Recursive Least Squares Estimator

Estimate model coefficients using recursive least squares (RLS) algorithm

• Library:
• System Identification Toolbox / Estimators

• Description

The Recursive Least Squares Estimator estimates the parameters of a system using a model that is linear in those parameters. Such a system has the following form:

$y\left(t\right)=H\left(t\right)\theta \left(t\right).$

y and H are known quantities that you provide to the block to estimate θ. The block can provide both infinite-history  and finite-history  (also known as sliding-window), estimates for θ. For more information on these methods, see Recursive Algorithms for Online Parameter Estimation.

The block supports several estimation methods and data input formats. Configurable options in the block include:

• Sample-based or frame-based data format — See the Input Processing parameter.

• Infinite-history or finite- history estimation — See the History parameter.

• Multiple infinite-history estimation methods — See the Estimation Method parameter.

• Initial conditions, enable flag, and reset trigger — See the Initial Estimate, Add enable port, and External Reset parameters.

For a given time step t, y(t) and H(t) correspond to the Output and Regressors inports of the Recursive Least Squares Estimator block, respectively. θ(t) corresponds to the Parameters outport.

For example, suppose that you want to estimate a scalar gain, θ, in the system y = h2θ. Here, y is linear with respect to θ. You can use the Recursive Least Squares Estimator block to estimate θ. Specify y and h2 as inputs to the Output and Regressor inports.

Ports

Input

expand all

Regressors input signal H(t). The Input Processing and Number of Parameters parameters define the dimensions of the signal:

• Sample-based input processing and N estimated parameters — 1-by-N vector

• Frame-based input processing with M samples per frame and N estimated parameters — M-by-N matrix

Data Types: single | double

Measured output signal y(t). The Input Processing parameter defines the dimensions of the signal:

• Sample-based input processing — Scalar

• Frame-based input processing with M samples per frame — M-by-1 vector

Data Types: single | double

External signal that allows you to enable and disable estimation updates. If the signal value is:

• true — Estimate and output the parameter values for the time step.

• false — Do not estimate the parameter values, and output the most recent previously estimated value.

Dependencies

To enable this port, select the Add enable port parameter.

Data Types: single | double | Boolean | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32

Reset parameter estimation to its initial conditions. The value of the External reset parameter determines the trigger type. The trigger type dictates whether the reset occurs on a signal that is rising, falling, either rising or falling, level, or on level hold.

Dependencies

To enable this port, select any option other than None in the External reset dropdown.

Data Types: single | double | Boolean | int8 | int16 | int32 | uint8 | uint16 | uint32

Initial parameter estimates, supplied from a source external to the block. The block uses this inport at the beginning of the simulation or when you trigger an algorithm reset using the Reset signal.

The Number of Parameters parameter defines the dimensions of the signal. If there are N parameters, the signal is N-by-1.

Dependencies

To enable this port, set History to Infinite and Initial Estimate to External.

Data Types: single | double

Initial parameter covariances, supplied from a source external to the block. For details, see the Parameter Covariance Matrix parameter.The block uses this inport at the beginning of the simulation or when you trigger an algorithm reset using the Reset signal.

Dependencies

To enable this port, set the following parameters:

• History to Infinite

• Estimation Method to Forgetting Factor or Kalman Filter

• Initial Estimate to External

Data Types: single | double

Initial values of the regressors in the initial data window when using finite-history (sliding-window) estimation, supplied from an external source. The Window length parameter W and the Number of Parameters parameter N define the dimensions of this signal, which is W-by-N.

The InitialRegressors signal controls the initial behavior of the algorithm. The block uses this inport at the beginning of the simulation or whenever the Reset signal triggers.

If the initial buffer is set to 0 or does not contain enough information, you see a warning message during the initial phase of your estimation. The warning should clear after a few cycles. The number of cycles it takes for sufficient information to be buffered depends upon the order of your polynomials and your input delays. If the warning persists, you should evaluate the content of your signals.

