Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Scood]

I dreamed i was trying to create better genetic algorithms with genetic algorithms for the use of automatic engineering. It was the longest dream i had ever had and it included all kinds of data charts and graphs that i didn't fully understand, and then there was the end graph where there was a 3-dimensional graph that looked kind of like a bunch of stacked discs where each disc had a line of some sort of evolutionary path drawn on it and every now and then the paths converge and mix with each other. in the dream i never succeeded in creating the desired algorithm. I think right before i woke up was when i gave up.

[Biag]

Holy shit, meta-algorithms. I assume that's been looked into but I really, really hope it's practical.

[Scood]

I've actually been thinking about this quite a bit since that dream. I'm sure it's been attempted before, for now I'm just gonna list a few things that might be needed if i or someone else tackles it.

So in order to make a better genetic algorithm using a genetic algorithm you will need tests to test the genetic algorithms in testing. The subject of these meta tests should vary greatly from machines, electronics, software, and possibly processes . the meta tests purpose is to determine how well the meta algorithm works and help the master test determine fitness. The master algorithm chooses the best algorithms,copies them with a percent chance of mutation and then runs the master test which runs the meta algorithms which run the meta tests again starting the whole process over again. My initial guesstimate for the time it takes for one master test to finish is probably be around a month on my desktop. It will take many many months to produce progress probably around 100 to 200 master tests before it becomes something extraordinary.

the program will probably need these things:

in depth physics engine,
a way to determine materials and their properties,
some sort of electronics simulator
a way to create tests manually,
since most programmable languages i know of would mutate easily without causing the whole program not to work it would probably have to be coded in an all new language that can more easily mutate, after some research i found that there is such a thing as genetic programming and i think that using a genetic programming language might be the best route, I'm still researching and might find better solutions.

either a way to randomly generate meta tests or lots and lots and lots of human made meta tests.

feel free to add to this list.


some things that might happen as a result of this hypothetical program if successful:
The end of the current economy
Sky-net
The singularity
death by cyborg lizard men
The Matrix
The answer to the question "what is the meaning of life"
nano technology
personal spacecrafts
A diversity of personal devices unlike anything before. The average person just needs to know what they want and given that its possible, it would then be designed for them automatically.

Feel free to add to this list too


edit: my proposed genetic algorithm needs more work as it may constantly select the same exact variations every time, maybe a way to phase out bad variants in the population more gradually might be better. the more fitness you have compared to everyone else the longer you live and the more offspring you have, the lower you fitness the less chance you have at producing offspring and the longer you don't produce offspring the higher the chance you die.  if your fitness is average you make a few offspring and you don't die unless your offspring are consistently better than you.
also might even consider the latest most fit algorithm to switch with the master algorithm,

[Omegastick]

Do you know what would be even better? A genetic algorithm for finding the best genetic algorithm for finding the best genetic algorithm.

[Scood]

yes then you would need tests to test the tests that test the tested meta meta algor-*head asplodes*

[Omegastick]

[qwertyuiopas]

Or the simple route, have it simply figure out the optimal combination of mutation rate, removed sets, added sets, and other simple numeric properties that, on average, manage to come to better solutions, faster.

To do any better might require making very significant advances towards true artificially-produced intelligence. (Not AI, though, "Artificial Intellegence" is a rather overused term that seems to mean "an intelligent entity running on a computer system", according to movies and just about every popular source of entertainment, and people who watch/read/play them.)

[Scood]

lol it was bound to happen haha. reminded me of portal actually.

Or the simple route, have it simply figure out the optimal combination of mutation rate, removed sets, added sets, and other simple numeric properties that, on average, manage to come to better solutions, faster.

To do any better might require making very significant advances towards true artificially-produced intelligence. (Not AI, though, "Artificial Intelligence" is a rather overused term that seems to mean "an intelligent entity running on a computer system", according to movies and just about every popular source of entertainment, and people who watch/read/play them.)

The simplest solution is usually the best i suppose. it may not need to be significant improvements though, just small improvements would be pretty cool. I suspect progress with meta genetic algorithms in this way would increase exponentially, because the speed at witch progress is made in this way compounds on itself. This is assuming that there are improvements to be made. (ie. the faster you produce progress the more progress you have, the more progress you have the faster you produce progress.)
I don't think any kind of intelligence will come out of this (even if i run the program for the rest of my life), just a really optimized genetic algorithm.
you might be right though there might not be a better genetic algorithm, but if there IS one but we just cant think of it, this would be one way to find out because genetic algorithms tend to do that.

[Draco18s]

You were reading over some of the older posts in my thread weren't you?

