1 |
[quote]There must be a small probability to make a bad step ("tunnel effect") which increases the function E that should be minimized. Otherwise you will probably get stuck in local minima and never find global minimum.[/quote]
|
1 |
[quote]There must be a small probability to make a bad step ("tunnel effect") which increases the function E that should be minimized. Otherwise you will probably get stuck in local minima and never find global minimum.[/quote]
|
|
|
2 |
Yeah, the descent algorithm was more for the purposes of proof-of-concept than for being a finished product.
|
|
|
3 |
\n
|
2 |
A question to ask is whether a global minimum (in terms of the objective of elo balance) is even the most desirable outcome. In a 10v10 game a difference of 1 or 2 in average elo will make very little difference, but if the 2-difference game is closer to the original point (and therefore probably the elo distributions on each team are more similar) it might be a preferable game to play.
|
4 |
A question to ask is whether a global minimum (in terms of the objective of elo balance) is even the most desirable outcome. In a 10v10 game a difference of 1 or 2 in average elo will make very little difference, but if the 2-difference game is closer to the original point (and therefore probably the elo distributions on each team are more similar) it might be a preferable game to play.
|