Multi-Task Bandits and Shared Priors

Summary

This project studies how information should be shared across related bandit tasks rather than relearned from scratch.

Research Question

When multiple users or tasks are related, what forms of shared structure can be exploited to improve exploration without collapsing meaningful task-specific differences?

Methodology

Bayesian and structured-prior views of multi-task bandits
Information sharing across related users or tasks
Emphasis on sample efficiency and best-arm learning under limited data

Key Contributions

Frames multi-task bandits as a problem of prior construction rather than isolated per-task optimization
Connects transfer, shared structure, and exploration efficiency
Serves as a foundation for future work on priors informed by external knowledge

Outputs

Publication: (Mukherjee et al., 2025)

References

2025

Preprint

Co-Exploration and Co-Exploitation via Shared Structure in Multi-Task Bandits

Sumantrak Mukherjee, Serafima Lebedeva, Valentin Margraf, and 7 more authors

2025

Preprint

Bib

@misc{mukherjee2025multitaskbandits,
  title = {Co-Exploration and Co-Exploitation via Shared Structure in Multi-Task Bandits},
  author = {Mukherjee, Sumantrak and Lebedeva, Serafima and Margraf, Valentin and Hanselle, Jonas and Yamaoka, Kanta and Bengs, Viktor and Bhalwankar, Raj and Neigel, Peter and Hullermeier, Eyke and Vollmer, Sebastian Josef},
  year = {2025},
  note = {Preprint},
}