What the ■■■■ are you talking about? All kills are already tracked in the entire game for a variety of reasons, seriously what on earth are you even talking about? Like this is the most blatantly “I know nothing about what I’m saying” post I’ve seen in a while. “It would be impossible to optimize a counter that goes up by 1 when a mob dies”, dude do you have any idea the number of variables being tracked at any given time in an MMO? There are thousands of equations happening every second that are a lot more complex than “add 1”.
Thank for you for your time composing this detailed response. I only wish that this was your initial response to this thread, so that that needless confrontations would not have happened.
The only claims I made were these:
I have consistently insisted that I am not an expert in statistics. I even admitted up front that this may not be the correct way to do this. Perhaps I did not emphasize it sufficiently by providing a personal background paragraph. I’ve only had basic statistics a long time ago in college (note: this is not the equivalent of colleges in US, it is somewhat between high-school and undergrad). Therefore our conflict must have stemmed from the fact that you came in expecting a peer in statistical prowess; which isn’t the case.
Due to courtesy, omission or my lack of comprehension, your initial post in this thread appeared brief and deficient. That got me off on the wrong foot.
But what’s done is done. And I apologize for my part of the problem.
Thank you for the detailed explanation. This is the kind of critique that I was hoping for. I attempted to revise sampling techniques because as you rightly brought up, it seemed problematic. Even as I was going through the material this morning before heading out to work, I saw the inconsistencies in my sampling method. But now with your elaboration, it is even clearer the bias involved by sampling this way. For that I thank you for a well given lesson.
On the matter of this statistics professor, it is currently the end of summer holiday and he is indeed absent until next week. In the mean time I am seeking for advice in Cross Validated, a site of Stack Exchange. I do not know if you already use this site, but if you are willing to, I invite you to answer the questions I have (and will have more no doubt) posted there.
As for cluster analysis or goodness of fit you have suggested, I’ve previously stated that I did not know how to do them. And suggested that you to demonstrate it (with no ill intent).
And since you have declined to offer this to our community, I will have to do further reading over this week as I have never learned neither of those.
Now no need to be sarcastic here, I have met many professors in my life, though not many of them are math / statistics professors. Unfortunately, I chose a path that had nothing to do with math / statistics in my higher level of education. It is one of the things I regret, as you can probably tell I do have an interest in math / statistics; and I did pretty well in math back in high school / college (mostly algebra / calculus etc.).
In closing, I just want to clarify that I REALLY want to get this right. I do not necessarily know the correct way to do it now, but I will keep going until I find a solution. I thank you in advance if you are willing to help me on this journey, or if you choose to decline, I can accept that as well.
Too much math, no more.
I am on SE, but mostly on mathoverflow, but that’s for research level stuff. I’m sure you’ll get good info on Cross Validated - there are some excellent posters tthere (Whuber and John Cook come to mind - both excellent in the field, and from what I’ve seen patient, more than can be said for me).
Be sure to make it clear what your are doing now is comparing a random sample with a sequential one, and others will likely point out the issues and others I might have missed.
As I recommended above (and what my first inclination would be were someone to ask me about a similar test need), I’d test the fit of the distribution.
If it’s “RNG” (IID trials), the distribution of number of trials to success will be geometrically distributed. If it has the counter (or some ~equivalent mechanism like a truncated geometric distribution) with as gross a difference as you propose (>=100 between “drops”), a pretty small sample will suffice to have good power in something like a Chi-Square goodness-of-fit test (e.g., sample until 20 or so “drops”).
So, if you can get on the game (I don’t know the game, obviously), and do your thing while tracking attempts (without external interference - don’t know if that’s a confounding factor here) until you’ve had 20 or so successes, then take the results (numbers of attempts to each success) and test its fit against the null hypothesis’ distribution (the plain-vanilla geometric).
If you get a p-value <=0.05, you’ve got evidence to reject the null (that the mechanism is plain “RNG”).
Be careful with wording - that does not prove it’s not RNG, any more than it would prove it’s not - it’s always possible (at least in real-world tests like this) that you’ve just experienced a fluke (and “proof”, like “confidence” have very distinct meanings different from the colloquial use)… But a low p-value is strong evidence against the null.
If you don’t have it already, you can use, e.g., R to do the test - it’s quite good and free.
1 Like
While the sample size isn’t very big. I can see the trend you’re trying to show and to make sense about.
While this shows drop rate may not be “pure” RNG, it does not show that it is DPK. It only shows that there may be correlations between the kill count and the drop chance, and there can be many theories people can think of to explain such drop system.
1 Like
The only way to ‘proof’ is to see the source code, or make IMC staff go clear things out. Also notice that there is no real RNG in computers currently.