Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Clojure: Genetic Mona Lisa problem in 250 beautiful lines (npcontemplation.blogspot.com)
66 points by ynd on Jan 26, 2009 | hide | past | favorite | 31 comments


"Lambdas are not garbage collected.

Yes. That means lambdas can be the cause of memory leaks.

As described by Charles Nutter, each lambda in Clojure is an anonymous class in Java. The problem is that classes are never garbage collected. They reside in a special place called the PermGen."

I know very little about Clojure, but that seems like a huge problem for Clojure, and other functional-ish JVM languages. Is it?

I've used Rhino (JavaScript on JVM) extensively, and I don't think it suffers from this.


He means the compiled code of a lambda can't be GC'ed. He was creating new functions inside of an eval.

The individual closures created as code executes are GC'ed like everything else.

Since eval is rather specialized, and everyone recommends it be avoided in "normal" code, this is not a major problem.

Plus, it is fixed.


"As of r1232, lambdas created in eval can be GCed."

- Comment from cgrand at the bottom of the article.


Honestly, the JVM is pretty crappy right now for functional language support. A lot of stuff has to be implemented by hand in either the compiler or the program itself. Several features in Clojure take the latter track at the moment.

I advocate tracking down people from Sun and asking them in a personable, friendly tone of voice to please fix this kind of problem. On their front lawns.


In this thread:

http://markmail.org/message/axarlob7wwmnm2xe#query:clojure%2...

Rich Hickey claims that the PermGen issue alleged in this article doesn't actually exist. A new class is created every time a (fn) is compiled, but only a new instance of that class is created each time it is invoked.

The only thing that generates a new class each time it is called is (eval), which as we all remember from reading our PG, is generally regarded as suspect when used in production code. Obviously you can see this effect at the Clojure REPL, however.


Yes the PermGen issue in general seems a major weakness of the "Sun" VM (now OpenJDK) - although its not an intractable problem - just an irritation.

It seems an out dated idea today to have a separate space/special rules for generated code versus the rest of the heap - I think other VMs like JRockit didn't do that and didn't have that issue.

Although - I think the class/classloader relationship is in the spec, so to allow easy GC of generated classes you need to have 1 to 1 with classloaders (I hear that all this is going to change anyway to reflect how people really use things).


Genetic programming generally won't appear in production code, though it might be used to generate production code. Using eval seems reasonable in the context of executing code generated by your program. I'd be interested in hearing about alternatives.

In response to this problem, Clojure now uses an ephemeral classloader. I don't know the JVM that well, but Rich Hickey says this means the anonymous classes generated by evaling fn forms can be collected.


if all the classes in the classloader are no longer referenced, and there are no references to the classloader itself, then they will all be GCed (but often there is a ref to the classloader that causes the PermGen issue).


For those interested in GP, here is Koza's course page (pdfs & source):

http://www.genetic-programming.com/coursemainpage.html

Koza "invented" GP. However, parent article is really describing Evolutionary Programming, invented by Fogel in the 1960s. The difference? Ostensibly, no crossover operator in EP, only mutation.

http://www.scholarpedia.org/article/Evolutionary_programming


Incidentally, I wrote the same thing for a class recently in Common Lisp. See http://github.com/smanek/ga

The biggest portion of the code was the bitmap/image manipulation libraries. Excluding those, it was about the same size.


... how long did you let it run? The results in the presentation did not turn out very well.


Did you look at the videos, or just the static pictures? The static pictures are just screenshots taken halfway through the videos.

You can see my videos at: http://www.youtube.com/watch?v=S1ZPSbImvFE http://www.youtube.com/watch?v=-_VFZ_ON0A8 http://www.youtube.com/watch?v=VbyQomrJSeU

Due to time constraints (I underestimated the time it would take to write a fast bitmap/polygon library) I only let those simulations run for ~1 hr each.


I looked at the static pictures in the presentation PDF. Now, I've looked at the videos though. Great stuff.


I love the evolution of McCarthy!


Someone should do Paul Graham, so he can write a new lisp dialect to write another program to evolve another version of Paul Graham to write a new lisp dialect to write another program to evolve another version of Paul Graham to write a new lisp dialect to write another program to evolve another version of Paul Graham to write a new lisp dialect to...


In the informal algorithm description, it says "Create Children from the best programs by mating and mutating them". By mating, do you mean actually taking two (or more) "programs" and merging the contents? I looked over the code, but I don't know enough clojure to find any part that looked like it merged two programs.

Incidentally, now I want to learn clojure :).


I do know Clojure and I didn't see any mating either - just mutating. I haven't had any caffeine yet though.

I'm also working on genetic programming in Clojure, and with the PermGen fix in rev. 1232, it seems to be a great tool for the job.


I removed mating in the end because it made the program converge slower.

I'm glad that got you interested in Clojure!


As python is beautiful, lisp is parenthesis-y.

Has anyone thought about a Lisp dialect that uses white space, like Python, to get rid of the ()?



Why don't you make your editor to paint them white and indent them?


Some of us prefer the parens.


Make an editor which will randomly color-code each set of parentheses such that they or their background is in a set of colors which don't appear as other keywords (like function defs or strings or whatever) so as not to add confusion. You now have a way of more easily seeing every matching pair of parens without having to hover over each one. However, now you have to deal with a strange assortment of colors on your screen.

Oh wait, you can just use a gradient of colors instead of random ones, with the more-nested paren pairs having colors farther in on the gradient. Silly me.


"Oh wait, you can just use a gradient of colors instead of random ones, with the more-nested paren pairs having colors farther in on the gradient. Silly me."

Good idea! Is it doable in VIM now? :|


"Logo is a computer programming language used for functional programming. It is an easier-to-read adaptation and dialect of the Lisp language; some have called it Lisp without the parentheses."

http://en.wikipedia.org/wiki/Logo_programming_language


Parens let you have multiple expressions in one line, which indentation alone could not really express (except perhaps in a Haskell-like with currying, . and $ operators, etc).

For example, I don't see how (map (fn [x] (+ x 2)) (concat (take 5 my-series) etc...)) could be broken down with indentation alone without taking quite a few extra lines.


Dylan removes the parenthesis, but it doesn't go as far as using whitespace.

http://en.wikipedia.org/wiki/Dylan_programming_language


See I-expressions.


I like this, it gives a clear set of code with a simple non important purpose.

Neat.


This has already been done and much better at that. Don't give props to someone copying anothers work a half year after the fact.

Credit and originality goes to : http://rogeralsing.com/2008/12/07/genetic-programming-evolut...


The author of this article links to that same post and gives credit to Roger Alsing.

He was just using it as a demonstration of Clojure, nothing more.

Read the article before criticizing.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: