Salvage is BAD for skillups

[Raederle]

We all know that as skillups get rarer they seem to come in bursts. We'll go 20 30 even 40 combines with nothing then get 2-3 skillups over the next 10 to 15 combines. Thats just the way the random number generator table works. Sure theres the lone skillup here and there but by and large when the table has spikes it has a bunch (high AND low) of them together. Patience and a large bunch of prepped combines was rewarded by once churning through the mass arriving at the 'hot' area where 2 plus skillups were to be had.

Salvage changes all that. Salvage, in checking each of the items used in the combine, goes through additional entries on the random number generator table. That 'hot' area that had extra failures along with extra improves suddenly gets whipped through by an additional 3-6 checked per failure. Instead of 15 entries with 2 plus successes and 2 plus failures the sequence is: failure (5 entries gone one of them a would have been improve) success success failure (5 entries gone, another would have been improve) success improve success failure (another five entries gone) and now youre back into the 40-60 no failures no improves lots of ingredients gone territory again.

I'm 1750 and working on 2100. Been doing trade skills forever. I used to call that long stretch with no failures and no improves the "wide flats" as it went on a long time with no activity, then was punctuated by the volitile area. If I could give back salvage I would. If I could just plain lose salvage, destroying the 15 aa spent, I would. I see very little point in working on the last bunch of points given how salvage hoses my 'good combine' areas.

Meredeth, Amicus, Tunare

[Neebat]

Random Number Generators come in many forms. There's a wide variety of different formulas to back them up. It's actually quite hard to check that a random number generator is behaving in a desirably random way. In the case of EQ, we're interested in an RNG with the following effects:

1. an even distribution of results - As an example, /rand 100 should give each value nearly exactly 1% of the time on a large sample.
2. no predictability - In a large sample, numbers under 50 should be followed by numbers over 50 pretty much exactly have the time.

In fact, computer random number algorithms tend to fail the second criteria fairly often and what you're describing is consistent with that. Sony has check the random number generator several times, but I don't know if they've checked it for that "streaky" behavior.

A reasonable check is to run about a large sample, and track the results. if you're generating numbers in the range 1 to 100, a sample size of 10 million might do. Note how many times each number appears and how many times it follows each other number. On a large enough sample, when a given value turns up, the next number should STILL be evenly distributed.

Also, I don't think there's a table of random numbers stored anywhere, but a streaky RNG can act like that. I know the RNG's for EQ live on the server, and I assume that means the zone server. If so, you can switch RNG's at any time by simply zoning. Having said all that, if the "wide flats" are generally poor results, wouldn't salvage also burn through random numbers faster than they did, getting you to the areas of volitility faster?

To say the RNG is "streaky" is actually equivilient to saying it's behaving in a non-random, predictable way. Any predictability may be a GOOD thing if you can figure out how to know when you're going to get a bad vs good result. If the wide flats are worth avoiding, couldn't you skip them by consuming random numbers with /rand. Once you spot the volitile section, you could switch to your skill up combines to take advantage of it. Even if there's a separate RNG just for tradeskills, you could still do this with cheap, trivial combines, wasting components until the streak passes.

Do you think there's a separate RNG for each person in the zone?

Honestly, if I feel like I'm not skilling up, I go to bed and try another time. I don't track my combines closely enough to notice any of this, but I find the ideas about the RNG very cool to think about.

--
Programmer & major math person

[Phantron]

This reminds me of people who watch someone lose a lot on a slot machine (i.e. long strings of fails) and then expect to hit a 200 or 300 coin thing because there was bad luck before (i.e. skillup).

By that logic if you failed to salvage anything it's eating up some fails that you'd have got.

[Verdandi]

The streaks you are seeing are not a quirk of the design of the RNG. They are a fundamental quality of randomness. Randomness is inherently streaky, including very long stretches. Therefore, the only way salvage could make it less streaky is if salvage made the RNG less random. If anything, according to your theory, it makes it more random.

[Wyvernwill]

Is there any proof to back this?

I hate to be skeptical, but this is posted as if it's proven fact. I can not prove you right or wrong myself. Could it be possible salvage has an impact on tradeskills? Possibly. Who knows. I may be inclined to believe you actually, as I have 0 salvage AA's but in tailoring I still manage to get a skill up in less than 10 combines on average. But then again, I could just be insanely lucky.

