Maximin Action Identification: A New Bandit Framework for Games
Statistics Theory
2016-02-16 v1 Computer Science and Game Theory
Machine Learning
Statistics Theory
Abstract
We study an original problem of pure exploration in a strategic bandit model motivated by Monte Carlo Tree Search. It consists in identifying the best action in a game, when the player may sample random outcomes of sequentially chosen pairs of actions. We propose two strategies for the fixed-confidence setting: Maximin-LUCB, based on lower-and upper-confidence bounds; and Maximin-Racing, which operates by successively eliminating the sub-optimal actions. We discuss the sample complexity of both methods and compare their performance empirically. We sketch a lower bound analysis, and possible connections to an optimal algorithm.
Cite
@article{arxiv.1602.04676,
title = {Maximin Action Identification: A New Bandit Framework for Games},
author = {Aurélien Garivier and Emilie Kaufmann and Wouter Koolen},
journal= {arXiv preprint arXiv:1602.04676},
year = {2016}
}