Figure 4
From: Developmental asymmetries in learning to adjust to cooperative and uncooperative environments
Model of reinforcement learning with social preferences. (a) Individuals acquire monetary payoffs ($) from interactions in the games (cf. Figure 1). How they subjectively value each outcome can be influenced by their social preferences (Fehr & Schmidt, 1999), whose impact is proportional to the differences in payoff between interaction partners (δ). (b) Over the course of the games, individuals update their expectations (p) about the behaviour of the other players. The extent of updating is proportional to the prediction error (i.e., the difference between the expected and actual outcome) and learning rate λ71. (c) The probability that an individual chooses A (not B) is an increasing function (F) of the difference in weight of the two choice options (wA and wB; Methods). We use maximum likelihood methods to estimate learning rates λ for each age cohort and each game separately. In the baseline model, the weights reflect individuals’ monetary payoff from their interactions ($) and λ is constant over trials. Extended models also incorporate social preferences and learning rates that decay exponentially over game trials (Methods). (d) For both games, bars show BIC differences of three model variants with the best model, which includes both social preferences (SP) and decaying learning rates. For each model, BIC values were calculated by summing BIC values of all age cohorts. Solid bars reflect models including social preferences; void and hatched bars respectively show models excluding and including decaying learning rates. (e) Estimated learning rates from the best models as a function of the trial number, for each of the four age cohorts separately for the Trust Game, and f, for the Coordination Game. For the older cohorts (15–18 and 19–23), the estimated learning parameters were virtually identical, causing the plotted lines to coincide.
