From: timm@cs.pdx.edu   
      
   I've built anomaly detectors  in many ways before: association rules,   
   SVDDs, etc but the following  scheme is fastest and simplest I've ever seen.   
      
   Does this ring any bells for comp.ai-ers? Have others done this before me?   
   References or test data sets would be appreciated.   
      
   t.   
      
      
      
      
      
                                                                                    
   DETAILS:   
   ** A Bayes classifier is "eating" input data in "eras" of 100   
   instances. New data is classfied and then the frequency tables in the   
   classifier are updated.   
      
   ** If the classification accuracy hits a plateau, updating is disable.   
      
   ** If the accuracy falls off the plateau then updating is enabled again.   
      
   ** The learner watches data coming out of a simulator which mostly is   
   some nominal state but occasionally will move into an era or two of   
   off-nominal stuff.   
      
   ** Average max likelihood for each era acts like an anomaly detector.   
   When the data moves into a novel error cases, the likelihoods drop by   
   two orders of magnitude- a pretty clear indicator of "hey, this is new   
   stuf"   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)