BBD's
Noah Gift
--
Cheers,
Noah
Graham Higgins
Hash: SHA1
Some further detail is available in:
http://vesicle.nsi.edu/nomad/iros2003_jlk_gme.pdf
One slight drawback to the "selectionst" approach is:
"While performance improved with training, it never reached
perfection and occasional 'mistakes' were made. This unpredictability
is a general property of selectionist systems ..."
This is true and probably inevitable.
Cheers,
Graham
http://www.linkedin.com/in/ghiggins
-----BEGIN PGP SIGNATURE-----
iEYEARECAAYFAkoZQHEACgkQOsmLt1NhivwsjQCfVP3uI0cI+YzZsRzkX7xfKJRD
waEAn1xqGuY4b5lkvwqAQ/EQo3d8LTn5iQCVAgUBShlAcVnrWVZ7aXD1AQK6vAP/
We9Ko3SEzdR5HXRhiGnc6Pt1Uma28QnfJZDoH/EZkJr0f8R8BMxPSROIzR4Xxhgq
PepztZdtqvYEfUBdxIE4akYuait2WJlqaGOTpNIKzQE5cyIvCKWlV7FUPIxJjU70
yiPOdS47phSKK2GDmDfugWl2kbX5u9zLuP3QT9DFxQU=
=yqxP
-----END PGP SIGNATURE-----
Eric Wald
> One slight drawback to the "selectionst" approach is:
>
> "While performance improved with training, it never reached
> perfection and occasional 'mistakes' were made. This unpredictability
> is a general property of selectionist systems ..."
>
> This is true and probably inevitable.
That isn't always a bad thing. For example, a game AI needs some
unpredictability to avoid getting beaten by the same old strategy over
and over. A feed pre-reader that occasionally keeps something you might
not like is also more likely to surprise you with something good that a
simpler system would have thrown out.
- Eric
Graham Higgins
Hash: SHA1
On 24 May 2009, at 23:27, Eric Wald wrote:
> Graham Higgins wrote:
>> One slight drawback to the "selectionst" approach is:
>>
>> "While performance improved with training, it never reached
>> perfection and occasional 'mistakes' were made. This
>> unpredictability
>> is a general property of selectionist systems ..."
>>
>> This is true and probably inevitable.
> That isn't always a bad thing.
Apart from being a special hell for debuggers, it also raise the
question: if unpredictability is inherent, how is "error" to be i)
defined and ii) recognised by the system?
> For example, a game AI needs some
> unpredictability to avoid getting beaten by the same old strategy over
> and over.
Call me a GOFAI stick-in-the-mud if you like but isn't the "AI" in
"game AI" essentially just a marketing term? If the difficulty is an
inability to learn it would seem to be a long running jump and grab to
get to "AI" from there.
> A feed pre-reader that occasionally keeps something you might
> not like is also more likely to surprise you with something good
> that a
> simpler system would have thrown out.
I freely admit to having no experience at all with RSS pre-readers, I
keep my RSS on a very short leash. However, isn't it less to do with a
necessary inherent unpredictability in the approach and more to do
with the inherent unpredictability of the domain -- the meta-modelling
of my preferences? In this matter I'm an adherent of Langian
interpersonal psychology (c.f. Laing's meta-perception -- http://davidakenny.net/ip/meta_acc.htm)
and as people struggle with this degree of modelling complexity, I
have little confidence that a mechanical approach can offer much in
the way of a useful solution.
Cheers,
Graham
http://www.linkedin.com/in/ghiggins
-----BEGIN PGP SIGNATURE-----
iEYEARECAAYFAkoZ6xUACgkQOsmLt1Nhivy1fACfZzfTl30e7Snrv/SD0uCmNqxo
6b0AoOmy561uTBBEtaKdq9CekvUITYPSiQCVAgUBShnrFVnrWVZ7aXD1AQLwGQP+
Juqp0UVOi7iq08mxvdHDyjbVQ+n2x70NNPbRdiP7B+RvBPs2YpgNvhiZ3v3kMnCw
6qpjDsuhR2dlurhE1rgMqiKU19bGU5AaijWZXE5+Pc5KNZ4030cVOYReuT0KSsvl
SuakPNLg6uvSDh5uCVPxxdJUbO+QPJrAlFn9yhFK6FA=
=xLoq
-----END PGP SIGNATURE-----
Eric Wald
> On 24 May 2009, at 23:27, Eric Wald wrote:
>> That isn't always a bad thing.
>
> Apart from being a special hell for debuggers, it also raise the
> question: if unpredictability is inherent, how is "error" to be i)
> defined and ii) recognised by the system?
Very good questions, and probably domain-dependent. It seems that
BBDs are best in domains where certain experiences can have negative
consequences without being considered "errors" in the technical sense.
I get the impression that BBDs aren't meant to be "debugged" like a
regular program; instead, they are "trained" like an animal. However,
the design of the inputs and subnets can certainly affect results, and
may be subject to a redesigning phase for the next BBD.
If a feed pre-reader lets something through that you would prefer not
to have wasted your time on, that's a mistake that it could learn from.
If it filters out something that you would have wanted to see, though,
that's much harder to let it know about.
If a game-playing AI is choosing between moves A, B, and C with a best
strategy (in the game theory sense) of 80%, 20%, and 0% respectively, it
certainly wouldn't be an error to choose B once in a while. However,
would it be an error to ever choose C, particularly after the initial
learning phase? Would it be an error to always choose A? To choose B
more often than A?
>> For example, a game AI needs some
>> unpredictability to avoid getting beaten by the same old strategy over
>> and over.
>
> Call me a GOFAI stick-in-the-mud if you like but isn't the "AI" in
> "game AI" essentially just a marketing term? If the difficulty is an
> inability to learn it would seem to be a long running jump and grab to
> get to "AI" from there.
The "AI" in "game AI" as used by the DAIDE is more a goal. There's a
reason we tend to use "bot" instead of "AI" to describe the current
computer clients. Granted, I did recently get a book titled "AI for
Game Developers" which had very little in the way of true AI, instead
focusing on ways to make NPC movement seem vaguely natural.
>> A feed pre-reader that occasionally keeps something you might
>> not like is also more likely to surprise you with something good
>> that a
>> simpler system would have thrown out.
>
>
> I freely admit to having no experience at all with RSS pre-readers, I
> keep my RSS on a very short leash. However, isn't it less to do with a
> necessary inherent unpredictability in the approach and more to do
> with the inherent unpredictability of the domain -- the meta-modelling
> of my preferences? In this matter I'm an adherent of Langian
> interpersonal psychology (c.f. Laing's meta-perception -- http://davidakenny.net/ip/meta_acc.htm)
> and as people struggle with this degree of modelling complexity, I
> have little confidence that a mechanical approach can offer much in
> the way of a useful solution.
Entirely possible, particularly given that individual preferences are
subject to change over time. Even worse, this domain includes the
aspect that you might not want to see articles that are significantly
similar to previous ones, or that simply link to a previous one without
significant commentary.
Such complex domains may in fact be where BBDs or other neural-net
programs excel over more mechanical programming techniques. After all,
isn't that why we're considering them as flagship AI problems?
- Eric