Proceedings of the Annual Meeting of the Cognitive Science Society

Human-like systematic generalization remains a challenge for artificial systems. Recent work suggests hippocampal replay as a mechanism for online planning that facilitates generalization in novel, dynamic, and complex environments. Recent models, based on spatial tasks and rodents, present a serial notion of planning. In contrast, recent work on humans finds replay sequences traversing multiple trajectories at the same time on a compositional generalization task. Such a possibility for parallel execution does not interface well with serial planning. Here, we present a novel model of replay as planning with parallel search capabilities, extending an existing serial model of replay. Our model displays human-level generalization faster than the serial baselines. Furthermore, we find that models actively using parallel search produce internal dynamics qualitatively resembling neural replay sequences in humans performing the task. Our findings present search as a powerful general computation leveraged during replay for quickly learning good policies from little data.