Asymmetrical Learning of Win and Loss Associations: Individual Differences and Task Effects

Abstract

Learning that particular objects or actions are associated with rewards and punishment (i.e., value) can impact many aspects of human behavior, including what we attend to, what we desire, and what we learn. A recent report (Lin et al. 2020) documented an asymmetry in the learning of win versus loss associations in a widely used value learning task (VLT; Raymond and O’Brien, 2009). In this task people learn, via trial and error, to associate otherwise neutral stimuli with win or loss outcomes of varying probability. However, a number of questions about the potential boundary conditions of this learning asymmetry remained unanswered. The goal of the present dissertation was to investigate whether better learning of win associations than loss associations can be attributed, in part, to individual differences that are consistent over time, whether the asymmetry depends on the context in which people learn win and loss associations, and whether the asymmetry depends on the magnitude of the win and loss outcomes. To this end, participants completed the VLT (1) on two separate occasions, (2) when win, loss, and no change trials were presented in separate trial blocks, and (3) when the magnitude of the win/loss outcomes varied. The results showed that while individuals differed in the extent to which they learned win associations better than loss associations, there was no strong evidence for intraindividual consistency in learning patterns or that learning patterns correlated with other individual difference variables known to affect decision behavior. However, evidence that learning context influenced the win-loss asymmetry emerged across all experiments. First, when participants completed the VLT in two sessions, the learning asymmetry was reduced during the second session. Second, when separating win and loss trials, we found that the learning asymmetry only emerged when loss trials directly preceded win trials. Third, we demonstrated that including multiple outcome magnitude conditions in the same context affects the learning pattern for win and loss associations. Taken together these findings reveal that while the learning asymmetry is robust, there are conditions that improve loss learning and reduce the asymmetry. Thus, loss learning is particularly malleable to the context in which value learning takes place. When examining explicit knowledge of the learned associations, we found that whether or not a learning asymmetry emerged in the VLT, explicit knowledge of the outcomes differed for win and loss scenes. We suggest that the memory representation of stimulus-outcome associations may depend, at least in part, on choice history. Lastly, when examining the association between reinforcement learning parameters and the learning asymmetry, we found that a higher learning rate was associated with a smaller asymmetry. Whereas the results for the inverse temperature (balance between exploration vs. exploitation) were mixed. We suggest that the conditions that affect the presence or absence of the learning asymmetry may also influence whether exploring versus exploiting is beneficial for learning the optimal choice. Future work can examine whether selectively increasing the loss value would equate wins and losses, thus reducing the asymmetry. Furthermore, future studies examining subsequent cognitive consequences of acquired win and loss associations of otherwise neutral stimuli should consider the learning context and how choice history plays a role in subsequent effects of learned value.