Traveling Salesman: the most misunderstood problem

http://www.nomachetejuggling.com/2012/09/14/traveling-salesman-the-most-misunderstood-problem/

by alter8 • 14 years ago

115 43 14 years ago

nomachetejuggling.com

tgflynn5• 14 years ago

The TSP optimization problem is polynomial time reducible to the TSP decision problem by binary search.

Suppose we have an algorithm D(TSP,C) which returns 1 if the TSP has a path of cost <= C.

Then the basic idea for an algorithm for the optimization version is :

  C = sum( costs of all paths in graph )
  while D(TSP,C) == 1:
    C = C / 2

Probably the greatest thing about having a polynomial time algorithm for an NP complete problem would be that it would allow finding the global optimum for any finite optimization problem in polynomial time.

chmike• 14 years ago

A priori it would be a computer science revolution if this would have been missed by everyone until now. So forgive me to be particularly cautious with your proposal. What I think is missing in your analysis is that we are lacking a polynomial method to build an hamiltonian path with cost <= C. If we had, then we would apply your algoritm to optimize in polynomial time.

alextp• 14 years ago

The above construction is a not a breakthrough and it is how I was taught in my approximate algorithms class. Also the way the optimization problem is usually defined is not as "give a circuit with minimal cost" but "what is the cost of the minimal-cost circuit?", which is easier to reason about in an approximate way.

tgflynn5• 14 years ago

No revolution intended. This is well known. The revolution would be finding a polynomial time D.

cperciva• 14 years ago

More precisely, binary search (with a correct termination condition) takes "is there a circuit with cost less than X?" and answers "what is the minimum circuit cost?".

Going from there to "what is a minimum-cost circuit?" is a bit more work, but still polynomial:

  G' = G
  X = TSP-minimum-circuit-cost(G)
  for E in edges(G):
    if (exists-circuit-with-cost(G' minus E, X))
      G' = G' minus E
  return (G')

konstruktor6• 14 years ago

What you are describing isn't even a proper binary search, it just gives you a lower bound that is guaranteed to be no more than 50% smaller than the solution.

Let's assume doing an actual binary search. What you are getting, in every step of binary search using your oracle, is an upper bound U (D(TSP, U) ==1), for which we know that there is a solution smaller or equal that U, and a lower bound L (D(TSP, L) == 0) for which we know there isn't. This is where your confusion between finite and integer-valued comes into play: If the problem were integer valued, you would have the optimal solution S as soon as L_i=U_i (implying U_i is S), which would be guaranteed after a polynomial number of steps. With real valued weights, all you get is a sequence of intervals (L_1, U_1], (L_2, U_2], ... where the size of the intervals is halved in every step. What you actually want is the exact value of the solution. For that, you would need to show that there is no hamiltonian path between L_i and U_i, which isn't something your oracle can determine.

Note that this probably doesn't show that a polynomial oracle for the TSP decision problem will not yield a polynomial solution to the optimization problem in general, it just deals with your proposition. Edit: Not -> Note

Peaker2• 14 years ago

The set of all possible costs is the set of sums of all N-sized subsets of the edges.

You only need to binary search that set, not the real numbers' axis. The set's size is exponential to N, but binary searching is logarithmic, so it becomes polynomial in N.

jb173• 14 years ago

As I mentioned somewhere else in this thread, the binary search is not a reduction, because you need to use TSP-decide a polynomial number of times (and for a reduction you reduce to one problem instance).

What you show is that the problem of finding the lowest cost is in FP^NP (the class of function problems solvable in polynomial time with the help of an NP oracle). However, by taking your idea a bit further one can show that TSP is in FP^NP. In fact, it is known that TSP is FP^NP-complete, i.e. it is one of the hardest problems in that class (and thus likely not in FNP, the NP for function problems).

qiemem• 14 years ago

I didn't immediately see what guaranteed that the binary search would run in polynomial time. In case anyone else doesn't see it:

There are O(V!) possible paths. Assuming binary search actually halves the search space each iteration, it will run in O(log(V!)) = O(log(n(n-1)...*1)) = O(log(n) + log(n-1) + ... + log(1)), which is clearly polynomial.

scythe• 14 years ago

You assume that in order to verify that a circuit exists, the algorithm for TSP-DECIDE must provide one. However, in a number of cases in mathematics -- notably the proof of Lagrange's four-square theorem -- this is not the case. It is possible that an algorithm could verify that a circuit exists without being able to find it.

tgflynn5• 14 years ago

It's true that my formulation isn't completely specified but the basic idea still works because if you have a decision function for an NP problem you can find a solution (assuming at least one exists) with one call to D per bit of the solution.

So to actually solve the optimization version you would need two iterations/recursions around the call to D. The inner iteration finds a solution with the desired property, ie. cost <= C and the outer one does the binary search over C. Overall it still reduces to a number of calls to D that is polynomial in the number of bits used to represent the problem, including those used to represent the weights.

