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

Lean module · EXP3

BanditRLProof.Exp3PredictableHedge

# Almost-sure Hedge control for predictable sampled EXP3 This module discharges the pathwise nonnegative-feedback premise of `Exp3SampledHedge` from the generated predictable `[0,1]` reward law. It first aggregates the time-zero and successor reward identifications into one finite-horizon almost-sure event, then applies the concrete pathwise Hedge bound on that event. The endpoint remains an almost-sure inequality. Integrating it and comparing the pure Hedge distribution with the exploration-mixed sampling distribution are separate downstream steps.

Module map

Declarations
4
Placeholders
0

Imports

BanditRLProof.Exp3SampledHedge

Imported by

BanditRLProof, BanditRLProof.Exp3ExplorationBias

Declarations

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

theorem BanditRLProof.Exp3.sampledPredictableTrajectoryMeasure_reward_nonneg_ae Compiled

Every observed scalar reward is nonnegative almost surely under the generated predictable sampled-EXP3 trajectory law.

theorem sampledPredictableTrajectoryMeasure_reward_nonneg_ae {Env : Type u} {Action : Type v} [MeasurableSpace Env] [StandardBorelSpace Env] [MeasurableSpace Action] [MeasurableSingletonClass Action] [StandardBorelSpace Action] [Nonempty Action] [DecidableEq Action] (prior : Measure Env) [IsFiniteMeasure prior] (arms : Finset Action) (harms : arms.Nonempty) (eta gamma : Real) (hgamma_nonneg : 0 <= gamma) (hgamma_le_one : gamma <= 1) (loss : PredictableLossVector Env Action) (t : Nat) : let mu := prior ⊗ₘ sampledImportanceWeightedTrajectoryKernel arms harms eta gamma hgamma_nonneg hgamma_le_one loss.environment ∀ᵐ sample ∂mu, 0 <= (sample.2 t).2
theorem BanditRLProof.Exp3.sampledPredictableTrajectoryMeasure_finiteHorizon_reward_nonneg_ae Compiled

One common almost-sure event supplies reward nonnegativity at every time strictly before a finite horizon.

theorem sampledPredictableTrajectoryMeasure_finiteHorizon_reward_nonneg_ae {Env : Type u} {Action : Type v} [MeasurableSpace Env] [StandardBorelSpace Env] [MeasurableSpace Action] [MeasurableSingletonClass Action] [StandardBorelSpace Action] [Nonempty Action] [DecidableEq Action] (prior : Measure Env) [IsFiniteMeasure prior] (arms : Finset Action) (harms : arms.Nonempty) (eta gamma : Real) (hgamma_nonneg : 0 <= gamma) (hgamma_le_one : gamma <= 1) (loss : PredictableLossVector Env Action) (horizon : Nat) : let mu := prior ⊗ₘ sampledImportanceWeightedTrajectoryKernel arms harms eta gamma hgamma_nonneg hgamma_le_one loss.environment ∀ᵐ sample ∂mu, ∀ t, t < horizon -> 0 <= (sample.2 t).2
theorem BanditRLProof.Exp3.sampledPredictableTrajectoryMeasure_hedge_regret_le_ae Compiled

The concrete finite-horizon Hedge inequality holds almost surely on the generated predictable sampled-EXP3 trajectory.

theorem sampledPredictableTrajectoryMeasure_hedge_regret_le_ae {Env : Type u} {Action : Type v} [MeasurableSpace Env] [StandardBorelSpace Env] [MeasurableSpace Action] [MeasurableSingletonClass Action] [StandardBorelSpace Action] [Nonempty Action] [DecidableEq Action] (prior : Measure Env) [IsFiniteMeasure prior] (arms : Finset Action) (harms : arms.Nonempty) (eta gamma : Real) (heta : 0 < eta) (hgamma_nonneg : 0 <= gamma) (hgamma_le_one : gamma <= 1) (loss : PredictableLossVector Env Action) (horizon : Nat) (comparator : Action) (hcomparator : comparator ∈ arms) : let mu := prior ⊗ₘ sampledImportanceWeightedTrajectoryKernel arms harms eta gamma hgamma_nonneg hgamma_le_one loss.environment ∀ᵐ sample ∂mu, (Finset.range horizon).sum (fun t => mixedLoss arms eta (sampledTrajectoryObservedLoss arms eta gamma sample) t) - cumulativeLoss (sampledTrajectoryObservedLoss arms eta gamma sample) horizon comparator <= Real.log arms.card / eta + eta * (Finset.range horizon).sum (fun t => mixedSquaredLoss arms eta (sampledTrajectoryObservedLoss arms eta gamma sample) t)
theorem BanditRLProof.Exp3.sampledPredictableScoreHedge_ae Compiled

The same almost-sure inequality with the comparator cumulative estimator exposed as the inclusive concrete sampled history score.

theorem sampledPredictableScoreHedge_ae {Env : Type u} {Action : Type v} [MeasurableSpace Env] [StandardBorelSpace Env] [MeasurableSpace Action] [MeasurableSingletonClass Action] [StandardBorelSpace Action] [Nonempty Action] [DecidableEq Action] (prior : Measure Env) [IsFiniteMeasure prior] (arms : Finset Action) (harms : arms.Nonempty) (eta gamma : Real) (heta : 0 < eta) (hgamma_nonneg : 0 <= gamma) (hgamma_le_one : gamma <= 1) (loss : PredictableLossVector Env Action) (n : Nat) (comparator : Action) (hcomparator : comparator ∈ arms) : let mu := prior ⊗ₘ sampledImportanceWeightedTrajectoryKernel arms harms eta gamma hgamma_nonneg hgamma_le_one loss.environment ∀ᵐ sample ∂mu, (Finset.range (n + 1)).sum (fun t => mixedLoss arms eta (sampledTrajectoryObservedLoss arms eta gamma sample) t) - sampledHistoryScore arms eta gamma n (Preorder.frestrictLe n sample.2) comparator <= Real.log arms.card / eta + eta * (Finset.range (n + 1)).sum (fun t => mixedSquaredLoss arms eta (sampledTrajectoryObservedLoss arms eta gamma sample) t)