The paper describes and evaluates three methods
of modelling earthquake accelerograms using artificial
neural networks and linear predictors. These are a radial-Gaussian
neural networking system, a linear predictor, and a hybrid
of these two. Two methods of using these models to predict
ground acceleration are adopted. The first is based on
a direct prediction of ground acceleration at some point
in time in the future from a series of ground accelerations
that occurred earlier in the earthquake event. The second
is a recursive approach whereby a model predicts a sequence
of future ground accelerations by feeding back its predictions
to its inputs. The performances of the models are tested
using the 1985 Mexico earthquake and aftershock. The linear
predictor and hybrid models perform best at direct prediction
while none of the models perform particularly well at recursive
prediction. The paper concludes with an outline of some
areas for future work.