We consider the problem of multi-step ahead prediction in time series analysis using the non-parametric Gaussian process model. k-step ahead forecasting of a discrete-time nonlinear dynamic system can be performed by doing repeated one-step ahead predictions. For a state-space model of the form yt = f(yt 1 ;:::;y t L), the prediction of y at time t + k is based on the estimates ^ yt+k 1 ;:::; ^ yt+k L of the previous outputs. We show how, using an analytical Gaussian approximation, we can formally incorporate the uncertainty about intermediate regressor values, thus updating the uncertainty on the current prediction. In this framework, the problem is that of predicting responses at a random input and we compare the Gaussian approximation to the Monte-Carlo numerical approximation of the predictive distribution. The approach is illustrated on a simulated non-linear dynamic example, as well as on a simple one-dimensional static example.
[1]
A. O'Hagan,et al.
Curve Fitting and Optimal Design for Prediction
,
1978
.
[2]
Dale E. Seborg,et al.
Adaptive nonlinear control of a pH neutralization process
,
1994,
IEEE Trans. Control. Syst. Technol..
[3]
Carl E. Rasmussen,et al.
In Advances in Neural Information Processing Systems
,
2011
.
[4]
Geoffrey E. Hinton,et al.
Bayesian Learning for Neural Networks
,
1995
.
[5]
David Mackay,et al.
Gaussian Processes - A Replacement for Supervised Neural Networks?
,
1997
.
[6]
Geoffrey E. Hinton,et al.
Evaluation of Gaussian processes and other methods for non-linear regression
,
1997
.
[7]
Daniel Sbarbaro,et al.
Nonlinear adaptive control using non-parametric Gaussian Process prior models
,
2002
.