Well, you did the work, and your results show that the success rate is reasonably close to 85%. So I suppose there's nothing wrong. My sample size just wasn't large enough.
Then Troodon80 went on to say (I only quoted the part I found relevant):
Clearly, the frequency of success is not 85% with a character of int 18. The evident, or empirical, rate (from what I can see,) is a median of 77% (using only the highest and lowest values that I have been able to get from the game and my program, actual median is about 84%, with deviation of 3%) with a deviation of ~16%.
The average person won't learn any more than about 150 spells during a fresh playthrough (I'm not even sure what the maximum number of spell scrolls there are (without mods)), so working with a sample higher than that is pointless when trying to determine if what you are experiencing is normal behaviour for the game.
I'm actually not quite sure what he meant by "actual median". The median is the median, there is no actual or not actual. I'm guessing he meant "mean". But whatever. Then I went on to say:
@Troodon80 It is reasonable to assume a low sample rate for the reason you've given, that throughout the course of the game you're not going to write more than a certain number of spells. But if the median you obtained from multiple trials is 77% with an 85% chance of success, then there is something wrong. The system clearly favors failure more than the % chance it's using. Whether or not the system is working properly programming-wise, it's not working as intended in practice. If the system really works as intended, you might get a lot of deviation due to the small sample size, but you should get more or less equal occurrences of both high and low, with a median much closer to 85%. Random "luck" should work both ways. From what you said from your trials, you're getting results that favor low. Maybe the mechanics work properly as programmed, but it was poorly designed. Or as moopy said, maybe it's just not a very good RNG that's been implemented.
In which Troodon80 agrees with my statement. I meant "sample size", btw, not "sample rate". That was a mistake.
You gotta keep up man. If you're too lazy to actually read all the posts in the thread, you probably shouldn't be trying to make arguments.
@Aosaw If he doesn't want to read them, that's fine. But he shouldn't be making arguments when he clearly doesn't know all the information. It just makes him sound like a fool.
@Awong124, I meant median for two different trials (one which consisted of several smaller tests of 50 and 150 since, like I said, there's no point doing anything higher if I want to find out whether or not the behaviour you were getting within the game was normal for that sample size), in which I got a numbers between 60% and 92%, and several number in-between, thus the median was ~77% (by getting the middle numbers of the set). The second set of trials consisted of every test I had done over the course of the debate, which is up in the millions (12,510,181) at this point, where the median is roughly 83%, with a margin of error of about 3% (since I can't exactly be that precise, I didn't create a split function to take all the numbers, so I can only go by what I think is the middle). The mean average is about 85% with 0.3% deviation. Key word being millions. I will note that I still didn't get anything as low as what your pointed out on the first page, but I did get lower than 85% success rate. I also got much higher.
If I take the results of six tests: { 60.0, 63.0, 71.5, 82.4, 84.0, 92.0 }; I get 77.0. But with a larger sample to gauge from, it is closer to 83%. But as I said before, it is highly unlikely that anyone is going to write over 12 million spells during the course of their interest in this game, so taking 83% with a higher or lower deviation of 3% (80% - 86%, ) as something that you're going to get when trying to learn seven or eight spells is a little pointless. In the case you provided, that being the Law of Large Numbers, a study of over 12 million tests prove that the working out is, at the very least, semi-accurate, and that using a smaller sample size you are going to get some large deviation. The graph in the wiki article even proves that theory when using a six sided dice, you get deviation up to about 425 trials. Since they could be using the logic of a 20 sided die, or perhaps they are working with stable 100, or possibly higher, then the number of trials needed before it reaches a regular mean is going to be higher. Perhaps if I did another 100,000,000 tests, the number would sit firmly on 85, but there's very little point in doing that.
On the flip side, they could be using the logic of a 4 sided die (which is highly unlikely), which would reduce the overall number of tests needed. If that is the logical number they are using, then something is definitely broken.
@Awong124, I meant median for two different trials (one which consisted of several smaller tests of 50 and 150 since, like I said, there's no point doing anything higher if I want to find out whether or not the behaviour you were getting within the game was normal for that sample size), in which I got a numbers between 60% and 92%, and several number in-between, thus the median was ~77% (by getting the middle numbers of the set). The second set of trials consisted of every test I had done over the course of the debate, which is up in the millions (12,510,181) at this point, where the median is roughly 83%, with a margin of error of about 3% (since I can't exactly be that precise, I didn't create a split function to take all the numbers, so I can only go by what I think is the middle).
Ah, I see what you mean. They were 2 separate medians of two different sets of results. I thought all your results were just lumped together, that's why I was confused as to why there were two medians.
But I don't understand why they would be using the logic of a 20 sided die when a computer can easily simulate a 100 sided die roll.
