[ Was sent only to participants, but I meant to send it to the list ] On Wednesday 07 April 2004 06:04 am, Brandon Michael Moore wrote:

Keith is talking about a "comitted choice" style of nondeterminism, where one of the arguments is picked and the computation continues from there.

If you want a computation with backtracking, or a list of all possibly results then you should use the list monad, or another monad that supports nondetermanism.

An important feature of amb is known as "bottom-avoidance", and you can get that with Keith's scheme (although it won't be terribly fair; see mailing list discussions about the fairness or lack thereof of the GHC RTS scheduler), but you can't get it with backtracking (indeed, backtracking is entirely deterministic). Of course, Keith's scheme forces you to work within the IO monad, as well it should. If we allowed amb to have the type amb :: a -> a -> a then all of the transformations carried out within GHC's mid-stage would need to be re-examined for their correctness (in the context of a semantics that understood sharing, as non-determinism makes sharing observable). It's possible that many would still be valid, but it's by no means clear. To my knowledge, no-one has developed a correct semantics for sharing and amb that enables proof of simple transformations like let-floating or lambda-floating or even call-by-need beta reduction, let alone some of the more complex ones like deforestation (which requires a theory that can handle recursion well). Cheers, Andy -- Andy Moran Ph. (503) 626 6616, x113 Galois Connections Inc. Fax. (503) 350 0833 12725 SW Millikan Way, Suite #290 http://www.galois.com Beaverton, OR 97005 moran@galois.com