jax-sysid 0.6.7

jax-sysid - A Python package for linear and nonlinear system identification and nonlinear regression using Jax.

A Python package based on JAX for linear and nonlinear system identification of state-space models, recurrent neural network (RNN) training, and nonlinear regression/classification.

Contents

• Contents
  • Package description
  • Installation
  • Basic usage
    • Linear state-space models
      • Training linear models
      • L1- and group-Lasso regularization
      • Multiple experiments
      • Stability
      • Static gain
    • Nonlinear system identification and RNNs
      • Training nonlinear models
      • Parallel training
      • flax.linen models
      • Custom output loss
      • Custom regularization
      • Static gain
      • Upper and lower bounds
    • Quasi-Linear Parameter-Varying (qLPV) models
      • Training qLPV models
    • Static models
      • Nonlinear regression
      • Classification
  • Contributors
  • Acknowledgments
  • Citing jax-sysid
  • License

Package description

jax-sysid is a Python package based on JAX for linear and nonlinear system identification of state-space models, recurrent neural network (RNN) training, and nonlinear regression/classification. The algorithm can handle L1-regularization and group-Lasso regularization and relies on L-BFGS optimization for accurate modeling, fast convergence, and good sparsification of model coefficients.

The package implements the approach described in the following paper:

[1] A. Bemporad, "Linear and nonlinear system identification under $\ell_1$- and group-Lasso regularization via L-BFGS-B," submitted for publication. Available on arXiv at http://arxiv.org/abs/2403.03827, 2024. [bib entry]

Installation

pip install jax-sysid

Basic usage

Linear state-space models

Training linear models

Given input/output training data $(u_0,y_0)$, $\ldots$, $(u_{N-1},y_{N-1})$, $u_k\in R^{n_u}$, $y_k\in R^{n_y}$, we want to identify a state-space model in the following form

$$ x_{k+1}=Ax_k+Bu_k$$

$$ \hat y_k=Cx_k+Du_k $$

where $k$ denotes the sample instant, $x_k\in R^{n_x}$ is the vector of hidden states, and $A,B,C,D$ are matrices of appropriate dimensions to be learned.

The training problem to solve is

$$\min_{z}r(z)+\frac{1}{N}\sum_{k=0}^{N-1} |y_{k}-Cx_k-Du_k|_2^2$$

$$\textrm{s.t.}\ x_{k+1}=Ax_k+Bu_k, \ k=0,\ldots,N-2$$

where $z=(\theta,x_0)$ and $\theta$ collecting the entries of $A,B,C,D$.

The regularization term $r(z)$ includes the following components:

$$\rho_{\theta} |\theta|_2^2 $$

$$\rho_{x_0} |x_0|_2^2$$

$$\tau \left|z\right|_1$$

$$\tau_g\sum_{i=1}^{n_u} |I_iz|_2$$

with $\rho_\theta>0$, $\rho_{x_0}>0$, $\tau\geq 0$, $\tau_g\geq 0$. See examples below.

Let's start training a discrete-time linear model $(A,B,C,D)$ on a sequence of inputs $U=[u_0\ \ldots\ u_{N-1}]'$ and output $Y=[y_0\ \ldots\ y_{N-1}]'$, with regularization $\rho_\theta=10^{-2}$, $\rho_{x_0}=10^{-3}$, running the L-BFGS solver for at most 1000 function evaluations:

from jax_sysid.models import LinearModel

model = LinearModel(nx, ny, nu)
model.loss(rho_x0=1.e-3, rho_th=1.e-2) 
model.optimization(lbfgs_epochs=1000) 
model.fit(Y,U)
Yhat, Xhat = model.predict(model.x0, U)

After identifying the model, to retrieve the resulting state-space realization you can use the following:

A,B,C,D = model.ssdata()