@Troodon80 Oh, another thing. You said you never got results as low as what I got. That's probably because I used Neera for my test. She has 17 INT with 75% chance of learning spells. You used someone with 18 INT and 85% chance right?
@Troodon80 Oh, another thing. You said you never got results as low as what I got. That's probably because I used Neera for my test. She has 17 INT with 75% chance of learning spells. You used someone with 18 INT and 85% chance right?
Ah, yes, I was using int 18 / 85% (my standard character), so that was the basis for my results across the board.
Imoen should be able to learn a spell 75% of the time, I believe. given her INT of 17. Perhaps I'm the unluckiest person in the world, but I noticed her having some problems learning spells, so I ran some tests (five scrolls tried, reload, try again) and she's actually hitting about 25% over about 40 attempts. Is anybody else seeing this? If she consumes a Potion of Genius all is well, but in the absence of that, I've actually had to resort to temporarily dropping down to "normal" rules from "core" rules when I need Imoen to learn a spell, it's that bad. (By contrast, Dynaheir is meeting correct expectations on memorization in my party.)
(Edit: I forgot to mention that I'm using spells that she should be able to cast, so no penalties involved with trying to learn high-level spells at a lower level.)
Folks, anyone who doubts that the pRNG is either broken or working with bad seeds should try the experiment with a potion of genius... You can get Dynaheir's chance to learn spell to 97% with one of these, and, even assuming that the pRNG simulates a 20-sided dice and thus all chances cap at 95%, there's no way this is the actual chance. If you keep reloading before the attempt, failures are about 1 in 3. It's slightly better when not reloading (thus I tend to blame bad seeding), but still about 1 in 5. Strange things also keep happening with attack rolls, having LOOONG (like, VERY unlikely assuming an even / random distribution) rows of extremely low or extremely high rolls that stay the same after a reload. And then there's the saving throw bias with enemies having about twice the chance to make a save compared to the party... (No, really, try this experiment: save right before an ancounter with an enemy group that will cast horror, confusion or hold person at the party, and cast the same at them. reload and repeat. On average, two out of 6 enemies are affected and 4 out of 6 party members. The more often you try, the more obvious.) Nah, this pseudo random number generator is very pseudo and not very random :-P
I played the original Baldurs Gate when it first came out back in the 90s, the BG collection on DVD with TotSC, and the GOG version. I also went through BG2 and ToB with a fighter/mage. And, trying to write spells in BG: EE feels off to me. It's so bad that I quicksave after every sucessful write so I don't lose the scroll. It really feels more like a coin toss than 75%+ it should be for Neera and Imoen. Even writing 1st level spells for mages that can cast 3rd level feels iffy.
I don't know if its in my head or if something is wrong with the seed or the BG engine handled it more favorably for players than the BG2 engine. But it should be looked at and possibly normalized to fit expectations better.
I've been trying to folllow some of these long posts to see if anything conclusive can be discerned from the results with *large* sample sizes (which are *really* needed for testing statistical chances), but it's quite difficult to follow withouth reading every post in full, which I don't have time to do, unfortunately...
Is there general conclusion, though, that whilst there *is* on the whole a higher chance of spell learning success with higher Int, it's not as high as perhaps should be expected and also possibly that the pseudo random number generator isn't as random as itshould be? Also has anyone requested that the random number generator for spell scribing be look at by the Devs in the Bugs forum? If the RNG is off, that's potentially a pretty serious bug. Or is it just that some people have had bad luck?
@Oxford_Guy check out @Troodon80 's post at the top of the second page. He did 3000 trials copying Color Spray and ended up with a success rate of 85.1% with an 18 INT (so expected 85% chance to learn spell), non-specialist mage. While that doesn't 'prove' spell learning is working as intended, it does suggest it.
Oh man, I just thought of a diabolical new mechanic:
The wild mage who spell surges when writing a spell, and permanently writes that surge's effect into his spellbook. Now and forever, his Magic Missiles will destroy 80% of the party's gold...
@ArminW Since this came up in another thread, recall that there is a +10%/level penalty for scribing scrolls that you can't currently cast. So if you're trying to scribe a level 5 spell when the highest you can cast is level 3 spells, you're failure rate is the default for your INT + 20%.
@Rann, @ArminW and others who are just joining the thread and asking stuff
1: Probability only becomes true as you approach infinity. 10 trials? 100 trials? 10 trials of 100 trials? Technically irrelevant, though the third option is the best of them. 2: There is the possibility of a bad seed. It's possible reloading hurts more than it helps.
@Troodon80: I divided up your data gathered from the game. I was going to try my darndest to make my quantitative psych teacher proud and voodoo my way into some kind of standard distribution of the means, using 60 trials of 50, instead of your 30/100s. It was inspired by the points you and @Awong124 raised about the sample size needing to remain relevant to the game...
