September 1998
Random Rob 3
I started writing this Random Rob several months ago. I note that the last one was in way back in June. Also, several topics in this Random Rob have been fairly recently covered by a Stu Chat. Who's copying who? Well, no-one actually, since AI and Car driving is something that has come up now and again, and ended with some very deep conversations indeed.
Why? There are several very nice elements about this problem, as a thought exercise. Nearly everyone can relate to the problem - even people who don't drive have ridden a cycle/push-bike, walked, rolled around in a wheelchair. This gives everyone some experience of the motion side of it - and driving a car, whilst at a higher speed is a much more limited problem in control terms than walking - you drive on a road, with markings, in a vehicle that has a limited steering capability. Also, this type of motion type control doesn't seem impossible to solve using standard technology.
So how would I do it. For a start, it really *has* to be a vision controlled system. An "along tracks" type system (even hidden in the road as electronic markers) is a bit like cheating. It would have to be, in my opinion, a multi-processor design. Of course it gains its advantages in this respect since it is a dedicated system, therefore doesn't suffer with load sharing problems (see below).
My "modules" would be as probably be as follows:
These are, of course, first pass guesses and maybe I'll take a proper look at it and present a more in-depth analysis. From the above you can see that I don't consider that Car Driving has an AI core. In fact, I would say it would be a "fairly simple" control system. Still not convinced? Have you seen the pictures of the missile in the gulf do I right-angle turn down a chimney. Now that's a control system. The biggest problem with this control system is the software design across multiprocessors, IMHO.
The whole subject of multiprocessor designs brings me to the question of "why are we still using single microprocessors in our desktop computers." Ok, there are things like keyboard processors, etc, taking the load off the main processor, but at the end of the day, the Macs and PC's on our desks are effectively single processor designs. There are of course some problems with the whole business of multiprocessor task sharing. There are two systems; asymmetric (where the one processor usually runs the system, whilst the other processor handles the user task) is OK, except for the fact that it isn't a general purpose solution. Symmetric multiprocessing, on the other hand is very good, but in actuallity it is difficult to achieve a high overall efficiency (i.e anywhere near a single processor at twice the clock frequency). Some of the problems are to do with hardware limitations, some are to do with information transfer limits, and some are to do with processor communications, but I expect most of it is based around current software design. People have been saying that multiprocessing will take off once the "wall" of technology has been reached. The problem is the wall keeps moving further and further away, and some believe we may never reach it.
AI - "brick on the road". This refers to a Stu chat comment on the car driving problem about avoidance and weighting of risks. I feel it is unfair to consider the problem of lack of knowledge in the control system, or AI if it was implemented that way. For a start, I probably wouldn't avoid the concrete in the road (even in Wales) for I don't know about it. Does that mean I shouldn't drive a car? No. What we must look at is the risk assessment, and "judgment" of what is safe.
The priority of a automatic car driving system must always be the for the safety of life, above all other goals. When faced with several different life threatening situations, the system must choose the safest course of action within the minimum amount of time. Anyone who has hit something on a Motorway (UK 70mph road/110kph road, where cars often do in excess of 80mph, although never me...), or had a blow out, knows you are acting on instincts. Knowledge of the way the car moves helps, but any movements are very close to a reflex. Having a lorry or cars next to you may make it a little more scary, but the main effect is a distinct feeling that you DO NOT want to hit it, thus modifying the reflex action.
Another subject I am very interested on is consciousness. There is a lot of information on the web about this. One which is quite good reading is "Why the Difference Between Quantum and Classical Physics is Irrelevant to the Mind/Body Problem" by Kirk Ludwig. I quote "
The problem is that from (2)-(4) we can deduce the negation of (1). That is the mind/body problem. (The consciousness/body problem is the mind/body problem restricted to mental properties the possession of which entail that their possessor is conscious or potentially conscious.)" The full article can be found at <http://psyche.cs.monash.edu.au/v2/psyche-2-16-ludwig.html> . It's published on Psyche <http://psyche.cs.monash.edu.au/> (volume 2), which is quite good for a look around.
