Mice alternate between inference- and stimulus-bound strategies during probabilistic foraging

Essential features of the world are often hidden and must be inferred by constructing internal models based on indirect evidence. During naturalistic foraging, animals must continually choose between trying to exploit a depleting food source at their current location and leaving to explore a new source at the expense of costly travel epochs. When is it best to leave the current site for another one? In a deterministic environment, the optimal strategy is to leave the current site when the immediate rate of reward drops below the average rate - a stimulus-bound strategy, assigning each action a value that is updated based on its immediate outcome. This strategy, however, is not optimal in a realistic foraging scenario, where rewards are encountered probabilistically and the optimal strategy is inference-bound, requiring the animal to infer the hidden structure of the world. Motivated by recent studies showing that mice alternate between discrete strategies during perceptual decision-making [1], here we tested the hypothesis that mice behavior during a probabilistic foraging task switches between inference- and stimulus-bound strategies within the same session. We developed a novel hidden Markov model with linear emissions (LM-HMM) to capture this switching dynamic and validated this model on ground truth data generated from a synthetic agent performing the task. When applied to mice engaged in the task, the LM-HMM revealed that mice switch between three strategies: a persistent inference-bound, an impulsive inference-bound, and a stimulus-bound strategy. Our results establish a new approach to modeling complex decision-making behavior involving accumulation of evidence and inference.