Now I'm sort of staring at a brick wall on how I'd go about it. But I noticed something else. I come up with a different success rate than you. You have 3000-222-225 successes, for 2553.
I come up with 2620 successes out of your data. I've double checked. And triple checked the sum. My horrible math says your trial of 3000 yields 87.333% success rate.
Like, here, you say YYYYYYNYYY 17 NYYYYYYYYY YYYYYYYYYY 12 YNYYYYYNYY
Shouldn't it actually be 18 and 18? I assumed those values in the center were the sum for success on that line. Am I mistaken? Either way, how does that address drastically different yields of two people counting the same data?
@ArminW Since this came up in another thread, recall that there is a +10%/level penalty for scribing scrolls that you can't currently cast. So if you're trying to scribe a level 5 spell when the highest you can cast is level 3 spells, you're failure rate is the default for your INT + 20%.
.
Has this penalty for trying to scribe spells higher than the level you can currently cast been confirmed? Also (if true) I think it would be clearer in the example you give to say that the success rate is the default for your Int - 20%, as it's normally success rate percentages that are provided, not negatives.
Also, does anyone know if the chance of scribing scrolls from your school of specialism is higher, or us that just wishful thinking?
@Kushu, looking back over it, something appears to have gone screwy in the copying. I remember far more instances where there were sets of three Ns in a line. No idea what happened there - I used a pretty simple tab delimited spreadsheet in order to get and count the original numbers to make it easy for myself, then used Notepad to remove the tabs in order to make it more readable when posted, so it could have been one point where I was using
Y N
as the search criteria and replaced it with a
YY
, or a different combination (which would be a tremendous goof on my part). Basically, the number at top represents the left side and the number at the bottom represents the right side.
Unfortunately, I didn't see any need to keep the original Y/N data since I thought everything was fine. Copying and formatting what is in my post, it would have looked something like this, just to give you an idea on how I did it. The original numbers in the middle are accurate, though, as I saved the final results to a separate text document without all the Ns and Ys to keep it tidy and less strenuous on my eyes.
I might redo the 3000 tests and then double check the copied data (maybe over this weekend if I get the time) and update my results. Then again, I suppose it doesn't really matter - the devs will either look into the issue and feel that something needs changing or they'll leave it alone.
@Troodon80 Ahh. Excellent! Glad your totals work then. Also? Holy crap, I wish I had thought to use a search function as opposed to just a line by line looksie. *facepalm*
I think that the only way to know for certain if the mechanic is working correctly is to look at the numbers used to determine success. Counting trials is anecdotal evidence, and it doesn't prove anything in terms of whether the mechanic is working as intended or not.
What? What they were doing here is empirical research. It's pretty far from anecdotal evidence. Anecdotal evidence is "I got a friend who once got max HP on every single level-up without reloading, lol something's wrong with the dice mechanics".
EDIT: Sorry, I just noticed this was already brought up. Still doesn't hurt to emphasize it once more
Comments
Then Troodon80 went on to say (I only quoted the part I found relevant):
I'm actually not quite sure what he meant by "actual median". The median is the median, there is no actual or not actual. I'm guessing he meant "mean". But whatever. Then I went on to say: In which Troodon80 agrees with my statement. I meant "sample size", btw, not "sample rate". That was a mistake.
You gotta keep up man. If you're too lazy to actually read all the posts in the thread, you probably shouldn't be trying to make arguments.
@Awong124, I meant median for two different trials (one which consisted of several smaller tests of 50 and 150 since, like I said, there's no point doing anything higher if I want to find out whether or not the behaviour you were getting within the game was normal for that sample size), in which I got a numbers between 60% and 92%, and several number in-between, thus the median was ~77% (by getting the middle numbers of the set). The second set of trials consisted of every test I had done over the course of the debate, which is up in the millions (12,510,181) at this point, where the median is roughly 83%, with a margin of error of about 3% (since I can't exactly be that precise, I didn't create a split function to take all the numbers, so I can only go by what I think is the middle). The mean average is about 85% with 0.3% deviation. Key word being millions. I will note that I still didn't get anything as low as what your pointed out on the first page, but I did get lower than 85% success rate. I also got much higher.
