CS885 Lecture 14c: Trust Region Methods

[@junchaoli1062]

very clear! thank you pascal!

[@Houd1ny]

This is great. Finally, understand TRPO

[@naifalkhunaizi4372]

Thank you Professor Pascal!

[@kamiladamsky]

Great lecture!

[@kadaj2k7]

Thank you for this video! Very helpful.

[@johntrash4962]

Thank you for your clear explanation!

[@hariprasad-qk8vs]

Very beautifully explained thank you

[@万杨阳]

I see several textbooks or blogs write that "trust region is dual method to line search", I can't agree with it. I think trust region is a superior method to line search because it's a quadratic approximation to the original function rather than a linear approximation, besides, it defines a region which parameters can't jump out of. Linear search actually has implicitly defined a region for step dt such that f(x+dt)>=0. If min{f(x)} < 0, we should define a trust region for linear search. Am I right or I have ignored something?

[@cucumcumber]

I guess the dual space of l1 in line search is l_infinity in trust region.

[@yehiamoustafa9801]

Great Lecture Thank you!

[@MiiiiiiTube]

really nice, thank you so much! :D

[@jonnychan785]

Good lecture thankyou!

[@BillHaug]

thank you

[@НиколайНовичков-е1э]

Thank you!

[@cyberpunk_edgerunners]

thank you ,sir

[@李凯-f1d]

Thanks prof, really helpful

[@AJ-et3vf]

great video sir! thank you!

[@mahsatech]

man that was great

[@alimaster5292]

how to answer this question What is the main motivation behind Trust region methods? What could be a signal to enlarge the Trust region in the next step? What is the Cauchy point?

[@金飘飘-o2o]

Do u have idea now？

[@jfigueroa789]

You can find all the information in the Nocedal’s book “Numerical Optimization”.

[@alexanderyau6347]

Good very intuitive