No abstract

No abstract

No abstract

Knowledge Selection @$) [I] is a completely new method which was developed at the department of Knowledge Engineering of the University of Vienna aiming at selecting relevant knowledge out of a knowledge base for a particular task This method will become airemely important for supporting the ejicient (re-)use of knowledge in knowledge management systems. KS is realised by three filters: Ident@cation selects knowledge items according to syntactical properties of the query, adaption uses background knowledge for the jiltering, and prediction inks to predict fiture queries for a small rtumber of time steps. In this paper neural network solutions for KS together with an Ks-implementation in the area of computer security b presented KS filtersIn the following the three filters identification, adaption and prediction are described in detail (figure 1). F 3 knowledge base Identifiction selects knowledge items according to syntactical properties of the query. Identification only filters strong irrelevant knowledge which is characterized by knowledge units that are never "touched" during the inference process. For realisation there exist pow& symbolic and neural techniques [l]. In the following identification with an associative neural net [2] is shown, filtering relevant knowledge items out of the foIlowing PROLOG knowledge base for the query ?-PO. R1: t(X) :-PO. R2: p o :-r m . F1: q(a). F2: I Q ) .F3: s(c). R3: p(x) :-q O . Input Layer RI R2 R3 F1 F2 F3 I R I R2 R3 F1 F2 F3 Output Layer Figure 2: Associative neural net for identification The associative neural net (figure 2) shows a possible representation of this small PROLOG knowledge base. This neural net mirrors the dependencies between knowledge units in the knowledge base. The query ?-PO. is coded in the input layer by the vector [0,1,1,0,0,0]. According to the connections the propagation results m the output vector [0,1,1,1,1,0]. After each Figure 1: Knowledge Selection filters 0-7803-5529-6/99/%10.00 01999 lEEE 2486

No abstract