PSO is quite closely related to genetic algorithms, and also Population Monte Carlo type of methods. In a computer simulation it probably doesn't matter all that much whether one "moves" existing particles in the search space or "generates" new ones based on the previous ones.
The choice of the objective function (for example, some kind of Bayesian posterior probability) is surely more important, and unless there is something very special in the structure of the optimization problem, I think that it's a bit silly to publish papers which test various optimization methods for a given problem and conclude that some of them perform better than the others. The results are typically dependent on a multitude of factors that cannot be controlled in the experiments (a significant one is the amount of effort one spends in finding the best optimization parameters), and do no progress science all that much, particularily if one does not have a hypothesis, and preferably also some analysis, _why_ some of the optimization methods are better than others.
By the way, here's a neat paper on the use of PMC in the global illumination (radiosity/raytracing/...) problem.
Not really that reasonable. If you look at the results of current optical flow and disparity estimation algorithms, they're really not that great. Discontinuities of the image (edges) are a huge problem, as is the whole top-down/bottom-up/gestalt-ordeal, and these have not been solved in any satisfactory manner.
To reconstruct the 3D scene generating the 2D images is effectively to solve vision, in its entirety. In real time, no less. So I would guess that they're doing something quite simple. I'd love to see it, but the information on the site is quite scarce. I'm just hoping that someone is not manually pulling the strings behind the scenes.
On the flipside - Megablocks are garbage - They are cheaper, but they do not stay together at all
Supposedly, Lego has a patent on plastic pieces that stick together with the little knobs on the top. Now, Megablocks and others do not infringe on that patent because, well, you know.:)
(disclaimer: my background in dynamical systems, much less this particular problem, is not that strong)
The second part of Hilbert's 16th problem deals with limit cycles, the way things will go on eventually in dynamical systems if they are not disturbed externally. The subproblem 2/3 of this problem (it's the indexing that makes math complicated..) asks if there exists an upper bound on the number of different limit cycles one can have in the system.
Oxenhielm attacks the problem by considering first a special case called the Lienard equation and approximating its solution by harmonic oscillation. The proof begins: "Noticing that the state variable x of the Lienard equation (1) behaves approximately like a sine function in simulations (see Fig.1),we assume -- in order to make a good approximation of x -- that both state variables are dominated by a harmonic term...." Now, to my engineer's eyes, the functions in Fig.1 seem more like triangular waves, with definitely more than one single frequency component. Yet the accuracy of the approximation has not been considered at all in the paper. Also, 'proof by looking at results of simulations' is not really valid if you don't have any other evidence.
Another bad part is on page 6, where it is claimed that "Note that the method of describing functions may be used in a similar manner as in the proof above,to find the upper bounds for the Hilbert number in any planar polynomial vectorfield. Thus, it is possible to completely solve the second part of Hilbert's 16th problem by using this approach." Wait a minute, how did that happen? What if the harmonic approximation fails on other than Lienard equations? It might just work, I have no idea, but this assertment hardly proves the fact.
Note however that this is very different from Andrew Wiles' proof of the Fermat conjencture. While very few people in the world could understand the odd-hundred pages of Wiles' proof, Oxenhielm's paper is just eight pages of much more accessible mathematics.
But I have a paper in the review process myself, and sure as hell would hate to see nonqualified people discussing its validity publicly, so maybe I'll just shut up now:)
I live in an old apartment built in the 1950s, and it has a cool cabient in the kitchen which is cooled by outdoor air. I put my computer there. Of course it's only useful in the winters, but then it provides enough cool air so that I can close the cabinet door without fearing overheating. The monitor, mouse and keyboard cables are about 5m each and run neatly inside cable ducting. With the door closed the computer is much quieter than my fridge, which is not that noisy either.
mean-time-to-failure of a hard drive is 15,000 to 20,000 hours
No, much more than that. A 120 GB WD Caviar for example is rated for 500 000 hours. While these MTBF figures do not mean that a single drive would last that long, it gives an estimate of the relative failure rate of new drives. When the drive gets older, it fill fail more often, but in the beginning of its life, the failure rate is much lower than your figures would suggest. So Google hardware guys are changing the failed disks only once a day, not once every hour. And when the drives do get old, they're better off changing also working drives for new ones.
