joshi@wuche2.wustl.edu (Amol Joshi) (03/07/90)
optimization i recently saw some inquiries about research in rule/knowledge extraction from neural nets. there was a paper presented at the recent AIChE (American Institute of Chemical Engineers) Annual Meeting in San Fransisco (Nov '89) which might be of interest to people investigating this area. the title and the authors: Application of Neural Network to Rule Extraction from Operation Data K. Niida, J. Tani, T. Hirobe and I. Koshijima. the only merits of the paper are - 1. it reduces the back propagation network to what they call as 'skeleton network' by using another optimization objective which accounts for the complexity of network and tries to reduce the redundancy in the network. in one example where they use the network to learn the rules from operation data from a felt mgf plant, they reduce no. of connections by 85% and therefore are able to 'read' the network. 2. it has been successfully applied in industrial scale problems. i don't think the paper addresses fundamental issues at all but might be useful to someone. :amol -- ------------------------------------------------------------ Amol Joshi | joshi@wuche2.wustl.edu Department of Chemical Engineering | Washington University in St. Louis.|