Abstract for:Combining machine learning and stocks to automatically generate structure
Building decent models is hard, and building decent dynamic models is harder still. While great strides have been made in the application of machine learning to real problems, the resulting models rarely have realistic (if any) dynamics associated with them. By explicitly incorporating the existence of stock variables, a neural net computational approach using a modified perceptron, is able to capture the structure necessary to generate observed dynamic behavior. This paper outlines an approach to do this using standard system dynamics software. The example presented demonstrates the viability of the approach by reproducing the bass diffusion and inventory workforce models using a limited amount of a priori structural information. Generalization of the methods outlined to less well-defined problems with higher complexity dynamics that is typical in real world problems remains to be done, but the approach seems promising.