Following the description on Wikipedia, reproduced here for completeness:
We try it on the six-hump camelback function:
Some sample trajectories during optimization, starting from uniformly sampled points:
The endpoint distribution is (counts indicate number of trajectories that end at that location from 100 trajectories):
Most points have converged to one of the local minima, but others are stuck at saddle points, in particular at (0,0).
We can add some Gaussian noise to the gradients (as is common in machine learning due to the use of a small batch size) to move away from the local maxima:
