Optimizer

DataDrivenDiffEq comes with some implementations for sparse regression included. All of these are stored inside the DataDrivenDiffEq.Optimize package and extend the AbstractOptimizer, if an explicit optimization is needed or the AbstractSubspaceOptimizer for an implicit problem (where the solution is within the nullspace).

Functions

STRRidge(λ = 0.1)

opt = STRRidge()
opt = STRRidge(1e-1)

ADMM()
ADMM(λ, ρ)

opt = ADMM()
opt = ADMM(1e-1, 2.0)

SR3(λ, ν, R)
SR3(λ = 1e-1, ν = 1.0)

opt = SR3()
opt = SR3(1e-2)
opt = SR3(1e-3, 1.0)

ADM()
ADM(λ = 0.1)

Implementing New Optimizer

Similarly to Algorithms for Estimation, the extension of optimizers is more or less straightforward. Suppose you want to define a new optimizer MyOpt, which should solve $A~X = Y$ for a sparse $X$.

mutable struct MyOpt <: DataDrivenDiffEq.Optimize.AbstractOptimizer
    threshold
end

To use MyOpt within SINDy, an init! function has to be implemented.

function init!(X::AbstractArray, o::MyOpt, A::AbstractArray, Y::AbstractArray)
    X .=  A \ Y
end

To perform thresholding - and use maybe for searching the right threshold - a setter and getter is required:

set_threshold!(opt::MyOpt, threshold) = opt.threshold .= threshold

get_threshold(opt::MyOpt) = opt.threshold

And, at last, the method which fits the data and returns the iterations needed:

function fit!(X::AbstractArray, A::AbstractArray, Y::AbstractArray, opt::MyOpt; maxiter, convergence_error)
    # Compute awesome stuff here
    return iterations
end