I can see the merits in 'organism testing'. For higher animals, perhaps
the tests applied to chimpanzees, parrots, and dolphins would be
appropriate?
Training in these tests would be less critical. I see a test consisting
of first demonstrating that the organism faces a new/unknown problem,
and then the organism manipulates the environment to solve the problem
(achieving a goal). It would then be satisfying, if the tool/method used
to solve this problem could then be used to solve an even more
challenging goal.
If we look to simple animals, the behavior is basically hard-wired, with
biological genetic evolution providing the 'solution' to the survival of
the organism. Here the tests are more toward the 4F's
fight/flight/food/fuck, with the organism itself being manipulated to
suite the environment.
I wonder if the environment could be manipulated to favor survival of
organisms that could add and subtract ;) ?
(Brett.
On 27/12/2012 7:29 AM, Anastasios Tsiolakidis wrote:
As much as I would love that list, or any other list, to act as a
benchmark of intermediate progress, I think there is something deeply
wrong with it, albeit it may be simply an "extension" of the cheating
concerns of Ben et al. The list does sound like a very reasonable way
to check the cognitive progress of children, and with the danger of
psychologizing Piaget style, I'd like to offer another candidate item
that occurred to me while watching the animated Tron: run-and-hide, or
hide-n-seek. Tron is an action-packed franchise which includes a lot
of fighting, and a lot of running. If you have any experience with any
pre-arithmetic, even pre-linguistic children, and of course pretty
much any animal all the way down to protozoa, you can see that they
all, occasionally, have to "choose" between the f-f (fight or flee),
with two more f's (food and fuck) being of interest.
Of course these strategies are cheatable, a couple of simple
heuristics and you are "in", except that one cheating strategy may be
extremely desirable: run-and-hide controlled by proper environmental
measures and simulations that balance energy expenditure in running
versus the risk of being caught/derezzed/eaten, all again in the
context of food gathering and intelligence gathering on the predators
(whether literally or ritualistically the grown ups etc). Which in a
way returns to my pet peeve, that intelligence is pretty much a
nonsensical word if not applied to organisms and even populations.
Sure, I can solve some algebra problems for you without the 4 f's
playing any major role, but in a sense I have also cheated, my
organism is much more suited to hunting gathering in the jungle and
after a lot of "brain hacking" I have acquired some algebra skills
superior to masses of hunter gatherers but arguably inferior to what
proper mathematicians would like them to be. Rephrased, I use a couple
of heuristics that were given to me by the giants of mathematics to
get by, and I don't see how this is different from coding a couple of
heuristics to pass the "early AGI childhood" test.
From a different point of view, my intuition is that a child is a
brilliant intellect, not a so-and-so iffy developing intellect.
Certainly a short while after passing this kind of basic capabilities
test it would be ready to pass most survival and IQ tests, including
the Turing test, as an "odd person" perhaps but a person nevertheless.
Generally, most people will not go past the "minimum required" model
building and experimenting with the world, ie they will suck at
algebra and dancing and poetry and..., and I am fully confident small
children cope just as well at a minimum level long before we teach
them any fancy mathematics and grammar. So, at the end of the day I
would favor open-ended free-form problem domains and model-building -
I have previously suggested the generalized game contests as suitably
free-form and deep, naturally language is also free-form enough. Also
programs that would learn to program the hard way are of interest,
let's say a program screen scraping http://www.tryfsharp.org/ and
trying to build interesting programs online, or even participating at
Topcoder, now that would be something! For internal use one could make
use of the occasional Brett-ian checklist, and vary some parameters,
for example one could throttle the bandwidth of the input channel(s),
also tweak the output channel, and see if learning speeds up or if it
falls apart, hopefully discovering internal architectures that can
zoom in on the environment's invariants quite reliably and
independently of common environmental transforms. Output channel would
be, for example, the ability to have a rich interaction with the
environment by "turning the head", "piling" objects etc. I expect very
little from architectures that are "observe-only", such as statistical
learning. Of course an SVM or similar could be made to pass the Brett
test, but I could only call such a pass cheating :)
AT
*AGI* | Archives <https://www.listbox.com/member/archive/303/=now>
<https://www.listbox.com/member/archive/rss/303/23142996-0ca616e2> |
Modify
<https://www.listbox.com/member/?&;>
Your Subscription [Powered by Listbox] <http://www.listbox.com>
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
