Conference on Cognitive Computational Neuroscience

When faced with a multi-step decision problem, humans and animals must balance flexible and accurate decision making with computational complexity. One prominent approach takes advantage of temporal abstraction of future states: by learning to predict long-run future trajectory independently of rewards, the successor representation (SR) can avoid the costs of full mental simulation, while retaining the ability to cheaply replan when goals change. Human behavior shows signatures of such temporal abstraction, but their trial-by-trial acquisition has not been elucidated and it remains an open question if reliance on such abstractions adapts to their usefulness, e.g. the predictability of long-run states. We developed a novel task to distinguish SR-based planning. Our results support findings that human behavior exhibits a mix of learning strategies, and crucially, we measure SR usage on a trial-by-trial basis. Further, by dynamically manipulating the task structure, we observe preliminary results suggesting that human reliance on temporal approximations is arbitrated by future predictability.