Edit on GitHub

Sparse Identification of Nonlinear Dynamics

Sparse Identification of Nonlinear Dynamics - or SINDy - identifies the equations of motion of a system as the result of a sparse regression over a chosen basis. In particular, it tries to find coefficients $\Xi$ such that:

\[\Xi = \min ~ \left\lVert Y^{T} - \Theta(X, p, t)^{T} \Xi \right\rVert_{2} + \lambda ~ \left\lVert \Xi \right\rVert_{1}\]

where, in most cases, $Y$is the data matrix containing the derivatives of the state data stored in $X$. $\Theta$ is a matrix containing candidate functions $\xi$ over the measurements in $X$.

Example

As in the original paper, we will estimate the Lorenz System. First, let's create the necessary data and have a look at the trajectory.

using DataDrivenDiffEq
using ModelingToolkit
using OrdinaryDiffEq

using LinearAlgebra
using Plots
gr()

# Create a test problem
function lorenz(u,p,t)
    x, y, z = u
    ẋ = 10.0*(y - x)
    ẏ = x*(28.0-z) - y
    ż = x*y - (8/3)*z
    return [ẋ, ẏ, ż]
end

u0 = [-8.0; 7.0; 27.0]
p = [10.0; -10.0; 28.0; -1.0; -1.0; 1.0; -8/3]
tspan = (0.0,100.0)
dt = 0.001
problem = ODEProblem(lorenz,u0,tspan)
solution = solve(problem, Tsit5(), saveat = dt, atol = 1e-7, rtol = 1e-8)

Additionally, we generate the ideal derivative data.

X = Array(solution)
DX = similar(X)
for (i, xi) in enumerate(eachcol(X))
    DX[:,i] = lorenz(xi, [], 0.0)
end

To generate the symbolic equations, we need to define a Basis over the variables x y z. In this example, we will use all monomials up to degree of 4 and their products:

@variables x y z
u = Operation[x; y; z]
polys = Operation[]
for i ∈ 0:4
    for j ∈ 0:i
        for k ∈ 0:j
            push!(polys, u[1]^i*u[2]^j*u[3]^k)
            push!(polys, u[2]^i*u[3]^j*u[1]^k)
            push!(polys, u[3]^i*u[1]^j*u[2]^k)
        end
    end
end

basis = Basis(polys, u)

A Basis consists of unique functions, so duplicates will be included just once

To perform the sparse identification on our data, we need to define an Optimizer. Here, we will use STRRidge, which is described in the original paper. The threshold of the optimizer is set to 0.1. An overview of the different optimizers can be found below.

opt = STRRidge(0.1)
Ψ = SINDy(X, DX, basis, opt, maxiter = 100, normalize = true)
Sparse Identification Result with 7 active terms.

Ψ is a SINDyResult, which stores some about the regression. As we can see, we have 7 active terms inside the model. To look at the equations, simply type

print_equations(Ψ)
3 dimensional basis in ["x", "y", "z"]
dx = p₁ * x + p₂ * y
dy = p₃ * x + p₄ * y + p₅ * x * z
dz = p₆ * z + p₇ * x * y

First, let's have a look at the $L2$-Error and Akaikes Information Criterion of the result

get_error(Ψ)
3-element Array{Float64,1}:
 6.7202639134663155e-12
 3.4645801800720905e-11
 1.2876598297504605e-11
get_aicc(Ψ)
3-element Array{Float64,1}:
 106.90381476450982
 102.34377824714254
 104.30267815708592

We can also access the coefficient matrix $\Xi$ directly via get_coefficients(Ψ).

To generate a numerical model usable in DifferentialEquations, we simply use the ODESystem function from ModelingToolkit. The resulting parameters used for the identification can be accessed via parameters(Ψ). The returned vector also includes the parameters of the original Basis used to generate the result.

ps = parameters(Ψ)
sys = ODESystem(Ψ)
dudt = ODEFunction(sys)

prob = ODEProblem(dudt, u0, tspan, ps)
sol = solve(prob, Tsit5(), saveat = solution.t, atol = 1e-7, rtol = 1e-8)

Let's have a look at the trajectory of $u_{1}(t)$.

