VISIT REPORT LBurnard

University of Surrey

IUSC Workshop ON 3-4 Jan 1984

Over 132 people attended this workshop on knowledge-based or expert systems, although no-one I spoke to appeared to have any very clear idea of quite what such systems were or might be used for Their chief components appear to be firstly a knowledge-base, which is not quite the same as a database because the entities represented in it include both vague or imprecise quantities and equally vague rules about their significance, secondly an inference engine, which is a processor capable of both using and adding to the knowledge base, and finally a fairly sophisticated man-machine interface, usually in natural language albeit within a fairly restricted semantic domain Such systems are VERY EXPENSIVE INDEED to construct, and thus are to be found only amongst the VERY RICH, e.g the oil industry, the US Dept of Defense and any UK Computer Science Dept which has got its hands on some Alvey funding.

The first speaker, from RACAL, appeared to be rather unnerved by the size of his audience and was also clearly under instructions not to reveal too much about his subject, a system for making sense of the vast amounts of data obtained during exploratory drillings in the North SeaSince exploitation of any resulting oil wells is usually a co-operative venture in which the cut each member of the consortium receives depends crucially on the size of reserves found at a particular place, it is rather crucial to get unbiassed and accurate information A 5% error could mean a difference of £350 millions The system was computationally unfashionable, being model based rather than rule based; that is, it works by constantly revising and improving on rules of thumb derived from observations rather than a priori knowledge There is no attempt to model causality in its knowledge-base which means among other things that the system has no way of reconciling contradictory conclusions reached from different premises It runs on a special purpose AI machine (the Symbolics LISP machine) which also supported numerical analysis in Fortran and a conventional database management system (unspecified).

This presentation, (which should have been more impressive given that it proved to be the only one describing a real fully functional system) was followed by two sales pitches, one for SAGE, an expert systems shell marketed by SPL, and

the other for REVEAL, a decision support system marketed by TYMSHARE Both are also available from ICL and cost an arm and a leg, except that there is a version of SAGE for VAX VMS currently on offer with a massive educational discount.

The SAGE knowledge base is created by the user as a set of rules and objects which the inference engine then uses to establish a goal (e.g "This user is liable to have a heart attack") by means of a dialogue and backward-chaining reasoning (i.e People under stress are more likely to have heart attacks, so I must establish how likely this user is to be under stress; to establish which I need to establish how many Advisory Sessions he has done in the last month; to establish which.etc It's called recursion!)Because objects can have a truth value (probably false, nearly true, etc.) rules involving combinations of such objects are said to exhibit fuzzy logic; thus, for SAGE, if p is 0.3 true, and q is 0.8 true, then p AND q is 0.3 true, and p OR q is 0.8 true Various smart alecks in the audience pointed out that this was a barely adequate fuzzy logic, with which the speaker had the good sense to agreeNevertheless, I think SAGE would be a good way to learn about expert systems and might even be useful for somethingREVEAL by contrast proved to be a souped-up financial modelling system with little to recommend it apart from the use of fuzzy logic in both its database system and its English language interface, so that the modeller can say things like "List all tall rich blondes with large bosoms' without having to specify what 'tall', 'rich' and 'large' mean exactly.

The conference dinner was satisfactorily sybaritic and was followed by the traditional drunken gossip in the bar, during which several people lurched up to me and said "Famulus?" in a menacing sort of way.

The following morning began with a presentation on Salford's new Prolog system (which was also available for hands-on use during the morning) This is (like Poplog) a hybrid in which all those bits that are difficult or impossible to do in pure Prolog (like reading from files or doing assignments) are hived off to another language, in this case LISP It also supports a better syntax for grammar rules, a Fortran interface and -yes- floating point arithmetic! But it is only available on Prime and is still under development.

The trouble with Prolog of course, the next speaker pointed out, is that it is really only practicable on machines that don't exist and are unlikely to for the next ten yearsSomething called "a given sector of my client base" was however attuned to it and so his company (Cambridge Consultants Ltd) were investigating its usefulness in

real time (but not as yet real) applications Their investigations had however thrown up nothing that could be communicated to us other than a list of the available versions of the language and some fairly superficial remarks about it.

John Baldwin from Bristol and Ron Knott from Surrey re-established the intellectual credibility of this workshop by the next two papers which described programming languages capable of building knowledge based systems Baldwin described his Fuzzy Relational Inference Language, which is a logic programming language incorporating uncertainty in a far more thorough and mathematically respectable way than SAGE Other buzzwords included the blackboard model, parallel architecture, a self organising knowledge base and a dataflow machine His paper was the only one that could properly be said to manifest state of the art knowledge, although Knott gave an interesting survey of the available functional programming languages, typified by LISP, variations of which are very much still alive and kicking.

After lunch, delegates were restored to good humour by Tim O'Shea (OU) who gave a good survey of Al-supported computer aided learning systems Apparently the crucial question to ask someone trying to sell you a computer tutor is "What sort of task difficulty model do you have?” and, if this fails to floor him (or it), 'Does it support dynamic student modelling?' The speaker was good enough to kindle enthusiasm for his subject which is saying a good deal in this case Apparently the process of producing CAL systems is generally known as "authoring" The people (or things) that do it are presumably known as "authorers".

Finally we were given an interesting account of the current structure of the Alvey directorate, and even some figures about how its huge funds were being split up Whether or not these large sums will succeed in reversing the process of rot supposedly created by that convenient scapegoat the Lighthill Report, or whether they will simply prove a useful way of cutting down funding for other academic research remains to be seen, though the fact that ALVEY expands to "All Large Ventures Except Yours” may be taken as some indication.