I updated the Newton solver to fail to gradient descent if the line search fails. This helps the optimization make progress even when the Hessian is ill-conditioned. The next step after this pull request is merged is to add a quasi-Newton solver (e.g. BFGS).
Exposed initialization of barrier epsilon in the UI.
I updated the Newton solver to fail to gradient descent if the line search fails. This helps the optimization make progress even when the Hessian is ill-conditioned. The next step after this pull request is merged is to add a quasi-Newton solver (e.g. BFGS).