Is there any optimal set? No. If there was some optimal set of algorithmic parameters, do you think there is any reason they would not have set it as the default? I can see them discussing things now - Lets let people guess! We will make them work for the right answer here, when we know the optimal way to drive this code. Be serious.
Is there any magical strategy to insure you get better answers? Not really.
Essentially, you need to force the algorithm to search longer, search harder, search broader. This improves the chances it will stumble on the best solution. If you know of constraints or bounds that should be pertinent, THEN USE THEM! You need to reduce the search space as much as possible.
But GA is just a stochastic search, one based on a genetic metaphor. As such, it is one we hope might work well, because genetics has done such interesting and useful stuff over billions of years. But there is no absolute assurance here. In fact, you might notice that the original genetic algorithm has not yet converged. Some might even argue we are the consequence of an evolutionary mistake - a locally suboptimal solution that may one day be replaced by a better one. :-)
Regardless, 100 data points, with 20 parameters? That is probably pushing the limits, since that is a huge search space to sift through. You may have barely enough useful information content in those 100 points to push the algorithm into the solution you want to see. And that is probably why you are seeing poor convergence behavior.
