Re: Humaniform robots... yea or nay?

Eivind wrote:

Still, it illustrates the dangers of thinking that such an evolved
algorithm does what you think it does. It's easy to think that "whatever
the algorithm does, it's obviously basically working", despite the
solution in this case being completely bananas wrong. (well, right
inside the universe we had created for it, but useless outside it)

Sure. Bad testing will do that.

The course left most of us unimpressed. Though in retrospect it
should've been obvious that the simplistic naive implementations we made
of various concepts where unlikely to do anything radical.

Well, at least the course left you with the realization that genetic algorithms/programmings are not to be trifled with if you have a lousy simulation environment/test case. Which goes for all other software processes, as well.

I think there's an obvious problem of generalizing this experience to genetic algorithms/programming in general. You were doing this for a school course, which means that invariably you're being asked to do something approachable (so the average student can be complete it in time) and that you can get results on. That _doesn't_ mean that the choice of applications is at all useful. This is routine (and not entirely inappropriate) at the level of education: You're being taught how to use a _technique_, not solving an amazing and impressive problem.

This is why, for instance, routinely students are given problems such as reversing singly-linked lists, computing Fibonacci numbers recursively, or things like "accomplish task X, but without using the obvious and most useful algorithm A." These are fantastically stupid things to do outside of academia (if you really were faced with them to the point it mattered you'd use different data structures of algorithms entirely), but the point isn't that this is the right way to solve this particular problem, but that you're trying to learn that particular approach (whether or not it applies to your current problem). The "X" and "O" recognition case is quite easy; that would be an easy solution for heuristics or neural networks trained with back-propagation.

I dunno. Are AI-research today turning out actually useful results
and/or programs ? Or are they still stuck at "this is kinda interesting
and almost useful, though nothin that a traditional algorithm wouldn't
do better at 1/10th the effort" ?

Well, artificial intelligence research hasn't hit upon an amazing solution to everyone's problems, or you would have heard about it by now :-). At this point people are still exploring different alternatives. But then, a lot of the things that would at one point have been considered the forefront of AI research are now becoming mainstream: facial recognition software, speech recognition and synthesis, and so on.

The more amusing -- but very revealing -- examples of this kind of
tomfoolery are when genetic algorithms/programming simulations end up
exploiting a _bug in the simulation_, rather than bad data. That's when
it really hits home that evolution works: It "finds" whatever ways it
can of getting ahead, and all without any plan in mind beforehand.

True enough. But that makes it dangerous to trust the results. Because
your evolved car-driving program may be able to avoid pedestrians in the
simulation because of a glitch in the simulation, not because it
actually knows how to drive.

Well, sure. And it's dangerous to trust software written the traditional way because people who have written it think it's right and it's gone through some basic testing. The class of potential problems might be different, but in terms of software development and quality assurance, this isn't anything new.

It's highly dangerous to trust any software anyone gives you on a silver plate, ever. In fact there's surely a high correlation between the likelihood that someone offering it exists it's "bug-free" and how many problems it has; only people who do not truly understand the software development process ever make such claims.

--
Erik Max Francis && max@xxxxxxxxxxx && http://www.alcyone.com/max/
San Jose, CA, USA && 37 18 N 121 57 W && AIM, Y!M erikmaxfrancis
Man is a hating rather than a loving animal.
-- Rebecca West
.

Relevant Pages

  • Re: Solution to the halting Problem?
    ... running on a computing machine or a simulation of a computing machine, ... simulated algorithm without disturbing its operation. ... >> subjects halt or not. ...
    (comp.lang.cpp)
  • Re: Variational calculus : A drop of water
    ... Simulation of a Dripping Faucet by Nobuko Fuchikami, ... By taking a stable equilibrium shape as an initial ... we describe a variational algorithm to examine ... Our algorithm of simulation of dynamics ...
    (sci.math)
  • Re: need a good implementation of pseudorandom generators
    ... use SPSS to generate the random mumbers. ... The Marsaglia algorithm is also available as an option. ... pseudorandom generator for it. ... But if i run the simulation 100 times then on all 100 times it gives ...
    (sci.stat.math)
  • =?utf-8?q?Q(=CE=BB)-learning_algorithm_question?=
    ... For Watkin's Q-learning algorithm; why would someone ... simulation and on real robot show that \lambda=0.5 gives a solution ... [ask your news administrator to fix the problems with your system. ...
    (comp.ai)
  • Re: Algorithm ideas
    ... > new problem that requires a non-trivial algorithm, ... Data fusion (fingerprint + voice, ... for example) might be useful for robust recognition. ... States of America never goes to war because we want to, ...
    (sci.crypt)
Loading