Re: AI - WHY?

Peter Tillotson (tillotso@liv.ac.uk)
Wed, 11 Nov 1998 06:03:48 +0100 (MET)

Hello,

By the sounds of it you are tackling a control problem in a real world
system. The first place to concentrate would therefore be to define the
problem acurately. Yes, I know this sounds obviouse, but in practice it is
far from trivial. Essentially any AI controller in real world systems must
take actions bassed on information. Usually the actions that can be taken
form a finite set. However, the inputs to a systems are a bit more tricky.
You have to decide what information is relavent and useful for the
development of your model. If you controller does not have the wright
information, how can it make accurate predictions.

Looking at the AI techniques that you have been discussed, I see no direct
reference to Reinforcement Learning. These are the techniques that Sebastien
Devaux was hinting at. They break an environment down into the states, using
probablistic mappings from the inputs. These states essentially describe how
a Learning automaton perceives its environment. Taking humans as an example,
we take many environmental measurments. If we look outside and the level of
illumination is very low we abstract this information using the
environmental state of night. Hence the information input into a Learning
automaton is abstracted into a more useful perception of the environment.
Indeed each level of abstraction reduces the amount of data to be handled at
each step.

Once the environmental states have been assigned, they are then mapped,
again probibalistically, to action. In the example above a sensible action
could be to turn on the light. Hence action change the state of the
environment we are trying to control. The change is then quantified against
our goal. In the scope of reinforcement learning these quantified measure of
how "good" an action was are fed back and used to update the probabilities
used when mapping. Hence actions that prooved effective in a given state
will be choosen more frequently.

Calling these systems intelligent, however, is a fallicy. They are not.
They're adaptive by the very nature that all the input, actions and internal
states are fixed. However using some of the techniques in AI this difficency
could be overcome. By generating new internal states and possible actions,
and removing older useless ones.

Peter Tillotson
Martin Sewell wrote in message <71qak3$vfc@news3.force9.net>...
>Hello (or perhaps: "Help!")
>
>I am currently doing a PhD in Intelligent Trading Systems (as in stock
>markets, etc.), and as the title suggests fully intended to utilise various
>aspects of Artificial Intelligence. As you'd imagine, the major goal is
>trying to predict an almost-but-not-quite-random time series.
>
>Neural networks, genetic algorithms, fuzzy logic - they all seemed so
>interesting (yes, even fun!), my only concern was which technique(s) were
>best applied to this sort of application.
>
>I now seem to be gradually eliminating all of the AI techniques which I set
>out to use.
>
>Goodbye fuzzy logic - why not do it properly using statistical inference
and
>probability?
>
>GAs? Just another search algorithm - adaptive simulated annealing (ASA)
may
>be better. (Okay, ASA could be considered AI too?)
>
>Neural networks? Classical statistical pattern recognition may be better
in
>many cases.
>
>I'm not aware of much work on the theoretical study of intelligent space.
>Maybe this should be tackled first?
>
>Any comments or suggestions would be appreciated. Either to rekindle my
>enthusiasm for AI, or to confirm my suspicions.
>
>Regards
>
>Martin Sewell
>martin@msewell.force9.co.uk
>
>

############################################################################
This message was posted through the fuzzy mailing list.
(1) To subscribe to this mailing list, send a message body of
"SUB FUZZY-MAIL myFirstName mySurname" to listproc@dbai.tuwien.ac.at
(2) To unsubscribe from this mailing list, send a message body of
"UNSUB FUZZY-MAIL" or "UNSUB FUZZY-MAIL yoursubscription@email.address.com"
to listproc@dbai.tuwien.ac.at
(3) To reach the human who maintains the list, send mail to
fuzzy-owner@dbai.tuwien.ac.at
(4) WWW access and other information on Fuzzy Sets and Logic see
http://www.dbai.tuwien.ac.at/ftp/mlowner/fuzzy-mail.info
(5) WWW archive: http://www.dbai.tuwien.ac.at/marchives/fuzzy-mail/index.html