Known horizon

Module: statrl.settings.bandits.stochastic.knownhorizon

This setting is the stochastic bandit problem again, but with the time horizon known in advance. That extra information lets algorithms budget exploration against a fixed number of rounds, so the agent’s reset receives the horizon:

class BanditAgent(ABC):
    def reset(self, timehorizon): ...
    def select_arm(self) -> int: ...
    def update(self, arm, reward): ...

The environment

StochasticBanditEnv is the anytime environment, re-exported unchanged: knowing the horizon changes what the agent may do, not what the problem is. Any environment built for the anytime setting therefore works here as is.

Bridging to the anytime setting

Because the two stochastic settings share the same environment/agent shape, the wrappers package adapts an object built for one setting to the other. Four adapters are provided in wrapper_anytime_knownhorizon:

Wrapper	Adapts …
AnytimeToKnownHorizonEnvironmentWrapper	an anytime environment for use in the known-horizon setting.
KnownHorizonToAnytimeEnvironmentWrapper	a known-horizon environment for use in the anytime setting.
AnytimeToKnownHorizonAgentWrapper	an anytime agent so it accepts a horizon at reset.
KnownHorizonToAnytimeAgentWrapper	a known-horizon agent (given a fixed horizon) for anytime use.

This lets you benchmark an algorithm designed for one setting against the environments and loops of the other without reimplementing it.

The interaction loop

BanditInteraction runs the same select_arm → pull → update cycle for horizon rounds and returns the cumulative-score array.