Link Adaptation Using Joint-Thompson Sampling

In wireless communication systems, link adaptation (LA) algorithms must dynamically select the optimal Modulation and Coding Scheme (MCS) to maximize throughput under varying channel conditions. Traditional methods often struggle with fast-changing channels or lack of reliable feedback. This paper explores whether modeling the MCS selection process as a multi-armed bandit (MAB) problem—while explicitly incorporating the physical constraint that lower MCS levels are inherently more reliable than higher ones—can improve performance and robustness.

Existing MAB-based link adaptation algorithms, such as standard Thompson Sampling (TS), typically treat each MCS level as an independent arm. The authors argue this ignores the physical reality that MCS success probabilities are monotonically ordered. To address this, they propose Joint-Thompson Sampling (Joint-TS), which replaces independent Beta priors with a Multivariate Ordered Beta (MOB) distribution. Because sampling directly from an MOB distribution is complex, the authors implement a Gibbs sampling technique to maintain the monotonicity constraint at every time step. They evaluate Joint-TS against state-of-the-art methods like OLLA, UTS, and Latent-TS (LTS) across static and high-mobility (Doppler) scenarios, both with and without Channel Quality Index (CQI) feedback.

Simulations demonstrate that Joint-TS provides consistent, robust throughput across all tested scenarios. While other algorithms like UTS or LTS may perform well under ideal conditions (e.g., perfect CQI or specific static environments), they often degrade significantly when CQI is unavailable or in high-Doppler environments. Joint-TS maintains stable performance without requiring the pre-existing lookup tables that make other methods like LTS vulnerable to inaccurate channel modeling. By enforcing the monotonicity of success probabilities, the algorithm converges effectively even in challenging, dynamic wireless environments.