I am simulating a non-rational expectation model, which is formed by the lagged variables only. Can I just put a tiny coefficient on the forwad variables, so that I can get a approximation of my model results?

Dynare can solve for the first order approximation of purely backward models like yours.

Second and third order approximation of such models are not yet implemented, they should be in the future.

Thanks a lot, you help me a lot!!

I forgot to mention, but of course such models can easily be solved without an approximation. You can solve them iteratively by using a non-linear solver at each period, going forward in time.