Just so you know back

(L) [2008/01/30] [toxie] [Just so you know] Wayback!

after a (too) loooong time, finally, here is the link to my thesis: [LINK http://vts.uni-ulm.de/doc.asp?id=6265]
comments, improvements, flames are appreciated!!
(and yeah, there is actually lots more information in it then published in the few RT-papers [SMILEY ;)])

(L) [2008/01/30] [Phantom] [Just so you know] Wayback!

Awesome! That's a huge doc btw. [SMILEY :)] Downloading it now.

(L) [2008/01/30] [toxie] [Just so you know] Wayback!

yup, did not want to compress the pictures, so full quality..

(L) [2008/01/30] [Phantom] [Just so you know] Wayback!

Aw man, that's going to be a lot of reading stuff... I thought you where going to focus on this a-priori memory bound stuff, but this is much better (sorry if I underestimate your effort in that area). Looks like an instant classic to me. [SMILEY :)]

(L) [2008/01/30] [davepermen] [Just so you know] Wayback!

hm.. i'd love to test out the quake2 implementation.. any chance to get hand on it? [SMILEY :P]
nice paper.. haven't really read it, as i'm at work. just looked trough it.
but the 70mb download over the proxy was a bit huge [SMILEY :)]

(L) [2008/01/30] [toxie] [Just so you know] Wayback!

i would love to give away the Q2-executable! but i doubt that this would be okay with the GPL??
cause i cannot give away the underlying RTRT source.. [SMILEY :(]

(L) [2008/01/30] [greenhybrid] [Just so you know] Wayback!

>> toxie wrote:i would love to give away the Q2-executable! but i doubt that this would be okay with the GPL??
cause i cannot give away the underlying RTRT source..
if you modify the source code then you have to distribute it with your executable, or at least make the source available somewhere / somehow (you may raise some money for it) + you are not allowed to change the license to a non-copylefted one (you will find gpl-compatible licenses on [LINK http://gnu.org/]).
edit: i am not sure if pm includes private use, but then and if so i want a GNU/Linux port at least [SMILEY ;)]
edit 10: oh, and of course thank you very much for your thesis!

(L) [2008/01/30] [tbp] [Just so you know] Wayback!

A RT paper that actually deals with Reality v1 and its gory details? Precision? Robustness? Constraints? That simply cannot be.
Please gently hand back your academic license.

(L) [2008/01/30] [derethor] [Just so you know] Wayback!

awesome work! I have something interesting to read tonight [SMILEY :)]
I have a deep respect for the work of a.keller and his qmc papers. you have a great teacher!

(L) [2008/01/31] [dshankar] [Just so you know] Wayback!

Would it be possible for one of the people who have successfully downloaded the thesis to upload it to a filesharing site that is faster? Or is that not allowed? Sorry for asking but I'm getting between 4 and 10 kbps download.

(L) [2008/01/31] [lycium] [Just so you know] Wayback!

did a diagonal read of the thesis just now, it's really excellent; the the markov chain quasi monte carlo methods in particular are very interesting, i'm going to have to study that in detail soon [SMILEY :)]
congrats mate, that's really a job well done!

ps. @dshankar, i downloaded it at 150KBs from that server (from faraway new zealand).

(L) [2008/01/31] [fpsunflower] [Just so you know] Wayback!

Thanks for sharing! Congrats on finishing your thesis [SMILEY :)]

Is the (t.m,s) net max/min distance paper (GHSK07) available online anywhere ?

(L) [2008/01/31] [toxie] [Just so you know] Wayback!

i fear that it's not available yet, but you could ask alex keller directly to receive a copy of it..
(and thanx for the positive feedback so far! [SMILEY :)])

(L) [2008/01/31] [derethor] [Just so you know] Wayback!

well, i got the inspiration to test some ideas in my architecture to speed up the multidimensional sample generation (I am thinking about an optimized memory manager to reuse buffers of the multidimensional length). And your work will be a valuable source of documentation (among a montecarlo couse that someone posted here few days ago)

(L) [2008/02/01] [slartybartfast] [Just so you know] Wayback!

