new papers on modular NN and DATA ALLOCATION

[Thanasis Kehagias]

NEW PAPERS

The following papers can be obtained from my WEB site. 

-------------------------------------------------------------
1. "A General Convergence Result for Data Allocation in Online Unsupervised
Learning Methods". (With V. Petridis). Poster Presentation in the Second
International Conference on Cognitive and Neural Systems, Boston
University, 1998. 
(http://skiron.control.ee.auth.gr/~kehagias/thn/thn02c05.htm)

2. "Identification of Switching Dynamical Systems Using Multiple Models".
(With V. Petridis).  In Proceedings of CDC 98, 1998. 
(http://skiron.control.ee.auth.gr/~kehagias/thn/thn02c04.htm)

3. "Unsupervised Time Series Segmentation by Predictive Modular Neural
Networks". (With V. Petridis). In Proceedings of ICANN 98, 1998. 
(http://skiron.control.ee.auth.gr/~kehagias/thn/thn02c03.htm)

4. "Data Allocation for Unsupervised Decomposition of Switching Time Series
by Predictive Modular Neural Networks". (With V. Petridis). Accepted for
Publication in the Proccedings of IFAC Conference on Large Scale Systems,
Theory and Applications, Patras, Greece, 1998. 
(http://skiron.control.ee.auth.gr/~kehagias/thn/thn02c02.htm)
-------------------------------------------------------------

All these papers deal with a common problem for which we use the term DATA
ALLOCATION. Briefly, the setup is the following: suppose a collection of
data y(1), y(2), y(3), ... is generated by more than one SOURCES. Namely,
at time t one of the sources is selected (perhaps randomly) and then the
selected source generates the next datum y(t). Now, it is required to build
a model for each source, or estimate some of its parameters and so on. NO A
PRIORI INFORMATION IS AVAILABLE regarding the number, statistical behavior
etc. of the sources. 

If the observed data were split into groups, each group containing the data
generated by one source, it would be relatively easy to train a model (e.g.
a neural network) for each source. But the problem is that the data are
UNLABELLED: no information is available as to which source generated which
datum. So the main problem is DATA ALLOCATION, i.e. the grouping of the data.

The problem is as described above; furthermore we consider an online
version of it. (Hence EM and iterative clustering approaches cannot be
used). The results are of great generality: we provide some sufficient
conditions (which can reasonably be expected to hold for a large class of
algorithms)  which guarantee CORRECT (in a precise sense) data allocation.

Our newly published BOOK (announced in a separate message) also deals with
the same problem, in greater detail (i.e. all the proofs are included).
More info can be found at my WEB site:

http://skiron.control.ee.auth.gr/~kehagias/thn/thn02b01.htm

I will also post a separate message with some additional thoughts and
biblio on the DATA ALLOCATION problem.
___________________________________________________________________
Ath. Kehagias
--Assistant Prof. of Mathematics, American College of Thessaloniki 
--Research Ass., Dept. of Electrical and Computer Eng. Aristotle Univ.,
Thessaloniki, GR54006, GREECE

--email:	 kehagias at egnatia.ee.auth.gr, kehagias at ac.anatolia.edu.gr
--web:	http://skiron.control.ee.auth.gr/~kehagias/index.htm