Cognitive graphs: Representational substrates for planning.

Making plans for upcoming actions is a computationally demanding process. To mitigate these demands, individuals can build extensive internal models of their environment-states, actions, and their sequential relationships-that allow for plans to be developed with minimal computational costs. Initially, these models reflect elaborate networks of learned associative relationships, which can be used to generate plans for reward through more iterative computations such as trajectory sampling. After sufficient experience, compressed forms of these models can efficiently capture long-range sequential structure, allowing them to be used for rapid planning even in pursuit of novel or changing rewards. Here, we review recent work on the multitude of representations that can support different forms of planning. We discuss how <i>cognitive graphs</i>, a framework with roots in both cognitive psychology and computer science, can provide a unifying view of these representations and their relationships to one another. Conceptualizing internal models as forms of graphs situates them on a spectrum where different kinds of structured sequences can be queried to support both planning and the formation of iteratively more compressed predictive representations. We discuss how each of these kinds of cognitive graphs are created during learning, and used to transfer and generalize knowledge across environments. Taken together, this review highlights the significant impact that the various associative structures of memory have on planning.