AW: AW: [ODE] numerical stability (sort of)

Marc marcrochel at yahoo.de
Wed May 5 20:47:34 MST 2004


Rehi!

>> I'm interested in this continuing deterministicly problem. A few days 
>> ago, there was a discussion about the use of rand() vs. dRand() in
stepfast.cpp.
>I forgot to say that i also re-seed step fast's RNG when inserting my 
>stuff into the world. Otherwise i wouldn't get deterministic/reproduceable
"behaviour" at all, even with the gcc version.
I assumed u did that for dRand(). But did u do it for rand() as well?

>> Until now, rand() is still present there. So maybe that's at least of 
>> your
>> problems: the joints are getting shuffled differently each run if u 
>> don't reset rand()-seed too. Or u replace rand() by dRand() yourself.
>> That's my approach. I'ld like to see this changed in cvs too.
>I've looked into that RNG when writing this thing, and i also have 
>doubts about its quality :) A mersenne >twister while a tad more expensive
would be a lot nicer. Or an external facility. I'm not sure, but i think the
jitter that RNG introduces is needed by design so totally removing it may
not be an option.

Yes maybe dRand isn't the best thing u can get. But then we should stick
some other RNG in it that's better instead of using rand() at only one
position, poluting all rand() calls in a project using ode. Anyway, I think
ode doesn't need such a good RNG since basicly it's outcome should be
deterministic and it doesn't use monte carlo methods or things like that.

ciao
	Marc



More information about the ODE mailing list