Mahout and online machine learning

[Marek Stefaniak <ma...@gmail.com>]

I got to solve a challenging task in the field of machine learning online,
and involving theconstruction of
system monitoring, controlling and optimizing the work of another device (eg
 aircraft).
Input:
Entrance system will stream data describing the work of the airplane and
śrdowiska (eg,strength and wind direction, precipitation, humidity).
About 60 parameters sampled at certain intervals (eg every 10 seconds)

Objective:
The system has two purposes:
1. Control settings to optimize the aircraft's flight
2. Warn of the possibility of a disaster (eg, engine overheating or too low
flight)

I assume that the target sytem should work on one computer (not in cluster).

The question is whether it can be achieved with the mahout.
If so, by which the algorithm?

Yours
Marek Stefaniak

Re: Mahout and online machine learning

[Marek Stefaniak <ma...@gmail.com>]

Thank you for your reply.
I put my comments in the text

2011/2/5 Ted Dunning <te...@gmail.com>

> Some answers,
>
> a) Mahout is not ready for use in applications that will kill people if
> they
> fail
>

All the time talking about unmanned aircraft, but I wonder maturity of the
project.Particularly in the context of high misclassification rates (which is
not a problem insystems like "we recommend you ...") and the type of input
data (numerical data instead of text). Is the mahout is an appropriate tool for
me?


> b) if you have a situation where you have training examples that you have
> classified and labeled, then Mahout online classifiers may help.
>

It is difficult to provide a broad collection of labeled examples. (difficult
to predict allpossible instances of emergencies). In my application will be
key features "fuzzylabeling"- which means the answer to the question. How "far
" I was at this time ofdisaster
I read about a similar system
http://www.nasa.gov/centers/dryden/pdf/104760main_H-2575.pdf, who
intensively uses various types of neural networks used for theoptimization
and minimization. Intuitively, it seems to me that neural networks are
particularly suited to working with numerical data. Is this true?
.

>
> c) if you have data that is not sparse, then Mahout's online learners will
> probably work reasonably well, but you will need to use them differently
> from the way that most people currently use them.  These learners were
> designed with very sparse data in mind.
>
> 2011/2/4 Marek Stefaniak <ma...@gmail.com>
>
> > I got to solve a challenging task in the field of machine learning
> online,
> > and involving theconstruction of
> > system monitoring, controlling and optimizing the work of another device
> > (eg
> >  aircraft).
> > Input:
> > Entrance system will stream data describing the work of the airplane and
> > śrdowiska (eg,strength and wind direction, precipitation, humidity).
> > About 60 parameters sampled at certain intervals (eg every 10 seconds)
> >
> > Objective:
> > The system has two purposes:
> > 1. Control settings to optimize the aircraft's flight
> > 2. Warn of the possibility of a disaster (eg, engine overheating or too
> low
> > flight)
> >
> > I assume that the target sytem should work on one computer (not in
> > cluster).
> >
> > The question is whether it can be achieved with the mahout.
> > If so, by which the algorithm?
> >
> > Yours
> > Marek Stefaniak
> >
>



-- 

Pozdrawiam,
Marek Stefaniak

Re: Mahout and online machine learning

[Ted Dunning <te...@gmail.com>]

Some answers,

a) Mahout is not ready for use in applications that will kill people if they
fail

b) if you have a situation where you have training examples that you have
classified and labeled, then Mahout online classifiers may help.

c) if you have data that is not sparse, then Mahout's online learners will
probably work reasonably well, but you will need to use them differently
from the way that most people currently use them.  These learners were
designed with very sparse data in mind.

2011/2/4 Marek Stefaniak <ma...@gmail.com>

> I got to solve a challenging task in the field of machine learning online,
> and involving theconstruction of
> system monitoring, controlling and optimizing the work of another device
> (eg
>  aircraft).
> Input:
> Entrance system will stream data describing the work of the airplane and
> śrdowiska (eg,strength and wind direction, precipitation, humidity).
> About 60 parameters sampled at certain intervals (eg every 10 seconds)
>
> Objective:
> The system has two purposes:
> 1. Control settings to optimize the aircraft's flight
> 2. Warn of the possibility of a disaster (eg, engine overheating or too low
> flight)
>
> I assume that the target sytem should work on one computer (not in
> cluster).
>
> The question is whether it can be achieved with the mahout.
> If so, by which the algorithm?
>
> Yours
> Marek Stefaniak
>