They are the same equation in different form. You can derive one from the other. Although the latter one is admittedtly nicer, because you can use it for particles with no rest mass (m0=0, v=c), when the first one gives just 0/0.
I have some very strong memories connected to smells, such as grinding metal (I grew up near a large construction site). I also sometimes smell in my dreams. Smell memories come to me often like flashbacks, like the one you describe. I guess this has to do with smell processing being more primitive and less conscious function of the brain.
While what you're saying has been true traditionally, it seems to me that more and more game creators are (re-)discovering that it's the physics and structure of real life that give rise to the best appearences.
That's only one side of the issue. For large 3D objects, you're correct. Billboarding a la Doom is definitely a fawing trend. But appearance models are very popular in current 3D and not going away anytime soon.
Think of texturing as an appearance model. Simulating the actual phenomenon of light hitting individual molecules is very heavy, but that's what you're going to need if you're trying to solve the true properties of a real-world material instead of just modeling its appearance.
The same goes for using programmable texture for modeling fur, for example. Modeling the individual hairs one by one with polygons is computationally much more intensive, and the results aren't necessarily better, unless you're doing raytracing to account for the scattered light in the fur. Computers must become a lot faster (millions, billions, gazillions) before appearance models are going away.
mathematical calculation of a reality is more expensive than the reality itself
That may well be true. Luckily humans do not perceive the reality directly, so most of the information contained in it is lost, and modeling just what is perceivable (i.e. appearance) continues to be a justifiable approach.
I do this too. Apparently your field is much unlike mine.
I work in computer vision research. I do all my rendering work on a P4-Geforce3 machine. We have SGIs too but PCs give a lot more bang for the buck (as a govermental organization we don't have that much to spend on equipment) and have more or less equal performance.
Second, no one wants to re-write the bulk of the code out there that runs really well on SGI's
What rewrites? My plain-C OpenGL code works just fine on both.
a researcher won't likely have his own sysadmin
I know some labs which are like that, but luckily mine isn't one of them.
it is now possible to walk around with an uncalibrated, handheld camera, and completely automatically get a decent 3-D model out of it
There are a lot of restrictions for scenes that can be modeled this way with current structure from motion algorithms. Occlusions usually cause severe problems, because the error model for the correspondence points does not take occlusions into account.
Structure from motion is a low level method which operates with pixels, not objects. In order to deal with occlusions, higher level object models need to be integrated somehow into the system, but this is far from easy. I'd like to see any references of such systems.
Re:Thinking about emergence
on
Emergence
·
· Score: 1
I think of emergence in terms of complex behavior resulting from simple rules
This is a nice way of thinking emergence, and a very useful one even in plain old classical physics. For example, the macroscopic properties of matter, such as temperature, are emergent in the sense that a lower level description of the system can be discarded. Temperature is not really a property of any of the system's parts, but one of the system as a whole. Temperature, pressure and such are a result of the collective behaviour of the molecules, and in systems where the interaction between the parts (molecules) is significant, the resulting behaviour is complex.
I think of emergence in terms of "the whole is greater than its parts"
I don't really like this phrase, or the variant "the whole is greater than the sum of its parts". The one I mention is more evil because it uses a very specific mathematical concept "sum" in a very vague manner. But it is clear that if I take the parts of a car, for example, and stuff them into a box, they're not a car. They are the parts of a car. I don't see what's the deep philosophical insight in a phrase like this. It's just stated in a mysterious way to conceal the fact that nothing has really been said.
On the other hand, "The whole is equal to its parts and their interactions" sounds a lot less flashy.
Re:Referrer tells you who's following the links
on
Emergence
·
· Score: 1
You can see the IP just fine in the logs, so you do know who has followed the link, though not by the referrer field, you're right there.
But the point is that you never know the links that exist, only the ones that are followed. In this sense linking is one-directional.
Searching google for links is another matter, but it's not a "glaring error" to claim that linking is one-directional. Debatable is a more appropriate word.
We would be literally playing god. The species died off because nature intended it to( even if it was hunted to death we are still a part of a larger cycle)
Huh? If we kill off a species, we're just a part of the nature. If we try to revive it (by whatever means), we're suddenly playing god. Where's the logic? I can't see any difference between killing and reviving in this respect. In either case we're stirring the balance in the ecosystem, which is bad for our own survival as a species.
