Difference between revisions of "Thread:Talk:RoboRumble/Participants/Is there a mistake in Rumble battle staging/reply"
(Reply to Is there a mistake in Rumble battle staging) |
m |
||
Line 5: | Line 5: | ||
Anyway, it's not a good idea to rely on LiteRumble to accurately measure how good each small change of your bot is. LiteRumble is designed to give everyone a ranking, well suited when you have a big change (proven to be effective) and want to publish the result. Small changes are prone to noises from the rumble environment (e.g. running on different computers for different versions, having updated opponents), thus may report unreliable result. | Anyway, it's not a good idea to rely on LiteRumble to accurately measure how good each small change of your bot is. LiteRumble is designed to give everyone a ranking, well suited when you have a big change (proven to be effective) and want to publish the result. Small changes are prone to noises from the rumble environment (e.g. running on different computers for different versions, having updated opponents), thus may report unreliable result. | ||
− | Rumble clients are also much slower than running battles locally (with | + | Rumble clients are also much slower than running battles locally (with RoboRunner), due to added network communication (which isn't parallelized), as well as running battles from other updated bots. One can also notice that once each bot reaches ~5000 battles, most battles are then nano battles instead of randomly distributed across the entire rumble, due to some "propagation" effect, further slowing down the stabilization of the general rumble. |
Revision as of 06:16, 21 February 2023
What do you mean by avoiding particular bots. From the code, LiteRumble selects missing pairings first, then pairings with less battles, etc.
Given enough time, each bot should have similar amount of battles.
Anyway, it's not a good idea to rely on LiteRumble to accurately measure how good each small change of your bot is. LiteRumble is designed to give everyone a ranking, well suited when you have a big change (proven to be effective) and want to publish the result. Small changes are prone to noises from the rumble environment (e.g. running on different computers for different versions, having updated opponents), thus may report unreliable result.
Rumble clients are also much slower than running battles locally (with RoboRunner), due to added network communication (which isn't parallelized), as well as running battles from other updated bots. One can also notice that once each bot reaches ~5000 battles, most battles are then nano battles instead of randomly distributed across the entire rumble, due to some "propagation" effect, further slowing down the stabilization of the general rumble.