Neural networks as expert systems

Abstract This paper shows that the high level decision-making function of expert systems, that depend upon many levels of logic, can be implemented in a neural network without the engineering of a detailed knowledge structure. Further, the neural network can interpolate and extrapolate a discrete set of associated input and output vectors, so that the output decision space is continuous. A drawback to the use of neural networks for decision making is that their training is universally problematic. We simplify the process with a new random optimization algorithm that consists of a global stage and a local stage. Unlike other methods, we also optimize the exponential rise parameters α and β in the sigmoids at the middle and output layers, which increases the speed of learning and decreases the minimum sum squated error.