Forecasting Chaotic Computational Ecosystems [1]

In this chapter, the authors investigate connectionist networks for short-term prediction of time series from computational ecosystems. They analyze a time series of resource allocation in computational ecosystems: data generated by Monte Carlo simulations of such systems were used to make future predictions which were then compared with experiments. The eventual forecasting function after several million iterations exhibits a frequency spectrum very similar to the original data. The authors use the error back propagation algorithm of David E. Rumelhart et al. to train the network: the parameters are changed by gradient descent on the cost surface over the space of weights and biases. They investigate the time series generated by the computational ecosystem of Kephart et al. for the use of resources in such a computational ecosystem. The authors begin their analysis with a review of some issues of deterministic chaos that are relevant to the prediction of time series.