Policy expectations converge when each decision eventually agrees
ProvedPandoraBO.Private20261023.policy_expectations_converge_of_eventual_decisionsoptimal-searchpandoras-box
If each first decision and each decision at a fixed history eventually agrees with a limiting policy, each finite rollout eventually agrees pointwise, and expected reward and cost converge. No regularity of the reward distributions beyond the original integrability assumptions is needed.
Preamble
import Definitions.Def_PandoraBO_Model open MeasureTheory Filter open scoped Topology set_option autoImplicit false
Formal statement
theorem PandoraBO.Private20261023.policy_expectations_converge_of_eventual_decisions {n : ℕ} (M : PandoraBO.Instance n)
(policies : ℕ → PandoraBO.Policy n) (π : PandoraBO.Policy n)
(hfirst : ∀ u : ℝ, ∀ᶠ k in atTop, (policies k).first u = π.first u)
(hnext : ∀ h : PandoraBO.History n, ∀ u : ℝ,
∀ᶠ k in atTop, (policies k).next (h, u) = π.next (h, u)) :
(∀ ω : PandoraBO.Outcome n, ∀ᶠ k in atTop,
PandoraBO.terminalHistory (policies k) ω = PandoraBO.terminalHistory π ω) ∧
Tendsto (fun k => PandoraBO.expectedCost M (policies k)) atTop
(nhds (PandoraBO.expectedCost M π)) ∧
Tendsto (fun k => PandoraBO.expectedReward M (policies k)) atTop
(nhds (PandoraBO.expectedReward M π)) := by sorry
Source
Auxiliary finite policy rollout and dominated convergence lemma for the author revised Pandora Bayesian optimization manuscript.