| 1 |
If you don't transform the whole elo scale, all players must start at 1500, because 1500 must always be the elo average.[spoiler]I would transform the whole scale from 1500 +- 400/ln(10) to 0 +- 1.[/spoiler]
|
1 |
If you don't transform the whole elo scale, all players must start at 1500, because 1500 must always be the elo average.[spoiler]I would transform the whole scale from 1500 +- 400/ln(10) to 0 +- 1.[/spoiler]
|
| 2 |
\n
|
2 |
\n
|
| 3 |
After I took a look at my mathematics from 2014, I realized that I only ran modified current systems and modfied (scalar) probability average systems, but neither any probability vector nor probability matrix systems yet. For the easy system types tested so far I can say that the modification of the final probability matters more than the fundamentals of the system.
|
3 |
After I took a look at my mathematics from 2014, I realized that I only ran modified current systems and modfied (scalar) probability average systems, but neither any probability vector nor probability matrix systems yet. For the easy system types tested so far I can say that the modification of the final probability matters more than the fundamentals of the system.
|
| 4 |
\n
|
4 |
\n
|
| 5 |
An
easy
quick
fix
(
but
not
the
best
system
I
have
tested
so
far)
that
doesn't
change
team
distribution,
but
only
probability
prediction
and
thus
possibly
results
in
more
adequate
elo
wins
and
losses,
would
be
modifying
the
!predicted
probability
p
to
(
p^D)
/(
p^D+(
1-p)
^D)
,
where
D
=
f(
(
#T_k
+
#T_l)
/2)
,
where
#T_k
is
the
number
of
players
in
team
number
k
and
f
a
function,
e.
g.
f=sqrt.
In
my
large
system
of
systems
I
call
this
system
"sqrt
sm"
(
f
sm
generally
for
Size
dependently
Modified
current
system
with
mod
function
f)
.
This
is
only
a
quick
fix
that
should
account
for
the
fact
that
probabilities
should
be
more
distinct,
when
there
are
more
players.
It
doesn't
change
anything
fundamental.
This
increased
"prediction
quality"
by
about
25%
in
all
considered
scoring
systems
in
those
1195
games
(
only
even
teams,
no
1v1,
no
ffa,
no
coop)
.
It
can
be
that
this
will
not
hold
in
a
larger
database.
If
it
should
be
implemented,
elo
change
should
account
for
it.
|
5 |
An
easy
quick
fix
(
but
not
the
best
system
I
have
tested
so
far)
that
doesn't
change
team
distribution,
but
only
probability
prediction
and
thus
possibly
results
in
more
adequate
elo
wins
and
losses,
would
be
modifying
the
!predicted
probability
p
to
(
p^D)
/(
p^D+(
1-p)
^D)
,
where
D
=
f(
(
#T_k
+
#T_l)
/2)
,
where
#T_k
is
the
number
of
players
in
team
number
k
and
f
a
function,
e.
g.
f=sqrt.
[spoiler]In
my
large
system
of
systems
I
call
this
system
"sqrt
sm"
(
f
sm
or
sm
f
generally
for
Size
dependently
Modified
current
system
with
mod
function
f)
.
[/spoiler]
This
is
only
a
quick
fix
that
should
account
for
the
fact
that
probabilities
should
be
more
distinct,
when
there
are
more
players.
It
doesn't
change
anything
fundamental.
This
increased
"prediction
quality"
by
about
25%
in
all
considered
scoring
systems
in
those
1195
games
(
only
even
teams,
no
1v1,
no
ffa,
no
coop)
.
It
can
be
that
this
will
not
hold
in
a
larger
database.
If
it
should
be
implemented,
elo
change
should
account
for
it.
|
| 6 |
\n
|
6 |
\n
|
| 7 |
[spoiler]There are some paramaters in those systems (like powers). The ideal ones can be calculated from the database itself. I just didn't do that yet, because 1195 games seem to be not enough for that and there's also the problem that elo depends on the system it is balanced with.
|
7 |
[spoiler]There are some paramaters in those systems (like powers). The ideal ones can be calculated from the database itself. I just didn't do that yet, because 1195 games seem to be not enough for that and there's also the problem that elo depends on the system it is balanced with.
|
| 8 |
\n
|
8 |
\n
|
| 9 |
The thread title is no more appropriate.[/spoiler]
|
9 |
The thread title is no more appropriate.[/spoiler]
|
| 10 |
If I should test and publish more advanced systems, I need 3 things: Suitable access to the database, LaTeX integration for the forum, a bit of time.
|
10 |
If I should test and publish more advanced systems, I need 3 things: Suitable access to the database, LaTeX integration for the forum, a bit of time.
|
Please banned player Firepluk