C&E

# Control & Estimation #

The control system provided by this library is comprised of a state estimator and a controller. The estimator is responsible for estimating the latent state of the system, given measurements up to and including the current time (i.e., filtering). At each time step, the controller then uses the resulting state feedback and an internal model of the system to update the inputs to the process being manipulated.

## State estimation #

In general, the filtering performed to estimate the underlying state proceeds recursively by first using the model dynamics to predict the state change at the next time step, followed by updating this prediction when a new measurement is available. For a LDS, this two-step process can be summarized by $\widehat{\mathbf{x}}_{t|t-1} = \mathbf{A}\widehat{\mathbf{x}}_{t-1|t-1} + \mathbf{B} u_{t-1} + \mathbf{m}_{t-1} \;,$

$\widehat{\mathbf{x}}_{t|t} = \widehat{\mathbf{x}}_{t|t-1} + \mathbf{K}^{\rm e}_t \left(\mathbf{z}_t - \widehat{\mathbf{y}}_{t|t-1}\right)\;,$

where $$\hat{\left(\cdot\right)}_{t|j}$$ indicates an estimate at time $$t$$ given data up to time $$j$$ inclusive, $$\mathbf{K}^{\rm e}$$ is the estimator gain, and

$\widehat{\mathbf{y}}_{t|t-1} = h\left( \widehat{\mathbf{x}}_{t|t-1} \right) \; .$

In the case of GLDS models, the estimator gain (called Ke in library) is calculated recursively by Kalman filtering, which requires knowledge of the process noise and measurement noise covariances (Q, R) in addition to the system matrices. For time-invariant GLDS models, the infinite horizon solution is often used, so this gain need not be time-varying. Users may instead set its pre-determined value with the lds::gaussian::System::set_Ke mutator.

In the case of PLDS models, there is an analogue of the Kalman filter developed for dynamical systems with point-process observations (Eden et al. 2004). This nonlinear filter recursively updates Ke at each time step and requires an estimate of the process noise covariance (Q) as well.

## Adaptive estimation of process disturbance #

Both the Kalman filter and point-process analogue are model-based; therefore, their performance can be sensitive to model mismatch, whether this be imperfect model fitting or true drifts in system behavior. A practical approach to improving robustness is parameter adaptation. To that end, this library provides dual state-parameter estimation. Specifically, an additive process disturbance (m) is adaptively re-estimated when the lds::System::do_adapt_m property is set to true. This effectively provides integral action on minimizing state estimation error that could either be due to model mismatch or a true disturbance.

When parameter adaptation is enabled, this process disturbance is assumed to vary stochastically on a random walk $\mathbf{m}_{t} = \mathbf{m}_{t-1} + \mathbf{w}^m_{t-1} \;,$ where $$\mathbf{w}^m \sim \mathcal{N}\left(0, \mathbf{Q}_m\right)$$ . Kalman filtering or the point-process analogue are then used to estimate this disturbance in parallel with the state.

## Control #

Given the estimated state, the controller updates the inputs to the system according to the following law: $\mathbf{u}_{t} = \mathbf{u}^{\rm ref}_t - \mathbf{K}^c_x \left( \widehat{\mathbf{x}}_t - \mathbf{x}^{\rm ref}_t\right)\;,$

where $$\left( \cdot \right)^{\rm ref}$$ correspond to reference/target signals and $$\mathbf{K}^c_x$$ is the state feedback controller gain. Recall that these controller gains are assumed to have been designed before the experiment using, for example, LQR.

If users are employing integral action for more robust tracking at DC and did not use the approach of augmenting the state vector and system matrices accordingly, there is an option to include the integral term as

$\mathbf{u}_{t} = \mathbf{u}^{\rm ref}_t - \mathbf{K}^c_x \left( \widehat{\mathbf{x}}_t - \mathbf{x}^{\rm ref}_t\right) - \mathbf{K}^c_{\rm inty} \sum_{j=1}^{t}\left( \widehat{\mathbf{y}}_j - \mathbf{y}^{\rm ref}_j \right) \;.$

An additional option available to users is a control law that updates the change in u,

$\Delta\mathbf{u}_{t} = -\mathbf{K}^c_u \left(\mathbf{u}_{t-1} - \mathbf{u}^{\rm ref}_{t-1} \right) - \mathbf{K}^c_x \left( \widehat{\mathbf{x}}_t - \mathbf{x}^{\rm ref}_t\right)\;,$ $\mathbf{u}_{t} = \mathbf{u}_{t-1} + \Delta\mathbf{u}_{t} \; .$

Notice that this takes the form of a first-order difference equation for updating control (i.e., $$\Delta\mathbf{u}_{t} = -\mathbf{K}^c_u \mathbf{u}_{t-1} + \epsilon_{t-1}$$ ), effectively low-pass filtering the input depending on the characteristics of $$\mathbf{K}^c_u$$ . This can be useful in cases where users have designed the controller gains by LQR to minimize not the amplitude of the input, but the change in input, by augmenting the state vector with the input during LQR design.

Integral action and the $$\Delta \mathbf{u}$$ control law can be combined. The library keeps track of the controller type by way of bit masks which can be bit-wise OR’d to use in combination.

## Calculating reference state-control from output #

In cases where an output reference is supplied and the goal is to track either a static or slowly varying output, users do not have to produce $$\mathbf{x}^{\rm ref}$$ and $$\mathbf{u}^{\rm ref}$$ . Methods are provided for calculating the state and control that would be required to reach the reference output at steady state (lds::Controller<System>::ControlOutputReference). This is achieved by linearly-constrained least squares. For single-output systems, it results in an exact solution; however, for multi-output problems it provides a least squares comprimise across outputs.