Given a new test sequence of inputs and outputs, an initial state that is compatible with the identified model can be reconstructed by running an extended Kalman filter and Rauch–Tung–Striebel smoothing (cf. [1]) and used to simulate the model:

x0_test = model.learn_x0(U_test, Y_test)
Yhat_test, Xhat_test = model.predict(x0_test, U_test)

R2-scores on training and test data can be computed as follows:

from jax_sysid.utils import compute_scores

R2_train, R2_test, msg = compute_scores(Y, Yhat, Y_test, Yhat_test, fit='R2')
print(msg)

It is good practice to scale the input and output signals. To identify a model on scaled signals, you can use the following:

from jax_sysid.utils import standard_scale, unscale

Ys, ymean, ygain = standard_scale(Y)
Us, umean, ugain = standard_scale(U)
model.fit(Ys, Us)
Yshat, Xhat = model.predict(model.x0, Us)
Yhat = unscale(Yshat, ymean, ygain)

L1- and group-Lasso regularization

Let us now retrain the model using L1-regularization and check the sparsity of the resulting model:

model.loss(rho_x0=1.e-3, rho_th=1.e-2, tau_th=0.03) 
model.fit(Ys, Us)
print(model.sparsity_analysis())

To reduce the number of states in the model, you can use group-Lasso regularization as follows:

model.loss(rho_x0=1.e-3, rho_th=1.e-2, tau_g=0.1) 
model.group_lasso_x()
model.fit(Ys, Us)

Groups in this case are entries in $A,B,C,x_0$ related to the same state.

Group-Lasso can be also used to try to reduce the number of inputs that are relevant in the model. You can do this as follows:

model.loss(rho_x0=1.e-3, rho_th=1.e-2, tau_g=0.15) 
model.group_lasso_u()
model.fit(Ys, Us)

Groups in this case are entries in $B,D$ related to the same input.

Multiple experiments

jax-sysid also supports multiple training experiments. In this case, the sequences of training inputs and outputs are passed as a list of arrays. For example, if three experiments are available for training, use the following command:

model.fit([Ys1, Ys2, Ys3], [Us1, Us2, Us3])

In case the initial state $x_0$ is trainable, one initial state per experiment is optimized. To avoid training the initial state, add train_x0=False when calling model.loss.

Stability

To attempt forcing that the identified linear model is asymptotically stable, i.e., that matrix $A$ has all eigenvalues inside the unit disk, you can use the following command:

model.force_stability()

before calling the fit function. This will introduce a custom regularization penalty that tries to enforce the constraint $|A|_2<1$.

Static gain

To introduce a penalty that attempts forcing the identified linear model to have a given DC-gain matrix M, you can use the following commands:

dcgain_loss = model.dcgain_loss(DCgain = M)
model.loss(rho_x0=1.e-3, rho_th=1.e-2, custom_regularization = dcgain_loss)

before calling the fit function. Similarly, to fit instead the DC-gain of the model to steady-state input data Uss and corresponding output data Yss, you can use

dcgain_loss = model.dcgain_loss(Uss = Uss, Yss = Yss)

and use dcgain_loss as the custom regularization function.

Nonlinear system identification and RNNs

Training nonlinear models

Given input/output training data $(u_0,y_0)$, $\ldots$, $(u_{N-1},y_{N-1})$, $u_k\in R^{n_u}$, $y_k\in R^{n_y}$, we want to identify a nonlinear parametric state-space model in the following form

$$ x_{k+1}=f(x_k,u_k,\theta)$$

$$ \hat y_k=g(x_k,u_k,\theta)$$

where $k$ denotes the sample instant, $x_k\in R^{n_x}$ is the vector of hidden states, and $\theta$ collects the trainable parameters of the model.

As for the linear case, the training problem to solve is

$$ \min_{z}r(z)+\frac{1}{N}\sum_{k=0}^{N-1} |y_{k}-g(x_k,u_k,\theta)|_2^2$$

$$\textrm{s.t.}\ x_{k+1}=f(x_k,u_k,\theta),\ k=0,\ldots,N-2$$

where $z=(\theta,x_0)$. The regularization term $r(z)$ is the same as in the linear case.

