questions 5 & 6

[geekoftheweek]

Once we find an average hypothesis, we have to compute and . In order to compute the expectation values of bias/var wrt x, I assume we need to generate a *new* set of points. Correct? How big should that set be?

[geekoftheweek]

...or are we just supposed to use the points generated in order to calculate g_bar? That would mean that bias and var have twice as many points to average over than the number of data sets used to calculate g_bar, because each data set had two two data points.

[sanbt]

So to calculate g_bar you used 2 points to get each hypothesis and average over them.

Now Bias and var should come from the entire range of the real line. I would say
about hundreds range from -1 to 1.

[geekoftheweek]

Originally Posted by sanbt

Now Bias and var should come from the entire range of the real line. I would say
about hundreds range from -1 to 1.

Are you saying generate 100 new points?

[sanbt]

Originally Posted by geekoftheweek

Are you saying generate 100 new points?

yes

[gah44]

All these are approximating integrals.

Many problems really are sums, but this one is, theoretically, continuous.

First you do 2D integrals to compute a, a 1D integral to compute bias,
and a 3D integral to compute variance.

(I think it would also work to compute bias+variance in the first place, and subtract bias to get variance, but I didn't try that.)

I used equally space points for all, but you could also use random points.

If I was in the right mood, I might have done Gaussian quadrature, or some other numerical integration method.