PWC121 - The Travelling Salesman


On with TASK #2 from the Perl Weekly Challenge #121. Enjoy!

The challenge

You are given a NxN matrix containing the distances between N cities.

Write a script to find a round trip of minimum length visiting all N cities exactly once and returning to the start.


Matrix: [0, 5, 2, 7]
        [5, 0, 5, 3]
        [3, 1, 0, 6]
        [4, 5, 4, 0]

        length = 10
        tour = (0 2 1 3 0)

BONUS 1: For a given number N, create a random NxN distance matrix and find a solution for this matrix.

BONUS 2: Find a solution for a random matrix of size 15x15 or 20x20

The questions

As it often happens, some questions are more… assumptions.

As we’re talking about distances between cities, the assumption is that they are greater than zero.

The NxN matrix needs not be symmetric, as also the example shows. This is consistent with one-way routes.

It’s a bit more difficult to understand whether the NxN matrix respects some triangular constraint. I mean, we are assuming that each city is directly connected to any other city, and also that this connection might be worse than passing through other cities. This is not an issue with the main problem, more for the bonus of generating a random arrangement.

Last, regarding the second bonus… is it 15x15 or 20x20?!?

The solution

The first solution, that is still good for low values of $N$ (say up to 10 or so) is to just brute force the problem. With N cities, we fix one of them (let’s call it $0$) and generate all possible permutations of the other ones (let’s call them with integers from $1$ up to $N - 1$). This gives us all possible paths by starting from $0$, going through the permutation, and going back to $0$.

Here’s the example in Raku:

sub tsp-bf ($dist-from-to) {
   my $n = $dist-from-to.elems;
   my ($best-distance, $best-path);
   for permutations(1 .. $n - 1) -> $perm {
      my $from = 0;
      my $sum = 0;
      for |$perm, 0 -> $to {
         $sum += $dist-from-to[$from][$to];
         $from = $to;
      ($best-distance, $best-path) = ($sum, $perm)
         if ! defined($best-distance) || $sum < $best-distance;
   return ($best-distance, (0, |$best-path, 0));

The $dist-from-to is assumed to be a bi-dimensional array whose first level contains each row of the input matrix. The permutations built-in comes handy here, allowing us to concentrate on the problem.

This is easily translated in Perl, leveraging our old friend Iterator-based implementation of Permutations:

sub tsp_bf ($dist_from_to) {
   my $n = $dist_from_to->@*;
   my ($best_distance, $best_path);
   my $pit = permutations_iterator(items => [1 .. $n - 1]);
   while (my @perm = $pit->()) {
      my ($from, $sum) = (0, 0);
      for my $to (@perm, 0) {
         $sum += $dist_from_to->[$from][$to];
         $from = $to;
      ($best_distance, $best_path) = ($sum, [@perm])
        if !defined($best_distance) || $sum < $best_distance;
   } ## end while (my @perm = $pit->(...))
   return ($best_distance, [0, $best_path->@*, 0]);
} ## end sub tsp_bf ($dist_from_to)

To address the first bonus, we just generate a random matrix with all positive values, except on the diagonal (the distance of a city from itself is 0). For simplicity we’re restricting to integers here, but it’s not a constraint.

sub generate_randoms ($n) {
   my @retval;
   for my $i (0 .. $n - 1) {
      my @row = map { 1 + int(rand 13) } 1 .. $n;
      $row[$i] = 0;
      push @retval, \@row;
   return \@retval;

The underlying model is a bit… crude, maybe it would be better to just put points on a rectangle and calculate their mutual distances. Anyway, as an input for the problem it should be fine.

To address the other bonus, a different algorithm is needed. I took a look at the Wikipedia page on the Travelling salesman problem and it does not give too much… except that branch and bound solutions should be better.

Anyway, in this case I settled for the Held–Karp algorithm because:

  • it has a much better complexity than brute-force
  • we’re limited to 20x20 anyway, so we don’t need anything too aggresive.

The page on the algorithm is a bit confusing, because:

  • it has a good textual explanation of the algorithm;
  • it has an example that is more or less in line with the explanation;
  • it has some pseudocode that I can’t relate 1-to-1 with the two above.

For this reason, I changed it to be in line with the explanation and the example, resulting in this:

function algorithm TSP (G, n) is
    for e := 2 to n do
        g({}, e) := d(1, e)
    end for

    for s := 2 to n−1 do
        for all S ⊆ {2, . . . , n}, |S| = s do
            for all e ∈ S do
               g(S\{e}, e) := min_{m ∈ S\{e}} [g(S\{m, e}, m) + d(m, e)]
               # & keep best predecessor (yielding the minimum)
            end for
        end for
    end for

    opt := min_{k = 2 to n} [g({2, 3, ..., n}\{k}, k) + d(k, 1)]
    return opt
end function

I will not put the implementations in Raku and Perl here, because they’re a bit long and are more or less an implementation of the algorithm above (with the small addition that allows to get the optimal path out of the calculation). They are available in the Perl Weekly Challenge repository, here and here.

The Perl implementation is a bit lower level, using only arrays, while the Raku implementation uses sets etc. I’m not sure, though, this is an advantage, at least computationally-wise, because the Perl solution is way faster.

Both solutions have an exponential behaviour where adding one city more or less means a factor of about $2.4$ on the time required to calculate it all. The 20x20 example case is solved by Perl in about 2 and a half minutes, while an extrapolation for Raku would require a handful of hours. I think that the Raku implementation can be Raku-ized much more, though, yielding an easily parallelizable algorithm that might take advantage of multiple cores to speed up the execution.

As hinted by the Wikipedia page, the memory requirements can be a bit taxing as the number of cities grows. I think it should be possible to remove lower level values of hash %g as we move up the ladder, although the %p hash tracking the predecessor still has to be kept until the end of the computation.

All in all it’s been… very interesting!

Comments? Octodon, , GitHub, Reddit, or drop me a line!