digitalmars.D - mir.random - my GSoC project
- Seb (67/67) Apr 23 2016 Hi all,
- Joseph Rushton Wakeling (54/62) Apr 23 2016 This is a very welcome contribution. Thank you for stepping up
- Martin Nowak (4/6) Jun 02 2016 I just happen to need a gaussian random number generator right now.
- Edwin van Leeuwen (3/9) Jun 02 2016 I tend to use rNorm from dstats:
- Seb (8/14) Jun 02 2016 My focus for the first six week is the transformed density
- HaraldZealot (11/17) Jun 03 2016 Is good workaround you can use:
Hi all, I am very proud to be selected as a GSoC stipend for the D foundation. Most of you already know me from github ( wilzbach) and IRC (greenify). In my GSoC project I will contribute to Dlang’s upcoming numerical library mir [1]. You probably heard about mir from the new ndslice [2] package in Phobos. It's also the development & testing spot for future math additions to Phobos. Ilya is working very hard to get more functionality to mir and he will also be my mentor for the mir.random package. [1] https://github.com/DlangScience/mir [2] http://dlang.org/phobos/std_experimental_ndslice.html mir.random ---------- This project is about adding non-uniform random generators to mir and hopefully eventually to Phobos. While it is intended to be similar in terms of functionality to C++’s <random> [3] and NumPy’s random [4], our main focus is it's performance. Hence I will do a lot of literature research. A simple example of achieving better performance is the normal distribution. In most implementations I looked at (<random> [5], NumPy [6]) the Box-Muller transform [7] is used, however there exists a newer, faster approach: the Ziggurat method [8, 9] which is about three to four times faster [9]. Moreover I plan to add a universal random generator to allow easy creation of arbitrary random distributions. It could be something like the Tinflex algorithm [10], but I still have to do more literature research on this topic. [3] <random> http://en.cppreference.com/w/cpp/numeric/random [4] numpy.random http://docs.scipy.org/doc/numpy/reference/routines.random.html [5] normal in <random> https://github.com/llvm-mirror/libcxx/blob/master/include/random#L4312 [6] normal in NumPy https://github.com/numpy/numpy/blob/master/numpy/random/mtrand/ istributions.c#L106 https://github.com/numpy/numpy/blob/master/numpy/random/mtrand/randomkit.c#L610 [7] Box-Muller transform https://en.m.wikipedia.org/wiki/Box%E2%80%93Muller_transform [8] The Ziggurat Method for Generating Random Variables http://www.jstatsoft.org/v05/i08/paper [9] An Improved Ziggurat Method to Generate Normal Random Samples http://www.doornik.com/research/ziggurat.pdf [10] Tinflex https://cran.r-project.org/web/packages/Tinflex/Tinflex.pdf Stay in touch ------------------ We will occasionally post updates to this newsgroup, but you can also follow us on Twitter ( libmir) [11] for more updates and for more general news DlangScience [12] is tweeting too! Of course you can also directly watch us on Github [13]. For discussions and questions, you are cordially invited to our Gitter chat room [14]. During the GSoC I will also regularly post articles to my blog [15] - it offers email, rss and atom subscription. Shortly before the GSoC starts, I will post the final time schedule here for tracking. As mentioned mir is quite young, so contributions are very welcome. Cheers, Seb PS: I will also be at dconf in Berlin, so maybe we can have a chat there :) [11] https://twitter.com/libmir [12] https://twitter.com/dlangscience [13] https://github.com/libmir/mir [14] https://gitter.im/libmir/public [15] https://seb.wilzba.ch
Apr 23 2016
Joseph Rushton Wakeling <joseph.wakeling webdrake.net> On Saturday, 23 April 2016 at 14:17:19 UTC, Seb wrote:I am very proud to be selected as a GSoC stipend for the D foundation.Congratulations, Seb!This project is about adding non-uniform random generators to mir and hopefully eventually to Phobos.This is a very welcome contribution. Thank you for stepping up to provide it. You obviously already have a very good handle on the literature related to RNG algorithms. What I'd advise, though, is that you also familiarize yourself with the problematic issues related to how random number generation relates to D's range framework. The broader scope of the problem is that both random number generators and