BanditRLlib
Lean gate passed before this site build; local proof declarations are shown as compiled.

Lean module · UCB

BanditRLProof.Algorithms.UCBBoundedFiniteArmSampledAsymptotics

# Sampled-pair UCB consistency for bounded finite-arm laws This module derives the direct centered sub-Gaussian contracts required by the stationary finite-arm sampled-pair consistency theorem from a common almost- sure reward interval and exact arm means.

Module map

Declarations
5
Placeholders
0

Imports

BanditRLProof.Algorithms.UCBFiniteArmSubGaussianSampledAsymptotics

Imported by

BanditRLProof

Declarations

Open an item to read its exact compact statement and source link. Detailed teaching notes are linked when registered.

def BanditRLProof.UCB.selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret Compiled

Exact sampled-successor expected pseudo-regret for stationary finite-arm laws bounded almost surely in a common interval. The genuine armwise proxy is the common Hoeffding proxy; the parent practical route pads its finite maximum.

noncomputable def selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret {K : Nat} (model : FiniteBanditModel K) (armLaw : Fin K -> Measure Rat) (hprob : forall arm, IsProbabilityMeasure (armLaw arm)) (lo hi : Real) (defaultAction : Fin K) (T : Nat) : Real
theorem BanditRLProof.UCB.selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret_isBigO_log Compiled

The exact common-bounded expected pseudo-regret family is logarithmic.

theorem selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret_isBigO_log {K : Nat} (model : FiniteBanditModel K) (armLaw : Fin K -> Measure Rat) (hprob : forall arm, IsProbabilityMeasure (armLaw arm)) (lo hi : Real) (hmeas : forall arm, AEMeasurable (fun reward : Rat => ((reward : Rat) : Real)) (armLaw arm)) (hbound : forall arm, Filter.Eventually (fun reward : Rat => Set.Icc lo hi ((reward : Rat) : Real)) (ae (armLaw arm))) (hmean : forall arm, integral (armLaw arm) (fun reward : Rat => ((reward : Rat) : Real)) = ((model.mean arm : Rat) : Real)) (defaultAction : Fin K) : (selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret model armLaw hprob lo hi defaultAction) =O[atTop] (fun T : Nat => Real.log (((T + 1 : Nat) : Real)))
theorem BanditRLProof.UCB.selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret_isLittleO_natCast_succ Compiled

The exact common-bounded expected pseudo-regret is little-o of `T + 1`.

theorem selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret_isLittleO_natCast_succ {K : Nat} (model : FiniteBanditModel K) (armLaw : Fin K -> Measure Rat) (hprob : forall arm, IsProbabilityMeasure (armLaw arm)) (lo hi : Real) (hmeas : forall arm, AEMeasurable (fun reward : Rat => ((reward : Rat) : Real)) (armLaw arm)) (hbound : forall arm, Filter.Eventually (fun reward : Rat => Set.Icc lo hi ((reward : Rat) : Real)) (ae (armLaw arm))) (hmean : forall arm, integral (armLaw arm) (fun reward : Rat => ((reward : Rat) : Real)) = ((model.mean arm : Rat) : Real)) (defaultAction : Fin K) : (selectedPolicySuccessorBoundedFiniteArmExpectedPseudoRegret model armLaw hprob lo hi defaultAction) =o[atTop] (fun T : Nat => (((T + 1 : Nat) : Real)))
def BanditRLProof.UCB.selectedPolicySuccessorBoundedFiniteArmExpectedAveragePseudoRegret Compiled

Expected common-bounded sampled-successor regret normalized by `T + 1`.

noncomputable def selectedPolicySuccessorBoundedFiniteArmExpectedAveragePseudoRegret {K : Nat} (model : FiniteBanditModel K) (armLaw : Fin K -> Measure Rat) (hprob : forall arm, IsProbabilityMeasure (armLaw arm)) (lo hi : Real) (defaultAction : Fin K) (T : Nat) : Real
theorem BanditRLProof.UCB.selectedPolicySuccessorBoundedFiniteArmExpectedAveragePseudoRegret_tendsto_zero Compiled

For stationary finite-arm reward laws bounded almost surely in a common interval, the expected pseudo-regret of the horizon-indexed canonical sampled-pair UCB family, normalized by `T + 1`, tends to zero. No nondegeneracy premise `lo < hi` is needed because the parent practical route pads the genuine Hoeffding proxy before using it as the UCB parameter.

theorem selectedPolicySuccessorBoundedFiniteArmExpectedAveragePseudoRegret_tendsto_zero {K : Nat} (model : FiniteBanditModel K) (armLaw : Fin K -> Measure Rat) (hprob : forall arm, IsProbabilityMeasure (armLaw arm)) (lo hi : Real) (hmeas : forall arm, AEMeasurable (fun reward : Rat => ((reward : Rat) : Real)) (armLaw arm)) (hbound : forall arm, Filter.Eventually (fun reward : Rat => Set.Icc lo hi ((reward : Rat) : Real)) (ae (armLaw arm))) (hmean : forall arm, integral (armLaw arm) (fun reward : Rat => ((reward : Rat) : Real)) = ((model.mean arm : Rat) : Real)) (defaultAction : Fin K) : Tendsto (selectedPolicySuccessorBoundedFiniteArmExpectedAveragePseudoRegret model armLaw hprob lo hi defaultAction) atTop (nhds 0)