these innovation patents are not such a bad idea at all... First of all, they do not grant the owner a monopoly, secondly it gets the idea into the 'innovation' office, maybe even the patent office.
Is it such a crazy idea to transform this 'innovation patent' into a prior art machine? I mean, a low cost way of getting the idea inside the patent office?
Assuming that the patent office is forced to search such a innovation patent database, this might get some of the more ridiculous quirks out of the system.
Just an idea. We might be better off with some more radical changes instead of smoothing things over a bit though.
And someone else has written a very nice article about how science cannot reject the fact that if God exists, he can interfere with all experiments and make them look like natural. It's actually a serious article.
The article might be serious, the hypothesis in it is simply not scientific as you can not devise an experiment to falsify it. Though it's a bit Popperian, falsifiable hypotheses is what science is all about, mystical hypotheses such as this one and religious stuff is only verifiable (show me this deity that does it and I believe you) and thus not useful in science.
One particular way to circumvent this is to actually _use_ the viral properties of the GPL to release the code. Just include readline or liboctave or whatever in your code that might be moderately useful and then you and the Uni are forced to release it under the GPL or not at all.
This is a great way of being able to take code with you when you hop from position to position. Universities are usually very reactive when it comes to licensing and patenting issues. They don't have a clue what to do with it, but they simply want to 0wn it just in case. This way you can argue that you cannot do that and this gives you the possibility of actually using the code you wrote yourself.
Ok, my problem with gcc stems from the time when I moved from Windows/MSVC to Linux/gcc as my development platform. Porting the code wasn't difficult, what was really bugging me was that compilation time for the project I was working on increased from 30 sec. on MSVC to more than 5 mins. using gcc. This was with full optimizations.
The libary I'm currently working on is heavily templatized and I need at least 256 Mb of RAM to be able to compile it using -O1, let alone -O2. Even with massive amounts of RAM (about 2 Gigs), the compile cycle takes ages. The same library compiles in a breeze using MSVC (factor 20 faster) and to top it off --- MSVC generates faster code!
Luckily we can discount MSVC because it also was routinely capable of generating extremely fast code that was absolutely wrong!
So, yes there is at least one optimizing compiler that is way faster than gcc. Maybe MSVC cut a few corners doing this (plus being far from standard compliant), but it sure seems possible.
About compilation speed, does anybody know whether g++ 3 supports the 'export' keyword now? That would help quite a bit in reducing compilation time.
And on a related note, does it still need approximately 2 Gigs of internal memory and the better part of the weekend to optimize heavily templatized code*?
"We've looked at the pluses and minuses," she said Thursday, and determined that overwriting would protect information on the computers while allowing the machines to be donated.
There must be one of a hell of a group behind this, looking at all those pluses and minuses on the hard drive to see if there's any protected information left!
If they would have put only minuses on the hard drive, the exercise may have been less hard. But then they might have concluded that half of the protected information was still there!
The guys pioneering these kind of techniques have put themselves in a subfield called evolution strategies (which is actually older than GA's). They use real valued vectors as the genetic material and co-evolve both the mutation rates (standard deviations) and the correlations between these rates. The system is then capable of performing directed mutations: all mutations that are correlated will change the original parameter vector together. This works really well on some really hard problems and is capable of outperforming gradient and newton methods for optimization. (this technique is developed in the early seventees, even before GA's become popular.)
In GA's the problem is known as linkage learning. How do you figure out, during a run, what are the important linkages in your problem. Thus how can you keep this conglomorate of genes that code for the eye shape together, while it is dispersed over the chromosome. Your looking for stuff that belongs together and start to combine such building blocks.
A simple GA cannot do this.
Enter the world of 'competent' GA's that are designed to handle this problem. There are now three types of competent GA's developed that are very different from each other. They all are capable of finding conglomerates of genetic material that constitute a building block. Building blocks are defined to be things that have a mostly linear effect on the fitness of the individual. One of these competent GA's actually examines the losers in Minsky's term to figure out what went wrong (so there goes that assertion).
It is quite funny that most people are still using the vanilla GA as it has been proven (by David Goldberg, one of the main figures in the field) that it cannot scale properly. Minsky is obviously using this vanilla GA as his strawman, while the research in GA's has moved on and improved the thing quite a bit the last 10 years.
So, yes, there are a lot of self-adapting genetic algorithms around, they use very different techniques to perform the self-adaptation. It goes from adding rules to changing rates, identifying building blocks (thus setting fuzzy markers), till building full fledged bayesian networks. We're still just scratching the surface, but there has been progress and more difficult problems can be tackled.
Didn't know that, thanx for the info. Just checked it out and it looks quite ok (the exported file when ran through latex looks quite nice:). Never seriously looked at kword before, will do that now. Kudos to them!
