Curse

tmakinen · Mar 28, 2008, 12:24 PM // 12:24

Quote:

Originally Posted by hallomik

If you have already made a thread with your findings, could you link to it? If you haven't done so, could you start a new thread in Riverside or the Farming subform with the details?

It was a side project that I started around September 2007 to find out the optimal method to farm collectables for Wintersday, and after that it kind of tanked because I got my max title and became interested in other things. I didn't write a detailed report on the findings because there were still some open questions that would have needed more data, including

are exempt items counted towards earnings (would think not but very hard to measure)?
exactly how many past drops are taken into account (the IDS result suggests 3, might be something else)?
how fine or coarse is the waste rate adjustment?
what are the actual values for adjustment thresholds and initial dummy drops as functions of party size?

Since I'm not an ANet dev, I can't say that this is the definite answer. It is a working hypothesis, and the simplest one of those that I'm aware of and which are consistent with the various observations. I won't be able to convince anybody right now, and the earliest time I'll be able to continue the project is towards the end of June this year. Of course, as a scientist I'd be perfectly happy to be disproved as well

Fril Estelin · Mar 28, 2008, 01:58 PM // 13:58

Thanks tmakinen for a lesson in a reverse engineering of a new kind

To really be accurate and conclusive, the study should have been done on totally unrelated accounts, because I guess there's a random factor that can shape things very differently. Who knows even if the various functions used in the algorithms are not randomised, thus preventing function profiling as you're trying to do?

I'd love to hear what Anet devs think of your post

.

SirSausage · Mar 28, 2008, 03:26 PM // 15:26

Quote:

Originally Posted by Fril Estelin

I'd love to hear what Anet devs think of your post

.

I SERIOUSLY doubt Anet will comment this post. They won't take our opinions to heart. Unless they will reply within 3 days. After that period I declare that Anet doesn't give a damn thing about it. The clock's ticking...

MithranArkanere · Mar 28, 2008, 04:27 PM // 16:27

The thing is that there's nothing else they could say that hasn't been said already.

Chthon · Mar 28, 2008, 05:33 PM // 17:33

Quote:

Originally Posted by tmakinen

There is still a lot of confusion about how LS actually works, so here's a rough description of the mechanism:

Every monster has one or more associated pieces of loot that may or may not drop when the monster is killed. If an item is on the LS exempt list, it will always drop. If it isn't there's a random chance of it getting withheld instead of dropping. The probability of that happening is the waste percentage. If the waste percentage of your team is 50% half of your potential LS relevant drops will never materialize.

However, there is an additional factor in play since the waste percentage is not fixed. It has a self-adjusting control mechanism that is designed to make sure that your rate of income (i.e., gold per hour) doesn't exceed a certain limit. The way that feedback loop works is the following: when a drop is pending, the program calculates the ratio between the monetary value of the previous drop that was granted and the time that has elapsed since then, and compares that to certain reference values which are determined by the number of party members and the maximal party size. If the ratio is below the lower threshold, you are earning too little and your waste percentage will be decreased by a fixed fraction. If the ratio is above the upper threshold, you are earning too much, and the waste percentage is increased by a fixed fraction. After the waste percentage has been modified, the program then throws a dice to see whether you get that drop or not. If it passes LS and gets assigned, its monetary value and assignment time are written down for the next call of the algorithm.

How does this then affect the actual game? If you kill several monsters at the same time with some AoE spell, the drops are not processed in parallel but sequentially. For the first drop the reference drop has happened some time before and everything works as described. If the drop passes LS something nasty happens. Its value and drop time become the reference values when the next drop in the queue is processed. Now, however, the effective time elapsed since the last drop is zero, and when you divide any finite value by zero you get infinite (in reality, there is a safeguard in the code to give a finite but big number instead to avoid an exception). Since the result is always above the upper threshold, waste is adjusted one notch up. This then happens with every subsequent drop until waste hits a hard cap of 100% and all the rest drops of that mob will be withheld. In practice 3 drops seem to get through regardless of mob size. This is known as the IDS bug but it isn't an actual bug, just a consequence of the way LS is implemented.

