Welcome to my site!

From time to time i will publish some of my data compression projects and experiments on this site. My work is mostly located in the area of statistical compression, context mixing (CM) in particular.

Data compression algorithms transform data streams, typically a sequence of symbols over a finite alphabet, into a less redundant representation. Statistical data compression breaks the compression process into two phases: modeling and coding. A statistical model generates probability estimations for upcoming symbols. Based on these estimations the coding step assigns unique encodings to the symbols. In contrast to classical methods, like Prediction by Partial Matching (PPM), CM combines multiple models prediction and often achieves outstanding compression performance. However it comes at its cost: astronomic CPU time and RAM usage.

I'm interested in defeating these shortcomings, i.e. creating efficient, universal statistical models and their automated fitting to training data. This approach mirrors engineering techniques for model development. The inefficiency which is related to wrong model parameters might mask out the superiority of a certain model structure and lead to wrong conclusions, i.e. parameter optimization can minimize this effect.