ijd@camcon.UUCP (02/17/87)
I'm currently engaged in a small research project looking at the problems
of updating AI databases in real time. [By database, I mean that collection
of information that I am currently reasoning against - not the large
commercial variety.]
A typical problem in the domain is that you have N (where large(N)) sensors
attached to a process plant all throwing lots of data with low information
content at an intelligent fault diagnosis system. The system has to cope
with contradictory data, have good coverage of the incoming signals, but
still be able to respond quickly to high-priority situations.
The particular issues that I am concerned with are:
(1) what are the representational inadequacies of current AI notations
that are suited to doing real time problems and/or handling noisy and
contradictory data?
(2) what are the computational costs of using such notations?
Primary choices for handling mucky, changing data are the RMS family
(Doyle, de Kleer etc) and other non-monotonic logics, so these are typical
of the notations that I am referring to. So the issues become: what can't
you do with them, and what would it cost anyway?
The questions I would like to submit to net.land are:
o anybody doing any work on extending non-monotonic notations in wierd
directions (eg integrating them with uncertain inference techniques)?
o anybody got any pet real-time AI problems?
o anyone else working in the real-time field?
Please mail responses directly to me, and I will post a summary for
discussion later.
Thanks in advance,
Ian.
--
!! Ian Dickinson Cambridge Consultants Ltd, AI group !!
!! Voice: (0223) 358855 [U.K.] Email: ijd%camcon.co.uk !!
!! uucp: ...!seismo!mcvax!ukc!camcon!ijd or: ijd%camcon.uucp !!
>> Disclaimer: All opinions expressed are my own (surprise!). <<