Open ericagol opened 4 years ago
It would be useful to add an acceptance rate to the sampler to see whether the a parameter should be adjusted.
a
If the acceptance rate is too small, I would adjust the a parameter closer to unity automatically via:
a_{new} = 1 + (a_{old} -1)*(acc_{measured}/acc_{target})
where acc_{measured} is the measured value of acceptance (over some number of steps), while acc_{target} is a target acceptance rate (usually 25%). Then, once the target acceptance rate is reached, the value of a stays constant.
acc_{measured}
acc_{target}
It would be useful to add an acceptance rate to the sampler to see whether the
a
parameter should be adjusted.If the acceptance rate is too small, I would adjust the
a
parameter closer to unity automatically via:a_{new} = 1 + (a_{old} -1)*(acc_{measured}/acc_{target})
where
acc_{measured}
is the measured value of acceptance (over some number of steps), whileacc_{target}
is a target acceptance rate (usually 25%). Then, once the target acceptance rate is reached, the value ofa
stays constant.