Cool, I hope they'll also start supporting importing and exporting to TeX. Maybe then the stuff will start looking professional*. Kidding aside, it's a shame that the word-processing crowd is ignoring the best type-setting system around. WYSIWIG documents just don't cut it compared with a doc prepared in LaTeX.
* professional as in 'professional publisher', not as in 'professional marketeer'
The point is that until we know what's going on in our own heads and how we learn we don't have a chance in hell of getting the same process on a computer.
You should have told that to the brother's Wright, although arguably they didn't produce the same process (a bird).
I don't know much about genetics, though a bit about genetic algorithms, and this is indeed a very important area of research. Lots of research in the GA-field is devoted to finding ways of automatically adapting the adaptation process during the evolution. The goal is indeed to learn from the mistakes (and successes) individuals make.
The basic intuition for this to work is of course the biological notion of fitness (which unlike the GA notion is the lifetime fecundity -- expected number of children that reach reproductive age -- of an individual).
Selfadaptation works as follows: suppose you have two solutions that encode for the same eye shape, but the adaptive rates of one individual make it more likely to change this shape than with the other. Now if changing the eye shape is detrimental, it is less likely that the offspring of the first one will reach reproductive age (not to mention the fact that they will inherit this wrong adaptation rate). It will thus have a lower fitness (less children) and given enough time these badly adapted adaptation rates will die out. Thus a genetic algorithm equipt with such a technique will learn from the losers, albeit in an implicit way.
So adapting the adaptation rates might seem a bit of a meta-meta hook, but it actually works.
Problem with this in GA's is that it is usually the best strategy not to change anything at all!
This is mainly because there usually is a static pay-off function (which the GA-people call the fitness function for crying out loud). Such a static function does not encourage adaptation. So in GA's you usually force change.
In real genetics there are a lot of strategies to do this: you mentioned repair and error correcting codes, but there are also modifier genes that modify which gene will become active. Evidence exists that some part of the chromosome are indeed more mutable than other parts. There are techniques using proteins that, once synthesized, activate a gene later on. It's great and it's messy.
Thus your contention is correct. Many mechanisms that ensure stability and ensure adaptability exists and the GA community is trying to learn how to bring that in an abstract model that is capable of solving problems. This is neccessary to be able to learn while optimizing.
It is really cool stuff and we have only scratched the surface of this.
Yep, though Marv Minskey would probably tell the Wright Brothers to first set up a theory of aviation before trying to build a plane.
Sometimes theory wins over practice, sometimes it's the other way around. It's much cooler to first have a theory and later figure out how to use it (viz. Quantum Mechanics and big bombs and computers), but it is foolhardy to insist on having this theory first.
Funny, made me remember an MS white paper I saw a couple of years ago, discussing how to port the Win32 API to Win64. The thing that really made me laugh out loud was the integer handling. They actually suggested a couple of macros:
INT64
but also
LONG64
both in signed and unsigned variants.
So the question could actually be:
Would you calculate MS's profits in int64 or in long64?
Just as a little off-topic conclusion I pose the question, where is Rockefeller's company now? Aren't they American Oil or something? No wonder they feared competition since they're all but gone now...
Well, they broke up into several companies. A tiny one then called Standard Oil of New Jersey is still now known as the marginal company called Exxon. But see http://www.us-highways.com/sohist.htm.
(apologies for the plain text link, I don't speak much html)
If you want to bring real probability theory in, you can even show that with even odds but differences in fortune (bank has infinity for all practical purposes, you have xxx USD), the bank is also guarenteed to win all your pennies. So even with even odds, your chances to win are smaller than the banks, let alone with several greens and maximal bets.
Microsoft's Chicago-based Midwest district office, which covers Illinois, Indiana and Wisconsin, is the tech giant's biggest moneymaker in the country, with more than 500 customers generating $500 million in revenue annually for Microsoft.
500 customers? Biggest moneymaker? That's 1 million in revenue per customer! Talking about expensive software!
The notion that the GPL model will do anything is laughable, considering how many companies are making money on GPL software (a handful at most),
You forget that the business world does not consist solely of software manufacturers. Consider all companies (the entire 'old' economy) that can save cost and possibly headache due to non-restrictive licensing from using GPL-ed software. They make more money because of the GPL.
The entire economy with the exception of software manufacturers is a bit larger than a mere handful.
The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Library General Public License as published by the Free Software Foundation; either version 2 of the...
Notice the 'Library' statement in there. It is LGPL, so linking against it is ok. About the socket and file access... you might be right, though I think the people at FSF might have an answer for that too. If this were the case maybe glibc is guilty of a GPL violation. Now that would be ironic
Slashdot Mirror
What's St. Patrick's Day????
Worldwide != AngloSaxonia
these innovation patents are not such a bad idea at all... First of all, they do not grant the owner a monopoly, secondly it gets the idea into the 'innovation' office, maybe even the patent office.
