Probabilistic Reversal Learning

The probabilistic reversal learning (PRL) task measures cognitive flexibility and reward-guided learning under uncertainty by presenting two response options with asymmetric probabilistic contingencies, typically 80% versus 20% reward delivery. Once the subject reaches a learning criterion on the initial discrimination, contingencies reverse so that the previously less-rewarded option becomes more rewarding. The task engages orbitofrontal cortex, anterior cingulate cortex, and dorsal striatal circuits that compute prediction errors and update action values. PRL is a direct translational analogue of tasks used to assess cognitive inflexibility in depression, OCD, and schizophrenia.

Performance is decomposed using computational reinforcement learning models that estimate learning rate (alpha), inverse temperature (beta), and separate parameters for positive and negative prediction error sensitivity. Win-stay probability quantifies the tendency to repeat a choice after reward, while lose-shift probability measures switching after non-reward. The number of trials to reach reversal criterion, perseverative errors (continued selection of the previously correct option after reversal), and total reversals achieved within a session are standard behavioral endpoints. These measures dissociate value updating from response selection processes.

ConductMaze automates PRL by delivering probabilistic reward outcomes on each lever press according to configurable probability matrices and detecting criterion-based reversals without experimenter intervention. The system logs every trial outcome, fits trial-by-trial reinforcement learning models, and exports win-stay/lose-shift matrices alongside raw event data. Reversal triggers, perseverative error counts, and session-wide learning curves are computed in real time. The platform supports multiple reversals within a single session to capture both initial learning and cognitive flexibility dynamics.