We're a part of the nature, and the nature does not intend anything. Our actions cannot be justified simply as "evolution in action".
Bill Hicks once said something like "Christianity would be a great religion if someone would actually practice it." I guess the same goes for Islam too.
5:69 Those who believe (in the Qur'an), those who follow the Jewish (scriptures), and the Sabians and the Christians,- any who believe in Allah and the Last Day, and work righteousness,- on them shall be no fear, nor shall they grieve.
Why would you want to reencode the files? That will just make them sound worse. It's not like all your audio files must be in a single format. I switched over to.ogg a few months ago, mainly because it promised better quality for the same size. All my older encoding is mp3. I can play both kinds just fine with xmms and Winamp.
Adding.ogg support to hardware players is a no-brainer for the manufacturers. The code is free and with just a bit more ROM they can add another acronym to the feature lists. A lot of the early adopters are geeks who will choose an ogg-and-mp3 player over a plain-mp3 one any day. Or at least I know I will.
OPL3 is 4 operators & 4(?) waveforms FM synthesis, I believe the same as in Yamaha's DX9, DX21, DX100 etc. (DX7 has 6 op, but only sine waveform. OPL2 is 2 op, no wonder it sounds bad)
The DX100 has been popular in jungle style basses. It's a cute synth with 4 octaves of mini sized keys. However, FM sounds become often very thin when the frequency ratio between the carrier and modulator is large. The spectrum has whole bands which lack harmonics completely. This is why a lot of people dislike FM synthesis. It's good for basses (where you don't need much high frequency content anyway) and metallic sounds like bells.
It's not that bad a resolution. The card does interpolation between the voxels too, so you're not seeing just a bunch of boxes on the screen.
I don't think you'll easily notice the voxelation (sp?) of reasonably smooth objects on a 1024x768 screen as the voxels are only few pixels apart from each other.
Slashdot Mirror
PSO is quite closely related to genetic algorithms, and also Population Monte Carlo type of methods. In a computer simulation it probably doesn't matter all that much whether one "moves" existing particles in the search space or "generates" new ones based on the previous ones.
The choice of the objective function (for example, some kind of Bayesian posterior probability) is surely more important, and unless there is something very special in the structure of the optimization problem, I think that it's a bit silly to publish papers which test various optimization methods for a given problem and conclude that some of them perform better than the others. The results are typically dependent on a multitude of factors that cannot be controlled in the experiments (a significant one is the amount of effort one spends in finding the best optimization parameters), and do no progress science all that much, particularily if one does not have a hypothesis, and preferably also some analysis, _why_ some of the optimization methods are better than others.
By the way, here's a neat paper on the use of PMC in the global illumination (radiosity/raytracing/...) problem.
Not really that reasonable. If you look at the results of current optical flow and disparity estimation algorithms, they're really not that great. Discontinuities of the image (edges) are a huge problem, as is the whole top-down/bottom-up/gestalt-ordeal, and these have not been solved in any satisfactory manner.
To reconstruct the 3D scene generating the 2D images is effectively to solve vision, in its entirety. In real time, no less. So I would guess that they're doing something quite simple. I'd love to see it, but the information on the site is quite scarce. I'm just hoping that someone is not manually pulling the strings behind the scenes.
Playboy for Thorne and Private Eye for Hawking, says here.
On the flipside - Megablocks are garbage - They are cheaper, but they do not stay together at all
:)
Supposedly, Lego has a patent on plastic pieces that stick together with the little knobs on the top. Now, Megablocks and others do not infringe on that patent because, well, you know.
(disclaimer: my background in dynamical systems, much less this particular problem, is not that strong)
...."
:)
The second part of Hilbert's 16th problem deals with limit cycles, the way things will go on eventually in dynamical systems if they are not disturbed externally. The subproblem 2/3 of this problem (it's the indexing that makes math complicated..) asks if there exists an upper bound on the number of different limit cycles one can have in the system.
Oxenhielm attacks the problem by considering first a special case called the Lienard equation and approximating its solution by harmonic oscillation. The proof begins: "Noticing that the state variable x of the Lienard equation (1) behaves approximately like a sine function in simulations (see Fig.1),we assume -- in order to make a good approximation of x -- that both state variables are dominated by a harmonic term
Now, to my engineer's eyes, the functions in Fig.1 seem more like triangular waves, with definitely more than one single frequency component. Yet the accuracy of the approximation has not been considered at all in the paper. Also, 'proof by looking at results of simulations' is not really valid if you don't have any other evidence.
