Goblin Squad Member
|
I did repeat Caldeathes experiment to see if knowledge local increases the drop rate. I did 2 tests of 100 kills each - the first one with a brand new 1000 XP char and knowledge 10 - the second one with Theodum - scholar - boosted for this experiment to knowledge local 73.
This should give him a 3.15% higher drop rate.
In addition I teamed both up and did 100 kills together. There was an irregularity in the distribution - so I only give the combined drop rate.
Bottom line:
Again the lower skilled character gets more drops. Together with Caldeathes data this gets more and more significant. I did some more detailed statistical tests on Caldeathes data alone (had to test some new stats software for work) and using T-test and TOST I got 81-95% confidence that knowledge is not increasing drop rate. Happy to do more statistical tests.
Won't do statistics on this set as it is late now. Statistics on their own will be less impressive as the ones on Cals dataset. But combining them makes an even stronger case.
Being suspicious I also checked if you truly get more (it should be 5% total) if you team up. This test showed the overall highest drop rate of all tests. So this seems to work.
There was an irregularity and I might have found a way to misuse this. Will be contacting the developers directly as I don't like to discuss possible unknown exploits openly here.
Raw data attached below.
Teseins Teseins Theodum Theodum Team
Alcolyte Battle Focus 1 1 0 1 0
Apprentice Staff 1 2 1 1 0
Apprentice Wand 0 0 0 0 2
Bag Bitter 0 1 0 2 1
Bag Dangerous 1 3 1 3 1
Bag Dreamy 0 0 2 2 0
Bag Fiery 1 1 0 2 0
Bag Healthy 3 2 1 0 0
Bag Itchy 1 1 2 3 3
Bag Smelly 1 0 1 1 0
Bag Sticky 1 1 0 1 0
Broken Goblin Weapons 0 1 0 3 2
Copper drops 17 22 14 19 18*
Goblin Armor Scraps 0 1 0 2 4
Introductory Holy Symbol 0 0 1 0 0
Introductory Trophy Charm 1 0 0 0 1
Introductory Rogue Kit 0 1 0 0 1
Introductory Spellbook 0 0 1 1 2
Lesser Awareness 0 3 0 0 3
Lesser Curing 1 1 1 1 3
Lesser Dodging 4 4 4 5 9
Lesser Freedom 5 2 4 2 3
Lesser Mind Blanking 1 2 3 0 1
Lesser Parrying 0 1 2 2 0
Lesser Riposting 3 5 3 4 7
Lesser Striking 6 5 3 4 4
Maneuver 0 0 0 1 1
Recipe 0 1 2 0 3
Pot Steel Plate 0 0 0 0 1
Quiet Iron Shirt 0 0 0 0 1
Pine Light Shield 1 0 0 0 1
Runespun Robes 1 1 0 0 0
Steel Battlaxe 0 0 1 0 0
Steel Dagger 1 1 0 0 2
Steel Great Sword 0 2 0 0 1
Steel Light Mace 0 1 0 0 0
Steel Longsword 0 0 1 0 0
Steel Short sword 0 0 0 1 1
Steel Spear 0 2 1 0 1
Stylish Padded Armour 0 0 0 1 0
Unresearched Chant 0 1 0 0 0
Unresearched Journal Entry 0 0 0 1 0
Unresearched Prophecy 0 1 0 0 0
51 70 49 63 59
Coppper total 102 144 79 103 107
6.0 6.5 5.6 5.4
*estimated drop - based on average 6 - likely higher but I lost track inbetween with 2 char active at the same time as copper drops have to be noted every single time as you get between 1 and 9 cp per goblin.
Goblinworks Executive Founder
|
Rough estimate of how much to update:
P(A|B)=P(B|A)*P(A)/(P(B|~A)P(~A)+P(B|A)P(A))
P(Knowledge makes drops worse, given that a small sample shows more drops for a low knowledge) is equal to P(a small sample shows more drops for a low knowledge, given that knowledge makes drops worse) divided by P(a small sample shows more drops for a low knowledge) times the prior probability that knowledge makes drops worse.
The key factor there is the P(B|A)/P(B); the update fraction.
P(B|A) is about .505, assuming that the hypothesis is that knowledge works exactly opposite the way it should. P(B|~A) is about .495,
the maximum update fraction is .505/.495, or about a 1.02 multiplier to the prior probability that knowledge works opposite the way it should.
When there have been 10 iterations of this experiment on this scale, I will have increased my belief that there is a major bug making knowledge work opposite the way it should by 22% of my original estimate, or an increase from about 1 in 100 to less than 1.22 in 100.
