Hacker News new | ask | show | jobs
by fadmmatt 5957 days ago
Herb Simon noted the optimal solution is often infeasible a long time ago.

He argued that economic agents rarely finding a satisfying solution to their constraints, and instead choose a "satisficing" or "good-enough" solution.

Simon also noted that the satisficing solution found by humans is often close to the optimal solution.

He got (and deserved) a Nobel prize for these two observations.

I find work of the flavor "X reduces to P = NP, so Nyah!" to be deeply unsatisfying and misleading.

For problems of human importance, good-enough solutions to NP problems can often be found in a reasonable amount of time.

In fact, state-of-the-art SAT solvers are mind-bogglingly fast (on average). They can find solutions in seconds, where the theoretical worst case solution time is many-fold the expected lifetime of the universe.

Generating a hard instance of SAT is actually hard to do. One of the few reliable techniques we know of is to encode prime factorization into a circuit. This is good news for crypto, but bad news for anyone looking to show the real-world intractability of a process by reducing it to P=NP. Few processes in nature actually mimic prime factorization or graph isomorphism.

4 comments

"X reduces to P = NP, so Nyah!" is a tightly formulated theoretical statement. Work of this form is very important in advancing our basic understanding of the fundamental mathematical foundations behind computation.

It only become deeply unsatisfying and misleading for people who do not understand how to translate this information to the real world.

It seems from your post as you do have that understanding, but not the respect for the importance of the underlying math.

> It only become deeply unsatisfying and misleading for people who do not understand how to translate this information to the real world.

Not really. The OP is arguing that translating this statement to the real world adds nothing we did not know before. In comparing markets with NP-complete problems an interesting (but still trivial after you get the idea) paper "Computational Complexity and Information Asymmetry in Financial Products" by Arora et al.

And, also, the point of this paper is not really interesting. Even if it takes exponential time for the market to reflect _all_ the available information about something, all the efficient market hypothesis says is that the market will incorporate the information as fast as the fastest agent in incorporating that information. And this has nothing to do with P and NP (since a quadratic market can still be beat by a linear agent, and this is a market that is definitely in P).

In the real world, problem \in P isn't that enlightening. The constant factors matter.

I haven't looked into modern computational complexity research a whole lot, but I suspect it gets a disproportionate amount of attention because of the historical successes and importance of the field. Kind of like particle accelerators and NASA.

I would just like to point that constants are rarely the important point. Exponents are a huge deal. A lot of things are possible and fast because FFT is N log N versus N^2.
Great commentary about the actual ramifications of NP problems.

This reminds me of a company I used to do some marketing work - CombineNet. They would basically take a large, multi-participant marketplace (for example, several hospitals buying a billion dollars in goods from pharma companies and medical device manufacturers).

CombineNet's technology wouldn't guarantee the optimal buy to the hospitals, but it usually could, and in the cases that it couldn't, it could get very close to optimal and specify how close the answer was. I'm not a mathematician, but I think they were solving a "clearing problem?"

Coincidentally, the founder of CombineNet was a professor at the same school as Herb Simon... Carnegie Mellon.

Sandholm also wrote a heads-up Texas Hold'em bot that derived its strategy from the rules, and beat a ton of other heads-up poker bots handily. I think articles about this showed up on HN a while back: http://portal.acm.org/citation.cfm?id=1402350

For those without acm access, also available at http://www.cs.cmu.edu/~sandholm/texas.aaai06.pdf
Yeah, this has led to the odd situation in AI that reducing to SAT is how you show a problem is easy--- there's a whole cottage industry of people speeding up algorithms by reducing them to SAT, and it's often seen as a promising result for your computational tractability if you can do so.

(And the biggest problems, when they arise, are usually actually in the reduction itself---it takes too damn long to write out the SAT problem. Once you've got the SAT problem written out, solving it is rarely an issue.)

Another factor is that the weak inefficient market hypothesis (as far as I know) isn't a claim that we can't find the method for analyzing past prices. It's a claim that there is no possible method because the future price of a security is for all purposes stochastically independent of past prices. No matter what algorithm you use, NP, P, or any other algorithmic space, you can never reliably predict the value of a variable that is independent of its previous values.

I'm not positive that this defeats the argument this guy was making (not a computer scientist, only a programmer), but from what I could gather it seems to.