Possible bug in Turing Gibbs of MH

[marcoct]

These two configurations give different results for the algorithmic model (inverse planning) problem (with noise and speed fixed to values near the ground truth value, to simplify the problem):

chain = sample(model(scene, times, xs, ys), Gibbs(iters, MH(1, :stop_x, :stop_y)))

chain = sample(model(scene, times, xs, ys), MH(iters, :stop_x, :stop_y))

See https://github.com/probcomp/pldi2019-gen-experiments/blob/master/algorithmic-model/turing/generate_turing_results.jl#L62

[marcoct]

I tried to reproduce this Turing problem on a small test case but have not succeeded yet:

https://github.com/probcomp/pldi2019-gen-experiments/blob/master/algorithmic-model/turing/turing_debug.jl