Stochastic bandits

Module: statrl.settings.bandits.stochastic

In the stochastic setting each arm has a fixed, unknown reward distribution, and pulling an arm draws an independent sample from it. statrl ships two variants, distinguished by what the agent knows about the time horizon:

Every stochastic-bandit agent implements the same three methods:

reset()

Start a new, independent run (clear all statistics).

select_arm() -> int

Return the index of the arm to pull next.

update(arm, reward)

Incorporate the observed reward for the chosen arm.

An environment exposes its arms through number_arms and means (the latter for evaluation and oracle construction only) and samples rewards through step(arm). The loop BanditInteraction().run(env, learner, horizon) ties them together and returns a cumulative-score numpy.ndarray — of the arms’ expected rewards, not the realized ones, so the regret curve carries no reward noise. See Core concepts.

• Anytime
• Known horizon