Artificial Intelligence

[fred18]

Hi all,

I've been reading a lot about A.I. and I started to think about it and to try to imagine how it can be somehow developed.

So, starting from the basic, when an A.I. can be defined as such?

And then, I can programme a thing to take decisions out of 10.000 different cases with a huge "switch" snippet, but those would be anyway preprogrammed choices, so not really decisions. When an A.I. makes decisions by itselves it's because it can choose between right and wrong, but what could be the ideas behind the programming of such a complicated task?

Thanks in advance to those who are willing to help me understanding that! :tiphat:

 

[Col_Klonk]

http://orbiter-forum.com/showthread.php?t=36680

There's tons of books on AI that all allude to the same ideas :-
- As of yet, we don't have the 'processing power' to match the brain's autonomous decision making.
- The Brain looks departmentalised (The Beano Book :lol: ), so maybe AI should be as well.
- etc..

What will be needed is a revolution of sorts in this department, before we can move forward.... but then again maybe that's not a good idea

 

[fred18]

Thanks!

sorry I lost those posts, didn't want to replicate the question.

Anyway still this is very difficult to imagine to me (and not only to me I guess). I can imagine how to programme a computer to learn correct physical parameters by simply trial and error. For example a robot that has to climb a stair and put his feet down until it has the feedback from the step etc. But to make decisions no... I cannot see how to would it be possible to do it without giving it massive quantity of examples to check.

Is that maybe how our brain works? since we keep learning we store massive examples quantities to check when we have to make decisions?

 

[Ripley]

I'm no expert, nor a programmer, but what is your context?
Among what "physical boundary" do you want your AI move/act/whatever?

Imagine a driving/racing simulator: there you have a limited number of possibile choices, dictated by track dimensions, speed limit, # of opponents, and defining what's wrong/right is somewhat an easier (still quite complex) task.

 

[Col_Klonk]

Thanks!
For example a robot that has to climb a stair and put his feet down until it has the feedback from the step etc. But to make decisions no...

Ja, that's an odd one.
Could a robot sit there and stare at it's humanoid parents, like a 3-6month old baby.. then suddenly crawl, get up and start walking. Can it learn from observation and is such a mechanism/algorithm available for silicon testing.

They have robots that can get up after falling down.. but can one start with a clean neuron slate, and then learn to climb stairs ?

This would be interesting.. and scary!!

 

[fred18]

I'm no expert, nor a programmer, but what is your context?
Among what "physical boundary" do you want your AI move/act/whatever?

Imagine a driving/racing simulator: there you have a limited number of possibile choices, dictated by track dimensions, speed limit, # of opponents, and defining what's wrong/right is somewhat an easier (still quite complex) task.

I have no specific context, I'm just wondering in general how can it be designed, not by me, but by experts.

Actually the racing simulator is perfect for what I mean: defining right or wrong is (let's say) simple because the possible cases are limited.

But point is: if you tell the machine what is right and what is wrong and then the machine just keeps checking if what is happening is included within the preprogrammed right/wrong settings would it be intelligence, or just comparison capabilities ?

Ja, that's an odd one.
Could a robot sit there and stare at it's humanoid parents, like a 3-6month old baby.. then suddenly crawl, get up and start walking. Can it learn from observation and is such a mechanism/algorithm available for silicon testing.

They have robots that can get up after falling down.. but can one start with a clean neuron slate, and then learn to climb stairs ?

This would be interesting.. and scary!!

If you think about it, brain has a series of preprogrammed capabilites that have to just build up: walking, talking etc

and during the first let's say 10 years it keeps absorbing every example it faces, 24 hours a day every day.

Now if I tell a computer to build up an example of right/wrong decisions database by watching what goes around for 10 years continously, would we understand if he's just comparing with its learnt cases or deciding by itself when we ask a question to him?

I have no position or answers to this, I'm just writing while thinking :hmm:

 

[Col_Klonk]

If you think about it, brain has a series of preprogrammed capabilites that have to just build up: walking, talking etc

That would be a good test of how good your neuro program is..
Say motor control.. You have the drive mechanisms, sensors, etc all connected, but no control program... this has to be learn't. ?
:thumbup:

 

[ADSWNJ]

One of the more interesting AI algorithms in recent years is IBM's Watson. You may remember it won the game show Jeopardy 3-4 years ago, which required it to be able to assimilate huge amounts of information and interpret meaning from questions that often use humor or hidden meanings.