The nature of consciousness. Some people have suggested that the human mind is based upon Quantum Mechanics, others have suggested that some other "unknown" science is involved. Whilst I don't believe either, at the moment, I will admit that the human brain is an extremely complex thing that we do not understand at the moment.
In the second case, I have a serious problem with inventing some new theories for something that doesn't appear to be any different to anything else we currently have, apart from complexity. If it were a case, for example, for an explanation of a puzzling, yet inconsistent (e.g. wave behavior as well as particle behavior) then fine. But no-one has yet clearly defined exactly what processes cannot be simulated using the "standard physics".
I also have problems with the application of Quantum mechanics in the brain. There are two related pro-points: (a) it can act as a way of expansion of parallism of the brain (increasing its' speed), and (b) it could be used to investigate multiple possibilities at the same time (by Quantum Superposition).
There are many negative points, however. Firstly, I'm not sure it fits in with evolution - where does the quantum theory get "put into" neurons? Only at human level? Personally I feel that the animals around us behave intelligently, and "process" a lot of information in a very similar fashion to the way we do. I do not really see a "quantum" jump in intelligence, just an incremental one. I will add, since humans have a habit of thinking themselves "better" than everything else, that, we share 98% of our DNA with chimpanzees, and our brains are the same size as dolphins. We are only a couple of million years different in evolution from our closest relatives, which is really nothing in evolutionary terms.
Noise is also a big problem when working with Quantum mechanics. Nature seems to work on the hundreds and thousands of atoms scale. This is usually outside the effects of Quantum ambiguity.
Lastly, I do not feel the ability to simultaneously evaluate multiple possibilities at a conscious level, so it must be hidden deeply if it exists at all. My point is why bury such a powerful tool so deep, rather that allow it access from the conscious brain - which would give a massive advantage to the species.
This does lead me to think, however, that my mind is probably computable and my consciousness isn't really as "real" as I'd hoped, and rather an illusion. I, myself now, appear therefore to be a series of almost pre-determined steps whose own consciousness is a creation of my own brain.
Mind you it still hurts when I touch the shelf when I'm getting my pizza out.
Something really cool is Quantum Computing. From my understanding, the first "quantum computing gates" were proved early this year using a Tunneling Electron Microscope. Because of the way quantum ambiguity works, you can actually explore whole data spaces simultaneously looking for a solution (e.g. key encryption spaces). Check-out the Center for Quantum Mechanics for more information.
Artificial Life is another subject which I'm very interested in. Basically it tackles the whole subject of creature "intelligent" (what ever that means) behavior by ground up modeling. Using such things as genetic algorithms, neural network, and much simpler technologies, gets things to move about a reasonable pace without massive amounts of computing hardware (like classical AI). A really good read is Steven Levvy's Artificial Life. Its a fairly old book, and I would give you the ISBN, but I've lent my copy to someone.
SOME REFERENCES: Center for quantum mechanics <http://www.qubit.org/>. Self organizing systems (Europe pages mirror) <http://www.calresco.force9.co.uk/> (American pages mirror) <http://www.calresco.org>. Also check out Yahoos' AI pages <http://www.yahoo.com/Science/Computer_Science/Artificial_Intelligence/>. Also these pages may be of interest <http://www.ed.ac.uk/~humphrys/philosophy.html> and <http://www.ed.ac.uk/~humphrys/eliza.html> Mind uploading? <http://sunsite.unc.edu/jstrout/uploading/MUHomePage.html> and <http://www.aleph.se/Trans/Global/Uploading/> (Of course, I have problems with peoples minds living forever, especially if we do not understand the mechanisms. For example - how does evolution occur? Does this mean we do not become any smarter? Who would mess with their own mind to improve themselves? This could be a bigger danger than anything else.). Interested in computing power in the mind? <http://www.hedweb.com/nickb/superintelligence.htm>
And lastly, Imac
Just a quick Imac note to say "you know when you've got a good product when everyone else copies you and comes out with the same thing 12 months later". Of course these copy-cats usually make the money, ironically enough. Don't believe everything you read about being first...
Till next time, Code On.
HIT YOUR BACK BUTTON or click Random Rob
Contact Rob at rob@lightsoft.co.uk