apgsearch v2.2

[thunk]

I strongly oppose the addition of badges on Catagolue and here's why. I am one of the staff on another site that has badges, an MLPFIM imageboard called Derpibooru. While some of the badges there are rather reasonable, a few are just plain ridiculous, such as the "Derpibooru Premium" badge which is nothing more than a trolling device (where a fake object is presented to you that purports to give privileges).

I don't think that's seriously going to be the case. conwaylife.com, and by extension, apgsearch, is well below one monkeysphere in size, and the badges are just small nods on userpages (and not forum space). As part of another, slightly larger, distributed computing project with badges/achievements of this sort, it's obvious nobody really pays attention to them that much. As long as this doesn't turn into Achievement Unlocked or something, it'll be fine.

The sensible badges there are awarded for actual human effort. Here on Catagolue, the work isn't being done by you, it's being done by the SHA-256 hash and the random number generator, both of which are beyond your control. So if one of the soups you ran contains a very interesting thing, you can't claim to have discovered it. For the same reasons, the person whose machine "found" an object first should not be recorded. Catagolue must be a census of soups and nothing more. 16-by-16 sounds too small now; we should try 32-by-32 soups.

Given >90% of new discoveries in CGoL are made through other automated search programs, it would be rather hypocritical to claim that apgsearch discoveries don't count. Someone ends up having to donate computer time, electricity, and other resources to find stuff with the program.

Also, I'm not sure why you're so hung up on larger soup sizes, but they really don't seem to make that much difference. Soups will either expand to a larger size anyway, collapse into small stable ash, or die out. In terms of censusing, it's irrelvevant. Though I will admit that allowing the apgnano code to handle 32x32 initial setups will make deep symmetric searches possible (which seem more productive in terms of finding new objects).

[Freywa]

Given >90% of new discoveries in CGoL are made through other automated search programs, it would be rather hypocritical to claim that apgsearch discoveries don't count. Someone ends up having to donate computer time, electricity, and other resources to find stuff with the program.

You can count discoveries made by Bellman and CatForce because you have to manually input the search parameters. With apgsearch it's just compile-and-run with absolutely no control over soup generation, so how can that count?

I don't think that's seriously going to be the case. conwaylife.com, and by extension, apgsearch, is well below one monkeysphere in size, and the badges are just small nods on userpages (and not forum space). As part of another, slightly larger, distributed computing project with badges/achievements of this sort, it's obvious nobody really pays attention to them that much. As long as this doesn't turn into Achievement Unlocked or something, it'll be fine.

The sad truth is that turning participation into a game only leads to a horrendous society. Derpibooru, in my opinion, is one of the least rational sites on the Internet, partly because of anonymity and partly because the people there are chasing for near-impossible badges. Then again, I was once part of the crowd-sourced protein folding game Foldit, which also has an achievement system, which led me into much insanity and social troubles there (I have Asperger syndrome and I didn't know at that time how to control it). Even more critically, the Chinese government (not Taiwanese) is going to assign a public score to each and every of its 1.4 billion citizens based on how well they obey the terrible regime there – a doublethink situation already.

We cannot let this corruption happen anywhere, not even on a database of experimental mathematics.

[M. I. Wright]

I don't think that's seriously going to be the case. conwaylife.com, and by extension, apgsearch, is well below one monkeysphere in size, and the badges are just small nods on userpages (and not forum space). As part of another, slightly larger, distributed computing project with badges/achievements of this sort, it's obvious nobody really pays attention to them that much. As long as this doesn't turn into Achievement Unlocked or something, it'll be fine.

The sad truth is that turning participation into a game only leads to a horrendous society. Derpibooru, in my opinion, is one of the least rational sites on the Internet, partly because of anonymity and partly because the people there are chasing for near-impossible badges.

Reread the post you quoted. Seems like you're taking all this wayyy too seriously.

Then again, I was once part of the crowd-sourced protein folding game Foldit, which also has an achievement system, which led me into much insanity and social troubles there (I have Asperger syndrome and I didn't know at that time how to control it).

FoldIt has a really mature and well-developed community, so I don't see (a) how it supports your claim about awards leading to 'horrendous societies', and (b) how you even managed to get in trouble with them in the first place.

