Contributions of statistical learning to learning from reward feedback

Abstract Natural environments are rich with patterns and regularities. Thus, it is not surprising that most animals have evolved neural mechanisms to detect and adapt to these regularities. Such detections and adaptations can, in turn, influence several cognitive functions including attention, perception, and memory, thereby enhancing survival. Here, we investigated whether detecting environmental regularities that are irrelevant to obtaining rewards can influence learning about multi-dimensional choice options––a process often constrained by the scarcity of reward feedback. To explore this, we trained human participants to perform a multidimensional reward- learning task alongside an orthogonal sequence-prediction task. We found that although feature- specific regularities in the sequence-prediction task were not predictive of reward, they incidentally biased participants’ behavior toward the feature with regularity during the reward- learning task. Fitting choice behavior with various computational models revealed that this effect was more consistent with modulations in learning rather than decision making, as evidenced by higher learning rates for this feature. This was particularly apparent for learning from chosen, rewarded options and unchosen, unrewarded options, demonstrating that environmental regularities can amplify confirmation bias in reward learning. Our results thus extend the notion of confirmation bias in learning about options and actions to their features. Furthermore, they suggest that temporal regularities can influence reward learning by biasing the association of reward with specific features and by enhancing confirmation bias. These effects help reduce dimensionality of the learning task to mitigate the curse of dimensionality in reward learning.