The problem with the original blog post is that while the author states correctly that the TSP optimization problem is NP hard rather than NP complete because it doesn't have the form of an NP problem, he doesn't mention that it is also NP easy and therefore NP-equivalent (see last line in this Wikipedia article):

http://en.wikipedia.org/wiki/NP-equivalent

gilgoomesh• 14 years ago

Perhaps I simply know different people but I'm accustomed to Travelling Salesman, without explaining which version is intended, being given as the canonical NP-Hard (implicitly the optimal solution problem).

I normally see SAT given as the NP-complete example.

jcd748• 14 years ago

Yes, I've always seen SAT as the canonical example given. Though, when I did some reading into phase transitions in NP-complete problems, 3-SAT was usually the example given (and used).

jawns• 14 years ago

Just a small quibble:

If you're going to provide a TL;DR: it's probably better to stick it up top. Otherwise, the people who DR because it's TL will likely never find it.

dsr12• 14 years ago

TL;DR: For now TSP-OPTIMIZE is in NP-Hard but not (necessarily) in NP, so its not in NP-Complete. TSP-DECIDE is both in NP-Hard as well as NP, and is therefore NP-Complete.

mistercow• 14 years ago

This is a major point of confusion for people who are new to the concepts, and I'm glad someone has taken the time to explain it in depth. And it's not just TSP that it happens with. Almost every popular example of an NP-complete problem I've seen is actually an NP-hard optimization problem corresponding to an NP-complete decision problem.

The bad part is that if you don't know better, you'll think "Wait, how the hell would you verify a correct answer to this without searching for an even shorter path?" And you'll be right, but you'll feel like you're missing something. Giving examples of CS that make people feel like they fundamentally don't understand CS is probably a Bad Thing.

jlebar• 14 years ago

Ironically enough, I think this guy misunderstands TSP. I wrote this comment on his blog:

Let's define a new problem, TSP-cost-optimize, which returns the cost of the best Hamiltonian. It's NP-Complete, because it reduces to TSP-decide: Just do a binary search on the space of possible scores; even if that space is exponential, we're fine, since binary search runs in the log of the size of the space.

TSP-cost-optimize's certificate can be the steps of the binary search; that will be polynomial in size, since it's a polynomial number of TSP-decide certificates, and you can verify it in polynomial time.

Given TSP-cost-optimize, I can in fact provide a polynomial-time-verifiable certificate for the result of TSP-optimize: The result of TSP-cost-optimize on the same graph! So under your definition of NP, TSP-optimize is in fact in NP. (This also makes sense if you use the non-deterministic Turing machine definition of NP.)

What's not clear to me is whether given the result of TSP-cost-optimize, you can actually find a shortest path in polynomial time. But I expect you probably can.

peterderivaz• 14 years ago

I guess you could find the shortest path by iteratively deleting an edge and checking whether the best cost has got worse. If it does get worse, then put that edge back in and try deleting another edge. If it does not get worse, then permanently delete that edge.

I think this should terminate (after trying all edges once each) with an example of the shortest path (that consists of all remaining edges).

jb173• 14 years ago

The binary search is not a reduction, because you need to use TSP-decide a polynomial number of times (and for a reduction you reduce to one problem instance). What you show is that your problem TSP-cost-optimize is in FP^NP (the class of function problems solvable in polynomial time with the help of an NP oracle).

jb173• 14 years ago

I think the author actually adds to the confusion about TSP by being unnecessarily vague.

Both P and NP are classes of decision problems. That means they only contain problems that ask for a yes/no answer. So because TSP asks for a route, it is not just a decision problem, and can't be in NP. However, as the author mentions, NP does in fact contain the TSP-DECIDE problem in the way he defines it.

Moreover, the problem TSP of finding a lowest cost route can be solved by a polynomial time Turing Machine that can use a TSP-DECIDE oracle (i.e. call a constant time function that solves TSP-DECIDE). It is in fact complete for that class FP^NP ("= function problems solvable in polynomial time with an oracle for NP problems"), which means it is one of the hardest problems there. The class FP^NP contains all of FNP (NP for functional problems) and is believed to contain more problems -- so there is a good chance that TSP is not in FNP.

drivebyacct2• 14 years ago

I'm so thankful. After 4.5 years of CS, the only class that taught me anything was my advanced algorithm class and I learned enough to retain an immense amount of the information a year later. This is fun stuff and I hated it two years ago.

zxcvbn• 14 years ago

Thank you. I finally get what NP-complete is.

drobilla• 14 years ago

Well, this is true, but it's also hyperbole. It's not really everybody misunderstanding the problem so much as simply not caring. People are sloppy with the terms "NP-hard" and "NP-complete" and use them interchangeably because it's usually not important. Usually they mean "NP-hard", and that is the significant thing.

Yes, it's wrong to use "NP-complete", but plenty of people who thoroughly understand the problem do so. Welcome to Computer Science, sloppy is par for the course.

IsTom• 14 years ago

And usually they just need to say that it's intractable, unless they're speaking about some kind of solver or approximation classes.

sherjilozair• 14 years ago

Since optimization and decision problems are equivalent, what someone means by saying that an optimization problem is in np-complete is that its equivalent decision problem is in np-complete.

Traveling Salesman: the most misunderstood problem

Related Stories