Re: Re: HA philosphy

["David Gumbrell" <david.gumbrell@xxxxxxxxx>]

I think you may be underestimating the training problem. 20q has the
benefit
of getting millions of people to train its' system and who probably don't
care that much when it gets it wrong.

The TiVo example wouldn't be any different with more sensors, it would just
require more training and still occasionally throw up misbehaviours that
will be impossible to debug because you won't know which factors led to
that
behaviour. If the "pain" of that misbehaviour outweighs the
general gain of
all the correct behaviours then the system has failed overall. So,
extending
the TiVo example, it gets 1 point for correctly recording, say, Neighbours,
because it wasn't that difficult to do with conventional video, but gets
-50
points for recording Big Brother instead of the film on the sex life of
otters on the other side which I would really like to have seen and is
unlikely to be repeated. Similarly, if a SmartHome puts the lights on full
blast in the middle of the night because you stayed up late the previous
night, it isn't going to get much credit for all the times it put the light
on when you wanted, because a dumb light switch could have got that
correct.


There's some interesting research done in a fully adaptive
house<http://www.cs.colorado.edu/%257Emozer/house/>.
Depending on the cost model associated with dissatisfaction (academic
equivalent of WAF), it doesn't always seem to outperform traditional
methods. Equally, some of the papers give plenty of anecdotal evidence for
the failure of programmed systems. The researcher admits that incorrect
behaviours can be irritating - I wonder if he lives alone ? His
observations
on what works well is interesting, eg heating control has much greater
acceptance of near misses since people are relatively insensitive to small
variations, whereas getting the lighting wrong creates much greater
dissatisfaction.

He has a bit of a downer on explicit user signalling  (eg "sleep
mode"), but
I think this  is a  "purity" problem.  For example,  very many
people  set
an alarm clock before going to sleep,  why should this not also be used (or
substituted) to indicate the  current user pursuit is  sleeping, rather
than  some other behaviour ? Similarly, one of his anecdotes was a couple
who couldn't handle putting a SmartHome into "I'm home" mode. But
how many
of us perfectly happily turn off a burglar alarm on returning home ? You
could argue that these are sensors, and yes, they could be treated as such
and augment the adaptive systems input. But they could also form an updated
operational UI in a programmed system.

Ironically, he notes that the adaptive system tends to work best for people
with regular routines, to the extent that occupants may actually regularise
their behaviour to fit in with the house. Talk about the tail wagging the
dog. And programmed systems are condemned for enforcing rigid routines :-)

Most likely will be some kind of hybrid approach based on some adaptive
behaviour within manually programmed constraint rules to avoid the
exceptional cases, especially those with very high dissatisfaction costs.

It's a hard problem !

Cheers,

David

On 10/4/06, Tim Fletcher < timfletcher@xxxxxxx> wrote:
>
>   I think if you have enough sensors almost anything is possible.
>
> Consider that 20 questions toy - choose anything (dog, knife,
Happiness,
> thunder) and it guesses it by asking 20 questions most of which seem
> irrelevant. The interesting thing is if you look at the website (
> http://www.20q.net/
) they say it doesn't matter what the real "truth"
> answer is so much as what people say it is - a similar sort of
"Learn"
> function from a network of sensors should be able to achieve whatever
> outcome was desired.
>
> My two cents.
> Tim.
>
>
>


[Non-text portions of this message have been removed]



** Sponsored by http://www.BERBLE.com **
all the Cool Stuff, in one Place