options.skipf for multisession has to be set up manualy

MBB-team / VBA-toolbox

The VBA toolbox

GNU General Public License v3.0

129 stars 67 forks source link

options.skipf for multisession has to be set up manualy #5

Closed vasilisa closed 8 years ago

vasilisa commented 9 years ago

Just a note but it might be worth mentionning it in the Wiki section on multisession inversion:

If you invert multisession learning models and want to skip the evaluation at state x0 (since you don't have a feedback for the first choice) you should set up options.skipfs mainly as a vector with ones at the beginning of each session. Puting options.skipfs = 1 will just skip the first observation for the first session and other estimates would be biased.

jdaunize commented 9 years ago

Dear Vasilisa, This is a very good point. I think this is due tot he way the multisession inversion has been implemented. Lionel: I believe this relates to the fact that session are concatenated over time. Can you to have a look and correct this? Ideally:

the default should be such that when only specifying one options.skipf for the whole set of sessions, then multisession inversion should apply this to each and every session.
users may be allowed to give an options.skipf per session, if this is required.

Cheers, Jean.

lionel-rigoux commented 9 years ago

The problem is how to deal with sessions of different sizes. What I can do:

if skipf has the size of one session and all sessions are similar: duplicate skipf
if skipf has the size of one session but sessions are dissimilar: throw an error
if skipf has the size of all sessions concatenated, assume it consists of the concatenation of session-wise skipf. I can also accept an array of skipf that will be automatically combined after a size chek.

jdaunize commented 9 years ago

Lionel, I think this can be simplified, under the assumption that skipf is useed only for the first(s) state updates. (This is usually what happens). I suggest the following slight modification of your suggestion:

[OK]
create 0-padded skipf for all sessions, with 1's in the same place than the user-specified skipf.
disallow 'wrong size' skipf, to prevent possible confusions (and issue an error).

Note: we should agree upon the 'full' multisession format of options.skipf. Why not a cell array? Note 2: a warning should be issued for cases 1 and 2, e.g.: "Warning [multisession inversion]: replicating options.skipf for all sessions!". Cheers, J.