Dependencies

To enable this port, set History to Finite and Initial Estimate to External.

Data Types: single | double

Initial set of output measurements when using finite-history (sliding-window) estimation, supplied from an external source. The signal to this port must be a W-by-1 vector, where W is the window length.

The InitialOutputs signal controls the initial behavior of the algorithm. The block uses this inport at the beginning of the simulation or whenever the Reset signal triggers.

If the initial buffer is set to 0 or does not contain enough information, you see a warning message during the initial phase of your estimation. The warning should clear after a few cycles. The number of cycles it takes for sufficient information to be buffered depends upon the order of your polynomials and your input delays. If the warning persists, you should evaluate the content of your signals.

Dependencies

To enable this port, set History to Finite, and Initial Estimate to External.

Data Types: single | double

Output

expand all

Estimated parameters θ(t), returned as an N-by-1 vector where N is the number of parameters.

Data Types: single | double

Estimation error, returned as:

• Scalar — Sample-based input processing

• M-by-1 vector — Frame-based input processing with M samples per frame

Dependencies

To enable this port, select the Output estimation error parameter.

Data Types: single | double

Parameter estimation error covariance P, returned as an N-by-N matrix, where N is the number of parameters. For details, see the Output Parameter Covariance Matrix parameter.

Dependencies

To enable this port:

• If History is Infinite, set Estimation Method to Forgetting Factor or Kalman Filter.

• Whether History is Infinite or Finite, select the Output parameter covariance matrix parameter.

Data Types: single | double

Parameters

expand all

Model Parameters

Specify how to provide initial parameter estimates to the block:

• None — Do not specify initial estimates.

• If History is Infinite, the block uses 1 as the initial parameter estimate.

• If History is Finite, the block calculates the initial parameter estimates from the initial Regressors and Outputs signals.

Specify Number of Parameters, and also, if History is Infinite, Parameter Covariance Matrix.

• Internal — Specify initial parameter estimates internally to the block

• If History is Infinite, specify the Initial Parameter Values and Parameter Covariance Matrix parameters.

• If History is Finite, specify the Number of Parameters, the Initial Regressors, and the Initial Outputs parameters.

• External — Specify initial parameter estimates as an input signal to the block.

Specify the Number of Parameters parameter. Your setting for the History parameter determines which additional signals to connect to the relevant ports:

• If History is InfiniteInitialParameters and InitialCovariance

• If History is FiniteInitialRegressors and InitialOutputs

Programmatic Use

 Block Parameter: InitialEstimateSource Type: character vector, string Values: 'None', 'Internal', 'External' Default: 'None'

Specify the number of parameters to estimate in the model, equal to the number of elements in the parameter θ(t) vector.

Dependencies

To enable this parameter, set either:

• History to Infinite and Initial Estimate to either None or External

• History to Finite

An alternative way to specify the number of parameters N to estimate is by using the Initial Parameter Values parameter, for which you define an initial estimate vector with N elements. This approach covers the one remaining combination, where History is Infinite and Initial Estimate is Internal. For more information, see Initial Parameter Values.

Programmatic Use

 Block Parameter: InitialParameterData Type: positive integer Default: 2

Specify Parameter Covariance Matrix as a:

• Real positive scalar, α — Covariance matrix is an N-by-N diagonal matrix, with α as the diagonal elements.

• Vector of real positive scalars, [α1,...,αN] — Covariance matrix is an N-by-N diagonal matrix, with [α1,...,αN] as the diagonal elements.

• N-by-N symmetric positive-definite matrix.

Here, N is the number of parameters to be estimated.

Dependencies

To enable this parameter, set the following parameters:

• History to Infinite

• Initial Estimate to None or Internal

• Estimation Method to Forgetting Factor or Kalman Filter

Programmatic Use

 Block Parameter: P0 Type: scalar, vector, or matrix Default: 1e4

Specify initial parameter values as a vector of length N, where N is the number of parameters to estimate.

Dependencies

To enable this parameter, set History to Infinite and Initial Estimate to Internal.