Another bad part is on page 6, where it is claimed that "Note that the method of describing functions may be used in a similar manner as in the proof above,to find the upper bounds for the Hilbert number in any planar polynomial vectorfield. Thus, it is possible to completely solve the second part of Hilbert's 16th problem by using this approach."
Wait a minute, how did that happen? What if the harmonic approximation fails on other than Lienard equations? It might just work, I have no idea, but this assertment hardly proves the fact.
Note however that this is very different from Andrew Wiles' proof of the Fermat conjencture. While very few people in the world could understand the odd-hundred pages of Wiles' proof, Oxenhielm's paper is just eight pages of much more accessible mathematics.
But I have a paper in the review process myself, and sure as hell would hate to see nonqualified people discussing its validity publicly, so maybe I'll just shut up now
I live in an old apartment built in the 1950s, and it has a cool cabient in the kitchen which is cooled by outdoor air. I put my computer there. Of course it's only useful in the winters, but then it provides enough cool air so that I can close the cabinet door without fearing overheating. The monitor, mouse and keyboard cables are about 5m each and run neatly inside cable ducting. With the door closed the computer is much quieter than my fridge, which is not that noisy either.
"Lies, damned lies and statistics. . ." is a quote from Mark Twain
Actually, Twain attributes the quote to Benjamin Disraeli in his autobiography.
mean-time-to-failure of a hard drive is 15,000 to 20,000 hours
No, much more than that. A 120 GB WD Caviar for example is rated for 500 000 hours. While these MTBF figures do not mean that a single drive would last that long, it gives an estimate of the relative failure rate of new drives. When the drive gets older, it fill fail more often, but in the beginning of its life, the failure rate is much lower than your figures would suggest. So Google hardware guys are changing the failed disks only once a day, not once every hour. And when the drives do get old, they're better off changing also working drives for new ones.
Try also Yaneer Bar-Yam's "Dynamics of Complex Systems". (it's freely available)
You might have more success with these origami masterpieces.
.
If not, well, at least you've done something with your hands
(caution: not work-safe)
>> E = m_0 c^2 / sqrt(1 - v^2/c^2)
> E^2 = c^2 * p^2 + m_0^2 * c^4
They are the same equation in different form. You can derive one from the other. Although the latter one is admittedtly nicer, because you can use it for particles with no rest mass (m0=0, v=c), when the first one gives just 0/0.
I have some very strong memories connected to smells, such as grinding metal (I grew up near a large construction site). I also sometimes smell in my dreams. Smell memories come to me often like flashbacks, like the one you describe. I guess this has to do with smell processing being more primitive and less conscious function of the brain.
While what you're saying has been true traditionally, it seems to me that more and more game creators are (re-)discovering that it's the physics and structure of real life that give rise to the best appearences.
That's only one side of the issue. For large 3D objects, you're correct. Billboarding a la Doom is definitely a fawing trend. But appearance models are very popular in current 3D and not going away anytime soon.
Think of texturing as an appearance model. Simulating the actual phenomenon of light hitting individual molecules is very heavy, but that's what you're going to need if you're trying to solve the true properties of a real-world material instead of just modeling its appearance.
The same goes for using programmable texture for modeling fur, for example. Modeling the individual hairs one by one with polygons is computationally much more intensive, and the results aren't necessarily better, unless you're doing raytracing to account for the scattered light in the fur. Computers must become a lot faster (millions, billions, gazillions) before appearance models are going away.
mathematical calculation of a reality is more expensive than the reality itself
That may well be true. Luckily humans do not perceive the reality directly, so most of the information contained in it is lost, and modeling just what is perceivable (i.e. appearance) continues to be a justifiable approach.
I do this too. Apparently your field is much unlike mine.
I work in computer vision research. I do all my rendering work on a P4-Geforce3 machine. We have SGIs too but PCs give a lot more bang for the buck (as a govermental organization we don't have that much to spend on equipment) and have more or less equal performance.
Second, no one wants to re-write the bulk of the code out there that runs really well on SGI's
What rewrites? My plain-C OpenGL code works just fine on both.
a researcher won't likely have his own sysadmin
I know some labs which are like that, but luckily mine isn't one of them.
