search

mobeets / nullSpaceControl

1 stars 0 forks source link

find hypothesis with best mean offset #13

Closed mobeets closed 8 years ago

mobeets commented 8 years ago

You can solve for the row space to get Zr. For Zn, what if (either for each kinematics condition, or just across the board) you minimize the errOfMeans (on block 1) by searching for some constant mu to always add. What would this score be?

Just estimating the mean factor from block 1, for example, gives you an errOfMeans of 1.2270, which is just as good or better than all the other hypotheses. So what if you actually looked for the best mean (while still satisfying the kinematics equations)?

mobeets commented 8 years ago

Just using fminunc to find "best mean".

Mean of block 1 latents: 1.2270 Best mean of block 1 latents: 1.2241 Mean of block 2 latents: 1.1343 Best mean of block 2 latents: 1.1237 [n.b. can overfit]

mobeets commented 8 years ago

Would be good to do conditional on kinematics condition as well...

Note that this is exactly what we are trying to predict: the mean point in the null space, conditional on each kinematics condition.