Parallelizing Thompson Sampling
amin karbasi, Vahab Mirrokni, Mohammad Shadravan
Abstract
How can we make use of information parallelism in online decision-making problems while efficiently balancing the exploration-exploitation trade-off? In this paper, we introduce a batch Thompson Sampling framework for two canonical online decision-making problems with partial feedback, namely, stochastic multi-arm bandit and linear contextual bandit. Over a time horizon $T$, our batch Thompson Sampling policy achieves the same (asymptotic) regret bound of a fully sequential one while carrying out only $O(\log T)$ batch queries. To achieve this exponential reduction, i.e., reducing the number of interactions from $T$ to $O(\log T)$, our batch policy dynamically determines the duration of each batch in order to balance the exploration-exploitation trade-off. We also demonstrate experimentally that dynamic batch allocation outperforms natural baselines.
BibTeX
@inproceedings{
karbasi2021parallelizing,
title={Parallelizing Thompson Sampling},
author={amin karbasi and Vahab Mirrokni and Mohammad Shadravan},
booktitle={Advances in Neural Information Processing Systems},
editor={A. Beygelzimer and Y. Dauphin and P. Liang and J. Wortman Vaughan},
year={2021},
url={https://openreview.net/forum?id=Kar8pVEtzeQ}
}