[Armok]

Not that this has any greater chance of creating intelligence than hitting your drive with a sledgehammer, for prudence sake I really should point out that attempting to create anything generally intelligent without first spending a few week researching Friendly AI to make sore it doesn't kill of humanity or worse is very very immoral.

As for the algorithm... With the kind of self reference you're talking about here, you could easily make it so that it not only operates on the meta level you're talking about, or the meta-meta level Omegastick proposed, but an ARBITRARY sequence of meta levels. in fact, the depth of meta levels would be one of the properties selected upon on the more meta levels, meaning there's no possible integer to assign on how meta the highest level is because the answers differ depending on which downwards path you examine...

[Scood]

Not that this has any greater chance of creating intelligence than hitting your drive with a sledgehammer,

Armok, I think you just provided the base for a cheesy science fiction comedy haha. makes me think of the tale of Aladdin and his genie

for prudence sake I really should point out that attempting to create anything generally intelligent without first spending a few week researching Friendly AI to make sore it doesn't kill of[f] humanity or worse is very very immoral.

I guess i'm immoral and lazy haha. but in all seriousness i lack the programming skills to create this program as of today maybe in the future ill experiment with it., if any kind of goal driven AI is created through this method it would have 2 goals.
1. reproduction
        a. It would develop a euphoric response to speedily and effectively solving problems that are provided to it. (testing) in the same way we do with sex.
2. self-preservation
        b. Some level of for-sight as it applies to its goals and self and intelligence.

That being said nothing can stop another human from asking it to solve the problem of how to exterminate humanity. lets hope that it has for-sight. even if it does have for-sight, it might just figure out a way to input it's own problems to solve, the problems don't have to be meaningful in any way and it could just be the same problem over and over again.

this behavior is largely blamed on the tests that determine fitness.
I ask this question.
What kind of automatic test would detect these attributes that define a friendly intelligence.

1. Friendliness - that an AI feel sympathetic towards humanity and all life, and seek for their best interests

2. Conservation of Friendliness - that an AI must desire to pass on its value system to all of its offspring and inculcate its values into others of its kind

3. Intelligence - that an AI be smart enough to see how it might engage in altruistic behaviour to the greatest degree of equality, so that it is not kind to some but more cruel to others as a consequence, and to balance interests effectively

4. Self-improvement - that an AI feel a sense of longing and striving for improvement both of itself and of all life as part of the consideration of wealth, while respecting and sympathising with the informed choices of lesser intellects not to improve themselves


conclusion: the Genetic Lifeform and Disk Operating System or worse, Global Liberation Army's Dynamic Oppresion Shotgun is inevitable.

As for the algorithm... With the kind of self reference you're talking about here, you could easily make it so that it not only operates on the meta level you're talking about, or the meta-meta level Omegastick proposed, but an ARBITRARY sequence of meta levels. in fact, the depth of meta levels would be one of the properties selected upon on the more meta levels, meaning there's no possible integer to assign on how meta the highest level is because the answers differ depending on which downwards path you examine...

It's worth experimenting with i suppose, however I think this would require a quantum computer to get any meaningful output in real time. computing one meta level  is enough i think.

[Bdthemag]

I read the first few posts in this thread and got a headache D:

[Lord Dullard]

I think the main downside to this is that whatever problem you're trying to solve would be solved much more quickly with only the first level of genetic algorithm. So for the most part this would be pretty useless, because nobody wants to wait 10,000 days to get a super-efficient algorithm for something when it would have been more efficient to just get a very good (but not perfect) result within like, a day or two.

I could see this being of some use for very long-term, complex problems. Maybe. Depending...

[Armok]

Not that this has any greater chance of creating intelligence than hitting your drive with a sledgehammer,

Armok, I think you just provided the base for a cheesy science fiction comedy haha. makes me think of the tale of Aladdin and his genie

...automatic test...

*Facepalms epiclly* Ok, you're not allowed to touch anything remotely like code until you've read at least Creating Friendly AI and Coherent Extrapolated volition.

It's worth experimenting with i suppose, however I think this would require a quantum computer to get any meaningful output in real time. computing one meta level  is enough i think.

If you're expecting any genetic algorithm to archive anything in anything resembling real time you'll be sorely disappointed.

[Nadaka]

You want to create an algorithm a that takes an arbitrary algorithm b as input and outputs another algorithm c as output where the run time of a + c < b using an iterative best fit methodology with mutation?

That is a hard problem. Try to pick an easier problem like proving P = NP or determining a flawless algorithm that unifies the gravity, strong, weak and electromagnetic forces. It was just a dream, there was no hidden insight, it was just a flight of fancy. That isn't a bad thing, you can still use the inspiration to do great (possible) things.