- Created Sunday, January 18th 2015 @ 07:37:46
I've run the same version of my bot for the last 120 games and during that time it's been as low as 11th-ish place and as high as 1st, floating up and down in a 150 point elo range.
I think that amount of fluctuation is maybe unnecessarily high.
Maybe the elo Delta value should be lower in general or maybe it should decrease to an even lower value after the same version of a bot has been up for a while?
- Updated Sunday, January 18th 2015 @ 09:47:31
Considering ELO, I gave it some thought and I agree that K-factor should be weighted. Unlike with human players who learn over time, these bots are static (they unfortunately do not learn over time). However, you don't want bots to be stuck, so you don't want to have bots at high ELO to have a lower K-factor which is usual for chess human play.
In fact, you also want a possibly much higher K-factor for "up-and-coming" bots, and a lower K-factor when both bots are on the same version for a long time. You should calculate the K-factor for ELO based on both bots. Then again, what would we do with 2 "new" versions of bots?
I think it would also be a really good idea not to have this queue be created at a certain time. E.g. when you have a new great version, you will most likely still fight a lot with lower ranked (since you are slowly climbing). You might be placed 10th and still match up against #30. The delay is unnecessary. Instead of the current queueing system, I'd propose:
After each match:
Sample a player from the 120 bots (probably indeed deactivate some, but that's a different story)
Sample an enemy within 10 ranks
You'll also avoid the new players from wondering why it takes so long for them to be matched up.
Sure, there will still be relatively long waits with unlucky rolls, but at least the matchups will be optimal considering the current estimate of ELO.
- Updated Sunday, January 18th 2015 @ 20:04:33
Well, your bot went for an impressive win streak of 14 games in a row and this against bots with quite some higher ELO rating if you started from a lower rank. It's not like you winning 3 games in a row and the rating algorthm assumed you were an unstoppable monster. At present rank 1 has 2046 and rank 10 has 1910 so I don't find it that bad if rank 10 gets quite some points for winning against a bot against which he has an 11% chance. At present all bots have seriouls flaws without there being a bot totally dominating the board.
Edit: I believe you can also see that the rating works pretty well from my "Advanced Java Starterbot". I believe that at least the bots Perfection and Resistence (amongst other bots...) are identical clones of this bot and they are both at 1500 / 1507 rating.