harrispopgen / mushi

[mu]tation [s]pectrum [h]istory [i]nference
https://harrispopgen.github.io/mushi/
MIT License
24 stars 6 forks source link

Rank regularize #23

Closed wsdewitt closed 5 years ago

wsdewitt commented 5 years ago

Some exciting updates:

  1. accelerated proximal gradient descent with backtracking line search—wicked fast!
  2. spectral regularization with a nuclear norm penalty (controlled by parameter \beta) to encourage low-rank solutions (closes #14)*

* combining fused LASSO (\lambda * \alpha > 0) with spectral regularization (\beta > 0) is not implemented, since that will require a proximal operator for the combined fused lasso and nuclear norm term (neither being differentiable). The current implementation allows a (differentiable) l2 penalty on the first time derivative (\lambda > 0, \alpha = 0) in combination with spectral regularization (\beta > 0). Fused LASSO is still available without spectral regularization (\beta = 0).