Updated:

I felt your pain, people. I did HW3 today. I do not vouch for these results absolutely (you can look at what I did tomorrow and compare notes). I used no loops in the cost function and gradient; I multiplied matrices instead (you can use loops if you want to, but don't use a loop to calculate h_theta(x) --- use a dot product). 

Columns I used: Smog Rating, Greenhouse Gas Score, DisplMunged, CylMunged, TransNumber, DriveNumber, FuelNumber, ClassNumber, SmartNumber, Mpg (combined)

I performed feature scaling and mean normalization as described by Ng and given in lecture.

Cost before training thetas: 

866.57590361445796

Cost after training:

28.503855265508022

My learning curve:

My cost before training thetas was about 866.

I didn't. I was just asking if you did, because I also got 0.01 - 0.001 range for cost too, when I scaled y.

msargent wrote:

You are not supposed to scale y.

You are right, I did. A Plato dollar for you. See updated post for new values.