Johannes Kirschner

[Johannes Kirschner @j_kirschner]

The model generalizes linear bandits by decoupling reward and observation. It captures many known settings, including dueling, network and combinatorial bandits.

We also extend the algorithm to the contextual and the RKHS setting, allowing for many more applications! 2/2

[Johannes Kirschner @j_kirschner]

Excited to present our work with Tor Lattimore and @arkrause on stochastic linear partial monitoring!

We show that a single, simple algorithm based on information directed sampling achieves the optimal regret rate in all finite games. 1/2 👇

#COLT2020
https://t.co/4IBGYoPur0