For example, let us consider the following residual RNN model without input/output feedthrough:

$$ x_{k+1}=Ax_k+Bu_k+f_x(x_k,u_k,\theta_x)$$

$$ \hat y_k=Cx_k+f_y(x_k,\theta_y)$$

where $f_x$, $f_y$ are feedforward shallow neural networks, and let $z$ collects the coefficients in $A,B,C,D,\theta_x,\theta_y$. We want to train $z$ by running 1000 Adam iterations followed by at most 1000 L-BFGS function evaluations:

from jax_sysid.models import Model

Ys, ymean, ygain = standard_scale(Y)
Us, umean, ugain = standard_scale(U)

def sigmoid(x):
    return 1. / (1. + jnp.exp(-x))  
@jax.jit
def state_fcn(x,u,params):
    A,B,C,W1,W2,W3,b1,b2,W4,W5,b3,b4=params
    return A@x+B@u+W3@sigmoid(W1@x+W2@u+b1)+b2    
@jax.jit
def output_fcn(x,u,params):
    A,B,C,W1,W2,W3,b1,b2,W4,W5,b3,b4=params
    return C@x+W5@sigmoid(W4@x+b3)+b4

model = Model(nx, ny, nu, state_fcn=state_fcn, output_fcn=output_fcn)
nnx = 5 # number of hidden neurons in state-update function
nny = 5  # number of hidden neurons in output function

# Parameter initialization:
A  = 0.5*np.eye(nx)
B = 0.1*np.random.randn(nx,nu)
C = 0.1*np.random.randn(ny,nx)
W1 = 0.1*np.random.randn(nnx,nx)
W2 = 0.5*np.random.randn(nnx,nu)
W3 = 0.5*np.random.randn(nx,nnx)
b1 = np.zeros(nnx)
b2 = np.zeros(nx)
W4 = 0.5*np.random.randn(nny,nx)
W5 = 0.5*np.random.randn(ny,nny)
b3 = np.zeros(nny)
b4 = np.zeros(ny)
model.init(params=[A,B,C,W1,W2,W3,b1,b2,W4,W5,b3,b4]) 

model.loss(rho_x0=1.e-4, rho_th=1.e-4)
model.optimization(adam_epochs=1000, lbfgs_epochs=1000) 
model.fit(Ys, Us)
Yshat, Xshat = model.predict(model.x0, Us)
Yhat = unscale(Yshat, ymean, ygain)

Parallel training

As the training problem, in general, is a nonconvex optimization problem, the obtained model often depends on the initial value of the parameters. The jax-sysid library supports training models in parallel (including static models) using the joblib library. In the example above, we can train 10 different models using 10 jobs in joblib as follows:

def init_fcn(seed):
    np.random.seed(seed)
    A  = 0.5*np.eye(nx)
    B = 0.1*np.random.randn(nx,nu)
    C = 0.1*np.random.randn(ny,nx)
    W1 = 0.1*np.random.randn(nnx,nx)
    W2 = 0.5*np.random.randn(nnx,nu)
    W3 = 0.5*np.random.randn(nx,nnx)
    b1 = np.zeros(nnx)
    b2 = np.zeros(nx)
    W4 = 0.5*np.random.randn(nny,nx)
    W5 = 0.5*np.random.randn(ny,nny)
    b3 = np.zeros(nny)
    b4 = np.zeros(ny)
    return [A,B,C,W1,W2,W3,b1,b2,W4,W5,b3,b4]

models = model.parallel_fit(Ys, Us, init_fcn=init_fcn, seeds=range(10), n_jobs=10)

flax.linen models

jax-sysid also supports recurrent neural networks defined via the flax.linen library (the flax package can be installed via pip install flax):

from jax_sysid.models import RNN

# state-update function
class FX(nn.Module):
    @nn.compact
    def __call__(self, x):
        x = nn.Dense(features=5)(x)
        x = nn.swish(x)
        x = nn.Dense(features=5)(x)
        x = nn.swish(x)
        x = nn.Dense(features=nx)(x)
        return x

