§8.1.2, p. 73 — Thompson sampling matches the posterior law of the optimal action
ProvedTSTutorial.UCBDecomp.ts_prob_matchingbayesian-banditsp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1probability-matchingthompson-sampling
Let be a measurable optimal action in the Bayesian model, and let be Thompson sampling with the same tie-break. For every history of periods and every action ,
This is the probability-matching property used in equation (8.4): conditional on the observed past, the next Thompson-sampling action has the posterior distribution of the optimal action.
Formalization Note The action and outcome sets are finite; the parameter space remains general. The likelihood includes the policy's action probabilities and the outcome kernel. The posterior uses the prior and outcome likelihood, with no restriction on policy behavior at histories of zero evidence.
Preamble
import Mathlib import Definitions.Def_TSTutorial_UCBDecomp_Setting open MeasureTheory open scoped BigOperators
Formal statement
namespace TSTutorial.UCBDecomp
variable {X Y Θ : Type*} [Fintype X] [DecidableEq X] [Nonempty X]
[Fintype Y] [DecidableEq Y] [MeasurableSpace Θ]
/-- TS draws the next action from the conditional law of the optimal action (p. 73). -/
theorem ts_prob_matching (m : Model X Y Θ) (xStar : Θ → X)
(hx : IsOptimalSelector m xStar) (π : Policy X Y)
(hπ : IsThompsonSampling m xStar π) (s : ℕ)
(h : Hist X Y s) (a : X) :
(∑ y : Y, histProb m π (s + 1) (snoc h (a, y))) =
jointProb m xStar π s a h := by sorry
end TSTutorial.UCBDecomp
Source
Russo, Van Roy, Kazerouni, Osband, Wen, A Tutorial on Thompson Sampling, Found. Trends Mach. Learn. 11(1) (2018), §8.1.2 p. 73, sentence following (8.4)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.