Finding Optimal Arms in Non-Stochastic Combinatorial Bandits With Applications in Algorithm Configuration

MCML Authors

Viktor Bengs

Dr.

* Former Member

→ Group Eyke Hüllermeier
 Artificial Intelligence and Machine Learning

Eyke Hüllermeier

Prof. Dr.

Core PI

Artificial Intelligence and Machine Learning

Abstract

We consider the combinatorial bandit problem with semibandit feedback under finite sampling budget, where the action is to choose a set of arms in a non-stochastic setting with subset-dependent feedback. We propose an algorithmic framework to solve it, which, additionally, can be leveraged for the algorithm configuration problem, where the goal is to find an optimal parameter configuration for a given target algorithm. We showcase that our introduced algorithm requires significantly less computation time than other existing theoretically-grounded approaches while still yielding high-quality configurations.

inproceedings BSB+20

---

DA2PL 2020

Wokshop From Multiple Criteria Decision Aid to Preference Learning. Compiègne, France, Nov 17-18, 2020.

---

Authors

J. Brandt • E. Schede • V. Bengs • B. Haddenhorst • K. Tierney • E. Hüllermeier

---

Links

PDF

---

Research Area

 A3 | Computational Models