Source lean/Tests/RewardModel.lean
import Determinize.Finite.Reward.Exploreimport Determinize.Finite.Reward.Solveimport Determinize.Proof.RewardModel.Normalizationimport Determinize.Proof.RewardModel.Momentsimport Determinize.Proof.RewardModel.Soundnessimport Tests.Parsingnamespace Determinize.Testsopen Determinize.Finiteprivate def parallelRewards : Spec.RewardModel.Model where  size := 2  initial := 0  kind := fun i ↦ if i = 0 then .transient else .returned 0  edges := fun i ↦ if i = 0 then [⟨1, 1 / 2, 0⟩, ⟨1, 1 / 2, 2⟩] else [⟨1, 1, 0⟩]  nonnegative := by decide +kernel  normalized := by decide +kernelprivate def rewardCase (text : String) (mass first second : Rat) : IO Unit := do  let program ← IO.ofExcept (Frontend.compile text)  match Reward.explore program.source .source {maxStates := 100} with  | .complete candidate valid =>    let solution ← IO.ofExcept (Reward.solve (candidate.toModel valid))    check (solution.statistics == ⟨mass, first, second⟩) s!"additive equations: {text}"  | .incomplete _ _ _ _ => throw (IO.userError s!"additive exploration did not finish: {text}")  | .failed _ failure => throw (IO.userError s!"additive exploration failed: {failure.message}")def rewardModels : IO Unit := do  let parallel ← IO.ofExcept (Reward.solve parallelRewards)  check (parallel.statistics == ⟨1, 1, 2⟩) "distinct rewards with the same target"  rewardCase "let f = rec f u => if flip(0.5) then 0 else 1 + f u in f ()" 1 1 3  rewardCase "let f = rec f u => if flip(0.5) then -1 else 1 + f u in f ()" 1 0 2  rewardCase "let f = rec f u => if flip(0.5) then 0 else (-1) + f u in f ()" 1 (-1) 3  rewardCase "let f = rec f u => if flip(0.5) then 0 else (1 + 2*bernoulli[G](0.5)) + f u in f ()"    1 2 13  rewardCase "let f = rec f u => if flip(0.5) then 1 + f u else if flip(0.5) then 0 else (let _ = \      observe(false) in 0) in f ()" (1 / 2) (1 / 2) (3 / 2)  rewardCase "let d = rec d u => d u in let f = rec f u => if flip(0.5) then 1 + f u else if \      flip(0.5) then 0 else d () in f ()" (1 / 2) (1 / 2) (3 / 2)  rewardCase "let f = rec f u => 1 + f u in f ()" 0 0 0  rewardCase "1 + (-2)" 1 (-1) 1  rewardCase "if flip(0.5) then 1 + 0 else 3 + (let _ = observe(false) in 0)" (1 / 2)    (1 / 2) (1 / 2)  for text in [      "let f = rec f u => if flip(0.5) then 0 else 1 + f u in 2 * f ()",      "let f = rec f u => if flip(0.5) then 0 else f u + 1 in f ()",      "let f = rec f a => if flip(0.5) then a else f (a+1) in f 0"] do    let program ← IO.ofExcept (Frontend.compile text)    match Reward.explore program.source .source {maxStates := 100} with    | .incomplete .states _ _ _ => pure ()    | _ => throw (IO.userError s!"unexpected additive extraction across a barrier: {text}")example : Reward.normalize (.deliver (.bool true)    [.right .add (.number 3), .right .add (.number (-3))]) =      (0, .deliver (.bool true) [Reward.guard]) := by decide +kernelexample : ∃ failure, Reward.step (.deliver (.bool true) [Reward.guard]) = .error failure :=  ⟨.invalid "binary operand types", rfl⟩example : Reward.normalize (.deliver (.number 2)    [.right .add (.number 3), .right .mul (.number 4)]) =      (0, .deliver (.number 2) [.right .add (.number 3), .right .mul (.number 4)]) := by  decide +kernelend Determinize.Tests#print axioms Determinize.Proof.RewardModel.integrableMoments#print axioms Determinize.Proof.RewardModel.replay_matches#print axioms Determinize.Proof.RewardModel.solution_result#print axioms Determinize.Proof.RewardModel.solution_conditional_variance#print axioms Determinize.Proof.RewardModel.solution_termination