Inverse Markov Chain for a stochastic process with partial observations

Alexandros Sopasakis
Centre for Mathematics
Lund University
Lund


Abstract:

A novel microscopic stochastic lattice-free process is developed and presented. The stochastic process is considered in order to resolve shortcomings due to finite size effects when considering lattice based models. The effect of the dynamics on the model parameters can be considerable especially when equilibrium is not certain. We examine how information theory can be applied in resolving the dynamics and obtaining the needed parameter values from available data.

Specifically we develop two different tools, the Relative Entropy Rate and the Fisher Information Matrix whose task is to reconstruct the Markov Chain and emulate the dynamics from available but incomplete data through Monte Carlo simulations.

We demonstrate this process by simulating a problem in vehicular traffic. In that respect a look-ahead potential is employed and an anisotropic simple exclusion process (ASEP) is created which moves particles in only one direction. The process is implemented using a kinetic Monte Carlo method and contrasted against other similar "state of the art" lattice-based approaches in that area. We present concrete results under rush-hour, traffic data from highway US-101 in Los Angeles, California.