Now that there's a large pool of players, it might not be a bad idea to tweak the settings some.

Right now, new players start out with 1300 (350). This implies that the average player has a rating of 1300. I suspect this number is a little high, as most players with low uncertainty and ratings around 1300 are pretty good players. It's probably closer to 1200, so let's split the difference and say that 1250 would be a good starting point.

That uncertainty is definitely high too. It's saying that 95% of players are between 600 and 2000. That's much, much wider an interval than necessary. The starting uncertainty should probably be closer to 200 or so. So with that previous lower average and this lower uncertainty, that interval tightens to 95% of players being between 850 and 1650.

Another way of looking at those numbers is that if there are 40 people online, you'd expect to see 38 of them in that range, and maybe one above and one below it.

Of course since the ratings are all sitting in a database somewhere, these numbers would be pretty easy to check. Just take the average and standard deviation of all ratings (excluding say ones over 250 uncertainty), and that'd be a decent starting point.

So what would be the result of making those adjustments? Those wouldn't affect existing players' ratings, but it would change players just starting out. I'd expect their ratings to be "truer" to start. This would make similar matching much more accurate. They'd be ranked lower and their initial scores wouldn't swing so much on their first match.

Speaking of swings, I've got the feeling that player ratings don't change enough once the player hits 35. This can be both a good thing and a bad thing. I suspect upping that minimum uncertainty to 50 would be an improvement. This might help with the alleged actions of certain players tanking matches, locking in a low rating, then raking in ladder matches.

And speaking of similarity, I'm a little disappointed that it doesn't take the uncertainty into account. Two players 100 rating points apart with 35 uncertainty each, are much, much less likely to be of similar skill levels as two players 100 rating points apart with 100 uncertainty each. With the former, there's only a slight overlap in rating ranges but with the later, the majority of the ranges are overlapping.

Just some random thoughts that I've been mulling over. . .