| 1 |
Thanks everyone for taking time to explain this to me!
|
1 |
Thanks everyone for taking time to explain this to me!
|
| 2 |
\n
|
2 |
\n
|
| 3 |
@Dunno
I
appreciate
that
betting
example,
I
can
follow
that
:)
|
3 |
@dunno
I
appreciate
that
betting
example,
I
can
follow
that
:)
|
| 4 |
\n
|
4 |
\n
|
| 5 |
[quote]
|
5 |
[quote]
|
| 6 |
@Brackman The last adj score term actually punishes Trueskill arbitrarily.
|
6 |
@Brackman The last adj score term actually punishes Trueskill arbitrarily.
|
| 7 |
[/quote]
|
7 |
[/quote]
|
| 8 |
\n
|
8 |
\n
|
| 9 |
\n
|
9 |
\n
|
| 10 |
Yes, you're right. I'm not close to understanding anything about information theory, but I've heard enough to see that my arguments are false and I'm properly out of my depth :D
|
10 |
Yes, you're right. I'm not close to understanding anything about information theory, but I've heard enough to see that my arguments are false and I'm properly out of my depth :D
|
| 11 |
\n
|
11 |
\n
|
| 12 |
One additional mistake I made in comparing to a fixed probability was not to randomize the team order. This appears to be necessary because of the team 1 win bias - anyone know why this is? Does/did the balancer always assign the higher probability to team 1? Anyway, randomizing that stops a fixed probability from getting such good results.
|
12 |
One additional mistake I made in comparing to a fixed probability was not to randomize the team order. This appears to be necessary because of the team 1 win bias - anyone know why this is? Does/did the balancer always assign the higher probability to team 1? Anyway, randomizing that stops a fixed probability from getting such good results.
|
| 13 |
\n
|
13 |
\n
|
| 14 |
What I hadn't really understood is that Trueskill works with individual ratings. In fact the function I took (from https://github.com/sublee/trueskill/issues/1) isn't even in the library proper. It stands to reason that this knows nothing about how individual ratings combine to get a team win probability in Zero-K, so this is clearly a place for adjustment (as noted by @Aquanim). Halfing the difference from p=0.5 seems to yield a substantial increase in the score, so I guess this is closer to the true team win probability.
|
14 |
What I hadn't really understood is that Trueskill works with individual ratings. In fact the function I took (from https://github.com/sublee/trueskill/issues/1) isn't even in the library proper. It stands to reason that this knows nothing about how individual ratings combine to get a team win probability in Zero-K, so this is clearly a place for adjustment (as noted by @Aquanim). Halfing the difference from p=0.5 seems to yield a substantial increase in the score, so I guess this is closer to the true team win probability.
|
Evaluating rating systems