Prediction error minimization for emergence of altruistic behavior

Young infants have been shown to help others even in novel situations where no rewards are given or expected. One interpretation is that this altruistic behavior is motivated by an early form of empathy. On the other hand, infants have also been demonstrated to help nonhuman entities, which implies that another source of motivation may trigger an infant's altruism. In both cases, understanding others' intentions and predicting their future goals appear to be necessary for the infant to perform such behavior. In this study, we attempted to find the minimal cognitive abilities and possible behavioral motivations for infants to behave altruistically. Our hypothesis was that infants understand and predict others' behavior and then help them to minimize the prediction error (PE) they estimate for others' action goals. To verify this hypothesis, we designed a computational model for the emergence of altruistic behavior and performed two experiments in order to determine the minimal cognitive abilities required to estimate the PE and demonstrate that PE minimization is a possible motivation for altruistic behavior. Our experimental results showed that the ability to recognize and predict others' actions is indeed required to generate a PE and that its minimization can motivate altruistic behavior without any given or expected rewards.