Case Based Reasoning approach to automatic comparison of models

Timo Ala-Kleemola
Institute of Signal Processing, Tampere University of Techno

Marja Merta
Institute of Signal Processing, Tampere University of Technology

Ari Visa
Institute of Signal Processing, Tampere University of Technology

Pekka Johansson
Institute of Signal Processing, Tampere University of Technology

     Full text: Login to access
     Last modified: June 16, 2005

Abstract
In recent years pattern recognition techniques have been utilized in many applications. The terrorist threat has changed the situation and everything abnormal is considered as a possible threat. We present a method to address this problem by detecting repetitive patterns from incomplete, scattered and misleading data stream.

Sensors produce events which have been collected and preprocessed. We assume that events do not come in any particular order. Our goal is to detect known patterns from data stream online and to find often occurring behavior offline. There are only a couple of techniques, which are able to handle ill-ordered data streams.

Detectable patterns are vague and often full of exceptions. Therefore we model patterns as templates, which are stored in database. Paradigm is commonly known as a Case Based Reasoning (CBR) technique. Experts form original case base, and later on the system learns more of the surrounding environment. In our method we integrate hierarchical structures to the cases that results more robust performance with incomplete information. All numerical information are converted to fuzzy numbers in the cases. The system compares collected event information to cases in case base and retrieves best cases to the user investigation.

Method has been tested using simulated ship route data. Data were created with a simulator 30% were know events and rest was noise. We removed up to 50 % of data lines. Created cases were constructed to match the events. Test results are good; 80% of cases were correctly recognized and only 5 % were wrongly classified. The rest 15 % were classified to unknown class. We are encouraged to further develop this paradigm.

Comments on this paper
      No comments made on this paper.