Training KSIM models from time series data

• Published in: Technological forecasting & social change 1994-11, Vol.47 (3), p.293-307
• Main Authors: Black, Ronald L., Oldham, William J.B., Marcy, William M.
• Format: Article
• Language: English
• ISSN: 0040-1625; 1873-5509
• DOI: 10.1016/0040-1625(94)90070-1

The suitability of KSIM models derived from group participation strategies is critically evaluated in a comparison with models generated by a gradient descent learning algorithm. Two learning algorithms are described to train KSIM socioeconomic models. The algorithms are used to train KSIM cross-impact matrices from initially random weights to final values producing a model that will closely fit a given time series. The time series can be obtained by integrating a KSIM model or by using raw data from other sources. KSIM modeling previously relied on insight, intuition, or knowledge of KSIM modeling to find suitable parameters. The training algorithms provide an organized approach to the minimization of a suitable cost function. At the same time, any system knowledge can be incorporated into initial conditions with learning performed around solid physical foundations. Some limits of the dynamic performance of the KSIM model are noted, further establishing the unsuitability of the KSIM model for many real systems.