Even more critically, the Chinese government (not Taiwanese) is going to assign a public score to each and every of its 1.4 billion citizens based on how well they obey the terrible regime there – a doublethink situation already.

We cannot let this corruption happen anywhere, not even on a database of experimental mathematics.

You lost me... I just don't see how how a fun way to track peoples' participation in a game is in any way comparable to China's situation right now. (An exaggeration, of course - I understand what you're getting at, but it's just so absurd that I can't see how it would transfer to apgsearch.)

[Apple Bottom]

Another idea for a badge:

• Centurion: discovered 100 objects.

But a badge or two for folks who haven't (yet) discovered a lot would be good, too. Reward effort, and dangle carrots to encourage further contributions!

[Alexey_Nigin]

You can count discoveries made by Bellman and CatForce because you have to manually input the search parameters. With apgsearch it's just compile-and-run with absolutely no control over soup generation, so how can that count?

Disagree. If one is going to do an effective search of some weird rules, then he or she should go through the following list of tasks:

• Find some potentially fruitful rules, maybe by analysing and comparing databases like this one.
• Make a list of potentially interesting symmetries for each rule.
• Start the search for the first rule/symmetry combination.
• Regularly check the state of the search so as to finish it at just the right moment: too late, and you are wasting time and money; too soon, and you don't find the most interesting things.
• Publish the most interesting results.
• Remove that rule/symmetry combination from the list and goto 3.

Would you argue that the steps I marked with italic require deep knowledge of CA theory as well as good intuition?

On the other hand, if one is going to use apgnano, where everything is hardcoded, then the aforementioned challenges are no longer relevant. However, throwing one computer at b3s23/C1 is not a particularly good way to really help the search. One should find other ways, which include using supercomputers (like Tom Rokicki, right?), secretly getting control over the school computer lab (Brett Berger), or doing something else (Adam himself).

In all cases there is clear human effort involved.

--------------------------------

Even more critically, the Chinese government (not Taiwanese) is going to assign a public score to each and every of its 1.4 billion citizens based on how well they obey the terrible regime there – a doublethink situation already.

I don't see a connection between the current state of Catagolue and that situation, but OK, let's go to extremes.

The hypothetical extreme situation is, roughly speaking, the Matrix. There is a large group of people who don't see the reality behind the relaxing fabric of badges and achievments. They supply the small group of, in our case, people who live in, and thoroughly understand, real Life. That small group analyses the results of the soup search and moves the CA theory forward. Both groups are happy. Drawbacks?

Well, here are the potential ones.

The sad truth is that turning participation into a game only leads to a horrendous society.

The truth is that, in our hypothetical situation, the folks within the Matrix are not a part of the society.

It's unfair that all those people are not told the truth of Life!

Well, let's be egoistic. Most of people in the world cannot understand CA, because most of people cannot even understand how to work with fractions (own experience). Most of people who can understand CA are smart enough to compare lots of different distributed computing projects and choose one that seems most appealing, which, given the abundance of such projects, will probably not be apgsearch.

The corollary is that if we want to make a large project, we will in any case have to involve people not understanding what's going on. There are no variants better than the Matrix.

--------------------------------

To conclude, I agree that my hypothetical situation is stupid. Its probability is definitely negligible, although I do not dare not compute its exact value for that I not a gematr. However, it was indeed the worst consequence of badges on Catagolue I could think of. This serves to demonstrate that Adam is going in the right direction.

And if Morpheus is insulted, tell him to take it easy.

[gameoflifeboy]

One trillion soups!
https://catagolue.appspot.com/statistics
And if I'm not mistaken, the trillionth soup was the 13071604th soup from this haul.

[thunk]

25 trillion objects (billion if you prefer the long scale).

We now have an Astronomically Large Muck and Ash collection. Keep souping, everybody.

[Kazyan]

25 trillion objects (billion if you prefer the long scale).

We now have an Astronomically Large Muck and Ash collection. Keep souping, everybody.

I think we have enough blocks to build our secret volcano lair now, but let's collect a few hundred billion more just to make sure.

