Random Solutions are Often Good Enough

http://www.sultanik.com/blog/RandomSolutions

by luu • 12 years ago

147 39 12 years ago

swayvil• 12 years ago

Get the depth (distance from edge to deepest point) of an irregular polygon. At first I was doing all kinds of crazy geometry and the future looked bleak. Then I was "fuck it" and did a randomsample, reduce, randomsample, reduce thing. It's fast and precise.

timdumol• 12 years ago

In case you're still interested in an exact solution, I think computing the Voronoi diagram for the polygon (treating each vertex as a point) should help in finding the answer. The deepest point should lie on one of the edges of the Voronoi diagram.

lifeformed• 12 years ago

I had a similar experience. I was writing a polygon intersection algorithm, heavily optimized for the domain of inputs, but kept having issues when vertices were coincident or when edges were colinear. I spent days trying to fix it, until I decided to just truncate the vertex values to some good-enough level of precision, and add a random unique jitter to each value (past the truncated precision). This made sure that there were never coincident or colinear situations. Then I just truncated the resulting intersection points again to remove the jitter.

It was trivial to add, and it was faster than all the extra checks I'd have to do for the edge cases. Also, it gave me peace of mind, since I knew it worked in theory, instead of my other feeble attempts which worked most of the time, but always failed in some new edge case.

leeoniya• 12 years ago

i'm currently working on a box-packing algo that attempts to find likely-to-be-used boxes to pack an order for shipping cost estimation. (variant of the knapsack problem [1]). "close enough" is the name of the game here :)

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

swehner• 12 years ago

Another angle: totally inefficient solutions can be useful too. I put this sample up a while ago, Ridiculous Adding, https://gist.github.com/stephanwehner/11227766

ajtulloch• 12 years ago

This seems to be a specific case of the more general principle of "concentration of measure". Terence Tao at [1] has some useful notes. From a quick scan of the paper at http://arxiv.org/pdf/1402.0423v1.pdf, Theorem 1 seems to essentially be a CoM result for the order statistics of symmetric RVs.

This same principle can be applied in other contexts [2] - e.g. Johnson-Lindestrauss is essentially a statement that random projections are "often" nearly isometries.

[1] http://terrytao.wordpress.com/2010/01/03/254a-notes-1-concen...

[2] https://news.ycombinator.com/item?id=7971188

thanatropism• 12 years ago

Where is those people "programming isn't math" God now, eh?

ISL• 12 years ago

"Offered the choice between mastery of a five-foot shelf of analytical statistics books and middling ability at performing statistical Monte-Carlo simulations, we would surely choose to have the latter skill." -- Numerical Recipes in C++, 2nd ed. p.696.

brockrockman2• 12 years ago

a good reminder that the TSP optimization problem is only NP-hard, not NP-complete -- http://www.nomachetejuggling.com/2012/09/14/traveling-salesm...

sidww2• 12 years ago

"only" NP-hard? You know that NP-hard problems are equal to order harder than NP-complete ones, right?

Strilanc• 12 years ago

Probably referring to the definition, not the difficulty.

codeflo• 12 years ago

Yes, NP-complete is a subset of NP-hard. Grandparent states that NP-completeness only applies to decision problems (which is correct), but seems to think that NP-hardness applies to optimization problems (which is incorrect; NP-hard too is a class of decision problems).

brockrockman2• 12 years ago

didn't mean to imply "only" as a less-than comparison -- just a distinction.

pdonis• 12 years ago

I wonder about one statement in the article: he says that Sudoku is NP-hard. I wonder about this since Peter Norvig has written a Sudoku solver:

http://norvig.com/sudoku.html

Of course, this solver is for the standard Sudoku game on a 9 x 9 grid. Perhaps the claim in the article refers to "Sudoku" in the sense of the generalized problem of solving n x n Sudoku, for any n?

rach3• 12 years ago

NP-hard does not mean unsolvable. I don't see that solver claiming to be polynomial?

matchu2• 12 years ago

The solver doesn't obviously negate the puzzle's NP-completeness. It first reduces the solution space, then starts brute-forcing what's left. I don't have a proof on hand, but that just smells nonpolynomial.

