Re: [Tuxmath-devel] GSoC project on "Personalized, adaptive game-play to improve learning in Tuxmath"

[Siddharth K. <sid...@gm...>]

Hi everyone,

An update how far I have got:
* I have made a separate branch 'bbn' in tuxmath's git repo. where I am
committing all the changes.
* As for the progress, I have a Bayesian Network library which supports to
create, set initial probabilities and allows inference on bayesian networks.
It uses Pearl's message-passing
algorithm<http://en.wikipedia.org/wiki/Belief_propagation>for
inferencing. Initially, I wanted to use an existing library with a
compatible license to handle the inference part, but I had hard luck finding
C implementations and so decided to write one myself.
* My current implementation supports only tree topology, which is good
enough when I don't include global nodes in a topic cluster. Refer to this
figure <http://gscbbn.files.wordpress.com/2011/05/topic-cluster.png> for a
topic cluster's structure. So, while I work on implementing the inference
part for singly-connected networks, I can start testing the BBN in the game.


There are also a couple of things which I need help and suggestions with:
* I have added the new source files in src/bayesian. How do I include them
in the build? I have not worked with GNU autotools and cmake before.
* Earlier in this thread, we discussed about modelling acceptable
failure/challenge for a user and the solution was something on the lines of
the system giving user recommended options for lessons based on difficulty.
On the back-end part, I take this into account by making a global node for
"challenge level"
[Figure<http://gscbbn.files.wordpress.com/2011/06/global_nodes.png>].
On the interface part, this would require changes. I have some initial
thoughts on that and am write it in a separate mail.


As the mid-term evaluations are only a week now, my priority will be to add
documentation and error checks to the existing code, and update the blog.

Thanks,
Siddharth

On Wed, Jun 1, 2011 at 3:13 AM, Tim Holy <ho...@wu...> wrote:

> Hi Siddarth,
>
> I'd also add that perhaps it might be best to confine this to "training
> academy." The intent behind the arcade games is that everyone is on equal
> footing, so getting the high score for your class represents an "absolute"
> achievement rather than "you did very well given what we expected of you"
> (i.e., what the Bayesian algorithm decided to set as problems). But for the
> training academy, the latter is indeed probably more appropriate .
>
> Best,
> --Tim
>
>
> On Tuesday, May 31, 2011 11:55:44 am David Bruce wrote:
> > Hi Siddarth,
> >
> > > Also, I would like if the question generation happens for every wave
> > > instead of happening once in a game.
> >
> > The question list is generated at the beginning of the game.  The
> > waves affect things like the background image and the speed of the
> > comets, but not the questions.  The question-generation code is in
> > mathcards.[ch], which doesn't know anything about waves.
> >
> > Depending on a few config file variables (see e.g. ~/.tuxmath/options
> > or any of the files under data/missions/),  answered or missed
> > questions are handled differently (I apologize about some of the
> > variable names, which now seem rather confusing)
> >
> > PLAY_THROUGH_LIST (boolean) - if set to 1 ("true"), correctly-answered
> > questions are removed from the list, so that the game ends in victory
> > when there are no remaining questions.  If set to 0 ("false"), the
> > questions are reinserted in a random question and the game goes on
> > indefinitely ("arcade mode").
> >
> > REPEAT_WRONGS (boolean) - controls the equivalent behavior for
> > questions the player misses (i.e. the ones that hit the igloos).
> > However, for this one, 1 means the questions are reinserted, and 0
> > means they are discarded, the opposite of the behavior for
> > PLAY_THROUGH_LIST.  This is set to 1 for all of the bundled lessons.
> >
> > COPIES_REPEATED_WRONGS (int) controls how many copies of the missed
> > question get put back in.  For example, if this is set to
> > 2, if a player misses "3 + 3 = ?", he/she will have to answer it twice
> > correctly after that to "win" the game.  This provides a minimal form
> > of adaptive learning, but it could be extended to include questions
> > closely related to the question that was missed.
> >
> > > The reason is to account for learner's
> > > performance of the past waves for question generation of the present
> > > wave. This wouldn't affect a Math command training lesson, but on tux
> > > missions and for arcade games, this can make a difference in the
> > > question selection.
> >
> > All very valid goals!  It seems clear that our current scheme isn't
> > sufficiently flexible to do everything needed for your project.  If
> > you are going to have "intra-game/inter-wave" feedback, we need to add
> > some functions to mathcards to support generating and inserting
> > additional questions during the game.
> >
> > Best,
>