Re: some news / ideas

On Sunday 05 of October 2003 19:13, Peter Bienstman wrote:
> Yeah, that's right. I also think it's important to stress that users should
> be consistent in the grades they choose. I guess the NN will not be very
> good if the user gradually changes his/her meaning of e.g. grade 3 over
> time.

Yes. I now think about adding description, on the status bar, for every grade, 
for example:
grade 5: "you remember it very well - last interval was too short"
grade 4: "you remember it - last interval was fine"
grade 3: "you rember it, but has some difficultly - last interval was slightly 
too long",
etc.
I think it ("last interval was...)" will simply explain how to interpret 
grades and how ANN uses them as feedback.

> Now that we're on that subject, I'm not sure how useful it would be to
> train the NN on data from lots of different users, because my
> interpretation of the grades could be different from yours.

Yes, if there would be small amount of data this could have a bad effect.
But I think there will be sufficient amount of them to generalize.
On the same cases that will have different resulting interval for ANN to learn 
- I will take mean of it.
Futhermore: I have somewhere a simple program in C (I'm not sure if it's in 
extras/ on memaid package - might be that's not, as it's a quick hack without 
any instructions) that scans data files and show me which cases are "weird" 
for ANN. (weird == resulting interval differ much from the one currently 
trained ANN "thinks" should be) so I can correct them manually, or just 
delete.
And in extras/ there is a sorter that sorts me (initially converted data*.bin 
files to readable ASCII format) data by specific variable (e.g. grade, 
interval, etc.) so I am able manually clean/correct some cases that are 
clearly bad (and I see it with naked eye - using common sense and comparing 
with other cases) for general ANN knowledge.

All-in-all, I think collecting user_data.bin files and trains ANN to be more 
general can help.
But of course - most important is training ANN to specific user that takes 
place on user's own learning process. 


-- 
David Calinski
http://neodave.civ.pl
/* FreeBSD: When we say free, we mean it. */



-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf