Private AI development?

Doing more than reading? Are YOU developing your own AI system? A system where you can change anything you want any time you want? What it does, how it works and how it’s limited? I’m interested mainly in problem solvers. I want to say to an AI system: “here is some info about my (or any other) world”, “here are my objectives”, “What steps would you suggest to take in order to reach these objectives?”. The system should be open to any type of knowledge which can be described by words. Can we develop a powerful AI system which uses just regular text as its I/O? It’s hard to get meaning from regular English sentences but if we use some sort of dynamic form-based communication then it may not be as hard. I’m looking for a general way how to describe objects, their relationships and processes. Something like Chomsky’s thematic roles needs to be used to teach AI systems about particular processes. The system also needs to work with some sort of timelines when assembling solutions and exploring combinations of processes. The system can also get some knowledge from other systems (like CYC). Any comments? Are you developing something like that? Sorry if it's already discussed somewhere else on sciforums.. I'm a new user and I still did not have time to check all the other places here..

 

Log in or Sign up to hide all adverts.

You know about Cyc it seems, but here's some links in case others are interested:
http://www.cyc.com
http://www.opencyc.org

Can you expand on what you mean by form-based communcation? As far as your problem description, a planner can solve simpler versions of the problem. Not exactly the 'intelligent' solution I think you're looking for, but I'll mention it anyway. Two systems that I know of are UCPOP and Sensory Graphplan. You're limited to relatively small problems with these, though. If your world gets too large or complex, the amount of work you need to express it in the system becomes quite large, as does the amount of work the algorithm must do.
Again, probably not what you're looking for, but worth a look if you're interested.

 

Log in or Sign up to hide all adverts.

User could create sentences using "chat elements" which are briefly described here. It's an old page. I have better elements now but it's not online yet. Using a few clicks, user would create a little form for each sentence and then fill it with words. Not a coolest solution, but it may work OK. Thanks for the links. I'll explore it later today.

 

Log in or Sign up to hide all adverts.

Hmm...I kind of see where you're going, but I'm still not completely clear. Could you give me a couple examples of what you'd fill that form in with?

 

This is a huge system, and a tremendous undertaking. The best thing to do is to use as many systems that are already available.

The English language, like all languages has it's little quirks. You'd be better off moving to a meta-language:

http://www.c3.lanl.gov/~rocha/univgram.html

http://www.unl.ru/introduction.html

Not only would this solve certain problems with English, but it would allow you expand your search to other languages.

Is CYC built with English or meta-language?

Your AI could be a data miner built to use an available relational database. MySQL? Prolog works well with SQL.

Then once you get past Chomsky you may look to Levi-Strauss and Barthes. All objects carry several layers of meaning. A rose is not a rose. Semiotics might play a role in the final relational process.

 

malkiri: It could be a sentence like: [95% of] [people] [in Czech Republic] [Do Not] [Believe] [God] [Does] [Exist]. It's usually one or very few words per textbox. In order to understand the meaning of a sentence, the system needs to figure out the part of speech (noun, verb,..) for each word. That's a problem in English. I saw same examples where the same word was used (in a few sentences) as 5(!) different parts of speech. But it's pretty clear when using the mentioned "chat elements" (each textbox is used just for a single predefined part of speech).

Puppeteer: First, I need to make sure the system has good algorithms for it's thinking. I do not think I need lots of data in this stage. I have a 2 years old daughter. Her view of the world is of course very limited comparing to our view, but her brain always finds some kind of "solutions". We played Tomb Raider 3 today and there is a guy in that game who lost one leg (in jungle). I challenged my daughter saying something like: "He lost one leg, what is he going to do now". She immediately suggested he needs to go to SafeWay to buy a new leg. When her brain generates these "solutions", I can almost see lines of a code which could have been used to find such answer using her limited knowledge base. The solution is incorrect but the algorithm is IMHO valid. And that's what I want to focus in the near future - valid algorithms. When I'm happy with algorithms, I'll try to get lots of data. Yes, Cyc is using meta-language. The communication with the AI system I'm going to develop needs to be close to the regular English. It should be something easy to use for average American. I may need to spend more time studying English grammar and grammar generally. English is far from my native language and I never had time to really study English. Thanks for the links. I'll probably not use Prolog. I'm currently considering VB.NET and MS SQL Server 2000 as main dev tool because it's very convenient and very productive. I still work on a bunch of other interesting projects (one includes a code which will generate new, hopefully nice, MIDI songs using inspiration from existing MIDI files), so I'm just slowly preparing for the AI development now. I have many AI thoughts I'm planning to explore in details. Like algorithms for getting experience and new definitions from fairy-tale-like stories (written in a format the system can understand).. Thinking about Joke generators and players of any game after explanation of rules etc.. But the problem solving is one of my top priorities for my AI systems. BTW there are so many things to study.. If I'm gonna study all the related details then I'll probably never find time to actually develop something.

 

If the system cannot make a good guess about the meaning then it should be able to ask user, offering some alternatives.

 

Sorry, but this is just wrong.
For instance, small children have no idea about what is a noun and what is a verb. At the same time children understand English pretty well.

Systax analysis is one of examples of <a href="http://www.dennisgorelik.com/ai/FalseWaysInAISDevelopment.htm">False ways in AIS development</a>

 

Children have built-in learning capabilities that is far beyond a adults.