There is another little known effect caused by this method of self-adjustment. Since the algorithm needs to know the value and drop time of the previously assigned drop, it must initialize these variables at the creation of the instance. Thus, there is a 'dummy drop' recorded when you enter an explorable. Now, since the waste percentage is modified according to your rate of earning and this dummy value gives you a fake earning rate, your waste percentage will be very high just after entering an explorable, and get progressively better as time passes and you're not earning over the limit. A monster killed about 20 seconds after entering an explorable has a nearly zero probability of dropping LS affected loot but if you wait 2 minutes before killing the first monster the chance of a drop goes up to almost 100% (I actually have data to back this up).

This description is necessarily very simplified but should give you a basic idea of what is happening to your loot and why.

This sounds very plausible, but how do you account for the well-recorded empirical observations in this thread?

Kyp Jade · Mar 28, 2008, 05:41 PM // 17:41

MMMM, i have some cake, im going to eat it too

Lootscaling sucks tho, now my only monies comes from the HoH chest

IlikeGW · Mar 28, 2008, 06:00 PM // 18:00

Quote:

Originally Posted by tmakinen

It has a self-adjusting control mechanism that is designed to make sure that your rate of income (i.e., gold per hour) doesn't exceed a certain limit.

Nothing about that jives with my experience playing... I mean the norm is more likely someone in a group gets way more good loot than others, some people get nothing, and a few get in the middle... so if they're trying to balance drops the system is so poor it's unnoticeable imo.

MithranArkanere · Mar 28, 2008, 06:05 PM // 18:05

It's all look. When you play with other players, unless you have any kind of consensus to share loot, it's like playing with H/H. What others get you never see.

Since playing with others (many times) is easier than playing solo, and usually faster, you get more drops by killing faster in teams.
Since it's a multiplayer game, it's just logical to encourage multiplaying in many ways like that. While still allowing solo playing... just without encouraging solo.

Of course, there are other ways to encorage parties... like, for example, making bosses drop one item for each party member, like with reward chests.

Loviatar · Mar 28, 2008, 06:06 PM // 18:06

Quote:

Originally Posted by Chthon

This sounds very plausible, but how do you account for the well-recorded empirical observations in this thread?

i truly wish Ensign (mr super stat) were here to answer that.

basically he would say that is such a statistically small sample relative to the entire playerbase playing that it was worthless to base anything on

tmakinen · Mar 28, 2008, 06:24 PM // 18:24

Here's a graph demonstrating the entry effect, one of the best clues to the inner workings of Loot Scaling:

The rectangles show the probability of the first killed monster dropping something as a function of time elapsed since entering the explorable. The data were collected with a 1/4 party (solo) by entering Diessa Lowlands NM from Nolani Academy. Errors are 1-sigma.

I should collect more data with higher temporal resolution to see whether there is any granularity in the probability function that would indicate some particular adjustment method. The entry effect is ideal for experiments because you start with a known and reproducible 'clean' state.

About the sync farming experiment, I've treated it extensively elsewhere. Very shortly put, matching LS exempt drops show that the two players have indeed entered two separate instances with identical seeds. However, if LS is not adjusted on the fly based on the killing rate, the rest of the drops should be identical as well. Since they are not, LS is adjusted through player actions. The small amount of difference suggest that the thresholds for adjusting LS define a rather large range of non-adjustment and are only triggered by extreme values in either direction.

mr_groovy · Mar 28, 2008, 06:49 PM // 18:49

Quote:

Originally Posted by tmakinen

Very technical stuff I could never write.

It would be nice if one of the devs could comment on your findings

Esan · Mar 28, 2008, 07:38 PM // 19:38

Devs don't comment on guru. Ask them on wiki.guildwars.com, but even there I doubt they will confirm or deny anything.