Is it such a crazy idea to transform this 'innovation patent' into a prior art machine? I mean, a low cost way of getting the idea inside the patent office?
Assuming that the patent office is forced to search such a innovation patent database, this might get some of the more ridiculous quirks out of the system.
Just an idea. We might be better off with some more radical changes instead of smoothing things over a bit though.
Milosovic extradited, Bill Gates still free.
And someone else has written a very nice article about how science cannot reject the fact that if God exists, he can interfere with all experiments and make them look like natural. It's actually a serious article.
The article might be serious, the hypothesis in it is simply not scientific as you can not devise an experiment to falsify it. Though it's a bit Popperian, falsifiable hypotheses is what science is all about, mystical hypotheses such as this one and religious stuff is only verifiable (show me this deity that does it and I believe you) and thus not useful in science.
Free != Gratis
A great many people think they are thinking when they are merely rearranging their prejudices. -- William James
A great many researchers think computers are thinking when they are merely rearranging the researchers' prejucides. -- me
We call this Good Old Fashioned (symbolic) Artificial Intelligence.
One particular way to circumvent this is to actually _use_ the viral properties of the GPL to release the code. Just include readline or liboctave or whatever in your code that might be moderately useful and then you and the Uni are forced to release it under the GPL or not at all.
This is a great way of being able to take code with you when you hop from position to position. Universities are usually very reactive when it comes to licensing and patenting issues. They don't have a clue what to do with it, but they simply want to 0wn it just in case. This way you can argue that you cannot do that and this gives you the possibility of actually using the code you wrote yourself.
Ok, my problem with gcc stems from the time when I moved from Windows/MSVC to Linux/gcc as my development platform. Porting the code wasn't difficult, what was really bugging me was that compilation time for the project I was working on increased from 30 sec. on MSVC to more than 5 mins. using gcc. This was with full optimizations.
The libary I'm currently working on is heavily templatized and I need at least 256 Mb of RAM to be able to compile it using -O1, let alone -O2. Even with massive amounts of RAM (about 2 Gigs), the compile cycle takes ages. The same library compiles in a breeze using MSVC (factor 20 faster) and to top it off --- MSVC generates faster code!
Luckily we can discount MSVC because it also was routinely capable of generating extremely fast code that was absolutely wrong!
So, yes there is at least one optimizing compiler that is way faster than gcc. Maybe MSVC cut a few corners doing this (plus being far from standard compliant), but it sure seems possible.
Ok, forget the second question, someone above mentioned that it now only takes 1 Gig and a day, sigh.
About compilation speed, does anybody know whether g++ 3 supports the 'export' keyword now? That would help quite a bit in reducing compilation time.
And on a related note, does it still need approximately 2 Gigs of internal memory and the better part of the weekend to optimize heavily templatized code*?
Thanx
I really loved this one:
"We've looked at the pluses and minuses," she said Thursday, and determined that overwriting would protect information on the computers while allowing the machines to be donated.
There must be one of a hell of a group behind this, looking at all those pluses and minuses on the hard drive to see if there's any protected information left!
If they would have put only minuses on the hard drive, the exercise may have been less hard. But then they might have concluded that half of the protected information was still there!
The guys pioneering these kind of techniques have put themselves in a subfield called evolution strategies (which is actually older than GA's). They use real valued vectors as the genetic material and co-evolve both the mutation rates (standard deviations) and the correlations between these rates. The system is then capable of performing directed mutations: all mutations that are correlated will change the original parameter vector together. This works really well on some really hard problems and is capable of outperforming gradient and newton methods for optimization. (this technique is developed in the early seventees, even before GA's become popular.)
In GA's the problem is known as linkage learning. How do you figure out, during a run, what are the important linkages in your problem. Thus how can you keep this conglomorate of genes that code for the eye shape together, while it is dispersed over the chromosome. Your looking for stuff that belongs together and start to combine such building blocks.
A simple GA cannot do this.
Enter the world of 'competent' GA's that are designed to handle this problem. There are now three types of competent GA's developed that are very different from each other. They all are capable of finding conglomerates of genetic material that constitute a building block. Building blocks are defined to be things that have a mostly linear effect on the fitness of the individual. One of these competent GA's actually examines the losers in Minsky's term to figure out what went wrong (so there goes that assertion).
It is quite funny that most people are still using the vanilla GA as it has been proven (by David Goldberg, one of the main figures in the field) that it cannot scale properly. Minsky is obviously using this vanilla GA as his strawman, while the research in GA's has moved on and improved the thing quite a bit the last 10 years.
So, yes, there are a lot of self-adapting genetic algorithms around, they use very different techniques to perform the self-adaptation. It goes from adding rules to changing rates, identifying building blocks (thus setting fuzzy markers), till building full fledged bayesian networks. We're still just scratching the surface, but there has been progress and more difficult problems can be tackled.
