AutoGP.jl
This package contains the Julia reference implementation of AutoGP, a method for automatically discovering models of 1D time series data using Gaussian processes, as described in
Sequential Monte Carlo Learning for Time Series Structure Discovery.
Saad, F A; Patton, B J; Hoffmann, M D.; Saurous, R A; Mansinghka, V K.
ICML 2023: Proc. The 40th International Conference on Machine Learning.
Given observed time series data, AutoGP uses Bayesian structure learning to synthesize covariance kernel functions and parameters for modeling the data, which is in contrast to traditional machine learning packages that only learn the parameters of a fixed, user-specified covariance kernel function.
Installing
The package can be added using the Julia package manager. From the Julia
REPL (version 1.8+), type ]
to enter the Pkg REPL mode and run
pkg> add AutoGP
Tutorials
Please see https://fsaad.github.io/AutoGP.jl
Developer Notes
Building Documentation
$ julia --project=. docs/make.jl
$ python3 -m http.server --directory docs/build/ --bind localhost 9090
Building From Clone
- Obtain Julia 1.8 or later.
- Clone this repository.
- Set environment variable:
export JULIA_PROJECT=/path/to/AutoGP.jl
- Instantiate dependencies:
julia -e 'using Pkg; Pkg.instantiate()'
- Build PyCall:
PYTHON= julia -e 'using Pkg; Pkg.build("PyCall")'
- Verify import works:
julia -e 'import AutoGP; import PyPlot; println("success!")'
Citation
@inproceedings{saad2023icml,
title = {Sequential Monte Carlo Learning for Time Series Structure Discovery},
author = {Saad, Feras A. and Patton, Brian J. and Hoffmann, Matthew D. and Saurous, Rif A. and Mansinghka, V. K.},
booktitle = {Proceedings of the 40th International Conference on Machine Learning},
series = {Proceedings of Machine Learning Research},
fvolume = {},
fpages = {},
year = {2023},
publisher = {PMLR},
}