Programmatic Use

 Block Parameter: InitialParameterData Type: real vector Default: [1 1]

Specify the initial values of the regressors buffer when using finite-history (sliding window) estimation. The Window length parameter W and the Number of Parameters parameter N define the dimensions of the regressors buffer, which is W-by-N.

The Initial Regressors parameter controls the initial behavior of the algorithm. The block uses this parameter at the beginning of the simulation or whenever the Reset signal triggers.

When the initial value is set to 0, the block populates the buffer with zeros.

If the initial buffer is set to 0 or does not contain enough information, you see a warning message during the initial phase of your estimation. The warning should clear after a few cycles. The number of cycles it takes for sufficient information to be buffered depends upon the order of your polynomials and your input delays. If the warning persists, you should evaluate the content of your signals.

Dependencies

To enable this parameter, set History to Finite and Initial Estimate to Internal.

Programmatic Use

 Block Parameter: InitialRegressors Type: real matrix Default: 0

Specify initial values of the measured outputs buffer when using finite-history (sliding-window) estimation. This parameter is a W-by-1 vector, where W is the window length.

When the initial value is set to 0, the block populates the buffer with zeros.

If the initial buffer is set to 0 or does not contain enough information, you see a warning message during the initial phase of your estimation. The warning should clear after a few cycles. The number of cycles it takes for sufficient information to be buffered depends upon the order of your polynomials and your input delays. If the warning persists, you should evaluate the content of your signals.

The Initial Outputs parameter controls the initial behavior of the algorithm. The block uses this parameter at the beginning of the simulation or whenever the Reset signal triggers.

Dependencies

To enable this parameter, set History to Finite and Initial Estimate to Internal .

Programmatic Use

 Block Parameter: InitialOutputs Type: real vector Default: 0
Input Processing and Sample Time
• Sample-based processing operates on signals streamed one sample at a time.

• Frame-based processing operates on signals containing samples from multiple time steps. Many machine sensor interfaces package multiple samples and transmit these samples together in frames. Frame-based processing allows you to input this data directly without having to first unpack it.

Specifying frame-based data adds an extra dimension of M to some of your data inports and outports, where M is the number of time steps in a frame. These ports are:

• Regressors

• Output

• Error

Programmatic Use

 Block Parameter: InputProcessing Type: character vector, string Values: 'Sample-based', 'Frame-based' Default: 'Sample-based'

Specify the data sample time, whether by individual samples for sample-based processing (ts), or by frames for frame-based processing (tf = Mts), where M is the frame length. When you set Sample Time to its default value of -1, the block inherits its ts or tf based on the signal.

Specify Sample Time as a positive scalar to override the inheritance.

Programmatic Use

 Block Parameter: Ts Type: real scalar Default: -1

Algorithm and Block Options

Algorithm Options

The History parameter determines what type of recursive algorithm you use:

• Infinite — Algorithms in this category aim to produce parameter estimates that explain all data since the start of the simulation. These algorithms retain the history in a data summary. The block maintains this summary within a fixed amount of memory that does not grow over time.

The block provides multiple algorithms of the Infinite type. Selecting this option enables the Estimation Method parameter with which you specify the algorithm.

• Finite — Algorithms in this category aim to produce parameter estimates that explain only a finite number of past data samples. The block uses all of the data within a finite window, and discards data once that data is no longer within the window bounds. This method is also called sliding-window estimation.

Selecting this option enables the Window Length parameter that sizes the sliding window.

For more information on recursive estimation methods, see Recursive Algorithms for Online Parameter Estimation

Programmatic Use

 Block Parameter: History Type: character vector, string Values: 'Infinite', 'Finite' Default: 'Infinite'

The Window Length parameter determines the number of time samples to use for the sliding-window estimation method. Choose a window size that balances estimation performance with computational and memory burden. Sizing factors include the number and time variance of the parameters in your model. Always specify Window Length in samples, even if you are using frame-based input processing.

Window Length must be greater than or equal to the number of estimated parameters.

