Efficient Learning of Predictive Maps for Flexible Planning

Cognitive maps enable flexible behavior by providing internal representations of task structure that can be efficiently reused across different contexts. The successor representation (SR) offers a promising model for how these maps might be learned and computed in the brain, but its dependence on specific behavioral policies limits its ability to support flexible planning. Here, we introduce SR-IS, a novel model that combines temporal difference learning with importance sampling to construct policy-independent cognitive maps. We show that SR-IS learns representations that capture the underlying structure of the environment without being limited by the agent's current behavioral policy. These representations can be efficiently updated when the environment changes, enabling rapid behavioral adaptation without requiring extensive relearning. In a series of simulations, we demonstrate that SR-IS outperforms existing models in classic replanning tasks and provides a better account of both rodent and human behavior in spatial navigation experiments. The model uniquely explains a key behavioral finding: humans show greater flexibility in adapting to reward changes than to structural changes in their environment—an asymmetry that previous models failed to capture. Our findings bridge theoretical models of predictive maps with empirical observations of planning behavior, while providing new insights into how the brain might implement efficient yet flexible decision-making.