If I take the results of six tests: { 60.0, 63.0, 71.5, 82.4, 84.0, 92.0 }; I get 77.0. But with a larger sample to gauge from, it is closer to 83%. But as I said before, it is highly unlikely that anyone is going to write over 12 million spells during the course of their interest in this game, so taking 83% with a higher or lower deviation of 3% (80% - 86%, ) as something that you're going to get when trying to learn seven or eight spells is a little pointless. In the case you provided, that being the Law of Large Numbers, a study of over 12 million tests prove that the working out is, at the very least, semi-accurate, and that using a smaller sample size you are going to get some large deviation. The graph in the wiki article even proves that theory when using a six sided dice, you get deviation up to about 425 trials. Since they could be using the logic of a 20 sided die, or perhaps they are working with stable 100, or possibly higher, then the number of trials needed before it reaches a regular mean is going to be higher. Perhaps if I did another 100,000,000 tests, the number would sit firmly on 85, but there's very little point in doing that.
On the flip side, they could be using the logic of a 4 sided die (which is highly unlikely), which would reduce the overall number of tests needed. If that is the logical number they are using, then something is definitely broken.
But I don't understand why they would be using the logic of a 20 sided die when a computer can easily simulate a 100 sided die roll.
@Moomintroll
I used to love Tom Lehrer. Still do actually.
(Edit: I forgot to mention that I'm using spells that she should be able to cast, so no penalties involved with trying to learn high-level spells at a lower level.)
You can get Dynaheir's chance to learn spell to 97% with one of these, and, even assuming that the pRNG simulates a 20-sided dice and thus all chances cap at 95%, there's no way this is the actual chance. If you keep reloading before the attempt, failures are about 1 in 3. It's slightly better when not reloading (thus I tend to blame bad seeding), but still about 1 in 5.
Strange things also keep happening with attack rolls, having LOOONG (like, VERY unlikely assuming an even / random distribution) rows of extremely low or extremely high rolls that stay the same after a reload.
And then there's the saving throw bias with enemies having about twice the chance to make a save compared to the party...
(No, really, try this experiment: save right before an ancounter with an enemy group that will cast horror, confusion or hold person at the party, and cast the same at them. reload and repeat. On average, two out of 6 enemies are affected and 4 out of 6 party members. The more often you try, the more obvious.)
Nah, this pseudo random number generator is very pseudo and not very random :-P
I don't know if its in my head or if something is wrong with the seed or the BG engine handled it more favorably for players than the BG2 engine. But it should be looked at and possibly normalized to fit expectations better.
Is there general conclusion, though, that whilst there *is* on the whole a higher chance of spell learning success with higher Int, it's not as high as perhaps should be expected and also possibly that the pseudo random number generator isn't as random as itshould be? Also has anyone requested that the random number generator for spell scribing be look at by the Devs in the Bugs forum? If the RNG is off, that's potentially a pretty serious bug. Or is it just that some people have had bad luck?
The wild mage who spell surges when writing a spell, and permanently writes that surge's effect into his spellbook. Now and forever, his Magic Missiles will destroy 80% of the party's gold...
1: Probability only becomes true as you approach infinity. 10 trials? 100 trials? 10 trials of 100 trials? Technically irrelevant, though the third option is the best of them.
2: There is the possibility of a bad seed. It's possible reloading hurts more than it helps.
Now I'm sort of staring at a brick wall on how I'd go about it. But I noticed something else.
I come up with a different success rate than you. You have 3000-222-225 successes, for 2553.
I come up with 2620 successes out of your data.
I've double checked. And triple checked the sum. My horrible math says your trial of 3000 yields 87.333% success rate.
Like, here, you say
YYYYYYNYYY 17 NYYYYYYYYY
YYYYYYYYYY 12 YNYYYYYNYY
Shouldn't it actually be 18 and 18? I assumed those values in the center were the sum for success on that line. Am I mistaken? Either way, how does that address drastically different yields of two people counting the same data?
Has this penalty for trying to scribe spells higher than the level you can currently cast been confirmed? Also (if true) I think it would be clearer in the example you give to say that the success rate is the default for your Int - 20%, as it's normally success rate percentages that are provided, not negatives.
Also, does anyone know if the chance of scribing scrolls from your school of specialism is higher, or us that just wishful thinking?
Unfortunately, I didn't see any need to keep the original Y/N data since I thought everything was fine. Copying and formatting what is in my post, it would have looked something like this, just to give you an idea on how I did it. The original numbers in the middle are accurate, though, as I saved the final results to a separate text document without all the Ns and Ys to keep it tidy and less strenuous on my eyes.
I might redo the 3000 tests and then double check the copied data (maybe over this weekend if I get the time) and update my results. Then again, I suppose it doesn't really matter - the devs will either look into the issue and feel that something needs changing or they'll leave it alone.
Also? Holy crap, I wish I had thought to use a search function as opposed to just a line by line looksie. *facepalm*
EDIT: Sorry, I just noticed this was already brought up. Still doesn't hurt to emphasize it once more
I'm aware that it has to come up some time but stuff like that seems to happen a lot. Though maybe that's just my perception.