Multi-Armed Bandit project

The multi-armed bandit problem is named by analogy to the one-armed bandit machine. In the multi-arms case, the gambler has to decide which arm to pull in order to maximize his total reward in a series of trials.

SourceForge: project page.

Bandit library

Several bandit strategies have been compiled in a .Net library written in C#. Please have a look at the documentation for the list of implemented strategies. Version 0.2 is restricted to the stochastic multi-armed bandit problem, where the underlying reward distributions for each lever do not change over time. The library is provided under LGPL open source licence.

• Bandit library, version 0.2 (October 2005)

Datasets

The following provided datasets could be used to benchmark the various bandit strategies.

• Universities webpage latency (networking data, May 2004)

Publications

• Multi-Armed Bandit Algorithms and Empirical Evaluation, Joannès Vermorel and Mehryar Mohri, ECML'05 (PDF)

Links

• Multi-Armed Bandit article of the Wikipedia.
• Reduced Design of eXperiments (re-Dox), web-based application providing (generalized) multi-armed bandits.

Author: Joannès Vermorel
Last updated: Tuesday, 24 January, 2006