A artificial life form would need data to compare it too but learning verbs is not that hard, the problem is when the grammar does not follow the mainstream knowledge database since it would basically have to store every new occurance and then attach a percentage likliness of it being right.

 

Dennis,

They may not know the terms ("noun","verb") but they quickly get a good understanding of these concepts because of the types of data they receive from their senses. A text-based AI (of course) needs a different way how get a good sense of what is what.

That's not what I'm doing. I'm trying to use limited syntax and keep the rich expression power + a new syntax can be created in my system.

ANNs are good just for certain tasks (mainly various recognition and conversion tasks). The key AI algorithms IMO should not be based on ANN. They are not flexible and transparent enough (cannot easily explain its decisions), they are hard to use for things like complex planning and the needed teaching procedures may get very hard to design. ANNs are not as practical as many seem to think. Many say "let's follow the nature's way" (neurons), but the nature's way is not always the best way to go. Airplanes don't flap wings.

Really? Wouldn't you lose what makes ANN so great - the massive parallel processing power??

I don't think they need it. Feelings are goal generators. Thy systems should be getting goals from us, not from its own feelings which could potentially cause some conflicts. They just need to have certain understanding how feelings work.

 

2 nodes make a line, 3 nodes make a triangel. When using a 3d engine the maximum number of nodes to make a sphere were set to 90, although it did not look perfect. How many nodes do you need to make a perfect sphere? Size does not matter. (It has to follow the laws of physics and it has to be 3 dimensional.)

I doubt that all of you can answer that question correctly. Can you build an AI that can?

 

3D engines are always limited by resolution so there is nothing like a perfect sphere. On the other hand, all vision systems have limits on details they can see so if the resolution goes beyond the vision's limit then a sphere might be considered "perfect". But before you start to mess with it, ask your brain if it ever needed to imagine a perfect true-3D sphere.

 

G71, the simplest solution is always the best. You set the limit too high.
How many nodes does a perfect sphere have? (if size does not matter)

The common answer is infinite. It is right.
2 nodes? 1 in the middle and then the other as a radius.
Well why not just 1 node? 1 in the middle. Just if it is infinite small you might say but think about it the radius is the only thing that is dynamic in a "size does not matter" spear. The center is static so actually one node is all that is neccesary, right?
Just like earth, it could have gravity that node.

If the universe is infinite large, can't an object be infinte small then also?

 

Baal: I could argue but I currently do not see it as something important for practical AI development. I spent too much time with similar discussions already. What are you trying to do?

 

Please Register or Log in to view the hidden image!

just thinking.

To me it has much to do with AI because if my AI cannot provide an answer to those questions (even a false one as long at it has some rational logic behind it) then it is not universal.
I have one problem left, Counsiousness. I am currently re-thinking the AI from the begining. (Well, the concept will still be the same but i need to shift power from one place to another)

 

So are you seriously trying to develop AI to create a machine race to wipe out humanity so you can live on Mars? Once you get your machines going, you're going to have to figure out how to get to Mars, you know. And though you won't need groundbreaking tech to get there, it would cost a buttload of money. What are you going to do, sell your robots to the military for money so you can go to Mars and then activate some hidden programming or something?

You seemed pretty serious about the subject in other threads. But I don't know, you could have been joking.

 

I am building a machine race equal to man. Getting some to mars to start a new, better society is one of my goals, yes. If the robots ever realise that they perhaps should "wipe out humanity" as you say then sure, i only see it as the next step in evolution. Although they would not be developed to kill us.

Listen, if a robot would ever kill a human then it would be a humans fault. The robot does not think "hey a human, lets kill it", with my latest idea for Group recognition then it might think "hey, 5 humans. They are all holding baseball bats and are comming my way. Maybe they are attacking me"

Ofcourse i could prevent thoughts like "kill" but there will always be holes that is not blocked. So, as long as there are humans there will be a chance of human death. And do you know why? Because humans are stupid. Just as Einstein said that "human stupidity is static" i am saying that "human stupidity is static and therefore it will destroy itself"

 

Ahh, so you're just planning on building the robots and hoping for the best? Sounds a little better.

Please Register or Log in to view the hidden image!

 

Please Register or Log in to view the hidden image!

If you tell the robot "i will kill you" then it would probably kill you first.
Ofcourse i could block Kill although it could see a gun and think Shoot instead of Kill. Even though that Kill is a synononym to Shoot it is not 100% sure that a gunshoot kills you so there are always holes for it to take.

 

I am developing a cyberbrain, Telos is its name.

It can input anything defined set of symbols, text, sound, ....

The basis of this is a new type of mathematics, so all processing is direct.
It only uses boolean logic to determine truth or logical instructions.
I have a mini one written in basic on a slow computer (33MZ) and it processes faster than I can type. It basically breaks down the input directly, converts into a data read, that has all the information encoded a bit like DNA..... so all information is immediately available... there is no searching..... it is a direct mathematical system.

A world brain would be a good thing... one that you could talk to in any language.... one goal.

Because this is a mathematical matrix, the matrix points can be encoded as ROM... with no mathematical processor, so a non-learning brain could be sold and no one will be able to decoded it... no copying.....
Brains could be sold in various grades
but the learning world brain would be able to talk to all who accessed it... the learned the simple the children the tradesmen, the people.... and it would learn from everyone, and it will question to learn.... fill in the total matrix.

Now that would be nice, don't you think?