Research topics
- Heterogeneous Neural Networks
The
topic of my Ph.D. Thesis, which I finished in 2000.
These networks are special classes of functions based on generalized
neuron
models. These models are cast in the common framework of computing an
explicit similarity function and can naturally handle
heterogeneous information such as categorical, ordinal or fuzzy
data, as well as missing values. Then the neurons can be used to create artificial neural
networks.
» I am still working in new training methods, as well as
further generalizations and developments.
- Feature
Selection
Algorithms
These are
algorithms that
try to find a reduced representation of a problem description, with
less features than the original. In this new representation
redundant, irrelevant, poorly relevant or highly noisy features have
(hopefully) been removed. This may bring benefits as better
generalization (less overfit), faster computation and gained domain
knowledge.
» I am interested in both
theoretical
developments
(which
are
lacking in the literature) and in
sensible experimental comparisons
These are one of the key ingredients of
kernel-based
methods (such as support vector machines). One of the fascinating
aspects of these methods is that they can be applied to almost any kind
of data without distorting codings, if one is able to develop a valid
kernel function.
» I am specially interested in the design of kernels capable of
handling non-standard data (categorical, fuzzy, ...) as well as missing
values.
These
are
two-place
functions
very
commonly used in data analysis and
strongly related to dissimilarity functions and
metrics.
» I am interested
in developing an axiomatic formalization, as well as a body of results
than can both characterise the two sets of functions, their relation
and their properties.
» A long-term and (probably very ambitious) goal is
the the development of methods to relate a data set with a promising
similarity.
- Breeder
Genetic
Algorithms
The
Breeder
Genetic
Algorithm
(BGA)
is
in
midway
between Genetic Algorithms and Evolution
Strategies. It works directly in a continuous space and is mainly
recombination driven, though mutation has also a place. It keeps a
good compromise between simplicity and effectiveness. I developed new
genetic operators for the BGA and was interested in comparisons with
other methods (GAs, backprop, ...) to train multilayer perceptrons.
» I have
discontinued research on this topic although I keep a general interest
in evolutionary algorithms, specially in the rather new CMA-ES method.
I have always tried to find
opportunities for collaboration
with people having data (and a specific data analysis problem).
This
brings
many
benefits:
you
meet people with different background,
you can apply your methods on real data and you have an expert that
guides the analysis. Through the years I have been lucky enough to meet
scientists at the University of
Girona (wastewater treatment plant
data) and the University of Barcelona
(microbiology data), as well as
local industries (e.g., insullation panels) as well as hospitals
(e.g., brain
tumours).
» I am still very interested in this
applied side of the work.
For a complete list of papers, please have a look at this
page.
Visit the Home Page of our Research Group
on Soft
Computing Systems