Conference on Cognitive Computational Neuroscience

Learning from the past requires assigning credit to specific causes. Here, we explored human credit assignment strategies by asking people to predict the pain level of an avatar who performed activities with short- and long-term pain-related consequences. Human behavioral results suggest that while participants can learn short-term consequences, they are unable to learn their long-term ones, except when activities are blocked such that each activity is repeated multiple times in a row. Standard model-free algorithms (i.e., temporal difference (TD) learning) also fail to properly learn long-term consequences of activities in our task, whereas Bayesian models that take into account the causal structure of the environment effectively learn both short- and long-term consequences across repetition and non-repetition conditions. Thus, neither model matched participant behavior across all conditions. Our results demonstrate that credit assignment critically depends on the order in which actions are observed, with repetitions aiding the learning of long-term consequences. This raises questions as to whether people might intentionally repeat actions (i.e., perseveration) to improve long-term credit assignment.