TL;DR

Using our constraint programming solver to resolve a $S(2, 4, 28)$ design.

In Steiner design S(2, 4, 28) we saw that it’s possible to set up a tournament with the following characteristics:

• 28 participants overall
• 4 players at the table for each match
• 63 matches overall
• each participants competes in exactly 9 matches
• every pair of participants face each other exactly once
• matches can be arranged into rounds, so that all players can play in a round at the same time and the wait time is limited.

In mathematical terms, this means that we have a Steiner design S(2, 4, 28) that is also resolvable (this is the division into rounds). The post also includes the solution to this problem, so that you can start playing right away. Go play!

Then, in Sorting an incidence matrix we set the first step to actually implementing the division into rounds. The program was needed because there are a lot of $S(2, 4, 28)$ in the linked file, but only 6 of them seem to have proved to be resolvable. Additionally, the linked file did not have this division into rounds… so we had to find a solution anyway.

Here, we start looking into the code for finding a rounds arrangement, leveraging on the function described in More constraint programming.

# Preparing the data structure

The incidence matrix has a lot of merits, but for our quest is not the most straightforward representation - at least, it’s not with respect to the implementation that we will see. Maybe there’s a clever way to use it directly!

So, the first thing will be to arrange matches in a list of hashes, each containing two keys:

• participants will hold a list of the players in the match
• round will be either a simple scalar indicating the specific round of the match, or an anonymous array containing all possible candidate rounds to which it can be assigned.

This is done by the following function:

 1 sub incidence_to_matches ($incidence) { 2 my$n_rounds = grep {$_}$incidence->->@*;
3    my @matches   = map { {participants => [], round => [1 .. $n_rounds]} } 4 1 ..$incidence->->@*;
5    my $player_id = 0; 6 for my$player_data ($incidence->@*) { 7 ++$player_id;
8       for my $match_idx (0 ..$#$player_data) { 9 next unless$player_data->[$match_idx]; 10 push$matches[$match_idx]{participants}->@*,$player_id;
11       }
12    } ## end for my $player_data ($incidence...)
13    return ($n_rounds, \@matches); 14 }  The number of rounds is not known beforehand, but it’s easily counted by the number of matches of player number 1 (line 2). The loop in lines 6 to 12 make sure to fill in the participants anonymous array with the right player identifiers. We decided to stick with starting from 1 for identifiers, just to avoid that anyone feels a zero. # Resolution main function The main function where we resolve a design is the following; it assembles some of the functions we saw so far, and calls more that will be described in due time:  1 sub resolve ($incidence_text) {
2    my $incidence = lexi_parse($incidence_text);
3    my ($n_rounds,$matches) = incidence_to_matches($incidence); 4 my$n_matches            = $matches->@*; 5 6 # The incidence matrix is lexicographically ordered, which means that 7 # first player's matches are also the first matches. We can assign them 8 # to its own round and fix this symmetry for good. 9 my @pre_assigned_match_idx = map { 10 ($matches->[$_]{round} =$_ + 1) - 1;
11    } 0 .. $n_rounds - 1; 12 13 # We are now ready to start searching 14 my$state = solve_by_constraints(
15       start => {
16          unassigned            => ($n_matches -$n_rounds),
17          matches               => $matches, 18 last_chosen_match_idx => [@pre_assigned_match_idx], 19 }, 20 constraints => [ \&prune ], 21 is_done => sub ($s) { $s->{unassigned} == 0 }, 22 search_factory => \&search_factory, 23 ); 24 25 my %matches_for; 26 push$matches_for{$_->{round}}->@*,$_ for $state->{matches}->@*; 27 return \%matches_for; 28 } ## end sub resolve ($incidence_text)


Lines 2 to 4 do the initialization as we saw, and compute how many matches we should expect.

Lines 6 to 11 do a pre-computing that is a kind of preliminar application of a constraint. Thanks to the lexicographical ordering that we have after line 2, we are sure that the first $n_rounds matches are all played by participant 1. It makes sense to initialize them each with a different round identifier, because they surely all belong to different rounds and any possible resolution would be a permutation of this choice. For this reason, we set them to fixed values, getting rid of the other candidates. The identifiers of these matches are saved in array @pre_assigned_match_idx, which we will later use to optimize our pruning activities. Lines 14 to 23 contain the actual call to the solver. The start key points to our data structure, with the following keys: • matches is the match arrangement we saw in the previous section; • unassigned keeps track of how many matches we still have to assign to a specific round. It’s initialized to the number of total matches, minus the ones for player 1 that we pre-assigned; • last_chosen_match_idx allows the search function to communicate to the prune function which match was assigned a tentative value last, in order to optimize the pruning function. It is initialized with the list of identifiers that we set in the pre-assignment, so that the pruning will start from them. Function is_done for the solver is straightforward: if the unassigned count ever drops to 0 we are done (line 21). The constraints are actually condensed in a single function prune, and the search function (or better, a factory for iterating through the search alternatives) is encapsulated in its own search_factory function. We will see them in due time. Last, if the returned $state contains a solution, it is re-arranged (lines 25 to 27) in a way that makes it easy for the caller to figure out which matches belong to which round. Not surprisingly, this is the format that torneo would be happy with.

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