Intrinsic motivation allows agents to engage in behaviors that typically do not have explicit rewards, such as exploration and play. Intrinsically motivated behaviors (i.e., not goal-directed) help develop the “broad competence” needed for Generalization.
A related concept is intrinsic social motivation, which I understand as the kind that compels someone to choose collaboration in a social dilemma scenario like the “prisoner’s dilemma.” Jaques et al. (2019) implemented intrinsic social motivation in a multi-agent reinforcement learning framework by including an “influence” measure for rewarding agents implicitly; interestingly, agents developed