Animal learning is often conceived as a gradual process that develops over many trials and involves the incremental strengthening of associations among stimuli, responses, and outcomes, a view deeply rooted in behaviorist theory and inherent in most neural network learning models. However, potentially abrupt performance changes during the learning process may often go unnoticed as learning curves presented in the behavioral literature commonly represent group averages across animals and/or sessions. Recently we confirmed the presence of sudden behavioral performance jumps in a rule learning and switching task, and observed that these were tightly related to rapid transitions in the neuronal ensemble dynamics in rodent prefrontal cortex (PFC). The interpretation of these data was that rule learning is better described as evidence based decision process rather than by a gradual buildup of associations reinforced by the reward feedback. Here, by using a combination of multiple single-unit recordings, optogenetic tools, advancement of multivariate neural data analysis methods, and computational modeling, we will investigate the idea of animal learning as an active decision process that involves the sampling and accumulation of evidence for different behavioral strategies or 'hypotheses'. Based on computational models within the framework of reinforcement learning theories we will generate specific predictions, in the form of time-varying signals to be correlated with the multivariate spiking data, to elucidate the computational and neuronal dynamics underlying the observed sudden neuronal and behavioral transitions. Moreover, we will test whether the switching between behavioral strategies can be manipulated by optogenetic stimulation or inhibition of the dopaminergic input into the PFC.