There are a lot of restrictions for scenes that can be modeled this way with current structure from motion algorithms. Occlusions usually cause severe problems, because the error model for the correspondence points does not take occlusions into account.
Structure from motion is a low level method which operates with pixels, not objects. In order to deal with occlusions, higher level object models need to be integrated somehow into the system, but this is far from easy. I'd like to see any references of such systems.
I think of emergence in terms of complex behavior resulting from simple rules
This is a nice way of thinking emergence, and a very useful one even in plain old classical physics. For example, the macroscopic properties of matter, such as temperature, are emergent in the sense that a lower level description of the system can be discarded. Temperature is not really a property of any of the system's parts, but one of the system as a whole. Temperature, pressure and such are a result of the collective behaviour of the molecules, and in systems where the interaction between the parts (molecules) is significant, the resulting behaviour is complex.
I think of emergence in terms of "the whole is greater than its parts"
I don't really like this phrase, or the variant "the whole is greater than the sum of its parts". The one I mention is more evil because it uses a very specific mathematical concept "sum" in a very vague manner. But it is clear that if I take the parts of a car, for example, and stuff them into a box, they're not a car. They are the parts of a car. I don't see what's the deep philosophical insight in a phrase like this. It's just stated in a mysterious way to conceal the fact that nothing has really been said.
On the other hand, "The whole is equal to its parts and their interactions" sounds a lot less flashy.
You can see the IP just fine in the logs, so you do know who has followed the link, though not by the referrer field, you're right there.
But the point is that you never know the links that exist, only the ones that are followed. In this sense linking is one-directional.
Searching google for links is another matter, but it's not a "glaring error" to claim that linking is one-directional. Debatable is a more appropriate word.
We would be literally playing god. The species died off because nature intended it to( even if it was hunted to death we are still a part of a larger cycle)
Huh? If we kill off a species, we're just a part of the nature. If we try to revive it (by whatever means), we're suddenly playing god. Where's the logic? I can't see any difference between killing and reviving in this respect. In either case we're stirring the balance in the ecosystem, which is bad for our own survival as a species.
We're a part of the nature, and the nature does not intend anything. Our actions cannot be justified simply as "evolution in action".
When American Standard releases a new toilet, are we gonna hear about it here first?
If it has an integrated web camera, then yes.
*ducks*
And the tv shows.
Bill Hicks once said something like "Christianity would be a great religion if someone would actually practice it." I guess the same goes for Islam too.
5:69 Those who believe (in the Qur'an), those who follow the Jewish (scriptures), and the Sabians and the Christians,- any who believe in Allah and the Last Day, and work righteousness,- on them shall be no fear, nor shall they grieve.
Why would you want to reencode the files? That will just make them sound worse. It's not like all your audio files must be in a single format. I switched over to .ogg a few months ago, mainly because it promised better quality for the same size. All my older encoding is mp3. I can play both kinds just fine with xmms and Winamp.
.ogg support to hardware players is a no-brainer for the manufacturers. The code is free and with just a bit more ROM they can add another acronym to the feature lists. A lot of the early adopters are geeks who will choose an ogg-and-mp3 player over a plain-mp3 one any day. Or at least I know I will.
Adding
I found this today on the Analogue Heaven mailing list:
Paraller port SID synthesizer
The schematics are kinda sketchy but it's a neat idea. The guy who built it coded also a patch editor and step sequencer for the thing.
OPL3 is 4 operators & 4(?) waveforms FM synthesis, I believe the same as in Yamaha's DX9, DX21, DX100 etc. (DX7 has 6 op, but only sine waveform. OPL2 is 2 op, no wonder it sounds bad)
The DX100 has been popular in jungle style basses. It's a cute synth with 4 octaves of mini sized keys. However, FM sounds become often very thin when the frequency ratio between the carrier and modulator is large. The spectrum has whole bands which lack harmonics completely. This is why a lot of people dislike FM synthesis. It's good for basses (where you don't need much high frequency content anyway) and metallic sounds like bells.
Only 256^3 volume? Yuck.
It's not that bad a resolution. The card does interpolation between the voxels too, so you're not seeing just a bunch of boxes on the screen. I don't think you'll easily notice the voxelation (sp?) of reasonably smooth objects on a 1024x768 screen as the voxels are only few pixels apart from each other.