Optimal stopping under partial observations