Hacker News Reader

I think there are a few derivative-free global optimizers, curious what sets this one apart? see e.g. https://www.microprediction.com/blog/optimize . If the current scipy-native one is unmaintained and this one is superior it certainly makes sense to update.

View on HN

The proposal isn't a "new" algorithm, rather a rewrite (plus bug fixes) of the existing scipy implementation of "COBYLA" (which I don't think it tested in your linked comparison).

These algorithms work best on smooth functions with mid-to-low 100's of variables. They are very sample efficient, often obtaining superlinear convergence using just a linear number of evaluations of the objective function.

Reply on news.ycombinator.com