Accelerating Linear Bandits with Persistent Candidate Sets

Name: Accelerating Linear Bandits with Persistent Candidate Sets
Uploaded: 2026-10-19T01:18:17.393345+00:00
Duration: 340 s
Description: Persistent Candidate LinUCB accelerates linear bandits by maintaining a small set of active actions ($k$) and sampling challengers ($q$) from the rest. This reduces score evaluations per round from $K$ to $k+q$, significantly lowering computational cost. It establishes regret bounds separating learning costs from unavoidable partial access costs.

Persistent Candidate LinUCB accelerates linear bandits by maintaining a small set of active actions ($k$) and sampling challengers ($q$) from the rest. This reduces score evaluations per round from $K$ to $k+q$, significantly lowering computational cost. It establishes regret bounds separating learning costs from unavoidable partial access costs.

Open the full episode