Right now you've convinced me to change my position from 99% sure that it was programmed correctly (because about 99.9% of things programmed with clear test cases work correctly, as well as 90% of things without one, and there's a 90% chance that this programing was done with clear test cases) to ~98.7% chance that the loot drop system is working essentially as designed.
|
Goblin Squad Member
|
Okay - so 99.9% should work as intended.
Please give me an explanaten then of the two following 0.1% cases
Reactive feats mirroring back on attacker (changed after input from Spitfire)
Growth rate of resources not fit for purpose (changed after input from me)
P(X) = P(A)*P(B)
0.1% times 0.1% equals a 1 in a million chance that this should happen.
Now I make it
P(X) = P(A)*P(B)*P(C)
You certainly have proven me wrong as the chance that three algorithms don't work is intended would now be 1 in a billion.
One of us should start playing in the lottery.
PS: I send you some non-public data as well. Interested in your explanation.
PPS: I would start with a 90% threshold and not 99.9%. That is much more in line with what we see so far. So how much would I have changed your mind if you started at 90%
|
Goblin Squad Member
|
If P(a) P(b) and ( c) are absolutely independant, yes; but …
P(a) * (P(b|a)) * (P(c|a,b) but also further terms of P(a|c) and P(a|b) and P(b|c) which I can not actually remember
P(a) * P(b) * P© only works if there is no cross product.
Take heavy, green , and glows.
Not all heavy things have even distribution of green color (maybe heavy things are more likely to be grey to black) or glowing (maybe light things glow more),
Not all green things are evenly spread over weight (green things may average light) nor do green things average glow (green things are more likely to glow),
et cetera
|
Goblin Squad Member
|
And enetring into this equation enters the value-weight that Devs have assigned to stuff. If the drop rate is designed around everything being of equal value the equations above is relevant ... But if just one thing that is dropped is assigned a value differing from 1 everything get skewed and ugly non-linearites enters (as I assume those are in the upper end creating a exponential effect of Knowledges).
This together with my suspicion of a cap (hard or soft) on the droptable for goblins make me asking if these tests are really valid to make assumptions from.
The data is of course interesting for the Devs that know these parameters making the data collection valuable indeed, but I recommend a bit of caution in stating arguments based on only these facts.
(I do a lot of multi-variate analysis of cases not very different from these ... Hmmm next week I should perhaps run this through a PCA ...)
|
Goblin Squad Member
|
If P(a) P(b) and ( c) are absolutely independant, yes; but …
Oh - I'm fully aware that they are not independent. So the 1 in a billion is just a misuse of numbers - but so is the 99.9%.
Because it isn't independent is why I give the hypothesis that it is wrong a much higher base probability as Decius.
Some unpublished data:
In my tag team experiment char A got 87 copper pieces and char B got 20 copper pieces - which is 81.8 vs 19.2%. The formula should be 50:50 or maybe 51:49 as one had a higher skill.
Incidentially I just calculated drop count and it mirros the numbers at 81.3% for one of the two.
There are some flaws in my experimental setup and I would do it differently if I would do it again. So instead of discussing statistics it might be better to discuss better ways to show issues.
This would be likely one area where botters could help getting good data. Unfortunately I don't want to encourage their use or give them insights in possible issues that could be exploited.
|
Goblin Squad Member
|
Some unpublished data:
In my tag team experiment char A got 87 copper pieces and char B got 20 copper pieces - which is 81.8 vs 19.2%. The formula should be 50:50 or maybe 51:49 as one had a higher skill.
Incidentially I just calculated drop count and it mirros the numbers at 81.3% for one of the two.
That assumption is only valid if the copper drop is independent of the rest of the droppings, and I'm pretty sure it isn't.
|
Goblin Squad Member
|
I despair ...
ALL drops should be equally distributed in a team. This shouldn't be up for discussion. The drop should be independent from who is doing the killing blow, who is first in the team or other variables like that.
Putting multiple events into one bucket is where I take liberties. I justify this to generate numbers big enough to generate some statistics.
I can split it further and further - like 22 to 8 lesser tokens.
But I already seee patterns there that certain tokens are more common as others.
But if you have
0.1% to get A
1.2% to get B
0.2% to get C
...
0.3% to get Z
Then assuming that
Knowledge will increase all chances in the same way
In a team you should have the same chances (apart of knowledge)
you should see a signal in the sum of all drops more easily as in individual ones. I can't do this 1 million times - but I shouldn't have to.
Anyhow - if I'm wrong then not because of the statistics but because of human error. My book keeping is part manual. I do my best to keep it all correct. But this is the most likely reason my assumptions are wrong.