and goal-tracking is favoured. When both systems are mixed, i.e.
with an intermediate v, the behaviour is more likely to oscillate
between sign- and goal-tracking, representative of the intermediate
group.
These results rely on the combination of two systems that would
independently lead to ‘pure’ sign-tracking or goal-tracking CRs.
Three tested variants of the model could reproduce these
behavioural results as well (see Figure S1): a combination of
Feature-Model-Free systems and simple Model-Free system
(Variant 1); a multi-step extension of Dayan 2006’s model [16]
giving a Pavlovian impetus for the lever (Variant 2); and a symmetrical version of this last model with two impetuses, one for
the lever, and one for the magazine (Variant 3) (see Methods).
Interestingly, a combination of Model-Based and classical Model-
Free (not feature-based : Variant 4) fails in reproducing these
results (see Figure S8). This is because both systems are proven to
converge to the same values and both would favour pure goaltracking,
such that varying their contribution has no impact on the
produced behaviours.
Thus, at this stage, we can conclude that several computational
models based on dual learning systems can reproduce these
behavioural results, given that the systems favour different