[biggiemac]

Woot! I'm surprised I'm still the largest slice of the pie, Adam has been closing in slowly but surely..

Any word on efforts to a) support other B3S23 symmetries or b) incorporate GPU parallelization?

[calcyman]

Woot! I'm surprised I'm still the largest slice of the pie, Adam has been closing in slowly but surely..

Do you intend to stay in the lead (that is to say, resume your hundred unparallelised instances and/or 64-core HPC in order to match my best efforts to overtake)?

Any word on efforts to [...] incorporate GPU parallelization?

I need to learn CUDA first, whence I'll commandeer one of Amazon's g2.2xlarge processors (8 CPU cores and 1536 CUDA cores) and begin experimenting. Initial calculations estimate that this will be able to do roughly 100000 soups per second once I've rewritten the performance-intensive parts of apgsearch to target the GPU. I can imagine that the end result will be something like:

The CPU will spend nearly all of its time handling switch-engines

which are already an embarrassing slowdown (currently accounting for around 10% of apgnano runtime!).

I suspect that the amount of effort required to refactor everything for the GPU is comparable to that involved in the transition from apgsearch 1.x to 2.x (maybe slightly less, since all of the non-critical code can just be copied over from the C++ version). Certainly this will be written from scratch as a version 3.x designed exclusively for GPU-enabled machines. As before, Catagolue's peer-reviewing system means that it will be automatically tested against the previous version (as was the case when 2.x was released).

I should probably sort out the horrible switch-engine slowdown in the comfort of v2.x before embarking on this mission, though. For something that processes 1000 soups per second, it's almost laughable that a block-laying switch-engine costs 2 seconds and a glider-producing switch-engine costs 10 seconds. There's probably something incredibly stupid happening in the code somewhere, so I should really gprof it with an infinite-growth seed and examine the results.

[biggiemac]

Woot! I'm surprised I'm still the largest slice of the pie, Adam has been closing in slowly but surely..

Do you intend to stay in the lead (that is to say, resume your hundred unparallelised instances and/or 64-core HPC in order to match my best efforts to overtake)?

I have 16 cores from the HPC running on nice mode, and I think over holiday break I'll also start the unparallelized instances. It just seems that during class sessions they are getting shut down by other users after not much progress, so it hasn't felt worth the half hour of logging in and starting up.

[Apple Bottom]

Initial calculations estimate that this will be able to do roughly 100000 soups per second once I've rewritten the performance-intensive parts of apgsearch to target the GPU.

100,000 soups per seconds, that's crazy (in a good way). I take it that's an estimate for a typical off-the-shelf consumer GPU, right?

Certainly this will be written from scratch as a version 3.x designed exclusively for GPU-enabled machines.

At this rate 4.x will be running on custom-designed ASIC hardware.

[calcyman]

Initial calculations estimate that this will be able to do roughly 100000 soups per second once I've rewritten the performance-intensive parts of apgsearch to target the GPU.

100,000 soups per seconds, that's crazy (in a good way). I take it that's an estimate for a typical off-the-shelf consumer GPU, right?

It's my estimate for the Kepler GK104 included with the Amazon g2.2xlarge machines, which I believe is comparable to modern graphics cards.

[praosylen]

Certainly this will be written from scratch as a version 3.x designed exclusively for GPU-enabled machines.

At this rate 4.x will be running on custom-designed ASIC hardware.

That would be fun! 1,000,000,000 soups per second!

[calcyman]

I take it that's an estimate for a typical off-the-shelf consumer GPU, right?

Ahh, it appears that the GPU included with my g2.2xlarge instance had a launch price of $3600, so maybe exceeds 'typical off-the-shelf consumer GPU'...

http://www.techpowerup.com/gpudb/2312/grid-k520.html

On a more positive note, I've tested the machine (it works; see https://catagolue.appspot.com/haul/b3s2 ... e14b34345c for example) and installed the nVidia drivers, so should be able to begin experimenting with CUDA after sleeping.

[Scorbie]

Speaking of amazon, has anybody experimented apgnano with amazon free tier instances? When I did it before I remember that it seems to exceed the free tier limit. Just wondering if anybody else tried.

