| 1 |
Yes,
exactly.
It
is
based
on
the
assumption
that
probability
predictions
should
be
more
distinct
with
higher
player
numbers
due
to
the
law
of
large
numbers.
If
no
distinctivity
factor
is
used
nothing
changes
if
you
replace
every
player
by
2
clones
of
himself.
|
1 |
Yes,
exactly.
It
is
based
on
the
assumption
that
probability
predictions
should
be
more
distinct
with
higher
player
numbers
due
to
the
law
of
large
numbers.
If
no
distinctivity
factor
is
used
probabilities
don't
change
if
you
replace
every
player
by
2
clones
of
himself.
|
| 2 |
\n
|
2 |
\n
|
| 3 |
Unfortunately I don't have a really good mathematical derivation why exactly sqrt. I just tried different functions on game data (back then without teamstrength) and this scored best. Higher or lower distinctivity was not as good. A reason could be that standard deviation is proportional to sqrt(number of samples)*mean.
|
3 |
Unfortunately I don't have a really good mathematical derivation why exactly sqrt. I just tried different functions on game data (back then without teamstrength) and this scored best. Higher or lower distinctivity was not as good. A reason could be that standard deviation is proportional to sqrt(number of samples)*mean.
|
| 4 |
\n
|
4 |
\n
|
| 5 |
On the other hand it is against mathematical simplicity. I really don't know whether it is good to use it together with teamstrength. I did reproduce all players' elo progresses in my tests, but what I didn't do was:
|
5 |
On the other hand it is against mathematical simplicity. I really don't know whether it is good to use it together with teamstrength. I did reproduce all players' elo progresses in my tests, but what I didn't do was:
|
| 6 |
- Considering ZK's weighting system
|
6 |
- Considering ZK's weighting system
|
| 7 |
- Testing it for at least 2000 games (only ~300 with and ~1200 without elo progress)
|
7 |
- Testing it for at least 2000 games (only ~300 with and ~1200 without elo progress)
|
| 8 |
- Considering teamstrength
|
8 |
- Considering teamstrength
|
math bork 2nd try