random algorithms (i.e. algorithms whose popFront includes a call to an RNG; e.g. randomSample or randomCover) face a number of problems: * if they are not accessed via reference, there are lots of ways in which unintended correlations can result - one suggestion has been to simply disable copy-by- value and force them to be passed by ref or pointer, but in my experience that places some nasty limits on how readily they can be integrated into e.g. UFCS chains of range functionality - another option is to implement them as reference types; however, this creates some challenges w.r.t. memory allocation and the cost of creating multiple instances of e.g. a random algorithm in the inner loops of a program * if pseudo-RNGs are implemented as forward ranges, then again, many unintended correlations can be generated, in this case because library functionality will freely use the .save method to copy range state - this is simpler to address; just make all RNGs and random algorithms input ranges, and implement (say) a 'dup' method for pseudo-RNGs that the programmer can call when they're really sure they want to duplicate RNG state * more an aesthetic issue than a practical one, but note that typical range design (where the initial state of the .front property is determined upon construction) maybe sits a little oddly with random ranges, where the values ought ideally to be _truly_ lazy in their generation Some of this is touched on in my DConf talk from last year: https://www.youtube.com/watch?v=QdMdH7WX2ew (... which says something about the relative busyness of my time since then, that I haven't been able to make much progress on it ...) Note that I'm not suggesting you need to find a solution to the above issues (although it would be cool if you did:-), but just to be aware of them in order to understand how to offer good guidance on the usage of the functionality you develop.As mentioned mir is quite young, so contributions are very welcome.Minor aside: I think that was an odd choice of project name, given it was already the name of a very well known free software project addressing completely different interests ;-)PS: I will also be at dconf in Berlin, so maybe we can have a chat there :)Great, looking forward to it. :-) Good luck & best wishes, -- Joe
Apr 23 2016
On 04/23/2016 04:17 PM, Seb wrote:This project is about adding non-uniform random generators to mir and hopefully eventually to Phobos.I just happen to need a gaussian random number generator right now. Is there already some WIP code, or would you have an intermediate recommendation?
Jun 02 2016
On Thursday, 2 June 2016 at 10:56:36 UTC, Martin Nowak wrote:On 04/23/2016 04:17 PM, Seb wrote:I tend to use rNorm from dstats: https://github.com/dsimcha/dstats/blob/master/source/dstats/random.d#L266This project is about adding non-uniform random generators to mir and hopefully eventually to Phobos.I just happen to need a gaussian random number generator right now. Is there already some WIP code, or would you have an intermediate recommendation?
Jun 02 2016
On Thursday, 2 June 2016 at 10:56:36 UTC, Martin Nowak wrote:On 04/23/2016 04:17 PM, Seb wrote:My focus for the first six week is the transformed density rejection with inflection points algorithm (being able to generate any distribution based on it's CDF). Unfortunately that won't help you, but yeah there is WIP https://github.com/libmir/mir/pull/222This project is about adding non-uniform random generators to mir and hopefully eventually to Phobos.I just happen to need a gaussian random number generator right now. Is there already some WIP codeor would you have an intermediate recommendation?Yep I can also recommend the NumPy-port dstats: https://github.com/DlangScience/dstats
Jun 02 2016
On Thursday, 2 June 2016 at 10:56:36 UTC, Martin Nowak wrote:On 04/23/2016 04:17 PM, Seb wrote:Is good workaround you can use: ```d real normalrnd(real mu, real sigma) { import std.random: uniform; import std.mathspecial: normalDistributionInverse; return mu + sigma * normalDistributionInverse(uniform(0.0L, 1.0L)); } ```This project is about adding non-uniform random generators to mir and hopefully eventually to Phobos.I just happen to need a gaussian random number generator right now. Is there already some WIP code, or would you have an intermediate recommendation?
Jun 03 2016