Finally, let's investigate the error of the chaotic equations:

which resembles the papers results. Next, we could use classical parameter estimation methods or use DiffEqFlux to fine-tune our result (if needed).

Functions

DataDrivenDiffEq.SINDy — Function
SINDy(X, Y, basis, opt = STRRidge(); p, t, maxiter, convergence_error, denoise, normalize)
SINDy(X, Y, basis, lambdas, opt = STRRidge(); f, g, p, t, opt, maxiter, convergence_error, denoise, normalize)

Performs Sparse Identification of Nonlinear Dynamics given the data matrices X and Y via the AbstractBasis basis. Keyworded arguments include the parameter (values) of the basis p and the timepoints t which are passed in optionally. opt is an AbstractOptimizer useable for sparse regression, maxiter the maximum iterations to perform and convergence_error the bound which causes the optimizer to stop. denoise defines if the matrix holding candidate trajectories should be thresholded via the optimal threshold for singular values. normalize normalizes the matrix holding candidate trajectories via the L2-Norm over each function.

Typically X represent the state measurements and Y the measurements of the differential state. Since only the number of measurements (column dimension of the matrices) have to be equal, it is possible to augment X with additional data, e.g. external forcing or inputs.

If SINDy is called with an additional array of thresholds contained in lambdas, it performs a multi objective optimization over all thresholds. The best candidate is determined via the mapping onto a feature space f and an (scalar, positive definite) evaluation g. The signature of should be f(xi, theta, yi) where xi are the coefficients of the sparse optimization,theta is the evaluated candidate library and yi are the rows of the matrix Y. Returns a SINDyResult. If the pareto optimization is used, the result combines the best candidate for each row of Y.

source
DataDrivenDiffEq.sparse_regression — Function
sparse_regression(X, Y, basis, p, t, maxiter, opt, denoise, normalize, convergence_error)
sparse_regression!(Xi, X, Y, basis, p, t, maxiter, opt, denoise, normalize, convergence_error)
spares_regression!(Xi, Theta, Y, maxiter, opt, denoise, normalize, convergence_error)

Performs a sparse regression via the algorithm opt <: AbstractOptimizer. maxiter specifies the upper bound of the iterations of the optimizer, convergence_error the breaking condition due to convergence. denoise defines if the matrix holding candidate trajectories should be thresholded via the optimal threshold for singular values. normalize normalizes the matrix holding candidate trajectories via the L2-Norm over each function.

If the data matrices X, Y are given with a Basis basis and the additional information for parameters p and time points of the measurements t, it returns the coefficient matrix Xi and the iterations taken. This function is also available in place, which returns just the iterations.

If Xi, Theta and Y are given, the sparse regression will find the coefficients Xi, which minimize the objective and return the iterations needed.

Example

opt = STRRidge()
maxiter = 10
c_error = 1e-3

Xi, iters = sparse_regression(X, Y, basis, [], [], maxiter, opt, false, false, c_error)

iters = sparse_regression!(Xi,X, Y, basis, [], [], maxiter, opt, false, false, c_error)

Xi2 = zeros(size(Y, 1), size(X, 1))
iters = sparse_regression!(Xi2, X, Y, maxiter, opt, false, false, c_error)
source
DataDrivenDiffEq.SparseIdentificationResult — Type
SparseIdentificationResult()

Contains the result of a sparse identification. Contains the coefficient matrix Ξ, the equations of motion, and its associated parameters. It also stores the optimizer, iteration counter, and convergence status.

Additionally, the model is evaluated over the training data and the $L_2$-error, Akaikes Information Criterion, and the $L_0$-Norm of the coefficients is stored.

source