I'm copying and pasting Nikhil's email here as he was unable to login to the elog (but should now be able to in order to reply to any comments, and add more details about this test, motivation, methodology etc).
I did some post-processing after running the grid search. The following steps were carried out:
1) Selected those sets whose cost fun were less than a specific threshold (here 10000)
2) Next task was to see if the parameters of these good solutions had some pattern
3) I used a dimensionality reduction technique called t-SNE to project the 6 dimensional parameter space to 2 dim (for better visualization )
4) Made a scatter plot of these (see fig )
5) Used K-Means to find the clusters in this data
6) MarkerSize & Color reflect the cost fun. Bigger the marker size means better the solution.
7) Visual inspection implied cluster 5 had the best ranking points & more than any other cluster
8) These points had the following Parameter set: Workers {20,40}, SwarmSize {500}, MaxIter {500}, Self Adjustment {1}, Social Adjustment {1}, Tolerance {1e-3,1e-8}
See fig: for the box plot
9) It looks like is a particular set of values rather than individual values that gives the best results.