Watson has been working hard on the process of diagnosis in the medical field. I have spoken with the dev team and seen how they do some of this work, and it is mind-blowing. Watson ingests all medical research, theory, drug information, and training materials it can find (constantly, every day) and then builds an assimilation map for the material. One interesting thing I found was how the model tries to resolve what is often contradictory research, or contradictory symptoms: it builds its own 'opinion', weighing the body of evidence, and considering the people named on the material. (E.g. if a new researcher produces a paper that challenges conventional thinking, then Watson will go to great lengths to assess whether the paper has merit. Often, humans would be preconditioned to reject a new ideas because it does not match their world-view, or most likely they never heard about it. This is where a more objective cognitive algorithm is intensely interesting.

There's lots of other applications for Watson, but I like this medical one, as it opens up huge opportunities in the future - e.g. medicine in the third world, especially linked with 3D fabrication techniques. Imagine being able to diagnose the problem, and then fab a custom medication to resolve it. Amazing.

 

[TMac3000]

Artificial intelligence is not possible. It's an oxymoron.

Think about it...true AI is when a computer can take initiative on its own--that is, do something without being programmed to. But since they originally did not do that, the only way to get them to act on their own is to program them to. Catch 22.

And you should all be relieved to hear that. Machines that can out-think humans are self-evidently a bad idea.

 

[Ripley]

Indeed...

 

[fred18]

Artificial intelligence is not possible. It's an oxymoron..

I don't totally agree with this. It depends a bit on what do you think it is aritificial intelligence.

To some extent also modern computers are already artificial intelligence. How do you measure the intelligence of your son? you teach math to him and then you ask him to make calculus to see if he can reach the right result? well that's exactly what a computer does, it has been taught math and if you ask him he can solve calculus.

If you do a bit of coding you can tell your program "if this happens do this, else do that", isn't it like teaching to a child what to do?

Actually what I see as missing right now is the self awareness. Usually a being is considered as "living" when he has self awareness.

I think that the artificial intelligence is not necessarily wrong, but the point is that if we build it with a computer it will be without any feeling and totally flawless, and that's dangerous for humans. If we could build an artificial intelligence with other matter that would be another story.

 

[Urwumpe]

Artificial intelligence is not possible. It's an oxymoron.

Well, AI also involves less intelligent algorithms. Important is that a system is capable of learning from past information, regardless how limited. The typical AI in video games is not artificial intelligence for example - it is just following decision trees without really deciding about a situation by itself, a better form of a if-else-explosion.

In Orbiter, you could for example already have a AI, if you would have an autopilot, that knows multiple alternative guidance schemes, and decides on past flight experiences with a spacecraft class, which guidance scheme yields the best results. Its just a typical expert system then, but that's already a primitive AI.

 

[TMac3000]

If you do a bit of coding you can tell your program "if this happens do this, else do that", isn't it like teaching to a child what to do?

Well, yes...but the child can get bored with the lesson and walk away. No computer can do that.

Computers are automatons. They can only do what they are told. The day a computer can do something without being programmed would be a dark day for humanity. Fortunately, as I said, telling something to take initiative is contradictory, so we don't have to worry about Terminators anytime soon :lol:

I think that the artificial intelligence is not necessarily wrong, but the point is that if we build it with a computer it will be without any feeling and totally flawless, and that's dangerous for humans.

Exactly.

Well, AI also involves less intelligent algorithms. Important is that a system is capable of learning from past information, regardless how limited. The typical AI in video games is not artificial intelligence for example - it is just following decision trees without really deciding about a situation by itself, a better form of a if-else-explosion.

But it would still have to be programmed by humans to do that. That's not artificial intelligence--that's natural intelligence being used to guide a machine.

 

[fred18]

Well, AI also involves less intelligent algorithms. Important is that a system is capable of learning from past information, regardless how limited. The typical AI in video games is not artificial intelligence for example - it is just following decision trees without really deciding about a situation by itself, a better form of a if-else-explosion.

In Orbiter, you could for example already have a AI, if you would have an autopilot, that knows multiple alternative guidance schemes, and decides on past flight experiences with a spacecraft class, which guidance scheme yields the best results. Its just a typical expert system then, but that's already a primitive AI.

