Vivian McPhail wrote:

Hi,

...

From the very helpful posts of John Hughes and others concerning "Monadic Loops", I've been advised to re-implement my neural net simulator using lazy state threads to avoid crashes to which recursive (and tail-recursive) monads lead.

I had been using monad transformers to lift IO operations:

data NeuralNet = NN { ... }

data NNO a = NNO (NeuralNet -> (a,NeuralNet))

data NNT m a = NNT (NeuralNet -> m (a,NeuralNet))

following the example of a state transformer in "Functional Programming with Overloading and Higher-Order Polymorphism".

I am trying to reimplement with

data ST s a

from Control.Monad.ST.Lazy and "Lazy Functional State Threads"

however there is something I think I do not understand, that is the "s" in the ST dataype.

The s is a safety mechanism, to ensure that you don't share an STRef or STArray between two state threads. It is usually instantiated/substituted where the runST function is applied. You don't supply a substitution for s; runST does. (The memoisation library which uses ST () is unusual in that respect.) Normal use of ST for your neural net example would be: do nnRef <- newSTRef (NN { ... }) -- set the initial NeuralNet state --from here on, you pass nnRef around as a parameter Control.Monad.State (as opposed to ST) may be a closer match for the NNO and NNT types. Regards, Tom