An Aspect Oriented Framework To Applying Markov Chain Monte Carlo Methods With Dynamic Models

Both dynamic modeling and Bayesian Markov Chain Monte Carlo (MCMC) methods are established as increasingly popular approaches in their own domains. Dynamic modeling, although widely used to address complex situations, often suffers shortage of empirical data for model parameterization. Dynamic modelers thus use calibration to estimate parameters for which direct evidence is lacking. Unfortunately calibration suffers limitations in capturing the global (for multi-modal distribution) structure of parameter distributions, and a lack of a means of translating uncertainty in parameter estimates directly into uncertainty with respect to model outcomes. We present here a generic user-friendly aspect-based implementation of a theoretically grounded approach to address these limitations by combining Bayesian MCMC methods with dynamic models to estimate model parameters by sampling from joint posterior parameter distributions. The framework is enriched by a user interface to enable the parameter selection at run-time and an interactive run-time graphical visualization of parameter traceplots is generated during MCMC operation. To enable this, a probabilistic model - including a prior distribution and a likelihood function - needs to be specified within the dynamic model. The framework, when enabled, performs MCMC experiments using the dynamic and probabilistic models. We describe here the framework, experiments conducted, and the results obtained.