3 ms·

Push programming language and genetic programming system implemented in Clojure

38 points by amirouche 11y ago

jwr 11y ago
Incidentally, I implemented PUSH in Clojure myself, several years ago. The language is well designed, it's a good match for genetic algorithms — you need something which is interpretable after mutations, and very few languages provide that. I learned a lot about Clojure performance while doing that work. It's funny how most of that knowledge is now obsolete anyway (because both Clojure and the JVM evolved so much).
vosper 11y ago
I've always figured that GP ought to be able to solve any problem, given sufficient computing power. Yet it's never mentioned when people talk about machine learning. Why is that? Also, I'd love to hear anyone's experience with GP to solve a practical problem - whatever language it was in.
- lspector 11y ago
  Many of the "human-competitive" results that have received "Humies" awards (http://www.genetic-programming.org/combined.php http://www.genetic-programming.org/combined.php) used GP. Some colleagues and I surveyed the first 10 years of Humies results in terms of techniques used and other factors; preprint here: https://www.dropbox.com/s/tjpa6afxqibwnno/Anaylzing_the_HUMIES.pdf https://www.dropbox.com/s/tjpa6afxqibwnno/Anaylzing_the_HUMI.... In what I think is another exciting development (although I'm biased -- I was his advisor), Tom Helmuth just finished his PhD at UMass using GP (and Clojush) for automatic programming, automatically synthesizing programs for first-year computer science assignments (abstract: https://www.cics.umass.edu/speakers/thomas-helmuth/2015/jul/27 https://www.cics.umass.edu/speakers/thomas-helmuth/2015/jul/...). This is just the tip of an iceberg, but it's a big iceberg. As for why GP is rarely mentioned when people talk about machine learning, I think there are a couple of reasons, but one important one is that GP usually shines the brightest on different kinds of problems than other ML techniques are applied to. If you put them head to head on a problem to which other ML techniques can be easily applied then GP will probably lose. But GP can succeed on problems for which it's not at all clear how one would even begin to apply most other ML techniques.
  - DougMerritt 11y ago
    I agree with all of that, but it's also important to note that simulated annealing also fills in a lot of those voids, and when it fits, is typically far more efficient than GP. So that narrows the application space further. GP is fascinating on the right kind of problem.

38 points by amirouche 11y ago

jwr 11y ago

Incidentally, I implemented PUSH in Clojure myself, several years ago. The language is well designed, it's a good match for genetic algorithms — you need something which is interpretable after mutations, and very few languages provide that. I learned a lot about Clojure performance while doing that work. It's funny how most of that knowledge is now obsolete anyway (because both Clojure and the JVM evolved so much).

vosper 11y ago

I've always figured that GP ought to be able to solve any problem, given sufficient computing power. Yet it's never mentioned when people talk about machine learning. Why is that? Also, I'd love to hear anyone's experience with GP to solve a practical problem - whatever language it was in.

lspector 11y ago
Many of the "human-competitive" results that have received "Humies" awards (http://www.genetic-programming.org/combined.php http://www.genetic-programming.org/combined.php) used GP. Some colleagues and I surveyed the first 10 years of Humies results in terms of techniques used and other factors; preprint here: https://www.dropbox.com/s/tjpa6afxqibwnno/Anaylzing_the_HUMIES.pdf https://www.dropbox.com/s/tjpa6afxqibwnno/Anaylzing_the_HUMI.... In what I think is another exciting development (although I'm biased -- I was his advisor), Tom Helmuth just finished his PhD at UMass using GP (and Clojush) for automatic programming, automatically synthesizing programs for first-year computer science assignments (abstract: https://www.cics.umass.edu/speakers/thomas-helmuth/2015/jul/27 https://www.cics.umass.edu/speakers/thomas-helmuth/2015/jul/...). This is just the tip of an iceberg, but it's a big iceberg. As for why GP is rarely mentioned when people talk about machine learning, I think there are a couple of reasons, but one important one is that GP usually shines the brightest on different kinds of problems than other ML techniques are applied to. If you put them head to head on a problem to which other ML techniques can be easily applied then GP will probably lose. But GP can succeed on problems for which it's not at all clear how one would even begin to apply most other ML techniques.
- DougMerritt 11y ago
  I agree with all of that, but it's also important to note that simulated annealing also fills in a lot of those voids, and when it fits, is typically far more efficient than GP. So that narrows the application space further. GP is fascinating on the right kind of problem.