Submitted: August 2022
Abstract
Modeling continuous probability distributions using Markov chains yields results that can predict the future very accurately, however the emerging models grow nonlinearly as a function of the number of states. To combat this, one can reduce these representations without altering the behavior but reducing the size in the process. For this the algorithm proposed in the paper Reduction of acyclic phase-type representations can be used. These compressed representations can then be rounded in a desired direction by converting exact fractions into floating point numbers and then applying rounding on them. This theses will show the compression algorithm and then propose an algorithm with which one can use floating point arithmetic to round the compressed results as well as show the implications for the memory consumption and an empirical approximation proximity.