Modèles Informatiques du Langage et de la Cognition - MILC
Artificial Intelligence Group
Computer Science Department
Français
Research
Papers
Home
Ph.D. Student
Biography
Born August 2nd 1970 at Talence (Gironde, France).
Education : Engineer at E.N.S.T. (1991-1994).
Five months internship in the AI department of the Regie Renault on the
forecasting of time series with neural networks (quarterly forecasting of the
number of car sales in France). For more details, you can have a look at my resume in postscript format (in french).
Thesis director : Alain Grumbach
Research Group: MILC
Laboratory : SNCF, Direction de la Recherche, Département Prospective
Intended examination date : November 1997
A connectionnist Approach to Time Series Forecasting
Nowadays, classical connectionnist techniques are very well-known, particularly in the
field of MultiLayers Perceptrons (MLP). Nonetheless, time series predictions using MLP
are quite difficult and results are not as good as they were supposed to be.
In fact, time series do not accomodate very well with neural networks because of :
- an important noise
- the notion of next and previous (that is to say the notion of time)
which are not present in MLPs
The aim of this Ph.D. is to design a methodology for this kind of neural
networks
based forecasting, to find architectures which are suitable for
some kind of prediction (FIR, TDNN, RST,... , hybrid ones), or design new
ones, and finally to design tests which may be used to compare our results
with statistical models.
Another part of our work is to find a new architecture dedicated to
time dependant series, that is to say a neural network architecture
where time will not be another spatial dimension. Opposed to TDNN, we
want time to be a built-in dimension. A first attempt in this
direction is the delta-NARMA neural network, which can be seen as a
connectionist extension of ARIMA and ARARMA statistical models.
Another topic of interest for us is the fusion between symbolic and
numerical processing in the particular case of time series
prediction. We study both the use of classical statistical methods
(Fischer statistic) and connectionist one to estimate relevance and
redundancy. We also study from a more general point of view how to
handle symbolic data with a classical connectionist network.
To implement delta-NARMA neural networks and to make experiments about
the fusion of symbolic and numerical processing, we have developped a
simulator: dnns.
A course for 3rd year ENST student I wrote on time series prediction
using neural networks is available here (in french, compressed
postscript, 683K).
Revues
- D.Bonnet, V. Perrault, A. Grumbach (1997)
"Delta
Networks: A Formalism to Design Local Extensions of Neural Network
Architectures"
submitted to Neural Processing Letters
(postscript)
- D.Bonnet, V. Perrault, A. Grumbach (1997)
"Using
symbolic data to improve connectionist forecasting: a methodology"
ENST technical report 97C002, Paris, France (postscript)
- D.Bonnet, V. Perrault, A. Grumbach (1996)
"Delta-NARMA neural networks: a new approach to signal prediction"
ENST technical report 96C008, Paris, France -- to appear in IEEE trans. on Signal Processing special issue on Neural Networks for Signal Processing (postscript)
Conferences
- D.Bonnet, V. Perrault, A. Grumbach (1997)
"Daily Passenger Traffic Forecasting using delta-NARMA Neural Networks"
to appear in the proceedings of the World Congress on Railroad Research (WCRR'97)
- D.Bonnet, V. Perrault, A. Grumbach (1997)
"Delta-NARMA
neural networks: a connectionist extension of ARARMA models"
Proceedings of the European Symposium on Artificial Neural Networks
1997 (ESANN'97), ed. M. Verleysen, D Facto, Brussels, Belgium, pp. 127-132 (postscript)
- D.Bonnet, V. Ancona, A.Grumbach (1995)
"Forecasting rail traffic using hierarchical mixtures of connexionist experts"
International Conference on Artificial Neural Networks 1995 (ICANN'95), EC2, Paris, France
- D.Bonnet, V. Perrault, A. Grumbach (1997)
"Utilisation
des modèles connexionnistes pour la prévision à court terme du trafic
ferroviaire"
to appear in the proceedings of the 3ème
journées industrielles Modulad, Applications industrielles de
l'analyse des données
- D.Bonnet, V. Ancona, A.Grumbach (1995)
"Modélisation du traffic ferroviaire par arbre d'experts connectionnistes"
Journée de sensibilisation aux méthodes neuronales, Observatoire Economique et Statistique des Transports, Paris, France
- D.Bonnet (1995)
"Approche temps-fréquence pour un prédicteur neuronal"
Technical Report, DR/RP/TSI/IA/DB-95-001, SNCF, Département Prospective, Paris, France
- D. Bonnet (1994)
"Prévision du M.T.M. par réseau de neurones"
Internship Report, ENST, Paris, France
Last changes March 7 1997
Denis Bonnet (bonnet@inf.enst.fr)
Denis BONNET
Research
Papers