# #2 ML … Gradient Descent

26.7.17 wednes

!! Getting stuck very early on may be avoided by taking different values for y=mx+c and then taking the min of all the diff. cases!! (???)

Q- What are the different methods to avoid stuck at local-minima? Is there a single good method which will never stuck at local-minima, given a good enough time?

A-

!1 first m = (last h — first h)/(last x — first x)

first c = sum(h(x)) / N

!!!2 in each step of optimization

take the best of a)m=m-alpha(fm) ; c = c-alpha(fc)

b)m = m-alpha(fm) ; c = c

c)m = m ; c = c- alpha(fc)

Q- Where do I get good, reliable and working data for Gradient Descent(to start) ?

A-