Random number generation?

[Chud]

Is the random number generation in Angband just the standard rand() from the C library?

A couple of things make me wonder... one from cheating, yes, it's true; I copied a save file for a character with 0 HP (note: still alive...) and poisoned... and since she had only one healing option, the actions from run to run were always the same (hit, miss, hit, breathe, etc) because my (healing) action was always the same. If the seed to rand() is stored in the save file, that makes sense.

The other thing that makes me wonder is that sometimes spell failure rates seem... oddly distributed, shall we say. Like, 45% failure fails 7 times in a row. Sure, it could happen... but it makes you go "hmmm..." :-)

http://www.random.org/analysis/dilbert.jpg

[Derakon]

Angband uses a custom RNG, and yes, the RNG state is saved to prevent savefile scumming. Someone did an analysis years ago comparing it to the Mersenne Twister, and it had comparable levels of randomness as far as I can recall. That said, it does seem odd how often I fail a 15% spell chance several times in a row, doesn't it?

[d_m]

Angband uses a custom RNG, and yes, the RNG state is saved to prevent savefile scumming. Someone did an analysis years ago comparing it to the Mersenne Twister, and it had comparable levels of randomness as far as I can recall. That said, it does seem odd how often I fail a 15% spell chance several times in a row, doesn't it?

If I remember, it's cycle is shorter than Mersenne Twister, but long enough that most people aren't going to notice (particularly with all the different unpredictable things we end up doing with entropy).

[Atarlost]

Not sure why Angband never switched to Mersenne Twister on the principle that nobody ever got fired for buying IBM.

[takkaria]

Angband uses a custom RNG, and yes, the RNG state is saved to prevent savefile scumming. Someone did an analysis years ago comparing it to the Mersenne Twister, and it had comparable levels of randomness as far as I can recall. That said, it does seem odd how often I fail a 15% spell chance several times in a row, doesn't it?

Angband's RNG is far far less random then MT. We could change it, but it always seemed to be part of the character of the game to me.

[fizzix]

Angband's RNG is far far less random then MT. We could change it, but it always seemed to be part of the character of the game to me.

I think with the current display where probability of failing is shown outright, it would be better to reexamine this mindset.

For example, a lot of times I make a calculation assuming independent events. I might say, I have a 5% chance of failing, so my chances of failing twice in a row is 1 in 400. This is an acceptable risk, so I'm going to attempt to cast the spell.

Now if the probability of failing twice in a row was actually 1/100 or even 1/50, I would definitely consider another course of action.

[Chud]

I think with the current display where probability of failing is shown outright, it would be better to reexamine this mindset.

I have to agree here -- especially given that all the hardware Angband is going to be running on these days is so much faster than the earliest versions needed to work with; implementing a computationally expensive (relatively) RNG is no longer a practical problem.

Or maybe it never was; I haven't looked at how fast the various RNG algorithms are these days--maybe this was never an issue, but in any case it can't possibly be today. And I'd really rather that a given percent chance was honestly presented in the game, if it's going to be presented at all. Bugginess to the contrary might add some character to the game, but in my opinion it's not the sort of character the game wants.

[takkaria]

I have to agree here -- especially given that all the hardware Angband is going to be running on these days is so much faster than the earliest versions needed to work with; implementing a computationally expensive (relatively) RNG is no longer a practical problem.

Or maybe it never was; I haven't looked at how fast the various RNG algorithms are these days--maybe this was never an issue, but in any case it can't possibly be today. And I'd really rather that a given percent chance was honestly presented in the game, if it's going to be presented at all. Bugginess to the contrary might add some character to the game, but in my opinion it's not the sort of character the game wants.

Alright, well, does anyone actually object to making the RNG closer to random? If not, we can make it happen...

[Magnate]

Alright, well, does anyone actually object to making the RNG closer to random? If not, we can make it happen...

How much work is it? If it's a simple drop-in replacement for z-rand.c then I'm all for it - but I wouldn't have thought this was a top priority for anybody, since z-rand.c has served us well enough for however many years. But you're the boss .... ;-)

[d_m]

I just received permission from Pierre L'Ecuyer to use his team's WELL1024a.c implementation in Angband. It uses 1024 bits of state and should perform well for us (better than Mersenne Twister).

It will be a change that could break savefiles, however I think it should be possible to detect and update people's saved seed (which has far too few bits).

So to answer Magnate's question, it should not require too much work at all

[buzzkill]

I've no idea what most of that means, but I'm excited about the change.

[Chud]

Someone asked a bit ago "how big a deal is this?" -- not very big, really, but *if* it's not difficult to fix (and it sounds like it's not at all) then I would definitely prefer that the RNG be more really random(-like). Not a huge deal in the big picture... but, just one vote.

[takkaria]

I just received permission from Pierre L'Ecuyer to use his team's WELL1024a.c implementation in Angband. It uses 1024 bits of state and should perform well for us (better than Mersenne Twister).

Can we releease that under the GPL?

[Tiburon Silverflame]

Looks like it.

http://www.iro.umontreal.ca/~panneton/WELLRNG.html

[d_m]

Looks like it.

http://www.iro.umontreal.ca/~panneton/WELLRNG.html

The page you're linking to actually uses a different license in the source files:

/* ***************************************************************************** */
/* Copyright:      Francois Panneton and Pierre L'Ecuyer, University of Montreal */
/*                 Makoto Matsumoto, Hiroshima University                        */
/* Notice:         This code can be used freely for personal, academic,          */
/*                 or non-commercial purposes. For commercial purposes,          */
/*                 please contact P. L'Ecuyer at: lecuyer@iro.UMontreal.ca       */
/* ***************************************************************************** */

I did write the authors and get specific permission to use it in Angband under the GPL, but the files they ended up posting to a different website under the GPL used version 3.

Since I specifically asked and got permission for Angband I am going to include it in my tree, but I'm also going to write to them for clarification and if they say GPL-3 is the only GPL they'll allow then we'll maybe have to do something else.