Theorem 7 — a pure policy read off an extreme optimal solution of the dual LP is average optimal
ProvedHordijkKallenbergLP.SingleLP.theorem_7average-rewardlinear-programmingmarkov-decision-processmultichainp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let , , with , and let be an optimal solution of the dual program
obtained by the simplex method, i.e. an optimal solution that is an extreme point of the feasible set. Let be any decision rule with such that
Then is average optimal: for every .
One linear program thus yields an average optimal pure stationary policy in every finite Markov decision chain, with no unichain assumption.
Formalization Note "The simplex method is used" is encoded as extremality of in the feasible set, the property of a basic feasible solution that the proof uses. The conclusion is asserted for every obeying the rule. Policies range over all history-dependent randomized policies.
Preamble
import Mathlib import Definitions.Def_MarkovDecisionProcesses_AverageReward import Definitions.Def_BlackwellDiscreteDP_NearOne_LimitMatrix import Definitions.Def_HordijkKallenbergLP_SingleLP_Model open MarkovDecisionProcesses BlackwellDiscreteDP.NearOne Matrix Filter Topology
Formal statement
namespace HordijkKallenbergLP.SingleLP
variable {S A : Type*} [Fintype S] [Fintype A] [DecidableEq S] [DecidableEq A]
/-- **Theorem 7.** If the simplex method is used to solve the dual problem (3)–(5) and an optimal
solution `(x, y)` is obtained, then the policy `f^∞`, where `f(i) = a_i` such that `x_{i a_i} > 0`
for `i ∈ E_x = {i | Σ_a x_{ia} > 0}` and `y_{i a_i} > 0` for `i ∉ E_x`, is average optimal.
Hordijk and Kallenberg, Linear Programming and Markov Decision Chains, Management Sci.
25(4):352–362 (1979), DOI 10.1287/mnsc.25.4.352, p. 357, Theorem 7.
**Formalization Note.** "The simplex method is used and an optimal solution is obtained" is
encoded as: `(x, y)` is an optimal solution of the dual program that is an extreme point of its
feasible set (a basic feasible solution, which is what the simplex method returns; the paper's
proof of Proposition 3 uses exactly this). Without extremality the statement is a different,
unproved claim. The conclusion holds for *every* `f` obeying the rule (Remark after Theorem 7:
"an arbitrary action"). Average optimality is Hordijk and Kallenberg's: `φ_i(f^∞) = φ_i` for all
`i`, against all history-dependent randomized policies. -/
theorem theorem_7 (M : StationaryMDP S A) [Nonempty S]
(β : S → ℝ) (hβpos : ∀ j, 0 < β j) (hβsum : ∑ j, β j = 1)
(z : (Pair M → ℝ) × (Pair M → ℝ)) (hext : z ∈ Set.extremePoints ℝ (dualFeasible M β))
(hopt : IsDualOptimal M β z)
(f : S → A) (hf : ∀ i, f i ∈ M.admissible i) (hsel : IsSelection M z.1 z.2 f hf) :
IsAvgOptimal M (stationaryPolicy M f hf) := by sorry
end HordijkKallenbergLP.SingleLP
Source
Hordijk and Kallenberg, Linear Programming and Markov Decision Chains, Management Sci. 25(4):352–362 (1979), DOI 10.1287/mnsc.25.4.352, p. 357, Theorem 7
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.