AI THAT DARE NOT SPEAK ITS NAME

Dick Pountain/05 November 1997/Idealog 39

I was sitting in a seminar room in Paris, observing some of the most impressive software I'd seen for a long time.  "Presumably KATE is an acronym for something that begins with Knowledge Acquisition?" I ventured. Michel replied rather sadly "Ah, I see there is no fooling you. Originally it was, but nowadays we mustn't mention Knowledge or AI because no-one would buy it".

The software in question is a highly successful fault-tracing system for Boeing 737 aircraft engines. You start by entering a symptom, in natural language, like "hesitant acceleration on throttle opening". Next the software asks you several questions to refine its understanding of the symptom - perhaps "Is there visible smoke in the exhaust?" - and then it presents you with a list of the best-fitting previously solved faults that exhibited those symptoms, and what the cure was then. Pick one that looks most like yours and click on it - zap, you go straight into the CD-ROM based workshop manual, showing a 'parts explosion' of the offending engine subsystem and a list of the part numbers you'll need to fix it. This program runs happily on a laptop PC, so a maintenance engineer can have it right on-site where it's needed. This is what computing was supposed to be about; the amplification of human ingenuity rather than dumb drudgery.         

During a week of software demonstrations I encountered this phenomenon several times over. The most interesting software - the smart software that did more than just look something up in a database - was all produced by ex-AI researchers who were exploiting their experience in advanced algorithms to produce vertical market applications. There was the system for routing telephone calls around failed exchanges; the system for managing credit card compensation claims; the system that builds you an insurance quote on-the-fly as you fill in a Web form; the system that deduces your interests as you browse the Web and builds a list of news pieces and links you might like.

These products all use algorithms and techniques that used to be the stuff of AI textbooks: logic programming, constraint solving, induction and decision trees, genetic algorithms, neural nets, linear adaptive filters. But in every case these algorithms are buried deep within a single-minded user interface that is geared precisely to a particular branch of business. And nowhere in their literature will you find the dreaded A-word, or E-word or K-word. These people are still traumatised by the memory of a sequence of great marketing disasters, of which they are the survivors. The disasters were the rising and breaking of three tidal waves of hype, the first called Artificial Intelligence, then Expert Systems, and then Knowledge Based Systems.
Each wave crashed to disaster not because the basic techniques were wrong, but because they were grossly overhyped. The very name Artifical Intelligence contained the seeds of its own destruction, because the word 'intelligence' positively invited people to anthropomorphise the machine, to imagine that it can think like us - something people are prone to without any prompting as the continued popularity of sci-fi testifies. Marketeers, who are quite like people in many ways, succumbed to this temptation too and made claims for the technology that it just couldn't deliver (any resemblance to object-orientation here is more than coincidental.)

That wasn't the whole of the problem though. The original AI products were all conceived as horizontal tools, their great role model being the spreadsheet which had just hit the jackpot via Lotus 1-2-3. All a product had to do was make a powerful algorithm like linear programming or induction available via a friendly spreadsheet-like user interface, and it would empower end users to develop their own smart systems. It didn't work that way though. Expert systems for example foundered because extracting the rules from human experts turned out to be far more difficult in the real world than in a university lab; a commodity trader may not even be conscious of any rules, he just follows a feeling in his guts or the pricking of his thumbs.  

By re-orienting their software to vertical market sectors the AI survivors take back control over these data sets, and over the tuning of their algorithms. That fault-tracing software that so impressed me works by Case-Based Reasoning using an induction algorithm called Quinlan's ID3, which is hardly new - I was writing about in Soft Magazine back in 1983. But as Michel Manago told me, over 80% of the work had nothing to do with the induction engine, but lay in the collecting of the case history data, cleaning it and massaging into suitable formats. The success of this application is precisely that it delivers the power of a well-known algorithm in a useable form, at last. Indeed the success of this application depends crucially on fault reports for 737 engines being stored in relatively few places. Making a similar system for say TV repairs would be an impossibility as the case histories only exist in millions of small independent repair shops, and probably on paper.

Of course my immediate thought was, wouldn't it be wonderful to have a fault-finding system this good for PCs, to sort out those maddening IRQ conflicts and missing-DLL horrors. And of course the case data does exist in relatively centralised and electronic forms, in Microsoft's and Compaq's and Dell's etc. help line records, and in magazines like PC Pro's Q&A pages. How about it chaps?