## Research

Our research focuses on natural inspired methods in machine
learning. Neural networks, evolutionary algorithms, neuroevolution,
etc. We apply our algorithms in data mining and artificial
intelligence. We study ensembling approaches and meta-learning in
predictive modelling, recommender systems or data clustering.

Disadvantage of black-box modelling results from its
multidimensionality and complexity. Where the traditional modelling
methods fail due to the “curse of dimensionality” phenomenon,
inductive methods are capable to build reliable models.

To overcome the black-box disadvantage we recommend using the
visualisation of models' behaviour. We are developing several
visualization techniques that can be directly used for the visual
knowledge mining. Pictures showing models' behaviour are very
useful especially for complex systems, where math equations are not
interpretable any more.

Our research is concentrated also in the field of knowledge
extraction. It means that useful information hidden inside a
black-box model is found and represented in a comprehensible manner
(e.g. in form of simple math formula, or a figure).

For time context based data the recurrent type of artificial neural
network can be used. Combined with the appropriate visualisation
technique or with method of the finite state machine description
from the recurrent neural network behaviour extraction, it could
give us the new view to the data structure in time (knowledge
mining, see above). Combination of this recurrent networks theory
and its behaviour visualisation methods with fuzzy logic will be
the part of the next research.

Many signals produced in the real environment have temporal
meaning. Such signal should be modelled using systems suitable for
processing of signals instead of system for static and steady data.
One can perform tasks such as recognition of signals and their
parts, prediction of time series, control and reasoning based on a
generated model. We use statistical models such as Hidden Markov
Models and recurrent neural networks such as extensions of
Categorizing and Learning Module. Various visualization techniques
help us to understand the model and a sense of the modelled signal
and help a user to exploit them directly in future industrial
applications.

Hardware support for fast simulation of adaptive system like neural
networks is the next objective of our research. Objectives are to
find methods for optimization of simulation models with respect to
fast hardware implementation. Various approximations simplificating
models but preserving their major behaviour are used.