[biggiemac]

I tried so far as to realize I needed the Linux executable to be able to run it on t2.micro but to make the Linux executable I would need to either make it there or copy a build from a different Linux computer (I use windows). After copying and chmod so that I could run it on the free instance I hit some library error and gave up.

It rations CPU time so I believe you can have one instance running nonstop and you'll be within free tier.. unless they don't enforce the rationing, so by running an intense program you use more than your allotted 20% or so of the CPU, meaning you get more than 1 t2.micro-CPU-second per second. Which would be lame given their description of allowed free tier usage.

[biggiemac]

And Adam is now the largest slice of the pie! Well, at the 11-12 23:00 update, at least, we both tie to 3 decimal places. 28.207%. Thought it was neat.

[Apple Bottom]

It's my estimate for the Kepler GK104 included with the Amazon g2.2xlarge machines, which I believe is comparable to modern graphics cards.

Ahh, it appears that the GPU included with my g2.2xlarge instance had a launch price of $3600, so maybe exceeds 'typical off-the-shelf consumer GPU'...

http://www.techpowerup.com/gpudb/2312/grid-k520.html

Fair enough, though the GK104 apparently also powers many consumer cards, like my own GTX 660 Ti. Makes you wonder what the price differential stems from; I hope it's not reliability/durability under 24/7 load, since I'd rather not be running apgsearch 3.x only to find out I don't have a GPU anymore some day. I kinda need it.

And Adam is now the largest slice of the pie! Well, at the 11-12 23:00 update, at least, we both tie to 3 decimal places. 28.207%. Thought it was neat.

Indeed, he's ahead of you now, at 7,444,408,256,840 vs. 7,444,204,086,137 objects contributed, giving him a 0.001 percentage point edge. Congrats!

(I remember how when I first started, you had contributed in excess of 80% of the census's objects. Things sure have changed! And then there's Tom Rokicki's meteoric rise. That kid's gonna go far!)

[danieldb]

Why is LongLife's page broken?

http://catagolue.appspot.com/census/b345s5/C1

[velcrorex]

Why is LongLife's page broken?

http://catagolue.appspot.com/census/b345s5/C1

I think that's what happens when the census upload is too big. Try running fewer soups per upload. Someone correct me if I'm wrong.

[Billabob]

10^5 hauls uploaded!

[Saka]

When I try to do command "make" on cygwin, I get:

Code: Select all

$ make
make: *** No targets specified and no makefile found. Stop

and when I try make apgnano:

Code: Select all

$ make apgnano
make: Nothing to be done for 'apgnano'.

When I try ./apgnano:

Code: Select all

$ ./apgnano
-bash: ./apgnano: Is a directory

So what should I do?

[Scorbie]

You have to run make in the directory where the makefile is in. You seem to be running Cygwin outside the apgnano directory. Go into apgnano directory, navigate to where the source file is in (e.g. if it's in apgnano do "cd apgnano". If it's in apgnano/src, do "cd apgnano/src" ...), and then you can run make.

[SuperJedi224]

I'm getting this error at compile-time:

life128.h183: error: unknown register name ‘xmm14’ in ‘asm’
life128.h183: error: unknown register name ‘xmm13’ in ‘asm’
life128.h183: error: unknown register name ‘xmm12’ in ‘asm’
life128.h183: error: unknown register name ‘xmm11’ in ‘asm’
life128.h183: error: unknown register name ‘xmm10’ in ‘asm’
life128.h183: error: unknown register name ‘xmm9’ in ‘asm’
life128.h183: error: unknown register name ‘xmm8’ in ‘asm’
life128.h183: error: unknown register name ‘xmm7’ in ‘asm’
life128.h183: error: unknown register name ‘xmm6’ in ‘asm’
life128.h183: error: unknown register name ‘xmm5’ in ‘asm’
life128.h183: error: unknown register name ‘xmm4’ in ‘asm’
life128.h183: error: unknown register name ‘xmm3’ in ‘asm’
life128.h183: error: unknown register name ‘xmm2’ in ‘asm’
life128.h183: error: unknown register name ‘xmm1’ in ‘asm’
life128.h183: error: unknown register name ‘xmm0’ in ‘asm’