from zero test

Hi!
First a good new, i succesfully run the new script from a zero stage
(totally empty database), and it handled 98 percent of the examined logs
correctly, only failed at incorrect log. And unlike the current script,
it recognized and ignored the log he couldn't interpret.

I've tested the new script w/ more games, especially teamgames. Here is
the method:

1, select all the players from the database who has 0 wins and more than
3 draws (so they're real team nicks). this resulted in 16 players.
2, get all the games from civserver that they played in. It resulted in
145 games, with a total of 314 players.
3, run both the current and the new ranking scripts w/ an EMPTY ranking
database for these games
4, make a graph from the distribution of the rankings in "ideal" case it
should shape a form of a bell (gaussian distribution).

See the attached files current.jpg and new.jpg. I also attached
current_full.jpg, which is the distribution for the real civserver.

Notes:

1, In the fromzero tests we have very few data, most player only has 1-2
games in the sample, and remember, we weight the info about the players
due to the nubmer of games they played. I'd need much more sample to
produce ranking that is closer to the reality. I'll work on it. 

2, in the current_full.jpg we much more data, it should be more
representive of the ranking -  in fact of the algorithm.

3, comparing current.jpg and new.jpg u can see the current method tends
to group results around the zero point (1000, where a new player
starts). it's not good, it means imho it can gain less info from the
system.

4, it even can be seen better in the current_full.jpg. i produced an
1600 height jpg, and it even doesnt fit, so very much players are around
the null point (which means either they're totally average, or we dunno
anything about them...mostly the 2nd).

5, note in the new script i currently doesn't examine chain of
alliances. Even in a team game it's possible if two player are in  the
same team and they never exchange techs (especially in 3+ member teams).
Now i don't consider them a team. I can alter it to recognize an ally's
ally as my ally. I think that produces more real result. What u think?

To tell the truth i don't know exactly why fromzero graphs differs that
much from the bell shape. First i need much more gamelog data, then i
can examine it. My guess is there are very much players/nicks who plays
only a few games (1-5) and we have very few data abou them and this
hurts the 'simmetricity'. In the current way there are many other
factors that made the result incorrect:

1, it doesn't recognize (some) incorrect gamelogs, and parse them
totally wrong.

2, it doesn't handle alliances at all.

3, usage of  multiple nicks.

4, all other type of cheats which affects the result (like a sudden
endy, eg when fredo virtually killed an ally of me+str1der, which has a
very little chance, and he couldn't do that in that game too w/o cheat).
hope u find it interesting.

bye, hirisov




-- Binary/unsupported file stripped by Ecartis --
-- Type: image/jpeg
-- File: new.jpg


-- Binary/unsupported file stripped by Ecartis --
-- Type: image/jpeg
-- File: current.jpg


-- Binary/unsupported file stripped by Ecartis --
-- Type: image/jpeg
-- File: current_full.jpg