AMCL converges to wrong pose and is unable to recover

Hi, I am working with AMCL together with laser_scan_matcher (as odometry source) for my simulated laser-only based virtual robot. I have a prior map and want to do localization only, without updating the map. I always call the *global_localization* service and avoid providing an initial pose to AMCL. In most of my experiments, AMCL does manage to converge to the correct pose. However, during long walks it diverges at some point (probably due to bad odometry from the laser_scan_matcher). Since at that point the particle cloud is already dense, AMCL will persist its wrong pose and won't recover. I played a little with AMCL parameters but couldn't find something which helped. I'm basically after a parameter which tells AMCL at what rate to expand its particle cloud when sensorial information doesn't not correspond with the map. Thanks!