Learning with variable number of attributes
setti...@gmail.com
Was just wondering if there is a good reinforcement learning algorithm
that is flexible in adding/removing attributes during the training
process. It seems that most ANN models has fixed number of inputs and
it is not easy task to add an 'ad-hoc' input component once the
training is under way. Any pointers for this is appreciated. Thanks!
John
Greg Heath
In general, it is highly improbable that one will find
the optimum combination of inputs. Therefore, I just
rely on backward elimination after I exclude my
favorite inputs from elimination.
Hope this helps.
Greg
Tomasso
"xch...@wisc.edu" <setti...@gmail.com> wrote in message news:1171497367....@l53g2000cwa.googlegroups.com...
TreeNet (boosted decision trees). Currently implementation requires all
attributes to be declared ahead of time, but values can be "missing" for
any of the input fields. A missing value is still a value (but the value is
"missing"). TreeNet combines a large number of simple decision trees.
T.
Greg Heath
If I was going to use a method which drops and adds
variables, I would translate the learning set to zero mean
and use a fixed input value of zero for dropped variables.
Hope this helps.
Greg
Gregory Ivahnenko
> Was just wondering if there is a good reinforcement learning algorithm
> that is flexible in adding/removing attributes during the training
> process.
Inductive GMDH algorithms gives possibility to find automatically
interrelations in data, select optimal structure of network and
increase the accuracy of existing algorithms. As input variables can
be used any parameters, which can influence on the process. Linear or
non-linear, probabilistic models or clusterizations are selected by
minimal value of an external criterion. GMDH algorithms are rather
simple and they get information directly from data sample.
This self-organizing approach is different from deductive methods or
networks used commonly for modeling on principle. It has inductive
nature - problems solution is based on sorting procedure by external
criterion. The effective input variables, number of layers and neurons
in hidden layers, optimal model structure are determined
automatically. It also solves the problem of inputs collinearity and
fixing of selected variables in output automatically. This is based on
fact that external criterion characteristic have minimum during
complication of model structure.
The Group Method of Data Handling (GMDH) is based on sorting-out of
gradually complicated models and evaluation of them by external
criterion on separate part of data sample. It was applied in many
countries for data mining and forecasting, systems modelling and
optimization, segmentation and pattern recognition. Since 1968 many
works and dissertations were devoted to investigations of GMDH in very
different fields. Until now this approach was implemented in several
commercial software products in USA and Germany.
The GMDH theory and source code of some algorithms was published in
"Self-Organising Data Mining"
Mueller, J.-A., Lemke, F. 2000, ISBN 3-89811-861-4, Libri, Hamburg,
http://www.knowledgeminer.net
"Inductive Learning Algorithms for Complex System Modeling",
Madala H.R. and Ivakhnenko A.G., 1994, ISBN: 0-8493-4438-7, CRC
Press
This approach is described at http://www.GMDH.net
GMDH books, articles and software can be found at http://www.GMDH.net/articles/
Best regards,
Gregory
Tomasso
An "unclamped" input (or output) can correspond to an absent or unset
value. You can consider a BM as a relation learning, but when a data
point is 1->M, it would flicker between estimates of the answers. The
mean field BM would be risky because of this (unless you ensured
functional data).
T.
--
"xch...@wisc.edu" <setti...@gmail.com> wrote in message news:1171497367....@l53g2000cwa.googlegroups.com...