On-line engine blitz tourney October

[Robert Pope]

Anyway, I like this system and I believe that it is more stable than swiss, in the sense that the results are more reliable and less a matter of luck. Of course a bit randomness occurs in every game even between (almost) equal opponents, but still.

I liked it too, except I was surprised that up to 6 points were handed out - that seems like a lot. Does it make sense to use fewer points? 2 points would be enough to force close pairings in the first two rounds, but would allow more movement up and down based on actual game results.

[hgm]

This certainly can be tuned. Only 5 points were handed out, btw. My initial idea was that this should be about half the number of rounds. Because the system was supposed to not award any quasi-free points to top players during the tourney (by pairing them with unworthy opponents), the top players should score on average about R/2 real points (R being the number ouf rounds), and end up at R. A strong player that was so misplaced that it started with 0 points would score R certain wins, and thus still end up in the top.

But that is really an unrealistic worst case. Top players got 5 free points, but that doesn't mean the other got 0 points. A player 75% down the list still got 2.5, and was thus only separated by 2.5 points from the number 1. If the initial seeding was good it is very unrealistic that a player so far down the list would stand any chance at all against the top players. So the purpose is to keep them well separated, and 2.5 points is not overkill to achieve that.

Perhaps the number of free points should be directly related to the rating, though. Like (rating-lowestRating)/100, but saturating when raiting >= higestRating - 300. That has the advantage that when there is a really large Elo gap, such as between the searchig and non-searching engines, this would be reflected in the gap of free points. So that the non-searchig engines indeed keep playing each other, no matter how 'well' they do, until they run out of opponents, and only then start to encounter the lowest searching engines of the ranking. (Which, at that time, should be pretty well established.)

[JVMerlino]

Perhaps the number of free points should be directly related to the rating, though. Like (rating-lowestRating)/100, but saturating when raiting >= higestRating - 300. That has the advantage that when there is a really large Elo gap, such as between the searchig and non-searching engines, this would be reflected in the gap of free points. So that the non-searchig engines indeed keep playing each other, no matter how 'well' they do, until they run out of opponents, and only then start to encounter the lowest searching engines of the ranking. (Which, at that time, should be pretty well established.)

I think this is an excellent idea. However, since in a typical tournament the difference between the top and bottom engines is at least 1000 points, it seems that ((rating-lowestRating)/100)*0.5 should be more than sufficient. Or, if you wanted larger initial groupings with the same score, you could use (rating-lowestRating)/200.

Just my thoughts.

jm

[hgm]

Well, in this tourney the lowest rating was ~800, and the highest ~2200. But the number 3 rating was ~2075. If we would keep a margin of 400 pts to have a chance to make top-3, that should make the number of free points saturate at a rating of 1675. So that leaves a rating range of 875 for the linear ramp. For 9 rounds we would like that to be 4.5 points. So indeed the factor should be a little smaller. I guess the formula would be

free points = rounds/2*(rating-lowestRating)/(nr3rating-400-lowestRating)

saturating at rounds/2. If nr3rating - lowestrating < 400 it makes no sense to do McMahon; even in the first round normal Swiss would pair the top players with those from the center of the table, which presumably then would be ~200 Elo below it. That would make a meaningful chance for an upset. And Swiss with accelerated pairing would make even closer pairings.

[JVMerlino]

Well, in this tourney the lowest rating was ~800, and the highest ~2200. But the number 3 rating was ~2075. If we would keep a margin of 400 pts to have a chance to make top-3, that should make the number of free points saturate at a rating of 1675. So that leaves a rating range of 875 for the linear ramp. For 9 rounds we would like that to be 4.5 points. So indeed the factor should be a little smaller. I guess the formula would be

free points = rounds/2*(rating-lowestRating)/(nr3rating-400-lowestRating)

saturating at rounds/2. If nr3rating - lowestrating < 400 it makes no sense to do McMahon; even in the first round normal Swiss would pair the top players with those from the center of the table, which presumably then would be ~200 Elo below it. That would make a meaningful chance for an upset. And Swiss with accelerated pairing would make even closer pairings.

Yep. I admit I was discounting the possibility of continuing to include random movers in the tournament. They can be fun to watch, and certainly shouldn't be excluded from the tournament if you (or the majority) want to keep them. But for the purposes of trying to solve this problem they do adjust the curve a fair bit.