Main Content


Prediction error minimization for refining linear and nonlinear models



sys = pem(data,init_sys) updates the parameters of an initial model init_sys to fit the estimation data in data. data can be a timetable, a comma-separated pair of matrices, or a data object.

The function uses prediction-error minimization algorithm to update the parameters of the initial model. Use this command to refine the parameters of a previously estimated model.


sys = pem(data,init_sys,opt) specifies estimation options using an option set.


collapse all

Estimate a discrete-time state-space model using n4sid, which applies the subspace method.

Load the data and extract the first 300 points for the estimation data.

load sdata7 tt7;
tt7e = tt7(1:300,:);

Estimate the model init_sys, setting the 'Focus' option to 'simulation'.

opt = n4sidOptions('Focus','simulation');
init_sys = n4sid(tt7e,4,opt);

Show the estimation fit.

ans = 73.8490

Use pem to improve the closeness of the fit.

sys = pem(tt7e,init_sys);

Analyze the results.


Using pem improves the fit to the estimation data.

Estimate the parameters of a nonlinear grey-box model to fit DC motor data.

Load the experimental data, and specify the signal attributes such as start time and units.

data = iddata(y, u, 0.1);
data.Tstart = 0;
data.TimeUnit = 's';

Configure the nonlinear grey-box model (idnlgrey) model.

For this example, use dcmotor_m.m file. To view this file, type edit dcmotor_m.m at the MATLAB® command prompt.

file_name = 'dcmotor_m';
order = [2 1 2];
parameters = [1;0.28];
initial_states = [0;0];
Ts = 0;
init_sys = idnlgrey(file_name,order,parameters,initial_states,Ts);
init_sys.TimeUnit = 's';

setinit(init_sys,'Fixed',{false false});

init_sys is a nonlinear grey-box model with its structure described by dcmotor_m.m. The model has one input, two outputs and two states, as specified by order.

setinit(init_sys,'Fixed',{false false}) specifies that the initial states of init_sys are free estimation parameters.

Estimate the model parameters and initial states.

sys = pem(data,init_sys);

sys is an idnlgrey model, which encapsulates the estimated parameters and their covariance.

Analyze the estimation result.


sys provides a 98.34% fit to the estimation data.

Create a process model structure and update its parameter values to minimize prediction error.

Initialize the coefficients of a process model.

init_sys = idproc('P2UDZ');
init_sys.Kp = 10;
init_sys.Tw = 0.4;
init_sys.Zeta = 0.5;
init_sys.Td = 0.25;
init_sys.Tz = 0.01;

The Kp, Tw, Zeta, Td, and Tz coefficients of init_sys are configured with their initial guesses.

Use init_sys to configure the estimation of a prediction error minimizing model using measured data. Because init_sys is an idproc model, use procestOptions to create the option set.

load iddata1 z1;
opt = procestOptions('Display','on','SearchMethod','lm');
sys = pem(z1,init_sys,opt);
Process Model Identification
Estimation data: Time domain data z1         
Data has 1 outputs, 1 inputs and 300 samples.
Model Type:    

Algorithm: Levenberg-Marquardt search
                          Norm of      First-order    Improvement (%) <br> Iteration       Cost       step         optimality     Expected   Achieved    Bisections <br>------------------------------------------------------------------------------------------
     0        29.7194          -            260        2.57           -         -                                                                                                                                                                                                                                                                                                                               
     1        28.6801          6           98.9        2.57         3.5         0
     2        8.38196       4.91           42.2        2.72        70.8         0
     3         8.2138      0.704           41.3        1.37        2.01        12
     4        8.00237      0.528           48.3        2.89        2.57         9
     5        7.65577      0.588           73.1        2.02        4.33         9
     6          6.851      0.809            196        4.51        10.5         9
     7        5.72335       1.08            459        4.59        16.5         8
     8         3.3434       2.11       1.63e+03        11.4        41.6         7
     9        1.80724      0.701            504        14.2        45.9         0
    10         1.6812      0.122             12        4.24        6.97         0
    11        1.68092      0.014           1.11       0.309      0.0168         0
    12        1.68092    0.00179         0.0215         0.3    0.000101         0
    13        1.68092   0.000112        0.00634         0.3    8.26e-07         0
    14        1.68092   1.36e-05       0.000382         0.3    7.62e-09         0
    15        1.68092   1.18e-06       5.01e-05         0.3    7.28e-11         0
    16        1.68092   1.23e-07       4.29e-06         0.3       7e-13         0
    17        1.68092   1.17e-08       4.56e-07         0.3    2.64e-14         0
Termination condition: No improvement along the search direction with line search..
Number of iterations: 18, Number of function evaluations: 115                      
Status: Estimated using PEM                                                        
Fit to estimation data: 70.57%, FPE: 1.7379                                        

