29. Jan 2024

Efficient and robust reversal learning

Datum: 29. January 2024 | 11:30 – 12:30
Sprecher: Angelika Steger, ETH
Veranstaltungsort: Raiffeisen Lecture Hall
Sprache: Englisch

One of the defining characteristics of biological intelligence is its rapid adaptability to new environments, a trait explored in cognitive science through reversal learning tasks.  In this talk we present a novel approach  for the multi-armed bandit problem, the machine learning analogue of the reversal learning paradigm, that departs from the traditional method of estimating values. Instead, our method relies on continuously holding a belief about the best current option, coupled with a verification mechanism that periodically assesses this belief's validity. We demonstrate that our approach not only theoretically outperforms the best known algorithm but also explains well measured data across various experimental settings.