Talk:Offline batch ELO rating system

Quite interesting. I fail to see how this makes "less assumptions on results than APS" however, since I don't believe APS makes any assumptions whatsoever, being a simple averaging of all pairings a robot is involved in. I don't see this as being less biased. Similarly valid certainly, but not less biased.
One thing I think is important to note is, I'm pretty sure the fact that you're rounding the result to a win/loss/draw probably explains the *vast* majority of the difference between this and the APS/ELO/Glicko-2 on the rumble server. I suspect that APS or the iterative roborumble ELO result wouldn't be that different if they performed the same rounding.
--Rednaxela 13:12, 12 August 2011 (UTC)

APS assumes a bots strength is proportional to the proportional difference in scores.

If a bot scores 2 and the opponent scores 1 (67%/33%), it is stronger than if it scored 3 and bot B scored 2 (60%/40%). Both bots increased the score by the same amount, but the APS system assumes one is stronger than the other.

If a bot scores 100% against MyFirstRobot and 40% against DrussGT, it is stronger than if it scored 60% against MyFirstRobot and 60% against DrussGT. APS assumes that +40% against MyFirstRobot is worth more than the +20% against DrussGT, leading to king-making.

Arpard Elo original rating system inferred strength on frequency of wins/losses alone, not proportional score difference. If a bot beats another, it is only better, not a little better or a lot better. --MN 16:32, 12 August 2011 (UTC)

If it scores 100% against MyFirstRobot and 40% against DrussGT, it is stronger than if it scored 60% against MyFirstRobot and 60% against DrussGT. In my opinion anyway, getting 100% against any robot is more time consuming and bug hunting and hard work then getting an okayish score against many robots (which many rambots, mirror bots and and random targeters can do). Are you telling me you think a random targeter is stronger then a highly tuned statistical targeting? — Chase-san 17:36, 12 August 2011 (UTC)

Of course, applying a boolean stronger/weaker to those two results is subjective. The only thing you can say objectively is that each has different strengths, like comparing the intelligence of chimps to dolphins to dogs. I often wonder how different our Robocode landscape would look if we hadn't settled on APS / "crush the weak" style measurements of robot strength. What if the thing that separated #1 from #2 is that #1 actually beat #2, and everyone else, over hundreds of rounds/battles? Adaptive, anti-adaptive, anti-anti-anti-adaptive movement and targeting would be a lot further along than they are now, I reckon. I'd like to put some focus on that front myself, sometime. I'll check these results when I am somewhere that I can unRAR. =) --Voidious 17:50, 12 August 2011 (UTC)

The whole point of the Arpard Elo origignal rating system inferring strength based on frequency of wins/losses alone, is because in Chess wins/losses are the only measure of performance that the rules of the game provide. There is no (widely accepted) way of measuring the range between "barely won" and "obliterated" in Chess.

It makes assumptions too. Original ELO assumes that your chances of winning against an opponent are proportional to rank difference, and that any deviations are due to random chance during measurement. Essentially it assumes all competitors are generalists, which is relatively true with humans in Chess, but is relatively untrue in Robocode with a variety of different strategies with all sorts of different performance characteristics against different subsets of enemies. It's by no means assumptionless.

Really though, any method you pick has subjective aspects. Also, the "score" number itself is not really objective either, and thus a win/loss (or anything else) based on it is not objective either anyway.

--Rednaxela 18:21, 12 August 2011 (UTC)