Change point estimation by the mouse medial frontal cortex during probabilistic reward learning
There are often sudden changes in the state of environment. For a decision maker, accurate prediction and detection of change points are crucial for optimizing performance. Still unclear, however, is whether rodents are simply reactive to reinforcements, or if they can be proactive to estimate future change points during value-based decision making. In this study, we characterize head-fixed mice performing a two-armed bandit task with probabilistic reward reversals. Choice behavior deviates from classic reinforcement learning, but instead suggests a strategy involving belief updating, consistent with the anticipation of change points to exploit the task structure. Excitotoxic lesion and optogenetic inactivation implicate the anterior cingulate and premotor regions of medial frontal cortex. Specifically, over-estimation of hazard rate arises from imbalance across frontal hemispheres during the time window before the choice is made. Collectively, the results demonstrate that mice can capitalize on their knowledge of task regularities, and this estimation of future changes in the environment may be a main computational function of the rodent dorsal medial frontal cortex.