Larian Banner: Baldur's Gate Patch 9
Previous Thread
Next Thread
Print Thread
#766272 22/03/21 12:36 AM
Joined: Dec 2020
A
Adiktus Offline OP
member
OP Offline
member
A
Joined: Dec 2020
I turned off loaded dice, because I don't want an unfair/distorted gaming experience - que sera, sera, and all that. But nobody told the kuo-toa. Their rolls in my fight against them:

11, 17, 16, 19, 4, 7, 13, 14, 12, 19, 10, 14, 9, 14, 17, 16, 2, 19, 6, 4, 9, 17, 17, 7, 8, 18, 13, 16, 20, 19, 18, 16


That's an average roll of THIRTEEN POINT TWO.

My party all had AC 19 (or 17 if I forgot to equip a shield at end of turn). That meant the fishy tossers had to roll at least a 14 (or 13) to hit. Which you would expect them to do 30% of the time. NOT 60% OF THE TIME.

Yes, we survived, but both Wyll and Shadowheart were dead dead dead (for comparison, Shadowheart's saving throw vs Ironbound Pursuit: 5).

Adiktus #766281 22/03/21 02:13 AM
Joined: Sep 2020
veteran
Offline
veteran
Joined: Sep 2020
It has been shown (https://forums.larian.com/ubbthreads.php?ubb=showflat&Main=96760&Number=765262#Post765262) that Larian's unweighted dice produce non-uniform results. Notably, they produce an overabundance of 17s and underabundance of 6s.

Your 32 rolls alone aren't nearly enough to be a statistically valid sample (you'd need at least >300 rolls), but they do reinforce the frequent occurrences of 17.

Combined, yours+@DragonSnooz+@Saberem's 308 unweighted die rolls produce a chi-squared value of 36.15, which is greater than the critical value of 30.1, showing that the sample is not-consistent with uniformity with 95% confidence.
--17 appears 2x more often than it should, a 4-sigma difference
--1 appears 50% as often as it should, a 1.9-sigma difference
--6 appears 50% as often as it should, a 1.9-sigma difference
--The average is 11.5 (The standard deviation of a n=308 sample is 0.33, putting our average 3-sigma off)

Similarly, 747 weighted die rolls produce a chi-squared value of 14.7, extremely consistent with a uniform distribution.
--6 appears ~70% as often as it should, a 1.8-sigma difference
--The average is 10.7 (The standard deviation of a n=747 sample is 0.21, putting our average only 1-sigma off = within expected tolerance)

Notably, the weighted dice are not actually weighted. They produce a uniform distribution, so in fact it is NOT cheating to use them. It would be more accurate to say that using the unweighted dice is cheating, as you get a lot of 17s and infrequent 1s.

Last edited by mrfuji3; 22/03/21 04:53 AM. Reason: updated errors
mrfuji3 #766621 23/03/21 09:50 PM
Joined: Dec 2020
A
Adiktus Offline OP
member
OP Offline
member
A
Joined: Dec 2020
Thanks for that highly authoritative explanation!

Adiktus #766633 23/03/21 11:51 PM
Joined: Jun 2020
veteran
Offline
veteran
Joined: Jun 2020
If you're interested in a little more die-rolling data that has been collected and collated as a way of showing pattern analysis, there have been a number of threads on the matter, if you poke around.

I've written up a couple of reports in various of those, but for the life of me I can't locate the threads now - they got buried deeply after the discussions ran their course.

However, I can share with you the results of the largest test I performed just prior to the last patch, before 'Loaded Dice' were implemented. These results are an attempt to expose more of the raw RNG algorithm that the Larian game engine is using, and compare it to other games that also base their mechanics on a random d20 die roll system. Important in this was preserving the order in which the rolls were gathered and gathering them in single long contiguous strings without interruption.

https://docs.google.com/spreadsheets/d/1tdyBoQNS_vwEGZGBgFRQex7b-Ma8S6P7zvEMK5wh9n4/edit?usp=sharing

Adiktus #766638 24/03/21 12:28 AM
Joined: Sep 2020
veteran
Offline
veteran
Joined: Sep 2020
Thanks for remining me about your data @Niara. Adding it to the unweighted sample, the data looks decently more uniform, but still statically inconsistent with a uniform distribution.
n=508
average=11.16 (The expected standard deviation of a n=508 sample is 0.256, putting our average 2.6-sigma off from the expected average of 10.5)
chi^2=30.89 (still greater than the critical value of 30.1, but significantly less so than the n=308 data)
Biggest Offenders
--1 appears 60% as often as it should, a 2.1-sigma difference from expectation
--6 appears 70% as often as it should, a 1.5-sigma difference
--17 appears 1.77x more often than it should, a 3.9-sigma difference

This goes to show the importance of sample size. Going from n=308 -> n=508:
-the data went from 99% confidence of inconsistency with uniform (chi^2=36.15) to 95% confidence (chi^2=30.89)
-the average went from 11.5 (3-sigma off from 10.5) to 11.16 (2.6-sigma off).
-all the offenders were less offensive

I will probably continue this analysis if I get more data, so if anyone has robust* die rolls from patch 4 or earlier, weighted or unweighted, please give them to me. 1I'd like to get 1000 rolls.
*Robust meaning following as many as possible of: recording all rolls, without reloading the game in between each roll, without advantage or disadvantage, not a halfling, etc.
**Extra points if you record the order of the rolls, like @Niara
***If you see an error with my stats, please let me know. I'm like 80% confident I'm doing things correctly.

mrfuji3 #766686 24/03/21 12:18 PM
Joined: Oct 2020
S
apprentice
Offline
apprentice
S
Joined: Oct 2020
I've recorded my roll order: https://pastebin.com/BsNC9y47 if you need it (you already have my results) Copy and paste raw data into google sheets will format it back into a table.

I think streaks are more important than overall distribution. % chance to hit isn't helpful if we are going to miss a 50% chance to hit 6 times in a row at least once every 100 rolls which seems to be the case in both mine and Niara's results.

Last edited by Saberem; 24/03/21 12:19 PM.
Adiktus #766688 24/03/21 12:31 PM
Joined: Oct 2020
old hand
Offline
old hand
Joined: Oct 2020
This whole time I didn't click on this thread cause I thought you was actually talking about food again. Title is miss leading.

Adiktus #766700 24/03/21 03:02 PM
Joined: Sep 2020
veteran
Offline
veteran
Joined: Sep 2020
Thanks @Saberem. To confirm, your results are all from Patch 4, right?

Visually, your data doesn't show the obvious sine-wave pattern that @Niara's data did, but this could be due to randomness, the smaller sample size, or perhaps Larian did some secret fix to their rng between patches 3 and 4. Although, your unweighted rolls 30-75 do show a suspicious lack of results <5, which could be part of a sine-wave pattern with a period of ~50+, and thus not visible with only 100 data points.

I agree @fallenj. @Adiktus, any chance you could update the title to "Suspicious rolls" or "BG3 die rng" or something like that?


Moderated by  Dom_Larian, Freddo, vometia 

Link Copied to Clipboard
Powered by UBB.threads™ PHP Forum Software 7.7.5