This paper considers parameter estimation in the presence of unmodelled dynamics and "poor" excitation. A Variable Structure Estimation (VSE) algorithm is proposed which adjusts the complexity of the estimated ARMAX model to make it consistent with the current degree of excitation. The converged parameter estimates generally depend on the complexity of the estimated model, the complexity of the process model and the order of the underlying excitation. A VSE algorithm has advantages over on/off strategies in the presence of unmodelled dynamics and poor excitation. The idea is: Fit a model to the I/O behavior of the process rather than to the true process model.