On Wed, Feb 23, 2011 at 10:52 AM, Chung-chieh Shan wrote:

Mostly we preferred (as do the domain experts we target) to write probabilistic models in direct style rather than monadic style. Haskell's laziness doesn't help -- in fact, to avoid running out of memory, we'd have to defeat that memoization by sprinkling "() ->" throughout the types.

I don't think that "() ->" is even guaranteed to work...

I am confused about this comment:

Mostly we preferred (as do the domain experts we target) to write

probabilistic models in direct style rather than monadic

In the haskell implementation of the lawn model there are two different version of the grassModel ( https://github.com/rst76/probability/blob/master/src/Lawn.hs) grassModel :: PM Bool grassModel = let_ (flip_ 0.3) (\ rain -> let_ (flip_ 0.5) (\ sprinkler -> let_ (dis (con (flip_ 0.9) rain) (dis (con (flip_ 0.8) sprinkler) (flip_ 0.1))) (\ grassIsWet -> if_ grassIsWet rain (dist [])))) and grassModel = do rain <- flip_ 0.3 sprinkler <- flip_ 0.5 wetInRain <- flip_ 0.9 wetInSprinkler <- flip_ 0.8 wetInOther <- flip_ 0.1 let grassIsWet = rain && wetInRain || sprinkler && wetInSprinkler || wetInOther if grassIsWet then return rain else dist [] By domain expert preferring direct style do you mean that they prefer the first version over the 2nd version? thanks, Daryoush On Wed, Feb 23, 2011 at 7:52 AM, Chung-chieh Shan wrote:

Hello! Thank you for your interest.

Daryoush Mehrtash wrote in haskell-cafe:

...

Is the "Embedded domain-specific language HANSEI for probabilistic models and (nested) inference" described in: http://okmij.org/ftp/kakuritu/index.html#implementation available in Haskell?

The closest to that I know of is this one: http://d.hatena.ne.jp/rst76/20100706 https://github.com/rst76/probability

Or you can apply this monad transformer to a probability monad: http://sebfisch.github.com/explicit-sharing/

...

Is there a reason why the author did the package in Ocaml rather than Haskell?

Mostly we preferred (as do the domain experts we target) to write probabilistic models in direct style rather than monadic style. Haskell's laziness doesn't help -- in fact, to avoid running out of memory, we'd have to defeat that memoization by sprinkling "() ->" throughout the types.

-- Edit this signature at http://www.digitas.harvard.edu/cgi-bin/ken/sig 1st graffitiist: QUESTION AUTHORITY! 2nd graffitiist: Why?

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

-- Daryoush Weblog: http://onfp.blogspot.com/

I see the problem now. But I am confused as to why there are no Bool class (like Num, Fractional...) in Haskell. If I had such a class then the problem is solved, (by making the "pm a" an instance of it) right? Or are there still more issues that I am not seeing? thanks, daryoush On Mon, Feb 28, 2011 at 7:34 AM, Chung-chieh Shan wrote:

Daryoush Mehrtash wrote in haskell-cafe:

...

I am confused about this comment:

...

Mostly we preferred (as do the domain experts we target) to write probabilistic models in direct style rather than monadic

In the haskell implementation of the lawn model there are two different version of the grassModel ( https://github.com/rst76/probability/blob/master/src/Lawn.hs)... By domain expert preferring direct style do you mean that they prefer the first version over the 2nd version?

No, there is no way to write probabilistic models in direct style in Haskell, and domain experts prefer neither Haskell version you showed.

A symptom of direct style is being able to write something like

flip 0.3 && flip 0.5

where (&&) takes two Bool arguments.

-- Edit this signature at http://www.digitas.harvard.edu/cgi-bin/ken/sig 1st graffitiist: QUESTION AUTHORITY! 2nd graffitiist: Why?

_______________________________________________ Haskell-Cafe mailing list Haskell-Cafe@haskell.org http://www.haskell.org/mailman/listinfo/haskell-cafe

-- Daryoush Weblog: http://onfp.blogspot.com/