From: ted.dunning@gmail.com   
      
   My own experience is that full-blown expert systems are   
      
   a) easy to build into demo systems   
      
   b) incredibly hard to build into full production systems   
      
   c) lead to really bad and difficult test and deployment cycles   
      
   d) rapidly become completely unmaintainable.  This happens for two   
   reasons.  The first is that you reach bug-fix equilibrium where every   
   fix you make breaks other things.  With very good test hygiene you can   
   increase the level where this happens, but it will happen at   
   relatively small scales.  The second problem is that the world changes   
   in subtle ways.  This means that your rules will have system errors   
   built into them.  Fixing this leads to complete collapse because old   
   rules are sabotaging the new rules in subtle ways (and vice versa).   
      
   Having gotten the inflammatory stuff out of my system, it is also   
   important to note:   
      
   1) ALL large decision systems are hard to maintain in production   
   settings.  They aren't like code and you can't guarantee functionality   
   with regression tests.  Expert systems are just worse than the   
   alternatives (such as other kinds of classifiers).   
      
   2) ALL large decisions systems will require some rule-like   
   functionality in production settings.  A key example is a large neural   
   network fraud detection system that shall remain nameless which has a   
   requirement that a customer cannot be contacted about a fraud alert   
   more than once every 90 days.  That sort of business constraint really   
   has to be implemented as a rule of some kind.   
      
   3) Very simple rule systems (no more than a few dozen rules or so) can   
   be maintained fairly reasonably.  It is the larger systems that   
   inevitably crumble.  If you only have rules to build on, you wind up   
   with large rule sets and your system will fail.   
      
   My qualifications are simply experience.  I have replaced knowledge   
   bases in a few industries and was responsible for creating the largest   
   identity fraud detection system ever built.  I also have lots of   
   friends who work on real-world automated decision systems and they   
   have a very long history of going up against rule-based systems (and   
   completely dominating them).   
      
   On Apr 4, 3:04 am, iulia.do...@gmail.com wrote:   
   > Dear all,   
   >   
   > Expert systems are often claimed to be better suited for certain kinds   
   > of problems as they support rapid prototyping, lead to better   
   > maintainable, reusable and adaptable code. I'm interested in empirical   
   > studies investigating these issues and making the claims above hard.   
   >   
   > Could you recommend some literature on the topic?   
   >   
   > Thanks in advance,   
   > Iulia   
   >   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)