@stacchella found that in some applications the intrinsic scale factor tuning for rwalk can lead to a lot of failed likelihood evaluations. I think this might be due to the variation in the scale factor from update to update, leading to overly-large (and subsequently overly-small) ellipsoids, forcing repeated tuning to "fix" the problem and leading to inefficient behavior. I think switching over so that we scale by volume (rather than length-scale) should help resolve these problems by reducing the variability, although it might make some of the shrinking a little bit less efficient.
@stacchella found that in some applications the intrinsic scale factor tuning for
rwalkcan lead to a lot of failed likelihood evaluations. I think this might be due to the variation in the scale factor from update to update, leading to overly-large (and subsequently overly-small) ellipsoids, forcing repeated tuning to "fix" the problem and leading to inefficient behavior. I think switching over so that we scale by volume (rather than length-scale) should help resolve these problems by reducing the variability, although it might make some of the shrinking a little bit less efficient.