When a model is linear, one can use Kalman filter to get likelihood function and use ML techniques. While nonlinear Kalman filter exists, it is suboptimal. What is the standard way to get likelihood function and apply maximum likelihood(ML) methods when a model is non-linear and not linearized?

  • $\begingroup$ It would be useful if you include in your question the specific non-linear model that you are dealing with. In the non-linear world, in many cases there do not exist answers of full general applicability. $\endgroup$ – Alecos Papadopoulos Jun 10 '15 at 14:09