On Structured Initial Solution Generation for Phase Type Fitting with EM Method

  • Mindaugas Bražėnas Kaunas University of Technology
  • Eimutis Valakevičius Kaunas University of Technology
Keywords: phase type fitting, expectation maximization, structure generation, sparse representation


The choice of initial solutions when fitting with a phase-type distribution (PH), using the expectation maximization method (EM) is investigated. It is known that the EM method can converge to a local solution, especially when fitting with a general structure. The problem of how to choose an initial solution for which the EM method would converge to a global solution is still open. We contribute to the research of this problem by studying the use of structures for initial solution generation. The proposed approach is tested by fitting with four state phase-type distributions (PH(4)). Numerical results show that the EM method converges faster from initial solutions of various structures.

DOI: http://dx.doi.org/10.5755/j01.itc.47.2.18169