Lemma 3 (Dynamic vs Static adversary) — a stationary policy has the same value in the dynamic and the static model
ProvedRobustDP.Discounted.lemma3_dynamic_staticdynamic-programminggame-theoryp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1robust-optimization
Let be a discounted ambiguous MDP, let be a randomized Markov decision rule ( for every ), and let be the stationary policy. Let be its value against the dynamic adversary and its value against the static adversary, who fixes one per state–action pair. Then for every state ,
For stationary policies the adversary's best response is static, so the optimal stationary policy of the static model is obtained from (30).
Formalization Note The page states the lemma for "any decision rule" and proves it for deterministic rules, adding that "the same technique extends to randomized policies". It is stated here for randomized Markov rules, which include the deterministic ones as point masses.
Preamble
import Mathlib import Definitions.Def_RobustDP_Discounted_Model import Definitions.Def_RobustDP_Discounted_Value import Definitions.Def_RobustDP_Discounted_Bellman
Formal statement
namespace RobustDP.Discounted
theorem lemma3_dynamic_static {S A : Type*} [Countable S] [Countable A] (M : Model S A)
(q : RandRule M) (s : S) :
Vstatic M (stationaryRand M q) s = V M (stationaryRand M q) s := by sorry
end RobustDP.Discounted
Source
Iyengar, Robust dynamic programming, CORC Tech Report TR-2002-07 (rev. May 4, 2004), p. 14, Lemma 3; static model, p. 9
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.