Suitable window length is independent of whether you are using sample-based or frame-based input processing. However, when using frame-based processing, Window Length must be greater than or equal to the number of samples (time steps) contained in the frame.

Dependencies

To enable this parameter, set History to Finite.

Programmatic Use

 Block Parameter: WindowLength Type: positive integer Default: 200

Specify the estimation algorithm when performing infinite-history estimation. When you select any of these methods, the block enables additional related parameters.

Forgetting factor and Kalman filter algorithms are more computationally intensive than gradient and normalized gradient methods. However, these more intensive methods have better convergence properties than the gradient methods. For more information about these algorithms, see Recursive Algorithms for Online Parameter Estimation.

Programmatic Use

 Block Parameter: EstimationMethod Type: character vector, string Values: 'Forgetting Factor','Kalman Filter','Normalized Gradient','Gradient' Default: 'Forgetting Factor'

The forgetting factor λ specifies if and how much old data is discounted in the estimation. Suppose that the system remains approximately constant over T0 samples. You can choose λ such that:

${T}_{0}=\frac{1}{1-\lambda }$

• Setting λ = 1 corresponds to “no forgetting” and estimating constant coefficients.

• Setting λ < 1 implies that past measurements are less significant for parameter estimation and can be “forgotten.” Set λ < 1 to estimate time-varying coefficients.

Typical choices of λ are in the [0.98 0.995] range.

Dependencies

To enable this parameter, set History to Infinite and Estimation Method to Forgetting Factor.

Programmatic Use

 Block Parameter: AdaptationParameter Type: scalar Values: (0 1] range Default: 1

Process Noise Covariance prescribes the elements and structure of the noise covariance matrix for the Kalman filter estimation. Using N as the number of parameters to estimate, specify the Process Noise Covariance as one of the following:

• Real nonnegative scalar, α — Covariance matrix is an N-by-N diagonal matrix, with α as the diagonal elements.

• Vector of real nonnegative scalars, [α1,...,αN] — Covariance matrix is an N-by-N diagonal matrix, with [α1,...,αN] as the diagonal elements.

• N-by-N symmetric positive semidefinite matrix.

The Kalman filter algorithm treats the parameters as states of a dynamic system and estimates these parameters using a Kalman filter. Process Noise Covariance is the covariance of the process noise acting on these parameters. Zero values in the noise covariance matrix correspond to constant coefficients, or parameters. Values larger than 0 correspond to time-varying parameters. Use large values for rapidly changing parameters. However, expect the larger values to result in noisier parameter estimates. The default value is 1.

Dependencies

To enable this parameter, set History to Infinite and Estimation Method to Kalman Filter.

Programmatic Use

 Block Parameter: AdaptationParameter Type: scalar, vector, matrix Default: 1

The adaptation gain γ scales the influence of new measurement data on the estimation results for the gradient and normalized gradient methods. When your measurements are trustworthy, or in other words have a high signal-to-noise ratio, specify a larger value for γ. However, setting γ too high can cause the parameter estimates to diverge. This divergence is possible even if the measurements are noise free.

When Estimation Method is NormalizedGradient, Adaptation Gain should be less than 2. With either gradient method, if errors are growing in time (in other words, estimation is diverging), or parameter estimates are jumping around frequently, consider reducing Adaptation Gain.

Dependencies

To enable this parameter, set History to Infinite and Estimation Method to Normalized Gradient or to Gradient.

Programmatic Use

 Block Parameter: AdaptationParameter Type: scalar Default: 1

The normalized gradient algorithm scales the adaptation gain at each step by the square of the two-norm of the gradient vector. If the gradient is close to zero, the near-zero denominator can cause jumps in the estimated parameters. Normalization Bias is the term introduced to the denominator to prevent these jumps. Increase Normalization Bias if you observe jumps in estimated parameters.

Dependencies

To enable this parameter, set History to Infinite and Estimation Method to Normalized Gradient.

