|Thread title||Replies||Last modified|
|Why x2?||12||02:44, 16 December 2011|
I think I have asked this before somewhere (long time ago). But anyway, why is the PL score the same as the number of pairings won times two? A percentage would make much more sense to me. Or just skipping the multiply with two part.
May be because of tie (exactly 50% APS) - in this case both bots get by 1 point.
I'm not sure, but i think, that it is you invented PL and you was inspired by English football Premier League. So you must understanding better rules of PL, than i, since i'm not football fan:)
Lolz. I remember that I was very supporting of the idea to measure how "undefeatable" a bot was. But I can't recall being the one to use the PL analogy. I do remember fighting fiercely for focusing on plain Average Percent Scores and not all that ELO hoochie coochie, but lost that argument back then. It feels like a glitch in the matrix when you remind me of the old PL discussions, because one of the first questions I had when "returning" to Robocode this time was "What is APS?". =)
Funny, I feel like a bunch of stuff you used to advocate for came to fruition in your absence. APS is one. Another is that we now do a pretty good job separating out gun terminology like Visit Count Stats, GuessFactors, Pattern Matching, and Play It Forward, instead of using confusing names for multi-part systems, like “GuessFactor Targeting” == VCS + Segmentation + GFs or “TronsGun” == KNN + PIF.
PS - Can you see more than 3 replies now, or still having that problem?
I think it's very likely Darkcanuck just copied what was done on the old RoboRumble server, and I think the soccer (er, football) reference is accurate there. I guess I would also prefer to just see wins instead of (wins * 2). Not sure I've ever seen a tie anyway...
Here's the page from the original wiki: http://old.robowiki.net/robowiki?RoboRumble/PremierLeague The description of how it works is unsigned, but from the chat I always assumed it was suggested by PEZ.
The new rumble server is a reverse-engineering effort, largely supported by what I could glean from the wiki. For the PL score, I cheated a little bit: since the odds of having a perfect 50/50 APS split is vanishingly small, I just multiplied the number of won pairings by two and skipped the calculation of ties. I'm not sure how the old rumble server's code handled it.
Changing the display to # won pairings / total pairs would be trivial. But I don't really want to add more columns, so would retiring the "classic" PL score column be ok? I also have another change in the works to add APW: average percentage win rate over all pairings, where the win rate for each pairing is # wins / # battles for that pair. That's based on the discussion from a while back about alternative scoring systems.
You certainly have my vote to change it (and to add APW in the future!).
Wow, yes, that was indeed my suggestion to begin with. I remember it now. As I read the old discussion it seems that there is agreement around making it really transparent. When I originally suggested it I didn't realize just how seldom bots truly tie. I also find this
- Why, if you don't win against a bot you could as well have 0 points. 3 points for a win and zero for other outcomes might be a bit weird though, so make it 1 point for the win instead. -- PEZ
Seems like what I find weird today, I found weird back then too. =)
I suggest making that change and also to remove the old PL reference. Yes, it hurts a bit, but it'll make for one less legacy thing for newcomers to wonder about.
I would also prefer counting 1 point for a win and 1/2 for a tie in PL. Other formats for showing the score could be:
- - Leaving wins, ties and losses separated, like "840-0-1".
- - Or using Copeland score, which is (wins - losses), 840 wins and 1 loss would be "839". The difference is, missing pairings would be more easily counted as ties instead of losses.
As for APW, after reading a lot about voting/ranking systems, I realized PL is superior to APW in the fairness subject. If there would be a new ranking system, it would vote for a Schulze based one.