[Nolrog]

Wouldn't the same be true of a 'bad spot' in the RNG? Salvage would take you past that bad spot faster, thus allowing you to gain those skill ups. Since the 'bad spots' grossly out number the good spots, then by your logic, salvage should increase the chance of skill ups by getting us over them that much quicker.

[Phantron]

Streaky does not imply non-randomness. Randomness is exactly that: random. If you knew what kind of pattern a random pattern is supposed to have, it's already not random.

The average RNG only guaranteeds an even distribution of numbers over a large number of samples. Some scientific stuff requires more randomness than just that, but it's certainly not needed for simulating a 1/X chance of skilling up.

The problem is that people do not get that unlikely events are very likely if you do something long enough. If the chance of skillup is 1/30 and you make 1000 combines, the chance of failing 200+ in a row at some point is 60%. Now if I make 5 simulated runs of 1000 and 3 of them has a string of 200+ in a row you'll probably say something's got to be messed up or it's fake. Yet statistically this is actually the norm. If you've ever gotten 2 consecutive skillups (I sure have), their chance given the current model is (1/24)^2 = 0.17%. Seems pretty unlikely? But if you do 1000 combines, there's a 82% chance you'll skillup twice in a row (or more). The short way to look at this is, if something only has 1 in 1000 chance of happening but you do it 1000 times, then it becomes very likely this rare event will happen.

[chrisian]

salvage is in fact good for skilling up.
more combines -> more skillups

your perception of patterns in the randomness is just a natural human tendency... we are pattern seeking creatures.

[Neebat]

As some folks have said, over and over, if there's no streaks, it's not random. But the inverse is also true... if there's nothing but streaks, it's not random.

The average RNG only guaranteeds an even distribution of numbers over a large number of samples. Some scientific stuff requires more randomness than just that, but it's certainly not needed for simulating a 1/X chance of skilling up.

Imagine the random generator was asked for a random number between 1 and 10 a million times. Now imagine that the result was 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, repeated 100k times. Now, that has the most basic property of an RNG. The results are evenly distributed, because every number occurs exactly 10% of the time. It's got "Even distribution", but it's "predictable". If you ran the analysis I describe above on that silly stream, you'd find that following a "1" the next number is "2" 100% of the time.

According to Pantron, that's all that matters, except in scientific stuff. Is it good enough for 1/X chance of skilling up? In other words, if we know the RNG just gave us a 1, we could make a better than average guess what we'll get next. If skillups will occur on 1 in 10 combines, then this would give exactly 1 skill up every 10 combines. That is not good enough, even for use in EverQuest.

The RNG in EQ is not that predictable, but I would still love to see the results for EQ's /random. To get a decent sample size, even using /random on a hotkey, would take hours. Then you need a parser for those logs to do the analysis. It's a lot of work.

Would it show predictability? The original poster seems to say it would be predictable and a lot of folks seem to think it wouldn't. I don't know and I'm not totally sure Sony does.

[mewkus]

The only way this whole premise would hold true is if you were the only person generating these random numbers. There are all kinds of random events happening on each of the server computers (each computer serves multiple zones). If someone happens to be engaged in combat in one of the zones that's also housed on the same computer as the one you're on, they will be pulling many more random numbers from the RNG as you are (every swing they take pulls one, every swing the mob takes pulls one, every spell casters cast pulls several numbers).

[Neebat]

The only way this whole premise would hold true is if you were the only person generating these random numbers. There are all kinds of random events happening on each of the server computers (each computer serves multiple zones). If someone happens to be engaged in combat in one of the zones that's also housed on the same computer as the one you're on, they will be pulling many more random numbers from the RNG as you are (every swing they take pulls one, every swing the mob takes pulls one, every spell casters cast pulls several numbers).

This goes back to my above question. How many separately seeded RNG's do you think there are? I suspect there is at least one per zone. Beyond that...

Scope and visibility issues in software design make it tough to have a lot of different code use a single resource like an RNG. I wouldn't be at all surprised if there's a separate RNG for each type of activity, so for instance, /rand pulls from one pool, melee pulls (multiple values per swing) from another pool, spawns from another and tradeskills from it's own pool. If that's the case, and you're the only one in the zone doing tradeskills, then you've got a private RNG. (I'm sometimes the only one in ShadowHaven when I tradeskill. Low lag, high availability of resources.)

It's also possible that your session with the server has a private RNG for everything you do. Same result, you're the only one pulling numbers out.

