Skip to yearly menu bar Skip to main content


Poster

Thompson Sampling and Approximate Inference

My Phan · Yasin Abbasi Yadkori · Justin Domke

East Exhibition Hall B, C #45

Keywords: [ Bandit Algorithms ] [ Algorithms ] [ Variational Inference ] [ Probabilistic Methods -> MCMC; Probabilistic Methods ]


Abstract: We study the effects of approximate inference on the performance of Thompson sampling in the k-armed bandit problems. Thompson sampling is a successful algorithm for online decision-making but requires posterior inference, which often must be approximated in practice. We show that even small constant inference error (in α-divergence) can lead to poor performance (linear regret) due to under-exploration (for α<1) or over-exploration (for α>0) by the approximation. While for α>0 this is unavoidable, for α0 the regret can be improved by adding a small amount of forced exploration even when the inference error is a large constant.

Live content is unavailable. Log in and register to view live content