Conference on Cognitive Computational Neuroscience

Reinforcement learning requires associating rewards with one or more of the states or actions that preceded them. The question of exactly which states or actions to associate with each reward is referred to as the “credit assignment problem”, and better solutions result in more efficient learning. In humans, credit assignment is informed by knowledge of the causal structure of the world. Here, we adapt a “structured” credit assignment task from the human literature for use with head-fixed mice. In this task, rewards of one type (“controllable”) depend causally on the mouse’s actions, while another distinguishable type (“distractor”) is independent of those actions. We present behavioral evidence that mice, like humans, adopt a strategy that is partially structure-sensitive: they update their behavior based on rewards of both types, but they update more strongly to the controllable reward. This work opens the door to investigations of the neural mechanisms of structured credit assignment using the wide range of tools that are available in head-fixed mice.