Examine the model fit.

ans = 70.5666

sys provides a 70.63% fit to the measured data.

Input Arguments

collapse all

Uniformly sampled estimation data, specified as a timetable, a comma-separated matrix pair, or a data object, as the following sections describe. For multiexperiment data, data can also be a 1-by-Ne cell array of timetables or matrix pairs, where Ne is the number of experiments. Data objects accommodate multiexperiment data within the object.


Specify data as a timetable that uses a regularly spaced time vector. tt contains variables representing input and output channels. pem derives the input and output channel variables from the model passed as an input argument.

Comma-Separated Matrix Pair

Specify data as a comma-separated pair of real-valued matrices that contain input and output time-domain signal values.

  • For SISO systems, specify data as a pair of Ns-by-1 real-valued matrices that contain uniformly sampled input and output time-domain signal values. Here, Ns is the number of samples.

  • For MIMO systems, specify u,y as an input/output matrix pair with the following dimensions:

    • uNs-by-Nu, where Nu is the number of inputs.

    • yNs-by-Ny, where Ny is the number of outputs.

Data Object

Specify data as an iddata, idfrd, or frd object.

If the model passed as an input arguments is:

  • A parametric model such as idss, then data can be an iddata, idfrd, or frd model object.

  • A frequency-response data model (either an idfrd, or frd model object), then data must also be a frequency-response data model.

  • An iddata object, then data must be an iddata object with matching domain, number of experiments and time or frequency vectors.

For more information about working with estimation data types, see Data Domains and Data Types in System Identification Toolbox.

Identified model that configures the initial parameterization of sys, specified as a linear, or nonlinear model. You can obtain init_sys by performing an estimation using measured data or by direct construction.

init_sys must have finite parameter values. You can configure initial guesses, specify minimum/maximum bounds, and fix or free for estimating any parameter of init_sys:

  • For linear models, use the Structure property. For more information, see Imposing Constraints on Model Parameter Values.

  • For nonlinear grey-box models, use the InitialStates and Parameters properties. Parameter constraints cannot be specified for nonlinear ARX and Hammerstein-Wiener models.

Estimation options that configure the algorithm settings, handling of estimation focus, initial conditions, and data offsets, specified as an option set. The command used to create the option set depends on the initial model type:

Output Arguments

collapse all

Identified model, returned as the same model type as init_sys. The model is obtained by estimating the free parameters of init_sys using the prediction error minimization algorithm.


PEM uses numerical optimization to minimize the cost function, a weighted norm of the prediction error, defined as follows for scalar outputs:


where e(t) is the difference between the measured output and the predicted output of the model. For a linear model, the error is defined as:


where e(t) is a vector and the cost function VN(G,H) is a scalar value. The subscript N indicates that the cost function is a function of the number of data samples and becomes more accurate for larger values of N. For multiple-output models, the previous equation is more complex. For more information, see chapter 7 in System Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice Hall PTR, 1999.

Alternative Functionality

You can achieve the same results as pem by using dedicated estimation commands for the various model structures. For example, use ssest(data,init_sys) for estimating state-space models.

Version History

Introduced before R2006a