Pool play into a bracket without the rematch: crossover seeds, searched exactly

Pool play now seeds the knockout crossover-style. Seeding by record alone could put poolmates together in round one in 29 to 86% of random draws.

A weekend tournament usually runs in two parts: pools first, so every team gets a few games, then a knockout bracket for the teams that advance. The league schedule generator has built the pool round robins for a while. It also listed the play-off seeds: every pool winner first, best record first, then every runner-up, and so on.

That list had a problem. Put it into a standard bracket (seed 1 plays the last seed, 2 plays the second-last) and two teams from the same pool often meet in the first knockout game. Take two pools of three with two advancing. If Pool B's winner has the best record and Pool B's runner-up has the worst, the record order is B1, A1, A2, B2. That gives B1 v B2 in one semifinal, a rematch of a game played that morning. Tournament organisers avoid it with a crossover: A1 v B2 and B1 v A2. The tool now does the same thing.

What it does

Winners still come before runners-up, runners-up before thirds, and so on. Inside each place, the new poolPlayoffSeeds picks the order so poolmates meet as late as the bracket allows. It counts the number of poolmate pairs that could meet in round 1 and makes that as small as possible, then does the same for round 2, and so on. Among the orders that tie, it keeps the one closest to the record order, and moving a top seed costs more than moving a low one.

The pool card now shows which pool each seed came from. When the crossover changed the order it adds a line saying so, with the possible poolmate meetings before and after. It also draws the bracket itself, as single or double elimination with an optional third-place game, using the play-off settings the page already had. Typing play-off results fills the bracket in.

The first version found a worse answer

The first version swapped two teams within a place whenever the swap helped, and stopped when no swap did. I checked it against a brute force that tries every possible order, on random pool tables. Over 159 draws it lost twice. With 3 pools of 4 advancing and with 4 pools of 4, it got stuck on an order with one poolmate pair that could meet in round 2, when an order with none existed. No single swap leads from one to the other.

So now an exact search runs after the swaps. It places seeds one at a time and drops a partial order as soon as its counts are already worse than the best full order found. The counts only go up as more seeds are placed, so dropping it early never throws away a better answer. The search stops at 300,000 steps and keeps the swap answer if it gets there. It matched the brute force on all 530 random draws I ran, and it runs in the tests on five pool shapes. It takes under a millisecond for most fields and about a fifth of a second for 8 pools of 4.

How often record order went wrong

eval/crossover-seeds.mjs draws 500 random sets of pool tables for each common shape. Results from the file:

Pools x advancingRecord order: poolmates could meet in round 1Crossover
2 x 2245 of 5000
2 x 4383 of 5000
3 x 2161 of 5000
4 x 2287 of 5000
4 x 4428 of 5000
6 x 2145 of 5000
8 x 2314 of 5000
8 x 4430 of 5000

With three or more pools the crossover also keeps poolmates apart through round 2 in every draw. The record order allowed a meeting in round 1 or 2 in up to 499 of 500 draws. Two pools can't manage that: each half of the bracket has to hold at least two teams from one pool, so they can meet by round 2. The search proves that and settles for it.

What it doesn't do

Places stay in order: a pool winner is always seeded above any runner-up, even when the pools were uneven and one pool's runner-up had the better season. That is the usual rule, and an organiser who wants something else can still type the seeds into the main play-off box by hand.

All writing · More in Tools and projects · More on League Schedule Generator · All tools and research