[Freeciv] recounted rankings

[Horn Gábor <Horn.Gabor@xxxxxxxxxxx>]

Hi!

I've finished recounting of the rankings and draw a graph due to that.
Count went from #42936  to #233550 (older games doesn't contain info
about that a player is ai or not, and aren't compatible neither the
current nor the new ranking scripts), it means about 10000 games (hey
how much time we have :). Here's the graph:

http://aktiv.co.hu/~hirisov/civserver/compare2.jpg

Well what it shows? :)

It's the distribution of the points, x axis means ranking points, y
means the number of the players w/ the current points.

I also listed top and last 30 for each ranking. Top might be obvious,
but i also listed bottom for demonstrate role of tv's. With the old
algorithm they have some influence on the graph, which is not good i
think. In every game there are tv's, the players will drain some points
from them. I tried to eliminate it in the new (green). Here are some
details about them:

1, in current freeciv database:

tv 431.581637782297 
tv2 512.650348684783
tv1 610.461379573265
tv3 686.457189114323
tv0 846.258435321282

2, in the counted database w/ current algorithm:

tv 410.782964417564 
tv2  495.59388043789
tv1 589.065300831826
tv3 657.521630050272
tv0 839.145357157532

3, and w/ the new method:

tv 738.253596866002 
tv2 836.648764670111 
tv1  862.34078200143 
tv0 957.342733793997 
tv3 957.649184236716 

The closer the tv's are to the average (1000) the better we eliminated
their role. It got better, but i'll try to remove their effect
completely (well except the cases when a player plays under nick tv).

Now some word about the rankings:

1, counted and original current script result differs because of the
fact i couldn't measure games before #42936. It can be worked around i
guess, but not sure if worths the work. The few huge differences came
from that, eg (_._) had 67 percent of his games in the early times,
before #42936 (well jb's other nicks are similar in the calculated and
original ranking :).

2, DDD has better result w/ the new system. I think it shows now the
role of number of played games is less important, which is closer to the
ELO system. It's mainly because of lowering the effect tv's did (that
was a benefit for those who played more games).

3, the middle area is wider a bit, it's good again we have more precise
results from that area this way. 

4, seems the effect on gangbangs got smaller, eg i think this causes the
better result for Pille and Chaos (Warlock), they got ganged a lot.

5, well to tell the truth i don't know what couses the gap above 1000
pts. It got a bit smaller, but still there. Anybody has idea? Maybe the
multi nicks or other strange thing? I dunno yet, but i don't think that
is good there :) 

I used the following constants:

C1=0.4
C2=0.05
N=20
AI_DEFAULT_WEIGHT=0.5 (a constant, i use it when in an team or alliance
sy got ai'ed, so the system doesn't use his own point)
AI_DEFAULT_STRENGTH=(1-C2)/2 (same as above just for the used strength)

If u think any should be adjusted let me now, now it takes about 1 hour
to regenerate this graph from the 10000 games on my system, and we can
check the effect.

I'm extending it a bit now (so it will produce more detail information,
like the development of a player and such), but these changes doesn't
effect the ranking part. 

What u think ladies/gents? :)

bye, hirisov