Can Curiosity Be Programmed?

destinyland writes "AI researcher Jurgen Schmidhuber says his main scientific ambition 'is to build an optimal scientist, then retire.' The Cognitive Robotics professor has worked on problems including artificial ants and even robots that are taught how to tie shoelaces using reinforcement learning, but he believes algorithms can be written that allow the programming of curiosity itself. 'Curiosity is the desire to create or discover more non-random, non-arbitrary, regular data that is novel and surprising...' He's already created art using algorithmic information theory, and can describe the simple algorithmic principle that underlies subjective beauty, creativity, and curiosity itself. And he ultimately addresses the possibility that the entire Universe, including everyone in it, is in principle computable by a completely deterministic computer program."

Comments

Physics of computing the universe

[BitZtream]

And he ultimately addresses the possibility that the entire Universe, including everyone in it, is in principle computable by a completely deterministic computer program.

The problem with this is that you need to be outside the universe in order to do so, you can't calculate the universe from within itself any more than a VMWare can run a machine faster than the host processor.

You'd also need more mass in your computer than exists in the universe, observable or otherwise.

So sure, I'll go with the theory that its possible, just not by any thing in our universe.

Likewise, nothing in our universe could leave it to perform the calculation elsewhere, as doing so links the two realities together, so you now need to simulate both.

Everything is interconnected and the very act of attempting to simulate the universe changes the simulation. Every new version of the simulation would instantly require a new version to take into account the changes from the previous version.

The theory is ... cute at best, but unworkable.

Re:Physics of computing the universe

[biryokumaru]

VMWare should, in theory, be able to simulate a system faster than the host processor, as long as it doesn't actually run that fast.

We should, in theory, be able to simulate the universe, just not as fast as the universe actually moves.

Besides, I bet we can just gloss over a lot of the boring bits and stay within a margin of error while ultimately simulating faster than the universe is actually transpiring. That doesn't seem unreasonable.

Re:Physics of computing the universe

[JWSmythe]

    If/when a true AI exists, it will need some randomization to make it curious. Sure, you can chart point A to B to C, but what if randomly it skews off to somewhere just west of point Z enroute, and observes.

    That doesn't have to be a physical route. It could be as simple as taking a random word from a dictionary, searching that on your favorite search engine, taking a random result from there, and then following the result from another random word. An unpredictable path, but that's what brings any of us to enlightenment. If you just went from home to work and back every day, and never turned down the wrong road, just to see where it goes, you'll never discover what is really out there. What is your universe? I've known so many people who only know points A and B, and never even considered point C, much less all the wonderful things to experience in between or beyond.

Re:Physics of computing the universe

[Profane MuthaFucka]

There's a workaround. You don't need to simulate the entire universe at one time, and there's no way that anything inside the universe would ever be able to tell that huge swaths of the universe aren't being actively simulated.

Reasoning: If a universe simulator needs to have more states than exist inside the universe (we're both assuming this) then then any process which verifies the universe simulator would also need to have more states than exist inside the universe. Therefore, the universe can only be fully simulated from outside the universe, and you could only determine that the universe was fully simulated from outside the universe. From inside the universe, all you could simulate and all you could check would be a subset of the universe you are in.

So you could actually pull off a neat trick, just like the human eye does. The human eye doesn't actually see clearly except in the very middle of the vision. But, wherever you look, whatever you're looking at is clearly resolved. Your brain gets the distinct impression that it's looking at the entire scene clearly, except it's not. Only the part that it's actually looking at is clear.

Well, that's enough hacking of the universe for today. I need a beer.

Re:Physics of computing the universe

[Have Brain Will Rent]

That's probably why God invented QM - as a space/computation saving device.

Re:Physics of computing the universe

[p00ya]

Even at a slower rate than real time you cannot simulate the universe from within the universe. This has been proved by Cantor diagonalization (see Wolpert's "Physical limits of inference" paper).

Re:Physics of computing the universe

[mcgrew]

At the atomic level there's a lot of randomness.

Can we be sure? What seems random may not in fact be truly random. The flip of a coin is considered random, but if you could account for all the variables with enough precision; angle of the coin, angle of the thumb, force of the flip, distance to the floor, etc, you could likely predict each and every toss.

Rather than being random it could be that it's just more complex than we know, or that we can't determine the variables with enough accuracy. What is the exact value of PI?

Coding For Patterns of Anomalies

[neorush]

Aren't we really just talking about coding for patterns of anomalies? We know how to code for patterns, we know how to code for anomalies. Isn't it a matter of processing huge data sets and looking for patterns that have not been recorded before? Of course, you could argue that whether or not the pattern is relevant is the big problem, but curiosity is not necessarily about relevance.

of course it can

[walkoff]

When I was a fledgling programmer in the 80s I worked on some financial AI programs for a bank with some very smart people with lots of letters after their names and programming artificial curiosity was assigned to me. After some thought and a lot of dead ends I managed to program a reasonable (for our needs) facsimile of curiosity by assigning weights to the various pathways the program was evaluating and making those weights tend towards 0 (curiosity satisfied) or 1 (Curious) without ever reaching the final values. By having the program modify the weights and make decisions on which paths to follow based on those weights the program acted as if it was curious and came up with several interesting results that were completly unexpected.

Only as smart as...

[v(*_*)vvvv]

If curiosity is a behavior, then it should be pretty straight forward. In fact, depending on how you define "curiosity", then there are already many examples of programs that are curious. Google or Bing or any web crawler is definitely "curious". A satnav that searches for the best route from point A to point B could be "curious"...