# output function
class FY(nn.Module):
    @nn.compact
    def __call__(self, x):
        x = nn.Dense(features=5)(x)
        x = nn.tanh(x)
        x = nn.Dense(features=ny)(x)
        return x
    
model = RNN(nx, ny, nu, FX=FX, FY=FY, x_scaling=0.1)
model.loss(rho_x0=1.e-4, rho_th=1.e-4, tau_th=0.0001)
model.optimization(adam_epochs=0, lbfgs_epochs=2000) 
model.fit(Ys, Us)

where the extra parameter x_scaling is used to scale down (when $0\leq$ x_scaling $<1$) the default initialization of the network weights instantiated by flax.

Custom output loss

jax-sysid also supports custom loss functions penalizing the deviations of $\hat y$ from $y$. For example, to identify a system with a binary output, we can use the (modified) cross-entropy loss

$$ {\mathcal L}(\hat Y,Y)=\frac{1}{N}\sum_{k=0}^{N-1} -y_k\log(\epsilon+\hat y_k)-(1-y_k)\log(\epsilon+1-\hat y_k) $$

where $\hat Y=(\hat y_0,\ldots,\hat y_{N-1})$ and $Y=(y_0,\ldots, y_{N-1})$ are the sequences of predicted and measured outputs, respectively, and $\epsilon>0$ is a tolerance used to prevent numerical issues in case $\hat y_k\approx 0$ or $\hat y_k\approx 1$:

epsil=1.e-4
@jax.jit
def cross_entropy_loss(Yhat,Y):
    loss=jnp.sum(-Y*jnp.log(epsil+Yhat)-(1.-Y)*jnp.log(epsil+1.-Yhat))/Y.shape[0]
    return loss
model.loss(rho_x0=0.01, rho_th=0.001, output_loss=cross_entropy_loss)

By default, jax-sysid minimizes the classical mean squared error

$$ {\mathcal L}(\hat Y,Y)=\frac{1}{N}\sum_{k=0}^{N-1} |y_k-\hat y_k|_2^2 $$

Custom regularization

jax-sysid also supports custom regularization terms $r_c(z)$, where $z=(\theta,x_0)$. You can specify such a custom regularization function when defining the overall loss. For example, say for some reason you want to impose $|\theta|_2^2\leq 1$ as a soft constraint, you can penalize

$$\frac{1}{2} \rho_{\theta} |\theta|2^2 + \rho{x_0} |x_0|_2^2 + \rho_c\max{|\theta|_2^2-1,0}^2$$

with $\rho_c\gg\rho_\theta$, $\rho_c\gg\rho_{x_0}$, for instance $\rho_c=1000$, $\rho_\theta=0.001$, $\rho_{x0}=0.01$. In Python:

@jax.jit
def custom_reg_fcn(th,x0):
    return 1000.*jnp.maximum(jnp.sum(th**2)-1.,0.)**2
model.loss(rho_x0=0.01, rho_th=0.001, custom_regularization= custom_reg_fcn)

Static gain

As for linear systems, a special case of custom regularization function to fit the DC-gain of the model to steady-state input data Uss and corresponding output data Yss is obtained using the following commands:

dcgain_loss = model.dcgain_loss(Uss = Uss, Yss = Yss)
model.loss(rho_x0=1.e-3, rho_th=1.e-2, custom_regularization = dcgain_loss)

before calling the fit function. Note that this penalty involves solving a system of nonlinear equations for every input/output steady-state pair to evaluate the loss function, so it can be slow if many steady-state data pairs are given.

