Categorization allows humans to effectively structure the world and to predict decision outcomes using the features of the to-be-classified objects (e.g., edible and poisonous mushrooms based on color or shape). Often, such decisions promise more or less rewarding consequences, which, however, are often not perfectly informative about the true categories (e.g., edible mushrooms sometimes lead to gastric problems). Further, research suggests that category learning involves rule- and memory-based cognitive processes depending on the category structure (e.g., color predicts the categories vs. perceptually unstructured categories). Here, we study in depth how humans solve such problems under probabilistic reward-feedback in a multi-method approach, using fMRi, eye tracking and cognitive modeling. For this, we used the well-studied categorization problems known as Type II and VI (rule-based vs. unstructured, with eight objects, each containing three attributes), however, adjusted to reflect a value-based preferential choice. Participants learned via trial-and-error, which of two simultaneously presented objects predicted reward or no reward, after instructing that either a rule or memorization will allow to maximize their rewards. For analyses, we employed the novel Category Abstraction Learning model (CAL; Schlegelmilch, Wills & von Helversen., 2021), a hybrid model including interacting processes of rule learning and memorization. We show, that CAL not only explains individual differences in learning behavior (sudden vs. incremental in rule-based vs. unstructured), but also predicts, via cross-validation, (a) when humans start to maximize their rewards, (b) dissociable neural correlates (rule vs. memory strategies), and (c) selective stimulus attention during preference formation, as well as during reward-processing.