PiecewiseInference.jl is a library to enhance the convergence of dynamical model parameter inversion methods. It provides features such as
- a segmentation strategy,
- the independent estimation of initial conditions for each segment,
- parameter transformation,
- parameter and initial conditions regularization
- mini-batching
Taken altogether, these features regularize the inference problem and permit to solve it efficiently.
Open Julia REPL and type
using Pkg; Pkg.add(url="https://github.com/vboussange/PiecewiseInference.jl")That's it! This will download the latest version of PiecewiseInference.jl from this git repo and download all dependencies.
Check out this blog post providing a hands-on tutorial.
See also the API documentation and the test folder.
DiffEqFlux is a package with similar goals as PiecewiseInference, and proposes the method DiffEqFlux.multiple_shooting, which is close to PiecewiseInference.inference but where initial conditions are not inferred. PiecewiseInference further proposes several utility methods for model selection.
Boussange, V., Vilimelis-Aceituno, P., Schäfer, F., Pellissier, L., Partitioning time series to improve process-based models with machine learning. [bioRxiv] (2024), 46 pages.