@gdalle thank again! I don’t have any good reason to use ForwardDiff, I was simply trying to get something up and running. I changed over to ReverseDiff based on your recommendation, for now doing the simplest (and most inefficient thing). I also get an error with this AD type, but it tells me how to fix it ReverseDiff.value(forward_model(prev_states[:, n])) allows for
p, re = Flux.destructure(ffnn);
f = (w) -> sim_vsmc(w, re, y, X, 10, ones(4), 0.1; lower_bound_only=true)
ReverseDiff.gradient(f, p)
to run without errors. Many thanks!