search

alphaville / optimization-engine

Nonconvex embedded optimization: code generation for fast real-time optimization + ROS support
https://alphaville.github.io/optimization-engine/
Other
512 stars 53 forks source link

Fix for non-finite gamma #262

Closed korken89 closed 2 years ago

korken89 commented 2 years ago

In my search for the following issue (https://github.com/korken89/lbfgs-rs/pull/8) I also found what is in this PR.

Is there a better way to estimate gamma when there is no curvature? Right now I just made it based on the starting cost and gradient, which works well in my tests but I don't think its a good solution.

korken89 commented 2 years ago

This issue only seems to happen when you have a problem that uses L1 norm or robust cost functions. This popped up for me when I was using the Huber norm in an MHE, and got a lot of outliers in the data.

I remember we discussed to some length before how PANOC could better support problems which have a L1 norm in it, did we ever write out thoughts down somewhere on that?

korken89 commented 2 years ago

Nice additions, thanks for cleaning up my mess! xD