Programmatic Use

 Block Parameter: NormalizationBias Type: scalar Default: eps
Block Options

Use the Error outport signal to validate the estimation. For a given time step t, the estimation error e(t) is calculated as:

$e\left(t\right)=y\left(t\right)-{y}_{est}\left(t\right),$

where y(t) is the measured output that you provide, and yest(t) is the estimated output using the regressors H(t) and parameter estimates θ(t-1).

Programmatic Use

 Block Parameter: OutputError Type: character vector, string Values: 'off','on', Default: 'off'

Use the Covariance outport signal to examine parameter estimation uncertainty. The software computes parameter covariance P assuming that the residuals, e(t), are white noise, and the variance of these residuals is 1.

The interpretation of P depends on the estimation approach you specify in History and Estimation Method as follows:

• If History is Infinite, then your Estimation Method selection results in:

• Forgetting Factor — (R2/2)P is approximately equal to the covariance matrix of the estimated parameters, where R2 is the true variance of the residuals. The block outputs the residuals in the Error port.

• Kalman FilterR2P is the covariance matrix of the estimated parameters, and R1 /R2 is the covariance matrix of the parameter changes. Here, R1 is the covariance matrix that you specify in Parameter Covariance Matrix.

• If History is Finite (sliding-window estimation) — R2 P is the covariance of the estimated parameters. The sliding-window algorithm does not use this covariance in the parameter-estimation process. However, the algorithm does compute the covariance for output so that you can use it for statistical evaluation.

Programmatic Use

 Block Parameter: OutputP Type: character vector, string Values: 'off','on' Default: 'off'

Use the Enable signal to provide a control signal that enables or disables parameter estimation. The block estimates the parameter values for each time step that parameter estimation is enabled. If you disable parameter estimation at a given step, t, then the software does not update the parameters for that time step. Instead, the block outputs the last estimated parameter values.

• You can use this option, for example, when or if:

• Your regressors or output signal become too noisy, or do not contain information at some time steps

• Your system enters a mode where the parameter values do not change in time

Programmatic Use

 Block Parameter: AddEnablePort Type: character vector, string Values: 'off','on' Default: 'off'

Set the External reset parameter to both add a Reset inport and specify the inport signal condition that triggers a reset of algorithm states to their specified initial values. Reset the estimation, for example, if parameter covariance is becoming too large because of lack of either sufficient excitation or information in the measured signals.

Suppose that you reset the block at a time step, t. If the block is enabled at t, the software uses the initial parameter values specified in Initial Estimate to estimate the parameter values. In other words, at t, the block performs a parameter update using the initial estimate and the current values of the inports.

If the block is disabled at t and you reset the block, the block outputs the values specified in Initial Estimate.

Specify this option as one of the following:

• None — Algorithm states and estimated parameters are not reset.

• Rising — Trigger reset when the control signal rises from a negative or zero value to a positive value. If the initial value is negative, rising to zero triggers reset.

• Falling — Trigger reset when the control signal falls from a positive or a zero value to a negative value. If the initial value is positive, falling to zero triggers reset.

• Either — Trigger reset when the control signal is either rising or falling.

• Level — Trigger reset in either of these cases:

• Control signal is nonzero at the current time step.

• Control signal changes from nonzero at the previous time step to zero at the current time step.

• Level hold — Trigger reset when the control signal is nonzero at the current time step.

When you choose any option other than None, the software adds a Reset inport to the block. You provide the reset control input signal to this inport.

Programmatic Use

 Block Parameter: ExternalReset Type: character vector, string Values: 'None','Rising','Falling', 'Either', 'Level', 'Level hold' Default: 'None'

 Ljung, L. System Identification: Theory for the User. Upper Saddle River, NJ: Prentice-Hall PTR, 1999, pp. 363–369.

 Zhang, Q. "Some Implementation Aspects of Sliding Window Least Squares Algorithms." IFAC Proceedings. Vol. 33, Issue 15, 2000, pp. 763-768.

System Identification Toolbox Documentation Get trial now