List of research publications using Lilian Besson’s SMPyBandits project

I (Lilian Besson) have started my PhD in October 2016, and this project is a part of my on going research since December 2016.

1st article, about policy aggregation algorithm (aka model selection)

I designed and added the Aggregator policy, in order to test its validity and performance.

It is a “simple” voting algorithm to combine multiple bandit algorithms into one. Basically, it behaves like a simple MAB bandit just based on empirical means (even simpler than UCB), where arms are the child algorithms A_1 .. A_N, each running in “parallel”.

For more details, refer to this file: and this research article.

2nd article, about Multi-players Multi-Armed Bandits

There is another point of view: instead of comparing different single-player policies on the same problem, we can make them play against each other, in a multi-player setting. The basic difference is about collisions : at each time t, if two or more user chose to sense the same channel, there is a collision. Collisions can be handled in different way from the base station point of view, and from each player point of view.

For more details, refer to this file: and this research article.

3rd article, using Doubling Trick for Multi-Armed Bandits

I studied what Doubling Trick can and can’t do to obtain efficient anytime version of non-anytime optimal Multi-Armed Bandits algorithms.

For more details, refer to this file: and this research article.

4th article, about Piece-Wise Stationary Multi-Armed Bandits

With Emilie Kaufmann, we studied the Generalized Likelihood Ratio Test (GLRT) for sub-Bernoulli distributions, and proposed the B-GLRT algorithm for change-point detection for piece-wise stationary one-armed bandit problems. We combined the B-GLRT with the kl-UCB multi-armed bandit algorithm and proposed the GLR-klUCB algorithm for piece-wise stationary multi-armed bandit problems. We prove finite-time guarantees for the B-GLRT and the GLR-klUCB algorithm, and we illustrate its performance with extensive numerical experiments.

For more details, refer to this file: and this research article.

Other interesting things

Single-player Policies

Arms and problems

  • My framework mainly targets stochastic bandits, with arms following Bernoulli, bounded (truncated) or unbounded Gaussian, Exponential, Gamma or Poisson distributions.
  • The default configuration is to use a fixed problem for N repetitions (e.g. 1000 repetitions, use MAB.MAB), but there is also a perfect support for “Bayesian” problems where the mean vector µ1,…,µK change at every repetition (see MAB.DynamicMAB).
  • There is also a good support for Markovian problems, see MAB.MarkovianMAB, even though I didn’t implement any policies tailored for Markovian problems.
  • I’m actively working on adding a very clean support for non-stationary MAB problems, and MAB.PieceWiseStationaryMAB is already working well. Use it with policies designed for piece-wise stationary problems, like Discounted-Thompson, CD-UCB, M-UCB, SW-UCB#.