问题
On this paper, the author suggests a translation between lambda terms:
data Term = Zero | Succ Term | App Term Term | Lam Term
and interaction nets:
data Net = -- if I understood correctly
Apply Net Net Net
| Abstract Net Net Net
| Delete Net Net Int
| Duplicate Net Net Net Int
| Erase Net
Unfortunately I can not understand his compilation algorithm. It seems like the actual algorithm is missing, and I have no idea what he means with the images on the third page. I've tried understanding it by looking at the published source code, but the author defines it using his own graph rewriting DSL so I would have to learn it first. How exactly is the translation implemented as a normal Haskell function?
回答1:
I have an implementation of interaction net reduction in Haskell that uses STRef
s to represent the net graph structure in a mutable way:
data NodeType = NRot
| NLam
| NApp
| NDup Int
| NEra
deriving (Show)
type NodeID = Int
type Port s = STRef s (Node s, PortNum)
data Node s = Node
{ nodeType :: !NodeType
, nodeID :: !NodeID
, nodePort0, nodePort1, nodePort2 :: !(Port s)
}
Conversion from lambda terms is implemented in a separate module. It is not the nicest code I've ever written, because it's a straight transliteration of a Javascript implementation and I didn't really take the time to understand what the JS version is doing:
encodeLam :: Lam -> IntNet s (Node s)
encodeLam lam = do
nextTag <- do
ref <- lift $ newSTRef 0
return $ lift $ do
modifySTRef ref succ
readSTRef ref
let go scope up (Lam body) = do
del <- mkNode NEra
lam <- mkNode NLam
linkHalf lam P0 up
link (lam, P1) (del, P0)
link (del, P1) (del, P2)
bod <- go (lam:scope) (lam, P2) body
linkHalf lam P2 bod
return (lam, P0)
go scope up (App f e) = do
app <- mkNode NApp
linkHalf app P2 up
linkHalf app P0 =<< go scope (app, P0) f
linkHalf app P1 =<< go scope (app, P1) e
return (app, P2)
go scope up (Var v) = do
let lam = scope !! v
(target, targetPort) <- readPort lam P1
case nodeType target of
NEra -> do
linkHalf lam P1 up
return (lam, P1)
_ -> do
dup <- mkNode . NDup =<< nextTag
linkHalf dup P0 (lam, P1)
linkHalf dup P1 up
link (dup, P2) =<< readPort lam P1
linkHalf lam P1 (dup, P0)
return (dup, P1)
root <- asks root
enc <- go [] (root, P0) lam
linkHalf root P0 enc
return root
It also implements the reverse transformation:
decodeLam :: Node s -> IntNet s Lam
decodeLam root = do
(setDepth, getDepth) <- do
ref <- lift $ newSTRef mempty
let set node depth = lift $ modifySTRef ref $
IntMap.insertWith (\ _new old -> old) (nodeID node) depth
get node = lift $ (! nodeID node) <$> readSTRef ref
return (set, get)
let go depth exit (node, port) = do
setDepth node depth
case nodeType node of
NDup _ -> do
let (port', exit') = case port of
P0 -> (head exit, tail exit)
_ -> (P0, port:exit)
go depth exit' =<< readPort node port'
NLam -> case port of
P1 -> do
depth' <- getDepth node
return $ Var (depth - depth' - 1)
_ -> Lam <$> (go (succ depth) exit =<< readPort node P2)
NApp -> do
f <- go depth exit =<< readPort node P0
e <- go depth exit =<< readPort node P1
return $ App f e
go 0 [] =<< readPort root P0
来源:https://stackoverflow.com/questions/29178805/how-do-you-translate-from-lambda-terms-to-interaction-nets