You could argue that, if we consider "Sudoku" to be the instance of the puzzle where n=9, then the puzzle isn't NP-hard because it has the constant-time solution of trying all 6.67e21 solutions, but the result is so uninteresting that it's understood that the author means the more general problem.

jimmaswell• 12 years ago

Depends on the initial state of the puzzle. Sometimes there's always a deterministic set of correct next moves to make. I guess this refers to cases that aren't like that?

tlarkworthy• 12 years ago

its easy to construct a Sudoku that only has a few numbers in it, and such you have to try filling it in many times before discovering a contradiction. Puzzle authors know this is not fun so the ones used for entertainment avoid being pathological.

zwegner• 12 years ago

As a general rule, when anybody refers to a problem being NP-hard or whatever, they're talking about a generalized version of the problem, with at least one "size" variable--the time complexity has to be polynomial/exponential/whatever in something.

9x9 Sudoku has a constant size, which means that it can be evaluated in O(1) time, if such notation makes sense.

fenomas• 12 years ago

Perhaps useful to add that Sudoku reduces to the more general exact-cover problem[0], which is indeed NP-hard.

Also, if you've never run across Dancing Links[1], do yourself a favor and go take a look. It's a method (by Knuth) for implementing an algorithm (ditto) for generally solving any exact-cover problem (including Sudoku of course). Basically it amounts to constructing a matrix that encompasses all of the game's constraints and possible inputs, after which all solutions can be found through mechanical operations on the matrix. It's both beautiful and very fun to implement.

[0] http://en.wikipedia.org/wiki/Exact_cover [1] http://en.wikipedia.org/wiki/Dancing_Links

jwecker• 12 years ago

In the context of clever startup solutions, this brings to mind the quote from George S Patton: "A good plan, violently executed now, is better than a perfect plan executed next week." Even if they didn't approach optimal solutions, the fact that generating random feasible solutions can be much easier and result in more tractable mental models (and therefore easier to maintain code etc.) means you can be 'violently executing' in the field of battle much earlier.

The fact that they approach optimality is gravy.

bshimmin• 12 years ago

Slightly worthless comment, but I think that is the first web site I've ever seen typeset in Computer Modern. Strangely, it looks rather nice to me (not that I think Computer Modern is particularly an attractive font, but my massive exposure to it over the years has given it a certain quality of reassurance).

zwieback• 12 years ago

It's not very readable but lends instant gravitas to the material, sort of like the anti-Comic Sans

brucehauman• 12 years ago

The last line is priceless.

"So the moral of the story is: Sometimes quickly and mindlessly choosing a random solution isn't half bad!"

JacobEdelman• 12 years ago

Actually, with the constant of approximation being two, its almost exactly half good.

coldcode• 12 years ago

As long as it isn't half-ass.

iandanforth• 12 years ago

This page is awesome, I love it. The author presents the material clearly and carefully. So many authors dealing with mathematical concepts are lazy and just throw up equations. Yuk! Here we see a careful progression of complexity supported with complete visual examples and numerical values, using the equations, worked out. I wish this were not so unusual, but huge thanks for having done it!

nicholas734• 12 years ago

I built my sudoku site using a pseudo random puzzle generator. I don't know the best algorithm to generate puzzles, so with some basic parameters for good puzzles, I let the puzzles fill out randomly. I generate, then test them for unique solutions and difficulty level. If the puzzle is invalid, I toss it. What I lack in algorithmic knowledge, I make up for with server resources :)

http://sudokuisland.com

I'm not a CS guy so reading this article made me feel a bit better, especially since it mentions sudoku.

Terr_2• 12 years ago

My gut-feeling is that building Sudoku puzzles shouldn't be too hard to generate, because every random choice you make reduces the space you need to pull from for the next choice.

Whether they're "hard enough to solve", though....

nicholas734• 12 years ago

The difficulty comes from multiple solutions to your puzzles (which would force the player to guess). As it turned out, a purely randomly selected puzzle would 99.9%* of the time give you a puzzle with multiple solutions. I would have settled for 99%, but more was too much server resources to hog!

And then add the screening for difficulty levels.

*Not an exact figure =)

LeicaLatte• 12 years ago

Totally tangential but,

Marco points out in latest ATP how he randomizes the lists of podcasts every time the list is shown. Neat!

Random Solutions are Often Good Enough

Related Stories