Hi Christian,
> is there a way to process time series of potentially different length
> with MDP?
How do you store such a thing? I mean, numpy arrays are
Ndimensional arrays, so if you store the time series in numpy
arrays you'll have to either use zeropadding or NaN or masked
arrays. There's been a lot of discussion about introducing
missingdata and notavailable data structures in numpy, a bit like
R does, but as far as I know, nothing that can be used right now.
> So far it seems that MDP can only deal with vectorial data,
Yes, MDP right now only works with 2D numpy arrays, and changing that
would be a big effort. Possible but not easy.
> by zeropadding every time series to the maximum length and
> for nodes that actually care about different lengths also pass the
> lengths of the different time series. But I would be interested in
> better ways to deal with this.
Maybe you could still store the time series in numpy arrays, and insert
a NaN where the time series is interrupted. Your nodes would then
know the length of the time series without you needing to pass
additional arguments. Beware that even if numpy works with NaNs, it
makes everything much slower, so I am not sure this is a good idea.
Or maybe just use pandas? http://pandas.pydata.org/
Ciao,
Tiziano
