2 ms·

Somewhat related, I've just been reading a summary [0] of a lecture by physicist/mathematician Stephen Wolfram on his quest to find the Program That Runs The Un

Somewhat related, I've just been reading a summary [0] of a lecture by physicist/mathematician Stephen Wolfram on his quest to find the Program That Runs The Unverse. An excerpt: "Wolfram goes off onto something else. He talks about searching through the space of possible computer programs, for ones of a certain type that can solve problems, rather than the having to go through the labor of humans programming something for a specific task. His meaning is more than a little unclear here. Wolfram says we're constrained to follow historical patterns of engineering, to only create things such that we have designed them with an understanding of what they will do. Nature, he says, doesn't have this constraint. Alright, this make more sense. He's talking about mining all potential programs, including useful ones that we ourselves might not have thought of. That's quite an idea, and sounds like no small task to actually perform it. Now he's bringing it in the direction I thought he would — comparing the search among the space of possible programs to natural selection, which is the natural equivalent of searching the space of possible biological programs. Now Wolfram is talking about whether we can search for our physical universe among the space of possible universes, and he's going still another level "meta." -- etc etc But doesn't a search in general program space almost immediately result in a combinatorial explosion? How on Earth are we supposed to find these programs that Wolfram dreams about? [0] http://futurisms.thenewatlantis.com/2009/10/stephen-wolframs-alternate-dimensions.html http://futurisms.thenewatlantis.com/2009/10/stephen-wolframs...

throwaway316943 6y ago
There are lots of algorithms for optimizing the search of a space. If you have a goal in mind it’s possible to come up with a strategy that will get you there faster. Think breadth first vs A star
- jmmcd 6y ago
  The terminology is bad. Don't confuse "search and optimisation"-type search with "pathfinding"-type search. A-star is the latter, but GP (and other metaheuristics like GA) are the former. To be clear -- if you're looking for eg a good exam timetable, you just need a good timetable -- you don't need a path from some other point in the space to that timetable.
red2awn 6y ago
Well GP is not random search, but still, the space is too huge for it to find anything of significance yet. I remember reading a paper that trained GP to write a sorting program, that is around the most complex problem we can solve with it now. In my opinion, what makes natural selection work but not GP, is that the life time of Earth and the number of individuals existed vastly exceeds the typical parameters of a GP simulation. Biological programs are also less likely to be killed off by a mutation compared to GP programs, so in some ways, the gradient is a lot smoother.

throwaway316943 6y ago

There are lots of algorithms for optimizing the search of a space. If you have a goal in mind it’s possible to come up with a strategy that will get you there faster. Think breadth first vs A star

jmmcd 6y ago
The terminology is bad. Don't confuse "search and optimisation"-type search with "pathfinding"-type search. A-star is the latter, but GP (and other metaheuristics like GA) are the former. To be clear -- if you're looking for eg a good exam timetable, you just need a good timetable -- you don't need a path from some other point in the space to that timetable.

red2awn 6y ago

Well GP is not random search, but still, the space is too huge for it to find anything of significance yet. I remember reading a paper that trained GP to write a sorting program, that is around the most complex problem we can solve with it now. In my opinion, what makes natural selection work but not GP, is that the life time of Earth and the number of individuals existed vastly exceeds the typical parameters of a GP simulation. Biological programs are also less likely to be killed off by a mutation compared to GP programs, so in some ways, the gradient is a lot smoother.