Learning sampling distributions for randomized motion planning: role of history size