Description
of MATEDA
MATEDA comprises a set of matlab
programs that implement different
EDAs. Each program is commented and some examples of their
use are available from the code. Current implementation
includes EDAs
for discrete and continuous problems. EDAs based on Bayesian and
undirected graphical models have been included. The software is not
warranteed in any way.
There are three main classes of EDAs implemented:
- EDAs based on undirected graphical models for
discrete problems
- EDAs based on Bayesian networks for discrete
problems
- EDAs based on Gaussian models for continous
problems
EDAs
based on undirected graphical models
All the EDAs in this class do only parametric learning. The structure
of the probability model is fixed in all the iterations. This structure
can be given to the algorithm by the user. The model of choice is a
junction graph where every set of interacting variables (definition
sets of the function when it is additive function) is represented by a
clique.
RunUMDA.m
This is the simplest EDA. It uses a univariate probabilistic model
where all the variables are independent.
RunMarkovFDA.m
EDA that uses a junction graph where each variables depends on the
previous k variables (k-Markovian model). k is a parameter of the
algorithm. For k=0, RunMarkovFDA.m behaves as RunUMDA.m
RunFDA.m
Implementation of the Factorized Distribution Algorithm (FDA). The set
of cliques where variables are defined is given. The cliques has
to form a junction graph. This is a generalization of the FDA where
cliques are constrained to form a junction graph (i.e. a junction graph
without cycles).
EDAs
based on Bayesian networks
RunBNEDA.m
EDA that uses as a probabilistic model a Bayesian network. Different BN
structure learning algorithms are used. The instalation of the BNT
and the BNT
structure learning software packages are needed.
Current implementations include the maximum weight spanning
tree and K2 learning algorithms.
EDAs
based on Gaussian models
RunGaussianEDA.m
EDA that uses univariate and multivariate Gaussian models to
approximate probability distributions for problems with continuous
representation.