Foraging animals use dynamic Bayesian updating to model meta-uncertainty in environment representations

Foraging theory predicts animal behavior in many contexts. In patch-based foraging behaviors, the marginal value theorem (MVT) gives the optimal strategy for deterministic environments whose parameters are fully known to the forager. In natural settings, environmental parameters exhibit variability and are only partially known to the animal based on its experience, creating uncertainty. Models of uncertainty in foraging are well established. However, natural environments also exhibit unpredicted changes in their statistics. As a result, animals must ascertain whether the currently observed quality of the environment is consistent with their internal models, or whether something has changed, creating meta-uncertainty. Behavioral strategies for optimizing foraging behavior under meta-uncertainty, and their neural underpinnings, are largely unknown. Here, we developed a novel behavioral task and computational framework for studying patch-leaving decisions in head-fixed and freely moving mice in conditions of meta-uncertainty. We stochastically varied between-patch travel time, as well as within-patch reward depletion rate. We find that, when uncertainty is minimal, mice adopt patch residence times in a manner consistent with the MVT and not explainable by simple ethologically motivated heuristic strategies. However, behavior in highly variable environments was best explained by modeling both first- and second-order uncertainty in environmental parameters, wherein local variability and global statistics are captured by a Bayesian estimator and dynamic prior, respectively. Thus, mice forage under meta-uncertainty by employing a hierarchical Bayesian strategy, which is essential for efficiently foraging in volatile environments. The results provide a foundation for understanding the neural basis of decision-making that exhibits naturalistic meta-uncertainty.Author summary: The ethological approach to understanding how animals make decisions is to use tasks that they often face in their natural environments. One such task, canonically termed patch-based foraging in behavioral ecology, involves harvesting resources from spatially separated areas (termed “patches”) that deplete over time. While patch foraging, animals must choose when to leave each patch to find a new, replenished one. The marginal value theorem (MVT), describes the optimal behavioral strategy when environment statistics are stable and known to the animal. However, naturalistic settings are often noisy and uncertain, which limits the applicability of the MVT. Here, to understand how laboratory mice make ethologically relevant decisions, we implemented a patch-based foraging task in either a physical or virtual patch-based foraging environment. The tasks incorporate uncertainty in the richness of patches, the distance between patches, and, importantly, the randomness of reward timings within a patch. When randomness of reward timing was low, animals behaved in a manner consistent with the MVT. However, when reward-timing randomness was high, mice dynamically weighted average statistics and recent observations, captured in a Bayesian estimator. Our results thus lay the groundwork for studying how the brain solves tasks when presented with multiple levels of uncertainty.