What do you guys think of that General Learner program?

[Arislan]

Walter Fritz has been a proponent of software AI for many years, and he describes his general learner (GL) program as something that learns the way people do.  He talks about it in length on his robotics site.
http://www.intelligent-systems.com.ar/intsyst/presInv.htm

There are actual source snippets here:
http://www.intelligent-systems.com.ar/intsyst/GLT.htm

He doesn't seem to mention much about storage, but I would think that any such program would require a lot of hard drive storage.
He uses a Macintosh Quadra motherboard in his robot!

[Arislan]

No one? It would be nice to have a robot controlled by a program actually capable of learning stuff.

[Admin]

See the problem with people who claim to have solved AI is that they have never actually built a working AI 

probably why no one replied . . .

[Arislan]

Don't know about "solving" AI but I like the idea of his approach. The idea that the robot starts like a "baby" and starts learning step by step. In this case, the robot's brain is connected to the microcontrollers and rest of the body, and the robot is given motivation to start learning how to move about until he can actually stand up.

Even a mouse knows how to stand up. We have to be able to program a brain that is at least as smart as a mouse by now!

[Admin]

Many people have made this proposal, and I completely agree with them too.

You may be interested in reading Marvin Minsky's book, was very enlightening for me in AI 

[JesseWelling]

Animals (and Humans) have instincs...they don't learn everything. I don't buy the whole Learning from scratch idea partly because it's to slow to build effective AI. I think one of the most active areas of AI is the representation of knowledge. Once we have knowledge representation under controll (first order logic is hardly knowledge on a human scale in my book) then I think "learning ai" in a more human sense will apear. But untill then I believe we are going to have to keep telling AI what they can do and how they can learn in order for them to build specialized knowledge base to suit thier application.

[Admin]

because it's to slow to build effective AI

I agree . . . but all you need is one fully trained, then you can make instant copies of as many as you want . . . plus, unlike in biology, the chassis doesnt grow from scratch but instead is instant adult size.

an interesting paraphrased quote from an AI researcher I spoke with, "To program AI to solve a difficult problem, you need to solve 99% of it yourself." he is referring to the fact that the programmer solves/writes the algorithm, and the machine in the end just crunches numbers . . .

perhaps true AI is programming a robot that can reprogram itself (like humans do, when challenged) . . . but do we want that 

[JesseWelling]

an interesting paraphrased quote from an AI researcher I spoke with, "To program AI to solve a difficult problem, you need to solve 99% of it yourself." he is referring to the fact that the programmer solves/writes the algorithm, and the machine in the end just crunches numbers . . .

That's what I'm learning in my AI class....Sudoku is no cake walk...

[Arislan]

Animals (and Humans) have instincs...they don't learn everything.

Can we not program these instincts into robots? Instinct seems to me an easy enough thing. Hunger, fear, pain, thirst, pattern of walking, etc...

[JesseWelling]

hunger of what? fear of what? And once we have these things how do we resolve them in a robot brain, when we don't exactly know how they are resloved in a human brain......

Seems like an over simplification to me.....

[Admin]

instincts are directly related to some very important need . . . such as eating food or running away from predators or making babies . . . instincts are hardwired programs to drive the learning process and acheive important goals - motivations to make the robot do something, and possibly perserve itself from destruction.

robots have needs too . . . but they are different needs than what you and me have . . . so instincts must be different too . . . hunger instead might be 'batteries are low' . . . then that instinct will drive higher decision making functions like 'find light to recharge with solar panels,' 'slow down to conserve energy,' or 'find a mcdonalds,' etc.

if a robot had a human to care for it (ie recharge batteries, keep the evil cat away, save it from falling off the table, etc.) a robot wouldnt need instincts other than the ability to cry for mommy 

[Arislan]

hunger of what? fear of what? And once we have these things how do we resolve them in a robot brain, when we don't exactly know how they are resloved in a human brain......

Seems like an over simplification to me.....

