- di 27 mei 2014
- notebook
- Jason K. Moore
- #notebook, #system identification, #prediction, #simulation

I didn't really understand the difference in prediction and simulation in the system id book. These are some notes that help clarify it.

This is from Ljung:

We shall generally work with k-step ahead model predictions \(\hat{y}_k(t|m)\) as the basis of the comparisons. By that we mean that \(\hat{y}_k(t|m)\) is computed from past data

\begin{equation*} u(t-1),\ldots,u(1), y(t-k),\ldots,y(1) \end{equation*}using the model \(m\). The case when \(k\) equals \(\infty\) corresponds to the use of past inputs only, i.e., a pure simulation. [m1]

Matlab documentation:

Simulation provides a better validation test for the model than prediction. However, how you validate the model output should match how you plan to use the model. For example, if you plan to use your model for control design, you can validate the model by predicting its response over a time horizon that represents the dominating time constants of the model. [m2]

Random notes found online:

Minimizing the one-step ahead prediction error of an output error model is the same as minimizing infinite-step ahead prediction errors. [m3]

I tried a Matlab OE model on the belt acceleration and velocity data and it gets a much better fit of about 50% (simuluation and predicition are the same with OE):

>> model model = Discrete-time OE model: y(t) = [B(z)/F(z)]u(t) + e(t) B(z) = 4.775 - 4.545 z^-1 + 1.714 z^-2 F(z) = 1 - 0.9197 z^-1 + 0.6324 z^-2 Sample time: 0.01 seconds Parameterization: Polynomial orders: nb=3 nf=2 nk=0 Number of free coefficients: 5 Use "polydata", "getpvec", "getcov" for parameters and their uncertainties. Status: Estimated using OE on time domain data "data". Fit to estimation data: 50.09% (simulation focus) FPE: 1.552, MSE: 1.551

[m1] | Ljung 1999 page 499 |

[m2] | http://www.mathworks.com/help/ident/ug/simulating-and-predicting-model-output.html |

[m3] | http://arunkt.yolasite.com/resources/ch5230/lectures/predictions.pdf |