Variational Inference for Markov Jump Processes
Manfred Opper
T.U. Berlin


Guido Sanguinetti
University of Sheffield

Markov  Jump  Processes  (MJP):  stochastic  processes  for  integervalued  random  variables  in  continuous  time. They  are  common  in  many  applications:  single  cell  biology, ecological models, telecommunications, etc. They are widely studied through simulation (Gillespie's  algorithm);  inference  however  is  a  long  standing  problem. We  propose  a  novel  meanfield  algorithm  that  allows  practical inference. We  demonstrate  the  method  on  two  examples,  the  classic  LotkaVolterra  predatorprey  model  and  a  gene  autoregulatory network.