The effect of reasoning with state information

In this article, we describe a method for the incremental development of knowledge-based neural systems which incorporate state information about a problem domain. Human expert domain knowledge is encoded in the architecture of a computational network and data-driven techniques are used to refine the encoded knowledge. State information, in the form of state nodes and recurrent network connections, is then added in order to augment the refined knowledge even further. The methodology is tested in the domain of protein secondary structure prediction, where a set of domain rules, the Chou-Fasman algorithm, is implemented in phases as an expert network. The domain rules are then refined using a backpropagation-like learning algorithm. Finally, state information is added to the resulting network and is shown to increase prediction accuracy