We are faced with many complex dynamic systems that need to be observed and/or controlled. As they are usually not fully theoretically tractable, we have to make their approximate models. The soft-computing techniques involving neural networks, evolutionary algorithms, fuzzy logic, and other natural computing algorithms are the state-of-art methods for solving complex problems like recognition, classification, optimization, identification, and simulation. These models then present a basis for closed-loop control of the complex systems.

The choice of the relevant inputs to the model is an important task, that can be solved either with classical techniques (correlation, spectral analysis), with analysis of average mutual information between inputs and outputs or with transformation of input data (PCA, ICA). These algorithms are controlled by some cost functions that are usually related to the error of the actual and model outputs. In a case when the error distribution does not follow the normal distribution, one has to use other criteria functions, like information entropy and divergence. Then, the learning/evolving procedures are very demanding in terms of the processing power, which implies the use of a parallel programming/processing approach.

For some applications we shall look for the most appropriate synthesis methods for their modeling, by taking into account natural based algorithms, information related cost functions and tools for parallel programming and/or processing on different levels of hardware. Based on the preliminary analysis the chosen models will be implemented and tested.

Last activities from above areas of research can be divided into three actual works: (i) parallel implementations of information-theoretic processing on graphics processing units, (ii) development of new methods for data clustering based on ensemble approach, and (iii) design of information processing in living cells with artificial gene repressors.