But a robot brain is programmable and we can input hunger for charging batteries. Fear would require danger assessment programming, but it could be fear of destruction.  Here's an example,

1) robot's sensors pick up approaching garbage truck

2) robot moves out of the way so it doesn't get hit by truck

or

1) sensors show that robot will fall into the Grand Canyon if he moves ten feet forward

2) robot backs away from precipice

These instincts look mathematical to me. Now suppose we add modifiers to the situations. Let's say we want the robot to go pick a flower at the edge of the cliff. He'd "be afraid" but because we ordered him he would approach the flower with a set of calculations of the risk involved. He would calculate the safest speed for approaching without rolling off the cliff, backing off if vibration sensors picked up that the edge was about to fall off.

It's not that hard to come up with ways to program instinct. Weather too hot? Roll under a shade. Pothole on road? Roll around it. Coconuts falling from trees? Move out of the way. This kind of stuff is even used nowadays in programming enemy AI in video games, it can't be that hard to give this common sense to a robot. The idea is that the robot will use sensors for gauging the environment all the time.

If you get as far as implementing stereo vision, a robot could detect a burglar breaking into the house and calling 911. I'd just program the robot to castrate the burglar.

[JesseWelling]

These instincts look mathematical to me. Now suppose we add modifiers to the situations. Let's say we want the robot to go pick a flower at the edge of the cliff. He'd "be afraid" but because we ordered him he would approach the flower with a set of calculations of the risk involved. He would calculate the safest speed for approaching without rolling off the cliff, backing off if vibration sensors picked up that the edge was about to fall off.

Not to mince words but we would program it in this situation to calculate the safest speed (in one way or another).
Don't get me wrong I think fuzzy controll is a great way to moddel certain kinds of behavior in robots, and you provide
an believable demonstartion but think of this....Why is the robot affraid? How would it know to be affraid? How would
you program that?

I'm not saying that it's impossible, but it is very much nontrivial. As for the General learner program....It is suposed to 'learn' to be afraid? or is that an instict we give it?

[Admin]

It is suposed to 'learn' to be afraid? or is that an instict we give it?

An example would be fire. You dont program 'fear of fire' into the bot. You program fear of pain (sensors detect damage somehow). Making the connection, the AI based robot then learns on its own to not get burnt . . . Its like a child learning not to touch a hot stove after the child touches it . . .

The cliff is a different example, because it assumes the robot can predict the future - 'falling off a cliff will hurt, even though this has never happened to me before.' No robot to date has this predictive ability . . . it has only been preprogrammed by the programmer . . .

[Arislan]

Because "fear" is a HEIGHTENED AWARENESS. When you are in the dark during a blackout, you fear you will trip over something and get hurt, so you begin a more thorough "sensor scan" of your own.   If you are afraid of being eaten by cannibals, lions, etc, during a safari in Africa, you start scanning the area for signs of movement and you begin an analysis of any movements to determine whether it comes from cannibals, lions or an army of baboons, whatever.  If you get on a plane and fear terrorists, you start scanning around to make sure none of the other passengers is very arab looking. In all these cases, you take a concept (falling down, being eaten, getting blown up) from your surroundings and you go through a series of reality checks.
The other type of fear is that of unknown things. These are not always rational. For example, a person, who believes in ghosts, spending a night in an old mansion, is likely to start acting in ways that dont make sense, like keeping the lights on, hiding their head in the blanket etc. But a "ghost" is something no one knows how to deal with, because it doesn't exist. If you had some hologram to scare that person, pretending it was a ghost, the person would not know how to deal with it and resort to that old universal imperative: running away.

All these things can be programmed. Fear as an indicator that heightened awareness is needed.

Today it has been one week since I had major surgery. I left the hospital saturday. I was told to be careful of getting a post-op infection so I, being "afraid", have been relatively alert to keeping my bandages clean.

Walter Fritz has emailed me his GL program source!! I'll post that source here later, because he sent it on a .sit format, which is Mac, and I'm on the PC right now.

[Admin]

An interesting question . . .

Would emotions enhance a robots' ability to survive in the 'real world'? Or be a flaw?

[Arislan]

I think we sort of have that answered in science fiction. Data from Star Trek functions fine without emotions, while the boy from AI suffers because of them.

No point giving them emotions, leave those to humans.