Jump to content
  • Advertisement
Sign in to follow this  
Tertsi

Strong AI without Neural Networks

This topic is 4632 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

First of all this is not directly game AI related. As some of you know, today's computers, even the super ones, can't simulate the human brain using Neural Networks. I however think that it's possible with today's technology to make a computer program which is more intelligent than humans and works much faster overall, at least on a super computer. Even if it turns out that this is currently impossible, I'll at least have a bot which can do some grunt work for me. Genetic algorithm alone can only be applied to solve some specific problems, so it's not of much use unless the program determines itself how to calculate the fitness score and what the problem exactly is. This has lead me into brainstorming ideas on how to have a more effective training algorithm for the bot which uses both randomization and knowledge. Currently I'd just use a modified Genetic Algorithm. This Genetic Algorithm should later on be modified so that it uses the bot's knowledge (if any) to determine how to determine the fitness, how to best define the problem and it would use its knowledge to determine what mutation rate, cross over rate etc. is best suited for the problem. I think this wouldn't exactly be more efficient than a neural network though, so I'm looking for algorithms which are faster than a neural network, yet allowing the bot to solve a problem on its own so I could get some ideas off of the algorithms. If you have any ideas, I'd love to hear them also.. If you have already done lots of research on the matter and have come to the conclusion that Neural Networks in one way or the other is the way to go, I'm interested to hear about it.

Share this post


Link to post
Share on other sites
Advertisement
My brainstorm.

Problems:

Too few sencors
mixing sencor data
random time assosiation/ not in same order always.
equal assosiation
cause-effect association
memory for different sencor data sametime
needs
loop backs

Actually neural network is not bad if you randomice little bit your cell order always. And about human brain the problem is the layers which connect everything together same time. So different input + memory + needs -> output



Share this post


Link to post
Share on other sites
Quote:
Original post by Tertsi
I however think that it's possible with today's technology to make a computer program which is more intelligent than humans and works much faster overall, at least on a super computer.


You're going to have to define 'intelligent' - or at least what it is to be 'more intelligent than humans' - for the task you propose.

Quote:

Genetic algorithm alone can only be applied to solve some specific problems


A GA is merely an optimisation algorithm that tends to do well on 'blind search' problems (as compared to other optimisation algorithms). A GA is NOT a model of intelligence, but could be used as a tool to tune a model to display 'intelligent' behaviour.

Quote:

so it's not of much use unless the program determines itself how to calculate the fitness score and what the problem exactly is.


Most 'fitness' feedback humans use to judge the quality (utility) of their actions/states comes from their environment. The biggest problems for humans are (a) recognising this feedback signal; and, (b) translating it into a signal that is meaningful to our system of values.


Quote:
Currently I'd just use a modified Genetic Algorithm. This Genetic Algorithm should later on be modified so that it uses the bot's knowledge (if any) to determine how to determine the fitness, how to best define the problem and it would use its knowledge to determine what mutation rate, cross over rate etc. is best suited for the problem.


A GA though is specifically designed so that the algorithm bears NO relationship to the information encoded in the hypothesis population. If you start changing the GAs algorithm based on what is encoded, then you'll assuredly lose performance, because the algorithms performance will be bound to the region of the state space it has previously searched... which means that all of the algorithms performance will be dictated by the quality of the initial population. This, to me at least, looks like a very bad idea.

Of course, I could be wrong.

Cheers,

Timkin

Share this post


Link to post
Share on other sites
Yes it seems to me that the bot will use GA only on things it knows nothing about if it can be used there.

My next 50-years' goal is to create or participate in creating a bot which is able to make software, art (sketches at least), invest, research, chat etc. on its own.

So basically it would be more intelligent and eventually have more knowledge of using the computer than humans but the first version of it doesn't have to know anything about how to survive in the outer world so I guess you can't compare its intelligence with that of humans.

For now I'm mostly just researching the current limitations and obstacles and trying to design the bot's 'brain'. The software bot won't probably have emotions nor consciousness but it should later on be expandable to a robot by some people.

Share this post


Link to post
Share on other sites
If you can come up with a better way to store knowledge than either augmented slot & filler or Hopfield relaxation networks then that would be a good start. As far as my limited understanding of human neurophysiology goes, human memory is somewhat akin to a "Processor In Memory (PIM)" system in that memory is organised/associated with other memories that should elicit a similar response. Slot and filler structures work well for limited systems where the number of "memories" is relatively low (thousands at most I imagine). Hopfield networks I haven't looked into, but while I know they work well with images or shapes I have no idea how they perform with standard "rules".

Once you get that one solved then you can start looking at ways of "training" your memory. ;)
-Cam

Share this post


Link to post
Share on other sites
Having a GA is well and good (even though it's a fantastically inefficient calculation method, think of how many computing-hours it took nature to come up with humans).

But it's only one part of the solution. It doesn't really say *what* the GA is going to actually evolve. Is it evolving PASCAL programs? Protein sequences? It could even be a GA that evolves neural networks. Pretty much any goal-directed system can be augmented with a genetic element.

Share this post


Link to post
Share on other sites
Guest Anonymous Poster
put in google "genetic programming"

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement
×

Important Information

By using GameDev.net, you agree to our community Guidelines, Terms of Use, and Privacy Policy.

Participate in the game development conversation and more when you create an account on GameDev.net!

Sign me up!