With gradient descent you are walking in the direction of negative gradient. Stirring the pile implies that you are walking randomly from point to point.
To see this most clearly, read up on estimation of distribution algorithms (EDAs), which generalise the ways in which GAs work.