Policy gradient methods can reliably learn optimal decisions in continuous-time bandit environments with simple constant learning rates, achieving near-optimal sample efficiency with O(log T) regret.
This paper analyzes how policy gradient algorithms learn to make optimal decisions in multi-armed bandit problems where the environment evolves continuously over time (described by stochastic differential equations). The authors prove the algorithm converges to the best choice and achieves logarithmic regret—meaning it learns efficiently without wasting too many attempts on suboptimal options.