Proof of Lemma 13, first case — the unvisited transition region contains the simplex
ProvedNonstatRL.NoWidening.unvisited_region_is_simplexconfidence-regionmdpnon-stationaryp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Fix a finite state space with at least two states, nonempty allowed action sets, a horizon , and . For an allowed state-action pair with no visits in the chosen window, the transition radius satisfies
and every probability distribution on the states belongs to the unwidened confidence region .
This is the first case in the authors' construction of a transition model inside the region.
Formalization Note A zero count gives and empirical center . The hypotheses and are explicit pins needed for the printed radius claim. The window uses zero-based half-open indices.
Preamble
import Mathlib import Definitions.Def_NonstatRL_PerStep_Window open BanditAlgorithm
Formal statement
namespace NonstatRL.NoWidening
theorem unvisited_region_is_simplex {S A : ℕ} (hS : 1 < S)
(As : Fin S → Finset (Fin A)) (hAs : ∀ s, (As s).Nonempty)
(T : ℕ) (hT : 1 ≤ T) (δ : ℝ) (hδ : 0 < δ ∧ δ < 1)
(st : ℕ → Fin S) (ac : ℕ → Fin A) (lo hi : ℕ)
(s : Fin S) (a : Fin A) (ha : a ∈ As s)
(hzero : NonstatRL.PerStep.count st ac lo hi s a = 0) :
1 ≤ NonstatRL.PerStep.radP As T δ (NonstatRL.PerStep.countPlus st ac lo hi s a) ∧
∀ p : Fin S → ℝ, (∀ s', 0 ≤ p s') → (∑ s', p s') = 1 →
p ∈ NonstatRL.PerStep.Hp As T δ st ac lo hi 0 s a := by sorry
end NonstatRL.NoWidening
Source
Cheung, Simchi-Levi & Zhu, Non-Stationary Reinforcement Learning: The Blessing of (More) Optimism, arXiv:1906.02922v4, p. 50, proof of Lemma 13, first case
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.