[Python] I need some intuition about the "trust_radius" in scipy's trust-ncg minimization...

I'm trying to minimize a fairly complex and non-linear problem, and since I have analytic Jacobians and Hessians at my disposal, I'd like to use Newton Conjugate Gradient minimization within scipy. However, I'm not getting very good results with the Newton-CG method of scipy.optimize.minimize, and started tinkering with the trust-ncg method instead. I am getting better results, so I want to use this minimization method. My problem is that I really don't have any sense for what the 'trust radius' is, so I'm just blindly setting values for the initial_trust_radius and max_trust_radius parameters.

The only reference I've seen for this method in the scipy documentation is this paper, which gives some intuition on the eta parameter, but that's all I could glean. I'm really hoping someone can provide some intuition on what the size of the trust region means, in terms of the data that are actually being used and the minimization problem at hand.

Thanks!!

-b

Continue reading...