Training without data: Knowledge Insertion into RBF Neural Networks

Often, in real-world situations no actual data is available for training neural networks but the domain expert has a good idea of what to expect in terms of input and output parameter values. If the expert can express these relationships in the form of rules, this would provide a resource too valuable to ignore. Fuzzy logic is used to handle the imprecision and vagueness of natural language and provides this additional advantage to a system. This paper investigates the development of a novel knowledge insertion algorithm that explores the benefits of pre-structuring RBF neural networks by using prior fuzzy domain knowledge and previous training experiences. Pre-structuring is accomplished by using fuzzy rules gained from a domain expert and using them to modify existing RBF networks. The benefits and novel achievements of this work enable RBF neural networks to be trained without actual data but to rely on input to output mappings defined through expert knowledge.