We present an algorithm for training feed-forward neural networks on decisiontables. The most significant property of our algorithm is its cumulative behavior: the more computational effort is invested in training the more knowledge is encoded in the network. Thus, in contrast to conventional training algorithms, like backpropagation, our algorithm always produces a network that 'knows something' about the given decision table. Moreover, although the algorithm is based on the translation of the training problem into a mixed linear Programming problem, it provides a framework for applications of other techniques developed for handling problems of combinatorial optimization like Heuristic Search, Constraint Satisfaction, Genetic Algorithms or Neural Networks.
展开▼
