[ml] wiimote libsvm experiments

Michael C. Toren mct at toren.net
Wed Jun 17 18:19:42 PDT 2009


I sadly didn't have time this week to dive into Hidden Markov models, but
I was able to massage the data a little in order be able to run libsvm on
it directly.  I was really curious to see if just libsvm by itself would
be enough to make sense of the data.

I ended up with some perl code that sorted the data by time, then slotted
each datapoint into a fixed number of buckets.  For buckets that had more
than one datapoint, I took the mean.  For buckets with no datapoint, I
assumed a linear progression from its neighbors.  For each experiment, I
randomly selected 75% of the data to use for training, and used the
remaining 25% to test my accuracy against.  I experimented with changing
two knobs -- the kernel function used by svm-train (0..3), and the number
of buckets I placed data in (10, 50, 100, and 1000). 

The results:

                     Kernel 0   Kernel 1    Kernel 2    Kernel 3
                     --------   --------    --------    --------
      10 Buckets:    96.875%    54.6875%    87.5%       93.75%
      50 Buckets:    83.908%    60.9195%    97.7011%    100%
     100 Buckets:    98.5507%   71.0145%    98.5507%    100%
    1000 Buckets:    98.4615%   63.0769%    100%        93.8462%


Some of those results look *really* promising, although I wonder if svm
is going to hold up when we introduce more complicated gestures...

-mct


More information about the ml mailing list