When inferring multiple parameters (e.g. 3 params in the bottleneck model), samples from the MCMC for two of the parameters are highly autocorrelated (for at least the randomwalk MCMC kernel). The one parameter with low autocorrelation is inferred with reasonable accuracy, whereas the highly autocorrelated paramers are not. Sampling every 10 accepted MCMC steps reduces the autocorrelation, but does not affect accuracy of param inference.
We might try to specify a new_state_fn for tfp.mcmc.RandomWalkMetropolis(). It's not clear to me how the new state proposal function and the bijectors might interact. More investigation needed here.
When inferring multiple parameters (e.g. 3 params in the bottleneck model), samples from the MCMC for two of the parameters are highly autocorrelated (for at least the randomwalk MCMC kernel). The one parameter with low autocorrelation is inferred with reasonable accuracy, whereas the highly autocorrelated paramers are not. Sampling every 10 accepted MCMC steps reduces the autocorrelation, but does not affect accuracy of param inference.
We might try to specify a
new_state_fnfor tfp.mcmc.RandomWalkMetropolis(). It's not clear to me how the new state proposal function and the bijectors might interact. More investigation needed here.