Bocjo Bassannn · Mar 28, 2008, 08:01 PM // 20:01

Hero faction farm.... 150 gold per min... no where else in the game can you make gold that fast (without getting a lucky item drop) Lutgardis Conservatory (factions) is where this is done from....

Breakdown is like this if your farming the Kurzick title track

10k faction is 25 runs thats 3750 gold per 10k faction... takes about 25 to 30 mins for 10k faction.

To max this title is 10 mill faction donated... that is 25thousand runs.....

that is 3 mill 750k gold farmed...

Personaly I think this farm right here is just one more set of evidence that loot scaleing seriously needs to change. When you can make more gold killing NOTHING......something is seriously wrong with the drop rate of cold hard cash.

Allow me to ad I dont find anything wrong with this hero faction farm it is a nice thing for those like me that can not stand alliance battles... But... come on I've totaly stoped killing baddies for loot because quite frankly .... this is more profitable thanks to loot scaleing...

Chthon · Mar 28, 2008, 08:06 PM // 20:06

Quote:

Originally Posted by tmakinen

About the sync farming experiment, I've treated it extensively elsewhere.

Link, please?

Quote:

Very shortly put, matching LS exempt drops show that the two players have indeed entered two separate instances with identical seeds. However, if LS is not adjusted on the fly based on the killing rate, the rest of the drops should be identical as well. Since they are not, LS is adjusted through player actions. The small amount of difference suggest that the thresholds for adjusting LS define a rather large range of non-adjustment and are only triggered by extreme values in either direction.

Or, in other words, this feedback mechanism (if it exists as described) is so weak that you'd really have to try hard to get a non-trivial effect out of it.

Anyway, I don't think whether or not Loot Scale has a subtle feedback mechanism or not is particularly germane to the question of whether it should be retired if/when the RTM companies are driven off. It's the baseline loot reduction (what you'd call the default waste ratio for 1-man parties) that bothers people.

Esan · Mar 28, 2008, 08:12 PM // 20:12

Quote:

Originally Posted by Bocjo Bassannn

To max this title is 10 mill faction donated... that is 25thousand runs.....

that is 3 mill 750k gold farmed...

Divide by 2 if you are donating to your alliance. 1,875k gold in total, assuming you can handle the murderous rage the 250 odd hours of grinding the same mindless flags will put you in. And 30 minutes for 25 runs is a tad optimistic.

Also observe that your gold earning rate is 7.5k/h. Nothing to get ecstatic over. You can do better farming raptors in HM even with LS.

Fay Vert · Mar 28, 2008, 08:28 PM // 20:28

Quote:

Originally Posted by Esan

And 30 minutes for 25 runs is a tad optimistic.

Understatement. Its cloud bonkers IMO, you are doing very well if you get 15 runs in

people always exagerate about their farming runs, more e-peen I guess

Cab Tastic · Mar 28, 2008, 09:09 PM // 21:09

Quote:

Originally Posted by tmakinen

There is still a lot of confusion about how LS actually works, so here's a rough description of the mechanism:

Every monster has one or more associated pieces of loot that may or may not drop when the monster is killed. If an item is on the LS exempt list, it will always drop. If it isn't there's a random chance of it getting withheld instead of dropping. The probability of that happening is the waste percentage. If the waste percentage of your team is 50% half of your potential LS relevant drops will never materialize.

However, there is an additional factor in play since the waste percentage is not fixed. It has a self-adjusting control mechanism that is designed to make sure that your rate of income (i.e., gold per hour) doesn't exceed a certain limit. The way that feedback loop works is the following: when a drop is pending, the program calculates the ratio between the monetary value of the previous drop that was granted and the time that has elapsed since then, and compares that to certain reference values which are determined by the number of party members and the maximal party size. If the ratio is below the lower threshold, you are earning too little and your waste percentage will be decreased by a fixed fraction. If the ratio is above the upper threshold, you are earning too much, and the waste percentage is increased by a fixed fraction. After the waste percentage has been modified, the program then throws a dice to see whether you get that drop or not. If it passes LS and gets assigned, its monetary value and assignment time are written down for the next call of the algorithm.

