Your favorite color makes learning more precise and adaptable

Learning from reward feedback is essential for survival but can become extremely challenging when choice options have multiple features and feature values (curse of dimensionality). Here, we propose a general framework for learning reward values in dynamic multi-dimensional environments via encoding and updating the average value of individual features. We predicted that this feature-based learning occurs not just because it can reduce dimensionality, but more importantly because it can increase adaptability without compromising precision. We experimentally tested this novel prediction and found that in dynamic environments, human subjects adopted feature-based learning even when this approach does not reduce dimensionality. Even in static low-dimensional environment, subjects initially tended to adopt feature-based learning and switched to learning individual option values only when feature values could not accurately predict all objects values. Moreover, behaviors of two alternative network models demonstrated that hierarchical decision-making and learning could account for our experimental results and thus provides a plausible mechanism for model adoption during learning in dynamic environments. Our results constrain neural mechanisms underlying learning in dynamic multi-dimensional environments, and highlight the importance of neurons encoding the value of individual features in this learning.