A hierarchical Bayesian approach to assess learning and guessing strategies in reinforcement learning

In two-armed bandit tasks participants learn which stimulus in a stimulus pair is associated with the highest value. In typical reinforcement learning studies, participants are presented with several pairs in a random order; frequently applied analyses assume each pair is learned in a similar way. When tasks become more difficult, however, participants may learn some stimulus pairs while they fail to learn other pairs, that is, they simply guess for a subset of pairs. We put forward the Reinforcement Learning/Guessing (RLGuess) model — enabling researchers to model this learning and guessing process. We implemented the model in a Bayesian hierarchical framework. Simulations showed that the RLGuess model outperforms a standard reinforcement learning model when participants guess: Fit is enhanced and parameter estimates become unbiased. An empirical application illustrates the merits of the RLGuess model.