Simulation input mixture models Event

Event details

Multimodal data occurs frequently in discrete-event simulation input analysis occurring when an input sample stream arises from different sources. A finite mixture distribution is a simple input model for representing such data. However fitting this model to such data is not straightforward, with the problem well-known to be non-standard. Even though much studied, the two most common approaches, maximum likelihood (ML) and Bayesian reversible jump Markov chain Monte Carlo (RJMCMC), do not seem very satisfactory in our simulation application. We review the problems that arise with each approach, and then describe an alternative Bayesian approach, MAPIS, which uses maximum a posteriori estimation with importance sampling, showing it overcomes the main problems encountered with ML and RJMCMC. We demonstrate use of a publicly-available implementation of MAPIS, which we have called FineMix, applying it to practical examples from finance and manufacturing, comparing the results with those of ML and RJMCMC.