Re: [boost] Upcoming review of Time Series

From: Eric Niebler (eric_at_[hidden])
Date: 2007-07-27 12:20:01

Hugo Duncan wrote:
> Eric,
> 
>> OK, the time period isn't a constant, but what is it measured in?
>> Milliseconds? Then you can make milliseconds your discretization, and
>> then resample the data at a coarser discretization with the
>> coarse_grain() algorithm. Am I still not understanding your problem?
>> Perhaps it would help if you were more explicit about what data is being
>> pulled from the bus, exactly. Values and time? Time measured how?
> 
> Light finally goes on! I agree that this is a solution.  To do this I use  
> a sparse_series, and each point becomes a run, I believe.  (Just for the  
> record, pulling time and values.)  Thanks for your patient explanation - I  
> think it was the name "sparse" that was giving me problems, as I was  
> thinking of it as dense with variable discretisation (40-60ms) rather than  
> sparse at constant millisecond discretisation.

I don't think sparse series is the right one for you. A sparse series of 
value, offset pairs such as the following:

   (42, 0)
   (43, 9)
   (44, 21)
   ...

would have the value 42 at offset 0, 43 at offset 9, 44 at offset 21, 
*and zero everywhere else*. So at offset 1, the value would be zero. You 
probably want it to be 42 there also, right?

I think what you want is a piecewise constant series of (value, offset, 
end offset) tuples like this:

   (42, 0, 9)
   (43, 9, 21)
   (44, 21, 30)
   ...

The offsets are the (irregularly spaced) intervals at which you are 
sampling your bus. Then you can regularize the series by resampling to a 
coarser discretization; e.g., by taking the value at every 10th offset.

>> I'll have to get back to you about the rest. I'm leaving tomorrow for a
>> long weekend.
> 
> Have a good weekend.

Thanks! I'm off.

-- 
Eric Niebler
Boost Consulting
www.boost-consulting.com
The Astoria Seminar ==> http://www.astoriaseminar.com