OMG - that's not a thesis, it's a BOOK !!  [SMILEY :shock:]
Just scanned through it - looks very comprehensive and seems to cover all aspects of current thinking on the subject. Now I just need to find time to actually read and digest some of it .....  [SMILEY :roll:]
Well done.

(L) [2008/02/01] [derethor] [Just so you know] Wayback!

I am interested in the "Simultaneous Simulation of Markov Chains" chapter. First, i would like to full understand the concept behind the Array-RQMC.
Looking about it (on a great montecarlo book found here in the must read thread), i see this definition:
Suppose that we want to generate N steps of a Markov Chain X defined
over an ordered space, and such that d uniform numbers are required to generate
the next state of the chain given the current state. Instead of using an
s = Nd-dimensional point set to perform this type of simulation – assigning
each nd-dimensional point to one path of the chain – the idea of array-RQMC
is to use N i.i.d. randomized copies of a d-dimensional point set for each step
of the chain. Furthermore, at each step, the order in which the points are assigned
to the chain paths is determined by the current state. That is, we can
think of Array-RQMC as using r = N underlying point sets in the same way
as Latin Supercube Sampling, but where the permutations πl for l = 1, . . . , r
are determined in the following way.
[...follows a math method that i am unable to copy here]
I never read about them... but i found some references on internet (i dont have the book with the original paper)
* [http://www.iro.umontreal.ca/~lecuyer/myftp/papers/arrayrqmc.pdf A Randomized Quasi-Monte Carlo Simulation Method for Markov Chains]
* [http://portal.acm.org/citation.cfm?id=1225280 Rare events, splitting, and quasi-Monte Carlo]
* [http://www.irisa.fr/armor/lesmembres/Tuffin/Publis/mcqmc04-array.pdf Randomized Quasi-Monte Carlo Simulation of Markov Chains with an Ordered State Space]
* [http://www.math.uwaterloo.ca/~clemieux/papers/book.pdf Monte Carlo and Quasi-Monte Carlo Sampling]
* [http://www.cgg.cvut.cz/members/havran/phdthesis.html Heuristic Ray Shooting Algorithms]
* [http://vts.uni-ulm.de/query/longview.meta.asp?document_id=6265 Quasi-Monte Carlo Light Transport Simulation by Efficient Ray Tracing]
any brief explanation would be nice [SMILEY :)]

(L) [2008/02/01] [toxie] [Just so you know] Wayback!

the basic idea is the following:
as QMC looses any advantages for high dimensional markov chain simulations (in our case = the multiple bounces necessary during light transport to get indirect lighting effects) when compared to MC, one can re-introduce this advantage by sorting the states (in our case the photons or whatever you would call it) by some order before doing the transitions.
the problem so far (i.e. the papers you mentioned): why does this actually work and how can the sorting be used/improved/optimized? what kind of sequences are better suited than others? etcetc.
my answers [SMILEY ;)]: cause MC and QMC samples are basically designed to work on [0,1)^s. For markov chains, this is basically only the case in the first simulation step if you're lucky (in our case we are, as sampling the light source(s) and emission characteristic is pretty much [0,1)^4). In later steps, the states are actually scattered all over the place, but you use samples designed to solve a problem on [0,1)^s (in our case the problem can be decomposed into [0,1)^2 = sampling the hemisphere of each state). By sorting the states, you group neighboring states close together and so (ideally) "get the case of sampling [0,1) again" (at least when neighboring states are very very close together).
So the uniformity of the samples can become useful again (i.e. subsequence properties of (t,s) sequences) for neighboring states. To put it very very simple for our scenario: Neighboring states should sample different directions of the hemisphere instead of doing very similar stuff.
this was a very short and very braindead summary, but maybe it helps that you get the point of the whole thingie.