On the other hand, if there is a single RNG for the zone, it might even be a single RNG for the machine the zone runs on. My private Shadow Haven may well be sharing hardware with PoK, including the RNG.

In fact, none of that matters unless the RNG is actually predictable. I hope it's not. I hope Sony has done tests to be sure, because all these issues make it very hard for a user to run a test. I don't think the test I described is actually the best, just easy to describe. Here's a better one: http://www.fourmilab.ch/random/

Regardless, I think the original poster is definately out on a limb assuming he's pulling from a table of random numbers. We just don't know enough.

[Arfyak Reflux]

A reasonable check is to run about a large sample, and track the results. if you're generating numbers in the range 1 to 100, a sample size of 10 million might do. Note how many times each number appears and how many times it follows each other number. On a large enough sample, when a given value turns up, the next number should STILL be evenly distributed.

The RNG in EQ is not that predictable, but I would still love to see the results for EQ's /random. To get a decent sample size, even using /random on a hotkey, would take hours. Then you need a parser for those logs to do the analysis. It's a lot of work.

Would seem to me that a rather easy macro (but not allowed by SOE) could be set up to run a /random 1 100 hotbutton X amount of times and run the log through Excel (or any spreadsheet prog) with a script to pull the returned number from the appropriate line and then come up with %'s of the returned numbers rather easily. Looking at the variations of the numbers next to each other may take a bit more work, but I think it would be do-able.

[Neebat]

Would seem to me that a rather easy macro (but not allowed by SOE) could be set up to run a /random 1 100 hotbutton X amount of times and run the log through Excel (or any spreadsheet prog) with a script to pull the returned number from the appropriate line and then come up with %'s of the returned numbers rather easily. Looking at the variations of the numbers next to each other may take a bit more work, but I think it would be do-able.

You wouldn't need anything fancy enough to be forbidden. Make a hotkey and jam the button down with a toothpick for a couple hours. You *might* get pretty decent results from this if you've got the RNG to yourself. (An empty, non-instanced zone, would give you the best bet.)

As for the processing, I wouldn't choose Excel, but go for it. Bit more work than I want in my gaming. Also, refer to the tool I noted above. It's better than what I suggested. In that set of tests, the "streakiness" is described as "Serial Correlation"

[Twistagain]

You wouldn't need anything fancy enough to be forbidden. Make a hotkey and jam the button down with a toothpick for a couple hours. You *might* get pretty decent results from this if you've got the RNG to yourself. (An empty, non-instanced zone, would give you the best bet.)

As for the processing, I wouldn't choose Excel, but go for it. Bit more work than I want in my gaming. Also, refer to the tool I noted above. It's better than what I suggested. In that set of tests, the "streakiness" is described as "Serial Correlation"

Excel will only hold 65536 rows, btw...might not be enough to be statistically meaningful. Besides, the easier way to test both the distribution AND the predicatability would be to write a quick app that does both. I'll see if I can get a "large" string of randoms in a log file and do a study. Don't expect a result "soon".

[Phantron]

As some folks have said, over and over, if there's no streaks, it's not random. But the inverse is also true... if there's nothing but streaks, it's not random.



Imagine the random generator was asked for a random number between 1 and 10 a million times. Now imagine that the result was 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, repeated 100k times. Now, that has the most basic property of an RNG. The results are evenly distributed, because every number occurs exactly 10% of the time. It's got "Even distribution", but it's "predictable". If you ran the analysis I describe above on that silly stream, you'd find that following a "1" the next number is "2" 100% of the time.

According to Pantron, that's all that matters, except in scientific stuff. Is it good enough for 1/X chance of skilling up? In other words, if we know the RNG just gave us a 1, we could make a better than average guess what we'll get next. If skillups will occur on 1 in 10 combines, then this would give exactly 1 skill up every 10 combines. That is not good enough, even for use in EverQuest.

The RNG in EQ is not that predictable, but I would still love to see the results for EQ's /random. To get a decent sample size, even using /random on a hotkey, would take hours. Then you need a parser for those logs to do the analysis. It's a lot of work.

Would it show predictability? The original poster seems to say it would be predictable and a lot of folks seem to think it wouldn't. I don't know and I'm not totally sure Sony does.

I don't think it needs to be said that any criteria of randomness obviously requires the sequence to be not predictable, but if you need it spelled out, here it is.

Assuming the sequence of numbers is unpredictable, all that matters is what I said earlier, that the distribution works out 1/X over large samples.