Talk:Rolling Averages
|
Time Weighted Rolling Average
Err, Nat, Paul Evans's version, and "Time Weighted Rolling Average" are essentially the same. The way you calculate the numeric constant is different, but the end result is the same. About the statement "The old value that exceed the n limited will be removed from averaged value like a magic!" isn't really anything magic, it's just weighting old values less. --Rednaxela 03:28, 17 February 2009 (UTC)
Thanks! I haven't realize that, just copy from old wiki! I'll change this page and move "Time Weighted Rolling Average" to sample implementations instead. About that statement, the value that exceed n will weight less enough that has really no side effects. The statement "like magic!" is one of statement on the old wiki, I want to keep it.
- [History↑]
Contents
| Thread title | Replies | Last modified |
|---|---|---|
| Rolling Depth | 3 | 03:18, 30 June 2019 |
In Basilisk, the only change I made from 3.2 to 3.3 was increasing the rolling depth by multiplying my bins by .99 rather than .95 every time a wave hits, and it gained me 2 ranking places!(From #8 up to #6). So, I was curious, what rolling depths do other people use in their guns? Also, are 1000 round battles good for testing rolling averages? I like 1000 rounds since it lowers the margin of error, but it might distort my results because it's long term learning, not fast 35 round learning. Thanks!
For non-adaptive movement, not using rolling average works as well, and if added a little, it performs better against multi-mode movement. But for adaptive movement, very fast rolling average is generally used.
1000 rounds battles are certainly not as accurate as running 30 * 35 rounds battles (1050 in total), that's why roborunner is invented, which you may have a try.
Btw, I'm using the whole rumble (1100+ robots) for gun testing now, ~5000 battles * 35 rounds = ~175000 rounds in total, which takes me ~2 hour on i9 8 core.
You might also try to weight non-bullet waves (the one where you don't actually fire a bullet) with a factor 0.1 or 0.2. For non-adaptive movements this should not impact scoring, but for adaptive movements it really increases accuracy. For mini Grimmig and micro I don't use rolling depths, for GresSuffurd I use two same guns, one without rolling depth and one with 0.9 (I think) to handle the wavesurfers. Currently my testing is just a handful of battles against my old self to tackle bugs. I try to only make changes that are logical (to me) and don't change to much at the same time. And don't forget, watch battles !! Especially your bots behaviour near walls/corners, at close range or against opponents where you score relatively low against, can give you insight on things to improve or fix.
OK, I think I'll keep my rolling average for now, but weight the non-bullet waves less like GrubbmGait said.
@Xor, I'll also install RoboRunner, I remember I planned to install it earlier but I keep forgetting :-P On my desktop at home, I have a Ryzen 3 1200 4 Core. 175000 rounds sounds like a lot! I'll probably not test against the entire rumble, but 30 seasons of 35 rounds sounds good!
@GrubbmGait, I try to only change one thing at a time, so that I can tell what change did what. And I always like watching the last 5 rounds of a battle at the end =)
Thanks for the help, both of you!
