Minimax Dynamic Regret Under Time-Varying Arm Sets

Non-stationary Linear Bandits | Wang et al., arXiv:2601.01069

0.877
Weighted LS Exponent
0.858
Restarting Exponent
0.667
Theoretical Optimal
d=5
Feature Dimension
10K
Max Horizon

Regret Scaling with Horizon (Log-Log)

Arm Variation Impact (T=1000)

Non-stationarity Budget

Scaling Exponents

Experiment Summary

AlgorithmExponentType
Weighted LS0.8771.000Adaptive
Sliding Window0.8771.000Adaptive
Restarting0.8580.999Adaptive
MASTER0.8781.000Static