Theoretical and empirical research on human and non-human animal learning
How do humans learn when the world around them changes constantly? Because of the ubiquitous presence of risk prediction errors in the brain, we have recently proposed a novel, robust model of learning that could explain both behavioural "anomalies" (e.g., satisficing behaviour, boredom-induced exploration) and neural signals. This model is inspired by robust control theory in engineering.
We focus on a particularly nasty type of uncertainty, namely, leptokurtosis. There, outliers are both frequent and salient. Leptokurtosis typifies financial risks. Control that is optimised for a Gaussian world (e.g., the Kalman filter) is inadequate here, as are traditional machine learning techniques such as TD learning. Besides unravelling how humans deal with leptokurtosis, we have been adapting distributional reinforcement learning to provide more effective (faster, more robust) learning experiences.
Peter Bossaerts, Elizabeth Bowman, Shijie Huang, Nitin Yadav