This answer.
Use the scoring function of Autodock. If you can (not sure if Autodock allows that) cluster your solutions and look at the average score of the clusters (or at the average of the top 5 or 10 solutions in each cluster), as this might filter some awkward high scoring isolated solutions.
If you have two clusters or solutions that are close in score, have a look at the structures yourself. Do they make sense? Do they match what they should from the biological role of the protein-ligand complex?