Problems with amcl in solving the problem of kidnapped robot [closed]

edit

I am trying to solve the problem of kidnapped robot using amcl. Sometimes the method returns a very good estimate of the position of the robot (+/- 15 cm) and sometimes there is a difference of 4 metres in the estimated and actual positions of the robot. The map was generated in 8 metres*8 metres room. Hence, an error of 4 metres is too large. Also,the estimate is very good in certain parts of the room. The particle filter has a maximum of 3000 particles. Can anyone suggest me the optimum number of particles I can use for good accuracy of localisation without slowing down the process too much? Are there other ways to support the amcl to improve localisation? I am using ROS Fuerte on Ubuntu 12.04 and my robot is a turtlebot.