Belief Updating Recurrent Fuzzy Rules for Ordered Dataset Modelling

J.F. Baldwin, T.P. Martin, J.M. Rossiter
1999 International journal of biomedical soft computing and human sciences  
We present a new method for modelling ordered datasets using Baldwinls mass assignment, 71his method generates a simplijied memory-based juizy belief updating model, Tlhe predicted class fuzzy set naturally describes the current state of belief and the previous class fuzzy set dqfines how previous beliojis colour future beliefe. Zhe modet is implemented using liVil evidential logic rules. Results are given in qpplication to particle classofication and focial feature detection. 1:he models
more » ... ted using this method are concise and linguistically clear glass box models. In this paper we will describe how mass assignment can be the enabling factor in generating clear, concise and descriptive glass box models of ordered data. These models are derived from simple analysis of human behaviour. Our simple memorybased model of ordered datasets aims to use high level features and mechanisms based on human behaviour. Results are given for gas particle classification and for facial feature extraction from digital images where the image is treated as an ordered database in pixel coordinate space. Memory-based modelling using soft computing does not focus on the features being manipulated, such as the perception-based features described in [6], but on how the computing model captures human belief and memory, Implicit in this model is a representation of belief and a method of updating beliefs. Our tool for this soft computing research is the fuzzy set theory based on Baldwin's mass assignment [4]. in classifying data in particle streams. The fo1lowing example outlines the particle classification problem, The particle stream classification problem 11his problem involves the detection of ha2ardous particles in a stream of gases. 71his problem is important in both ctosed spaces, such as in chemical plants and in mine shafts, and in open spaues, such as during toxic chem{cal leaks and chemical and biological wambre. A detector is needed to determine when poisonous gases are present and to raise an alarm. The gas to be analysed fiows past a sensor. J:his sensor generates a 5 featttre tuple < El,E2,E3,E4,T > where E1"",E4 are cont{nuous domain foatures and T is some measure of time. Figure 1 shows this mere clearly. Fleatures El, . . . , E3 measure the tight deflected by a particle in three directions when a laser is fired at it. E4 is a measure of the laser tight that is not doflected at all. tfi,?gyiszllp --1i --ms lnputPardcles 11 OrderedDataset Oututvalues EiE2ElmTInputc]ass 2Justification for a fuzzy belief updating model The belief updating method proposed in this paper was developed in response to problems encountered 'supported by EPSRC and CASE awards, in cooperation with the UK DERA A system is needed which takes each tuple in the ordered dataset in tum and generates a cowfidence that the cuTvent gas being detected is of a pariicular class. This corijidence can be used to cGlculate concentration of each gas in a mixture, Problems exist in such ordered datasets where class boundaries are indistinct, Indistinct classification is clearly an area where soft computing can have an impact. Given that the datasets we are dealing with contain some (however loose) ordering, it is Manuscnpt reccived July l2. 1999. revised September 22. ]999 NII-Electronic Mbrary
doi:10.24466/ijbschs.5.1_73 fatcat:p6afokq5cnhurnmb6y25cbyqoy