Linear Bandits under Exact Sliding-Window Constraints
Exact feasibility is the hard constraint here, and the paper says learning can fail without reachability structure.
The authors study linear bandits where every sliding window of actions must remain feasible. They prove offline optimality results under convexity and cyclic-shift invariance, then show online sublinear regret is not guaranteed from geometry alone. Their rare-switching OFUL variants add reachability measures, including transition diameter and history-state diameter, to recover regret bounds while preserving exact feasibility. ArXiv · AI/CL/LG's note
The authors study linear bandits where every sliding window of actions must remain feasible. They prove offline optimality results under convexity and cyclic-shift invariance, then show online sublinear regret is not guaranteed from geometry alone. Their rare-switching OFUL variants add reachability measures, including transition diameter and history-state diameter, to recover regret bounds while preserving exact feasibility. ArXiv · AI/CL/LG's note
score 4