Didn't know that, thanx for the info. Just checked it out and it looks quite ok (the exported file when ran through latex looks quite nice:). Never seriously looked at kword before, will do that now. Kudos to them!
Cool, I hope they'll also start supporting importing and exporting to TeX. Maybe then the stuff will start looking professional*. Kidding aside, it's a shame that the word-processing crowd is ignoring the best type-setting system around. WYSIWIG documents just don't cut it compared with a doc prepared in LaTeX.
* professional as in 'professional publisher', not as in 'professional marketeer'
You should have told that to the brother's Wright, although arguably they didn't produce the same process (a bird).
I don't know much about genetics, though a bit about genetic algorithms, and this is indeed a very important area of research. Lots of research in the GA-field is devoted to finding ways of automatically adapting the adaptation process during the evolution. The goal is indeed to learn from the mistakes (and successes) individuals make.
The basic intuition for this to work is of course the biological notion of fitness (which unlike the GA notion is the lifetime fecundity -- expected number of children that reach reproductive age -- of an individual).
Selfadaptation works as follows: suppose you have two solutions that encode for the same eye shape, but the adaptive rates of one individual make it more likely to change this shape than with the other. Now if changing the eye shape is detrimental, it is less likely that the offspring of the first one will reach reproductive age (not to mention the fact that they will inherit this wrong adaptation rate). It will thus have a lower fitness (less children) and given enough time these badly adapted adaptation rates will die out. Thus a genetic algorithm equipt with such a technique will learn from the losers, albeit in an implicit way.
So adapting the adaptation rates might seem a bit of a meta-meta hook, but it actually works.
Problem with this in GA's is that it is usually the best strategy not to change anything at all!
This is mainly because there usually is a static pay-off function (which the GA-people call the fitness function for crying out loud). Such a static function does not encourage adaptation. So in GA's you usually force change.
In real genetics there are a lot of strategies to do this: you mentioned repair and error correcting codes, but there are also modifier genes that modify which gene will become active. Evidence exists that some part of the chromosome are indeed more mutable than other parts. There are techniques using proteins that, once synthesized, activate a gene later on. It's great and it's messy.
Thus your contention is correct. Many mechanisms that ensure stability and ensure adaptability exists and the GA community is trying to learn how to bring that in an abstract model that is capable of solving problems. This is neccessary to be able to learn while optimizing.
It is really cool stuff and we have only scratched the surface of this.
Yep, though Marv Minskey would probably tell the Wright Brothers to first set up a theory of aviation before trying to build a plane.
Sometimes theory wins over practice, sometimes it's the other way around. It's much cooler to first have a theory and later figure out how to use it (viz. Quantum Mechanics and big bombs and computers), but it is foolhardy to insist on having this theory first.
Funny, made me remember an MS white paper I saw a couple of years ago, discussing how to port the Win32 API to Win64. The thing that really made me laugh out loud was the integer handling. They actually suggested a couple of macros:
INT64
but also
LONG64
both in signed and unsigned variants.
So the question could actually be:
Would you calculate MS's profits in int64 or in long64?
Well, they broke up into several companies. A tiny one then called Standard Oil of New Jersey is still now known as the marginal company called Exxon. But see http://www.us-highways.com/sohist.htm.
(apologies for the plain text link, I don't speak much html)
If you want to bring real probability theory in, you can even show that with even odds but differences in fortune (bank has infinity for all practical purposes, you have xxx USD), the bank is also guarenteed to win all your pennies. So even with even odds, your chances to win are smaller than the banks, let alone with several greens and maximal bets.
Microsoft's Chicago-based Midwest district office, which covers Illinois, Indiana and Wisconsin, is the tech giant's biggest moneymaker in the country, with more than 500 customers generating $500 million in revenue annually for Microsoft.
500 customers? Biggest moneymaker? That's 1 million in revenue per customer! Talking about expensive software!
The notion that the GPL model will do anything is laughable, considering how many companies are making money on GPL software (a handful at most),
You forget that the business world does not consist solely of software manufacturers. Consider all companies (the entire 'old' economy) that can save cost and possibly headache due to non-restrictive licensing from using GPL-ed software. They make more money because of the GPL.
The entire economy with the exception of software manufacturers is a bit larger than a mere handful.
From the copyright statement of glibc:
...
The GNU C Library is free software; you can redistribute it and/or modify it under the terms of the GNU Library General Public License as published by the Free Software Foundation; either version 2 of the
Notice the 'Library' statement in there. It is LGPL, so linking against it is ok. About the socket and file access... you might be right, though I think the people at FSF might have an answer for that too. If this were the case maybe glibc is guilty of a GPL violation. Now that would be ironic
And what about the children?
We have to protect the children!
>O, by the way, I speak Dutch, so where I'm from?
:-)
Vriesland, obviously