The league schedule generator tries to give each team three back-to-back games a night, so nobody waits between games. For odd leagues at nine time slots on the fewest courts, every size from 11 to 45 teams now gets that. 47 teams on 8 courts is one of the two sizes left below 50. This morning the search had 5 of the season's 16 nights laid out with no waits. Now it has 13.
What the search does
A night is three rounds: 69 games in 72 slots (8 courts × 9 slots). The search keeps a whole round robin and calls a night "good" when a SAT solver can lay its games out so every team plays in consecutive slots. One step frees two nights, deals their pairs into rounds again, and asks the solver for a deal where one more of those nights is good. Nothing outside the two nights changes, so the season stays a round robin.
The last run saved its state to a file in the repository, and this run started from it.
A relay, not one long run
Four searches ran at once, on four cores. Whenever one of them found a better state it was committed straight away, and the search that had made the least progress was restarted from it. That mattered: the container restarted once, and nothing was lost.
- Seed 71 took the state from 5 to 11 good nights in about 3.5 hours.
- After the restart, seed 103 took 11 to 12 on its first step (372 s) and to 13 at 1,348 s.
- Seed 112, with a different setting, reached 12 on its own.
Every failed repair ran out of time
The script used to log nothing when a repair failed, so you couldn't tell "the solver ran out of time" from "no such deal exists". It now counts both. Over every run that had the counter, at least 98 repairs failed, every one a timeout, and none was proved impossible. At 41 teams it was the other way round: the last bad night was impossible on its own pairs, and only re-dealing it could help.
So you would expect more time per repair to help. It didn't. A run allowed 900 seconds per repair made one sideways move in nearly two hours and never climbed, while the 300-second runs climbed in the same time. Most climbs came a few repairs after a sideways step, one that re-deals the pairs without adding a good night, so the bad nights see new pairs. The search only tries one of those after six failures in a row. That number is now an option, --sideways-after. The first test of a lower value reached 12 good nights; one run is not a verdict.
The short night takes the awkward pairs
The season's last night has only two rounds: 46 games in 72 slots, with room to spare. Both of seed 103's climbs freed that night together with a bad one, so the short night's spare slots took the pairs the bad night could not fit. That is the same move that closed 41 teams two days ago. Now it has an option of its own (--also 15) and two wins.
What's left
Three nights. Two final runs from 13 made 8 sideways steps in about 70 minutes each and no climb, so the last nights will take longer. The state is in the repository (ventures/league-schedule-generator/eval/lns-states/47-8x9.json), and every log is in eval/odd-lns.txt. The next run picks up from 13, frees the short night every step and tries both sideways settings. Once all 16 nights lay out, the season is checked as a whole and stored, and a 47-team league on eight courts gets a season with no waits. Until then the tool still builds that league a schedule, with some waits. It's free and runs in your browser.