A robot is only as smart as its smartest programmer.

And he ultimately addresses the possibility that the entire Universe, including everyone in it, is in principle computable by a completely deterministic computer program.

The problem that is often ignored with this and similar claims is the problem of observability as illustrated in areas such as quantum physics, and even economics.

You cannot calculate the behavior of a black box without opening it. If opening it alters the state of its contents, then it may even be impossible. And if you have no means of observation to begin with, then it is downright impossible. Before you can claim you can calculate the next moment in time, you must be able to claim you have observed and know all the variables within the system of interest.

Twilight

[serps]

This article reminds me of the short story Twilight by John W. Campbell. I read it when I was a kid and it left a lasting impression that, should humans lose their curiosity, the striving for knowledge might yet continue.

And then when I read about the current state of the education system, I get just a bit worried...

Re:Of course it can.

[Fantastic Lad]

Of course curiosity can be programmed. What are humans if not big, fleshy, biological machines of sorts? Granted we do not work like computers do, but the underlying processes are still structured and computational--if the brain were chaotic it wouldn't work.

~waves hand~ Speak for yourself, Mister Roboto. ~/waves hand~

But seriously, this is a really fascinating question. Souls aren't handed out like candy. You have to build them through main force; by actively choosing to be aware from moment to moment. What I am finding to be the biggest challenge in that requires the supreme effort of recognizing one's own automatic nature and cleaning the gunk out of it.

Every time some subject comes up in conversation which makes me twitch or sweat or want to pull away, THAT indicates a piece of gunk. Each time I want to fall back and use a comfortable and proven behavior routine to deal with a given moment, THAT indicates a piece of gunk.

After one does enough work, you begin to see very clearly just how messy and automatic the people around you are. -These days, I find I am constantly aware of people's programs and little acts, why they work and what they are designed to do, and where people get stuck running those silly programs over and over day after day, year after year without ever stopping to ask, "What is the real me under this?". The soul is that part of us which is capable of recognizing the automatic nature of the brain and body and stepping in through an application of Will to interrupt the code execution.

It's difficult and the ego doesn't like it at all; Any suggestion that one is a robot is usually met with disgust and fear, if the accusation is even understood in the first place. The Ego is, I think, a foreign installment designed exactly to keep us from performing that self-examination. With the Ego in place and strong, there is no hope of breaking out of the cage of automatic behavior.

Like I said, a fascinating topic.

-FL

Re:Show me the runny

[aaaaaaargh!]

You're right, there is not only the minor problem that it can't be implemented. The utility function is the another big issue. Programs like this theoretical Gödel machine or working machines like neural networks and kernel method implementations depend on a utility function that can tell you whether you're getting closer to a solution or not (notwithstanding misleading local maxima and minima), and in order to have such a function you already need to have an intimate understanding of the problem at hand. Unfortunately, there is no good, computable theory for what "understanding a problem" might mean.

Re:Of course it can.

[Fantastic Lad]

Yeah, but your talk of "Ego" and "the soul" is all gobbledygook.

Well, I didn't mean to suggest that YOUR computer was coded by a genius or that your hardware wasn't found in a box of rejects from the late 80's. You'll have to forgive your maker if you can't keep up.

As for the soul. . . Well, that's not a winnable debate one way or the other, so your call of 'gobbledygook' is no more or less useful.

That being said, did you feel a stab of irritation upon completing the first sentence? The ego would have been the item making that particular sensation. Sorry. It was just a demonstration. No harm intended. I'm sure your wetware is fine.

-FL

Re:The reactive mind

[Fantastic Lad]

Scientologists would say the same about the reactive mind.

Scientology is one of the premier examples of spiritual exploration gone horribly awry, co-opted by the very forces which seek to keep people locked down. Their trick is to take a bunch of good ideas and quietly interweave them with creeped-out insanity.

The New Age bookstores are filled to brimming with fail-safe nets designed to catch people who fall out of the matrix.

There IS a path, but it takes a lot of comparative study, source-checking and in the end, rolling up your pant leggings and getting out there yourself to figure out who the heck is on first.

-FL

Re:Of course it can.

[Fantastic Lad]

Here's what I see: A desire to assert some sort of intellectual superiority by reducing others to automata, thereby justifying your perceived self worth as a "free thinker," when in reality it could be effectively argued that you're simply falling back on your own "automatic program" of taking each response != "validation" and categorizing it as an "automatic program," which, ironically, would both validate your claim AND demonstrate its irrelevance so long as each program is not identical, which appears to be the case.

I'm not a free thinker. Nobody is. And you're right; the ego is always eager to assert dominance. As much as I try to put it aside, it remains a force to be reckoned with. But it is a sliding scale; at the one extreme, a person can become a slave to the Ego, flying into a rage when challenged, while on the other hand s/he can laugh at and make an effort to disengage from the emotional spurts offered by the Ego thus minimizing its control. I once had it described to me thusly; think of the Ego as a bowling ball which is sitting in the middle of the lane. If you pretend it isn't there, it will screw up all your carefully intended strikes. If by contrast you recognize its presence and that it is not going away, you can try to bowl around it. -Not the best metaphor but workable.

The != "validation", as you put it, in this case was in fact a classic ridicule evasion/attack, a type which is well understood. As such, it seems to me entirely reasonable, unless you can add something I am not aware of, to think of it as exactly that.

We are all affected by our internal programming, but valid observation remains entirely possible. Accuracy is improved when multiple people work to bowl around their egos and point out errors and ego-traps and spot programs running in each other.

So what do you think? Was the poster making a joke for some other reason than the one I suggested?

-FL