6-socket Orb of Fusing Community Log (done)

#3 is plausible. The variance of a binomial distribution is np(1-p), where n is the number of trials and p is the probability, so if the probability was 1/4, we would have a standard deviation of 10% of the answer in less than 400 trials.

#2, however, is contradicted by the data. If we take the entire dataset, 6L has a probability of 0.111%, which would imply p ~= 25.65% from the expected probability p^5. Under the independence assumption, 5L should have a probability 5 * p^4 * (1 - p) ~= 1.61%, which is way higher than the observed value.

I wish ggg would release the exact numbers already :P its not like it would deter people anymore so than seeing horror stories of 3000 + failed fusings.

@E4Z

My supposition is indeed 25%, or about 0.1% for 6L. (pre 0.11 patch)


Regarding 5L, you are counting 5+1, but also 4+2 and 3+3. You have to multiply by 2, not by 5 (the missing link must be either the first or the last). This is one of the reason for which i say that we need also other data.

With p=1/4 the resulting probability for exactly 5L is 0.586% (1 out of 170.7), for at least a 5L is 0.686% (1 out 146)

There is one problem with this system: Usually people will stop fusing when they get their 5L or 6L. To get accurate results, people would need to pre select a quantity of fusings and continue fusing no matter the result. Otherwise, we include a bias based on the result. Reporting and sampling must not be biased by the results and I'm not sure how to avoid that here.

For the math:

Yesterday
7 Fuse with 20% quality on item.

Prepatch:

107 Fusing.

"

patlefort3 wrote:

There is one problem with this system: Usually people will stop fusing when they get their 5L or 6L. To get accurate results, people would need to pre select a quantity of fusings and continue fusing no matter the result. Otherwise, we include a bias based on the result. Reporting and sampling must not be biased by the results and I'm not sure how to avoid that here.

Well the issue is with people stopping at 5L rather than anything with 6L.

Probabilities should be divided individually for people who are looking to get a 5L (or better), and people who are looking to get a 6L. The problem combining the two is that most people are stopping at 5L making 6L rates disproportionate (I think). I think (not very thoroughly) that it makes 5L chance seem higher, and 6L chance seem lower. Not sure by how much (obviously will vary)

"

HellGauss wrote:

2) I STRONGLY believe that, before patch, the probability was 1/4 for each link, as I discussed previously. In my opinion, a fast way to implement a prob gain with quality is to change the above ratio as a function of Q%.

If this was the case, then the probabllity of getting 0L would be 23,7%, which is definately not true. Probability of getting 0L is quite low.

"

Xapti wrote:

"

patlefort3 wrote:

There is one problem with this system: Usually people will stop fusing when they get their 5L or 6L. To get accurate results, people would need to pre select a quantity of fusings and continue fusing no matter the result. Otherwise, we include a bias based on the result. Reporting and sampling must not be biased by the results and I'm not sure how to avoid that here.

Well the issue is with people stopping at 5L rather than anything with 6L.

Probabilities should be divided individually for people who are looking to get a 5L (or better), and people who are looking to get a 6L. The problem combining the two is that most people are stopping at 5L making 6L rates disproportionate (I think). I think (not very thoroughly) that it makes 5L chance seem higher, and 6L chance seem lower. Not sure by how much (obviously will vary)

In both case, if people stop at 5 or 6L it will bias the results. If people pre selected a quantity and didn't stopped (one could continue on another piece if they achieve their target), there would be no bias and we could get accurate stats on both 5 and 6L.

There is no bias simply because people stop. All trials are independent.

The real problem people are having here is that they assume that "links" are independent... they're not. They've simply assigned arbitrary probabilities to the various outcomes. They don't decided links one at a time. It's simply not predictable theoretically. You have to look at the data.

"

There is no bias simply because people stop. All trials are independent.

I agree. But there is a little bias due to the fact that maybe people are not posting results until they get a 6L, ore maybe there are some correlation between the number of trials to get 6L an the fact that player are reading and posting in this thread.

"

The real problem people are having here is that they assume that "links" are independent... they're not. They've simply assigned arbitrary probabilities to the various outcomes. They don't decided links one at a time. It's simply not predictable theoretically. You have to look at the data.

I disagree. I remark again that

1) In this thread we have info only on the three complementary events 6L ; 5L ; <5L. If we suppose that the prob of each of the 5 links are independent and equal to p=1/4, we get

a) P(6L)=1/1024=0.098 %
b) P(6L or 5L)=7/1024=0.684 %

which is in EXCELLENT agreement with final data available for pre 0.11, and i suppose that these prob are unchanged in 0.11 if we set Q=0%.

All we need to do is collect dat for various Q% and answer the question : "It is convenient to spend AS/BSW to increment the chances to get 5/6L?".

I think we should discuss of a new standard to collect data. I propose, for a report something like:

Item: ES-Chest
Tests:
10-1001X
10-1100X
05-0001X
05-0010X
-etc....

The example is for a 5S chest

First two characters is Q%, next five characters are links (On=1, off=0) in the order: Up, right, middle, left, bottom. The X is because a 5S has no bottom link. X are always at the end. ( A report for a Q=8% 3S shield is like 08-01XXX).

If POE community is in agreement, i will start a new thread ONLY to collect data (discussion can be continued here).

HG