Upper and lower bounds

To include lower and upper bounds on the parameters of the model and/or the initial state, use the following additional arguments when specifying the optimization problem:

model.optimization(params_min=lb, params_max=ub, x0_min=xmin, x0_max=xmax, ...)

where lb and ub are lists of arrays with the same structure as model.params, while xmin and xmax are arrays of the same dimension model.nx of the state vector. By default, each value is set equal to None, i.e., the corresponding constraint is not enforced. See example_linear_positive.py for examples of how to use nonnegative constraints to fit a positive linear system.

Quasi-Linear Parameter-Varying (qLPV) models

Training qLPV models

As a special case of nonlinear dynamical models, jax-sysid supports the identification of quasi-LPV models of the form

$$x_{k+1} = A(p_k)x_k + B(p_k)u_k$$

$$y_k = C(p_k) x_k+D(p_k)u_k$$

where the scheduling vector $p_k$ is an arbitrary parametric nonlinear function (to be trained) of $x_k$ and $u_k$ with $n_p$ entries

$$p_k = f(x_k,u_k,\theta_p)$$

and

$$A(p_k) = A_{\rm lin}+\sum_{i=1}^{n_p} A_i p_{ki},~~B(p_k) = B_{\rm lin}+\sum_{i=1}^{n_p} B_i p_{ki}$$

$$C(p_k) = C_{\rm lin}+\sum_{i=1}^{n_p} C_i p_{ki},~~D(p_k) = D_{\rm lin}+\sum_{i=1}^{n_p} D_i p_{ki}$$

For both LTI and qLPV models, jax-sysid must enable the feedthrough term $D(p_k)=0$ by specifying the flag feedthrough=True when defining the model (by default, no feedthrough is in place). Moreover, for all linear, nonlinear, and qLPV models, one can force $y_k=[I\ 0]x_k$ by specifying y_in_x=True in the object constructor.

Let's train a quasi-LPV model on a sequence of inputs $U=[u_0\ \ldots\ u_{N-1}]'$ and output $Y=[y_0\ \ldots\ y_{N-1}]'$, with scheduling parameter function $f(x,u,\theta_p)$ = qlpv_fcn, initial parameters $\theta_p$ = qlpv_params_init, regularization $\rho_\theta=10^{-2}$, $\rho_{x_0}=10^{-3}$, running the L-BFGS solver for at most 1000 function evaluations:

from jax_sysid.models import qLPVModel

model = qLPVModel(nx, ny, nu, npar, qlpv_fcn, qlpv_params_init)

model.loss(rho_x0=1.e-3, rho_th=1.e-2) 
model.optimization(lbfgs_epochs=1000) 
model.fit(Ys, Us, LTI_training=True)
Yhat, Xhat = model.predict(model.x0, U)

where Us, Ys are the scaled input and output signals. The flag LTI_training=True forces the training algorithm to initialize $A_{\rm lin},B_{\rm lin},C_{\rm lin},D_{\rm lin}$ by first fitting an LTI model.

After identifying the model, to retrieve the resulting matrices $A_{\rm lin}$, $B_{\rm lin}$, $C_{\rm lin}$, $D_{\rm lin}$, ${A_i}$, ${B_i}$, ${C_i}$, ${D_i}$, $i=1,\ldots,n_p$, you can use the following:

A, B, C, D, Ap, Bp, Cp, Dp = model.ssdata()

where Ap, Bp, Cp, Dp are tensors (3D matrices) containing the corresponding linear matrices, i.e., Ap[i,:,:]=$A_i$, Bp[i,:,:]=$B_i$, Cp[i,:,:]=$C_i$, Dp[i,:,:]=$D_i$.

To attempt to reduce the number of scheduling variables in the model, you can use group-Lasso regularization as follows:

model.loss(rho_x0=1.e-3, rho_th=1.e-2, tau_g=0.1) 
model.group_lasso_p()
model.fit(Ys, Us)

Each group $i=1,\ldots,n_p$ collects the entries of $(A_i, B_i, C_i, D_i)$ and tau_g is the weight associated with the corresponding group-Lasso penalty.

