Initialising neural networks with a priori problem knowledge

In general problem knowledge can be used to formulate rules as an aid to finding solutions to specific problems. The rules need not be complete and may be contradictory in some details. This paper develops a number of schemes that map rules to the weights in a special network architecture (FuNN). These weights are used as the initial state for the training of the network. A mapping scheme is also given for a general MLP network. Results of our experiments show that with only a small set of rules, networks used to solve complex problems can converge more reliably and often to a better solution