How does this then affect the actual game? If you kill several monsters at the same time with some AoE spell, the drops are not processed in parallel but sequentially. For the first drop the reference drop has happened some time before and everything works as described. If the drop passes LS something nasty happens. Its value and drop time become the reference values when the next drop in the queue is processed. Now, however, the effective time elapsed since the last drop is zero, and when you divide any finite value by zero you get infinite (in reality, there is a safeguard in the code to give a finite but big number instead to avoid an exception). Since the result is always above the upper threshold, waste is adjusted one notch up. This then happens with every subsequent drop until waste hits a hard cap of 100% and all the rest drops of that mob will be withheld. In practice 3 drops seem to get through regardless of mob size. This is known as the IDS bug but it isn't an actual bug, just a consequence of the way LS is implemented.

There is another little known effect caused by this method of self-adjustment. Since the algorithm needs to know the value and drop time of the previously assigned drop, it must initialize these variables at the creation of the instance. Thus, there is a 'dummy drop' recorded when you enter an explorable. Now, since the waste percentage is modified according to your rate of earning and this dummy value gives you a fake earning rate, your waste percentage will be very high just after entering an explorable, and get progressively better as time passes and you're not earning over the limit. A monster killed about 20 seconds after entering an explorable has a nearly zero probability of dropping LS affected loot but if you wait 2 minutes before killing the first monster the chance of a drop goes up to almost 100% (I actually have data to back this up).

This description is necessarily very simplified but should give you a basic idea of what is happening to your loot and why.

One of the best examples I came across for the IDS was with the Turai procession farm.

With this farm you would bascially insta death a group of about 20 - 25 poor margonites.

Before loot scaling you would have a stack of loot filling your inventory on every run. With the advent of LS this was reduced to 1-2 common drops and maybe the odd gold,dye,tome etc.

You may remember that splinter weapon( around which this farm is based) got nerfed and it was no longer possible to insta kill the margonites. The process was now a lot slower but still possible. The margonites now dropped in the region of 8-10 common drops.

This farm alone showed me the effect of IDS.

Fay Vert · Mar 28, 2008, 11:37 PM // 23:37

Which bright spark came up with silly name of Icy Dragon Sword for the RoK bug?

Sleeper Service · Mar 29, 2008, 01:28 AM // 01:28

Quote:

Originally Posted by tmakinen

Here's a graph demonstrating the entry effect, one of the best clues to the inner workings of Loot Scaling:

The rectangles show the probability of the first killed monster dropping something as a function of time elapsed since entering the explorable. The data were collected with a 1/4 party (solo) by entering Diessa Lowlands NM from Nolani Academy. Errors are 1-sigma.

I should collect more data with higher temporal resolution to see whether there is any granularity in the probability function that would indicate some particular adjustment method. The entry effect is ideal for experiments because you start with a known and reproducible 'clean' state.

About the sync farming experiment, I've treated it extensively elsewhere. Very shortly put, matching LS exempt drops show that the two players have indeed entered two separate instances with identical seeds. However, if LS is not adjusted on the fly based on the killing rate, the rest of the drops should be identical as well. Since they are not, LS is adjusted through player actions. The small amount of difference suggest that the thresholds for adjusting LS define a rather large range of non-adjustment and are only triggered by extreme values in either direction.

Maybe i got this wrong but does that graph mean that waiting for 60 seconds before killing anything increases the potential drop rate on the first monster.

are you saying that: If you wait, kill a monster and get a drop, the rest of the instance will benefit from higher drop rates.

JeniM · Mar 29, 2008, 01:48 AM // 01:48

Another fine example of skill>time

I think the numbers in the poll speak for themselves. The majority of the customer base (that have voted, i'm not making a generalisation) want Loot Scaling out, so as that would actually make us happy and is what we have asked Anet for, I somehow bet it doesn't happen.