Restrictions on goal-driven searches using rigorously inverted artificial neural network models

Artificial neural networks (ANNs) and expert systems are on opposite ends of the artificial intelligence (AI) spectrum. Goal-driven and data-driven search strategies are commonly used in expert systems. The equivalent of a data-driven search strategy is commonly employed with ANNs, both in chemical engineering and in other industrial applications. The only version of a goal-driven strategy now used with artificial neural networks takes the form of inverse plant dynamic models. In industrial applications, these models have proven inadequate. Using matrix pseudoinversion, a mathematically rigorous method of inverting a feedforward neural network has been developed. This inverted network offers a new approach to goal-driven searches with neural networks. Limitations to this approach are posed by network topology, by matrix colinearity, and by differences among neural network output functions