model with three timescales (τFast−1 = 2 trials, τFast−2 = 20 trials, and τSlow = 1000 trials) is fitted to the data by tuning the weights wFast−1 (black), wFast−2 (brown), and wSlow (red) of the different integrators for each session independently (wFast−1 + wFast−2 + wSlow = 1). b The weights of different timescales change over consecutive experimental sessions. The short timescales are dominant in early sessions, but the longer timescale becomes progressively more influential. The opposite trend is observed around session 160 of monkey F, probably due to the shortening of experimental sessions and longer inter-experimental-intervals (see Supplementary Figure 16 for more details). c Deviation from the matching law is correlated with the weight of the reward integration on a long timescale. Undermatching, computed over the last 50 trials of each block to ignore transients, is plotted against the fitted value of wSlow, the weight of the longest reward integration timescale. Both monkeys show significant correlations between undermatching (1 − slope) and wSlow. We found, however, this model-based analysis inconclusive, as the correlation is expected from the model structure (please see the text)