So, the point is: let it absorb as much experience it can, and listen to as much teaching it can and then it will be what we consider "intelligent" exactly?

So, back to my point initially: is it enough to let it learn a build a huge database of what is right and wrong to do (not only right or wrong in absolute, but for example in order to solve a particular issue) to have an intelligent machine?

 

[Urwumpe]

So, the point is: let it absorb as much experience it can, and listen to as much teaching it can and then it will be what we consider "intelligent" exactly?

So, back to my point initially: is it enough to let it learn a build a huge database of what is right and wrong to do (not only right or wrong in absolute, but for example in order to solve a particular issue) to have an intelligent machine?

Well, would you call an ant intelligent? A cat is never doing what you tell it to, is it intelligent? Or just a unreliable form of a mousetrap? What about the intelligence of dogs.

As much as you can appraise the intelligence of supercomputers for playing Go or Chess... they fail epically when having to solve basic problems that animals solve every day. You can't yet build an AI to hunt a mouse. And lets not get started about rats.

And just take other problems of AI: Could we write an AI to understand the TTIP agreement?

 

[fred18]

I agree, but a human child is exposed to learning for years before being capable of critical reasoning.

Aren't we trying to build it directly finished instead of programming just its very basic functions and then let it learn by itself? if we raise a child just teaching him on how to play chess he will not know how to do simple math, right?

I'm just "thinking out loud", I'm trying to understand how this could work.

PS: I live with two cats, one is very intelligent, you can easily notice, while I call the other "pentium" because when she has to decide something I think I could see the "please wait..." sign in her eyes :lol:

 

[TMac3000]

A cat is never doing what you tell it to, is it intelligent?

Frighteningly so:lol:

What about the intelligence of dogs.

It depends on how you separate "intelligent" from "self-aware". When most people speak of "true" AI they are talking about self-awareness, which does not by itself require a great store of knowledge.

Well, would you call an ant intelligent?

An ant is part of a hive mind. It can be debated whether they are even self-aware, let alone intelligent.

 

[Ravenous]

So, back to my point initially: is it enough to let it learn a build a huge database of what is right and wrong to do (not only right or wrong in absolute, but for example in order to solve a particular issue) to have an intelligent machine?

I don't think so; you have artificially imposed your view of what is the right and wrong thing to do, based on the circumstances you are living under.

In reality, if the machine later leaves you and goes to live in a new place, with new owners, new goals, perhaps even on a different planet, how can it work intelligently based on the old set of rules?

If you think about it, brain has a series of preprogrammed capabilites that have to just build up: walking, talking etc

Human babies are sometimes born deaf; in that case they tend to learn to lip-read automatically. This seems (in my opinion) to violate the idea of pre-programmed capabilities.

AI is an interesting subject. I think the list-based method you've mentioned is more like the old expert systems, etc. (Not truly intelligent, but still useful.)

 

[Urwumpe]

I agree, but a human child is exposed to learning for years before being capable of critical reasoning.

Aren't we trying to build it directly finished instead of programming just its very basic functions and then let it learn by itself? if we raise a child just teaching him on how to play chess he will not know how to do simple math, right?

I'm just "thinking out loud", I'm trying to understand how this could work.

Well, there is always a core system of reflexes that you can build on. :lol:

But generally, I think proper AI like in this case requires the ability to expect less. We are applauding children already if they do things approximately right.

Also, I know quite many adults who are unable of critical reasoning, even here on this forum. Don't expect too much there as well.

PS: I live with two cats, one is very intelligent, you can easily notice, while I call the other "pentium" because when she has to decide something I think I could see the "please wait..." sign in her eyes :lol:

My cat has no noteworthy talents, but he can calculate precisely when to drop from a tree to fall directly on top of a quickly running dog... rendezvous problem solved.

 

[Quick_Nick]

It has already been possible for a long time to toss an artificial neutral network into a robot and end up with emergent deceptive behavior in the pursuit of the robot's own gain. The only real human involvement there is defining 'gain' (fitness) to the robot and confining the robot to some environment. What it ends up doing is a product of evolution and randomness, same as with the rest of us. We just have a couple billion years head start and access to more neurons.