Design decentralized learning protocols for heterogeneous asynchronous multi-player bandits

Develop efficient decentralized learning policies for heterogeneous asynchronous multi-player bandits that adapt collision-based or communication-based protocols to asynchronous activity and heterogeneous player roles, while balancing rapid information exchange among frequently active players with multiple exchanges conducted in parallel.

Background

The paper’s approximation results provide centralized oracle guarantees for heterogeneous asynchronous multi-player bandit allocation and suggest that centralized learning methods such as CUCB could be adapted to this setting. However, the decentralized setting is substantially more difficult because players cannot rely on free communication through a central controller.

In an asynchronous network, heterogeneous activation probabilities mean that players are active at different rates and no longer play symmetric roles in communication. The authors identify the development of efficient protocols that accommodate these asymmetries and support parallel information exchange as a major unresolved research direction.

References

Extending these results to the decentralized setting remains a major open challenge. Without a central controller, communication is no longer free, and adapting communication protocols such as those introduced by \citet{boursier2019} becomes more challenging in an asynchronous environment. Heterogeneous activation probabilities introduce an additional difficulty: players no longer play symmetric roles in the communication process. Efficient protocols must therefore balance rapid information exchange between frequently active players with the ability to conduct multiple exchanges in parallel.

— Approximating Optimal Policies in Heterogeneous Asynchronous Multi-player Bandits  (2610.18514 - Arroues-Paykin et al., 13 Oct 2026) in Section 5, “Conclusion”