Parallel training from different initial guesses is also supported for qLPV models. To this end, you must define a function qlpv_param_init_fcn(seed) that initializes the parameter vector $\theta_p$ of the scheduling function for a given seed. For example, you can train the model for 10 different random seeds on 10 jobs by running:

models = model.parallel_fit(Y, U, qlpv_param_init_fcn=qlpv_param_init_fcn, seeds=range(10), n_jobs=10)

Static models

Nonlinear regression

The same optimization algorithms used to train dynamical models can be used to train static models, i.e., to solve the nonlinear regression problem:

$$ \min_{z}r(z)+\frac{1}{N}\sum_{k=0}^{N-1} |y_{k}-f(u_k,\theta)|_2^2$$

where $z=\theta$ is the vector of model parameters to train and $r(z)$ admits the same regularization terms as in the case of dynamical models.

For example, if the model is a shallow neural network you can use the following code:

from jax_sysid.models import StaticModel
from jax_sysid.utils import standard_scale, unscale

@jax.jit
def output_fcn(u, params):
    W1,b1,W2,b2=params
    y = W1@u.T+b1
    y = [e-mail](y)+b2
    return y.T
model = StaticModel(ny, nu, output_fcn)
nn=10 # number of neurons
model.init(params=[np.random.randn(nn,nu), np.random.randn(nn,1), np.random.randn(1,nn), np.random.randn(1,1)])
model.loss(rho_th=1.e-4, tau_th=tau_th) 
model.optimization(lbfgs_epochs=500) 
model.fit(Ys, Us)

jax-sysid also supports feedforward neural networks defined via the flax.linen library:

from jax_sysid.models import FNN
from flax import linen as nn 

# output function
class FY(nn.Module):
    @nn.compact
    def __call__(self, x):
        x = nn.Dense(features=20)(x)
        x = nn.tanh(x)
        x = nn.Dense(features=20)(x)
        x = nn.tanh(x)
        x = nn.Dense(features=ny)(x)
        return x
    
model = FNN(ny, nu, FY)
model.loss(rho_th=1.e-4, tau_th=tau_th)
model.optimization(lbfgs_epochs=500)
model.fit(Ys, Us)

To include lower and upper bounds on the parameters of the model, use the following additional arguments when specifying the optimization problem:

model.optimization(lbfgs_epochs=500, params_min=lb, params_max=ub)

where lb and ub are lists of arrays with the same structure as model.params. See example_static_convex.py for examples of how to use nonnegative constraints to fit input-convex neural networks.

Classification

To solve classification problems, you need to define a custom loss function to change the default Mean-Squared-Error loss. For example, to train a classifier for a multi-category classification problem with $K$ classes, you can specify a neural network with a linear output layer generating output predictions $\hat y\in R^K$ and define the associated cross-entropy $\ell(\hat y,y) = -\sum_{k=1}^Ky_k\log\left(\frac{e^{\hat y_k}}{\sum_{j=1}^Ke^{\hat y_j}}\right)$ function as follows:

def cross_entropy(Yhat,Y):
    return -jax.numpy.sum(jax.nn.log_softmax(Yhat, axis=1)*Y)/Y.shape[0] 
model.loss(rho_th=1.e-4, output_loss=cross_entropy)

See example_static_fashion_mist.py for an example using Keras with JAX backend to define the neural network model.

Contributors

This package was coded by Alberto Bemporad.

This software is distributed without any warranty. Please cite the paper below if you use this software.

Acknowledgments

We thank Roland Toth for suggesting the use of Kung's method for initializing linear state-space models and Kui Xie for feedback on the reconstruction of the initial state via EKF + RTS smoothing.

Citing jax-sysid

@article{Bem24,
    author={A. Bemporad},
    title={Linear and nonlinear system identification under $\ell_1$- and group-{Lasso} regularization via {L-BFGS-B}},
    note = {submitted for publication. Also available on arXiv
    at \url{http://arxiv.org/abs/2403.03827}},
    year=2024
}

License

Apache 2.0

(C) 2024 A. Bemporad