MODELING THE INTERPLAY OF MOTIVATION AND LEARNING IN MOUSE PERCEPTUAL DECISION-MAKING

Decisions arise from multiple needs and drives, and motivational state shapes their relative contributions to behavior. Our prior work has shown that, in a two-whisker discrimination task in water-restricted mice, thirst-induced motivation markedly affects performance and modulates cortical sensory representations, and that excessively high or low drive impairs both. In the same work, we also showed that rapid acquisition of sensorimotor contingencies can be masked by slower, concurrent allostatic regulation of body mass, yielding apparently shallow learning curves.
Motivated by these experimental results, here we develop an interpretable model that integrates motivation, effort-reward trade-offs, and sensorimotor learning. Agents learn stimulus-action values through a Q-learning rule and select actions via a sigmoid decision function whose argument combines learned Q-values and a Pavlovian action bias term, both homeostatically discounted by thirst, together with a constant effort cost term. This formulation captures how internal need dynamically reshapes policy selection, allowing the model to recapitulate the behavioural dynamics and learning trajectories observed empirically.
We also develop a procedure to fit the model to individual animals by minimizing the negative log-likelihood of observed actions, combining differential evolution with a gradient-based method. Although validated in our two-whisker task, the framework is general and suitable for a wide range of goal-directed paradigms in systems neuroscience, enabling behavioural phenotyping and optimization of training procedures. By explicitly linking sensory evidence with learned value, effort, and motivation, the model provides a compact, interpretable account of mouse goal-directed behavior and a practical tool for explaining behavioural data and refining animal training.